Connect with us


Unlock near 3x performance gains with XGBoost and Amazon SageMaker Neo

When a model gets deployed to a production environment, inference speed matters. Models with fast inference speeds require less resources to run, which translates to cost savings, and applications that consume the models’ predictions benefit from the improved performance. For example, let’s say your website uses a regression model to predict mortgage rates for aspiring…



When a model gets deployed to a production environment, inference speed matters. Models with fast inference speeds require less resources to run, which translates to cost savings, and applications that consume the models’ predictions benefit from the improved performance.

For example, let’s say your website uses a regression model to predict mortgage rates for aspiring home buyers to see what type of rate they could expect, based on inputs they provide such as the size of the down payment, their loan term, and the county in which they’re looking to buy. A model that can send a prediction back in 10 milliseconds versus 200 milliseconds for every time an input is updated makes a massive difference in terms of the website’s responsiveness and user experience.

Amazon SageMaker Neo allows you to unlock such performance improvements and cost savings in a matter of minutes. It does this by compiling models into optimized executables through various open-source libraries, which can then be hosted on supported devices on the edge or on Amazon SageMaker endpoints. Neo is compatible with eight different machine learning (ML) frameworks, and in the context of gradient boosted tree algorithms such as XGBoost, Neo uses Treelite to optimize model artifacts. Due to the popularity of XGBoost and its unique categorization as a more classical ML framework, we use it as our framework of choice throughout this post. A near 3x speedup will be demonstrated for the optimized XGBoost model compared to the unoptimized one. The Abalone dataset from UCI will be used to train the model. Please feel free to use your own model and dataset, however, and let us know in the comments what type of acceleration was achieved.

This post will take a deeper dive into compiling XGBoost model artifacts using Neo and will show you how to accurately measure and test the performance gains of these Neo-optimized models in general. By the end of this walkthrough, you’ll have your own framework for quickly training, deploying, and benchmarking XGBoost models. In turn, this can help you make data-driven decisions on what type of instance configurations best fit your unique cost profile and inference performance needs.

Solution overview

The following diagram visualizes the services we use for this solution and how they interact with one another.

The steps to implement the solution are as follows:

  1. Download and process the popular Abalone dataset with a Jupyter notebook, and then run an XGBoost SageMaker training job on the processed data. We use a local mode SageMaker training job to produce the unoptimized XGBoost model, which can be faster and easier to prototype compared to a remote one.
  2. Deploy the unoptimized XGBoost model artifact to a SageMaker endpoint.
  3. Take the unoptimized artifact and optimize it with a Neo compilation job.
  4. Deploy the Neo-optimized XGBoost artifact to a SageMaker endpoint.
  5. Create an Amazon CloudWatch Dashboard from the SageMaker notebook to monitor inference speed and performance under heavy load of both endpoints.
  6. Deploy Serverless Artillery from the SageMaker notebook, which we use as our load testing tool. We set up Serverless Artillery entirely from the SageMaker notebook, and directly invoke your SageMaker endpoints from the internet through manually signed AWS Signature Version 4 requests—no need for Amazon API Gateway as an intermediary.
  7. Perform load tests against both endpoints.
  8. Analyze the performance of both endpoints under load in the CloudWatch dashboard, and look at how the optimized endpoint outperforms the unoptimized one.


Before getting started, you must have administrator access to an AWS account, and complete the following steps:

  1. Create an AWS Identity and Access Management (IAM) role for SageMaker that has the AmazonSageMakerFullAccess managed policy attached along with an inline policy that contains additional required permissions.

The following screenshot is an example of a properly configured role called NeoBlog.

The AdditionalRequiredPermissionsForSageMaker inline policy contains the following JSON:

{ “Version”: “2012-10-17”, “Statement”: [ { “Effect”: “Allow”, “Action”: “cloudwatch:PutDashboard”, “Resource”: “arn:aws:cloudwatch::*:dashboard/NeoDemo” }, { “Effect”: “Allow”, “Action”: [ “s3:CreateBucket”, “s3:GetBucketLocation”, “s3:GetObject”, “s3:ListBucket”, “s3:PutObject”, “s3:DeleteBucket”, “s3:DeleteObject”, “s3:DeleteObjectVersion”, “s3:PutLifeCycleConfiguration”, “s3:GetEncryptionConfiguration”, “s3:PutEncryptionConfiguration”, “s3:PutBucketPolicy”, “s3:DeleteBucketPolicy”, “s3:GetBucketPolicy”, “s3:GetBucketPolicyStatus” ], “Resource”: “arn:aws:s3:::serverless-artillery-*” }, { “Effect”: “Allow”, “Action”: [ “cloudformation:CreateStack”, “cloudformation:UpdateStack”, “cloudformation:DeleteStack”, “cloudformation:DescribeStacks”, “cloudformation:DescribeStackEvents”, “cloudformation:DescribeStackResource”, “cloudformation:DescribeStackResources”, “cloudformation:ListStackResources” ], “Resource”: “arn:aws:cloudformation:*:*:stack/serverless-artillery-*” }, { “Effect”: “Allow”, “Action”: [ “cloudformation:ValidateTemplate” ], “Resource”: “*” }, { “Effect”: “Allow”, “Action”: [ “iam:GetRole”, “iam:CreateRole”, “iam:DeleteRolePolicy”, “iam:PutRolePolicy”, “iam:DeleteRole”, “iam:PassRole” ], “Resource”: “arn:aws:iam::*:role/serverless-artillery-*” }, { “Effect”: “Allow”, “Action”: [ “sns:CreateTopic”, “sns:DeleteTopic”, “sns:GetTopicAttributes” ], “Resource”: “arn:aws:sns:*:*:serverless-artillery-*” }, { “Effect”: “Allow”, “Action”: [ “lambda:UpdateFunctionCode”, “lambda:ListVersionsByFunction”, “lambda:PublishVersion”, “lambda:InvokeFunction”, “lambda:GetFunction”, “lambda:CreateFunction”, “lambda:DeleteFunction”, “lambda:GetFunctionConfiguration”, “lambda:AddPermission” ], “Resource”: “arn:aws:lambda:*:*:function:serverless-artillery-*” }, { “Effect”: “Allow”, “Action”: [ “logs:DescribeLogGroups”, “logs:CreateLogGroup” ], “Resource”: “arn:aws:logs:*:*:log-group:serverless-artillery-*” }, { “Effect”: “Allow”, “Action”: [ “logs:DeleteLogGroup”, “lambda:RemovePermission” ], “Resource”: “*” }, { “Effect”: “Allow”, “Action”: [ “events:DescribeRule”, “events:PutRule”, “events:DeleteRule”, “events:PutTargets”, “events:RemoveTargets” ], “Resource”: “arn:aws:events:*:*:rule/serverless-artillery-*” } ] }

Our next step is to create a SageMaker notebook instance.

  1. On the SageMaker console, under Notebooks, choose Notebook instances.
  2. Choose Create notebook instance.
  3. For Notebook instance name, enter NeoBlog.
  4. For Notebook instance type, choose your instance (for this post, the default ml.t2.medium should be enough).
  5. For IAM role, choose the NeoBlog role that you created.
  6. In the Git repositories section, select Clone a public Git repository to this notebook instance only.
  7. For Git repository URL, enter
  8. Choose Create notebook instance.
  9. After the notebook has reached a Running status, choose Open Jupyter to connect to your notebook instance.
  10. Navigate to the neo-blog repository in Jupyter and choose the NeoBlog.ipynb notebook to start it.

You’re now ready to walk through the remainder of this post and run the notebook’s contents.

Notebook walkthrough

The code snippets in this post match the code in the NeoBlog notebook. This post contains the most relevant commentary, and the notebook provides additional detail. When extra information is provided in the notebook, it’s called out accordingly. Let’s get started!

First, we must retrieve the Abalone dataset and split it into training and validation sets. We store the data in lightsvm format.

  1. Run the following two cells in the Jupyter notebook:

from pathlib import Path import boto3 for p in [‘raw_data’, ‘training_data’, ‘validation_data’]: Path(p).mkdir(exist_ok=True) s3 = boto3.client(‘s3’) s3.download_file(‘sagemaker-sample-files’, ‘datasets/tabular/uci_abalone/abalone.libsvm’, ‘raw_data/abalone’) from sklearn.datasets import load_svmlight_file, dump_svmlight_file from sklearn.model_selection import train_test_split X, y = load_svmlight_file(‘raw_data/abalone’) x_train, x_test, y_train, y_test = train_test_split(X, y, test_size=0.20, random_state=1984, shuffle=True) dump_svmlight_file(x_train, y_train, ‘training_data/abalone.train’) dump_svmlight_file(x_test, y_test, ‘validation_data/abalone.test’)

Now that we have our data shuffled and prepared, we can train an unoptimized XGBoost model. Refer to the commentary in the Jupyter notebook for details related to the container framework version, hyperparameters, and training mode being used.

  1. Train the model by running the following code cell:

import sagemaker from sagemaker.xgboost.estimator import XGBoost from sagemaker.session import Session from sagemaker.inputs import TrainingInput bucket = Session().default_bucket() role = sagemaker.get_execution_role() # initialize hyperparameters hyperparameters = { “max_depth”:”5″, “eta”:”0.2″, “gamma”:”4″, “min_child_weight”:”6″, “subsample”:”0.7″, “verbosity”:”1″, “objective”:”reg:squarederror”, “num_round”:”10000″ } # construct a SageMaker XGBoost estimator # specify the entry_point to your xgboost training script estimator = XGBoost(entry_point = “”, framework_version=’1.2-1′, # 1.x MUST be used hyperparameters=hyperparameters, role=role, instance_count=1, instance_type=’local’, output_path=f’s3://{bucket}/neo-demo’) # gets saved in bucket/neo-demo/job_name/model.tar.gz # define the data type and paths to the training and validation datasets content_type = “libsvm” train_input = TrainingInput(‘file://training_data’, content_type=content_type) validation_input = TrainingInput(‘file://validation_data’, content_type=content_type) # execute the XGBoost training job{‘train’: train_input, ‘validation’: validation_input}, logs=[‘Training’])

When the local training job finishes running (it should only take a few minutes), the next step is to deploy the XGBoost model artifact to a SageMaker endpoint. The Jupyter notebook contains additional information related to why we use the c5 instance family class, along with how the model artifact is saved in Amazon Simple Storage Service (Amazon S3).

  1. Deploy the model artifact by running the following cell:

from sagemaker.xgboost.model import XGBoostModel # grab the model artifact that was written out by the local training job s3_model_artifact = estimator.latest_training_job.describe()[‘ModelArtifacts’][‘S3ModelArtifacts’] # we have to switch from local mode to remote mode xgboost_model = XGBoostModel( model_data=s3_model_artifact, role=role, entry_point=””, framework_version=’1.2-1′, ) unoptimized_endpoint_name = ‘unoptimized-c5′ xgboost_model.deploy( initial_instance_count = 1, instance_type=’ml.c5.large’, endpoint_name=unoptimized_endpoint_name )

After the unoptimized model is deployed (the cell has stopped running), we run a Neo compilation job to optimize the model artifact. In the following code, we use the c5 instance type family, choose the XGBoost framework, and include an input shape vector. The input shape is unused by Neo, but the compilation job throws an error if no value is provided. The compilation job also uses the 1.2.1 version of XGBoost by default, which again is why we specified the 1.2-1 framework version during model training.

  1. Run the Neo compilation job with the following code:

job_name = s3_model_artifact.split(“/”)[-2] neo_model = xgboost_model.compile( target_instance_family=”ml_c5″, role=role, input_shape =f'{{“data”: [1, {X.shape[1]}]}}’, output_path =f’s3://{bucket}/neo-demo/{job_name}’, # gets saved in bucket/neo-demo/model-ml_c5.tar.gz framework = “xgboost”, job_name=job_name # what it shows up as in console )

  1. When the cell stops running and the compilation job is complete, we deploy the Neo-optimized model to its own separate SageMaker endpoint:

optimized_endpoint_name = ‘neo-optimized-c5′ neo_model.deploy( initial_instance_count = 1, instance_type=’ml.c5.large’, endpoint_name=optimized_endpoint_name )

  1. Next, we validate that the endpoints are functioning as expected. When you run the following code block, you should see numerical predictions returned from both endpoints.

import boto3 smr = boto3.client(‘sagemaker-runtime’) resp = smr.invoke_endpoint(EndpointName=’neo-optimized-c5′, Body=b’2,0.675,0.55,0.175,1.689,0.694,0.371,0.474′, ContentType=’text/csv’) print(‘neo-optimized model response: ‘, resp[‘Body’].read()) resp = smr.invoke_endpoint(EndpointName=’unoptimized-c5′, Body=b’2,0.675,0.55,0.175,1.689,0.694,0.371,0.474′, ContentType=’text/csv’) print(‘unoptimized model response: ‘, resp[‘Body’].read())

With both endpoints up and running, we can create the CloudWatch dashboard that we use to analyze endpoint performance. For this post, we monitor the metrics CPUUtilization, ModelLatency (which measures how long it takes for a model to return a prediction), and Invocations (which helps us monitor the progress of the load test against the endpoints).

  1. Run the following cell to create the dashboard:

import json cw = boto3.client(‘cloudwatch’) dashboard_name = ‘NeoDemo’ region = Session().boto_region_name # get region we’re currently in body = { “widgets”: [ { “type”: “metric”, “x”: 0, “y”: 0, “width”: 24, “height”: 12, “properties”: { “metrics”: [ [ “AWS/SageMaker”, “Invocations”, “EndpointName”, optimized_endpoint_name, “VariantName”, “AllTraffic”, { “stat”: “Sum”, “yAxis”: “left” } ], [ “…”, unoptimized_endpoint_name, “.”, “.”, { “stat”: “Sum”, “yAxis”: “left” } ], [ “.”, “ModelLatency”, “.”, “.”, “.”, “.” ], [ “…”, optimized_endpoint_name, “.”, “.” ], [ “/aws/sagemaker/Endpoints”, “CPUUtilization”, “.”, “.”, “.”, “.”, { “yAxis”: “right” } ], [ “…”, unoptimized_endpoint_name, “.”, “.”, { “yAxis”: “right” } ] ], “view”: “timeSeries”, “stacked”: False, “region”: region, “stat”: “Average”, “period”: 60, “title”: “Performance Metrics”, “start”: “-PT1H”, “end”: “P0D” } } ] } cw.put_dashboard(DashboardName=dashboard_name, DashboardBody=json.dumps(body)) print(‘link to dashboard:’) print(f’{region}#dashboards:name={dashboard_name}’)

After you run the cell, you can choose the output link to go to the dashboard, but you won’t see any meaningful data plotted just yet.

Now that the dashboard is created, we can proceed with setting up the Serverless Artillery CLI. To do this, we install Node.js, the Serverless Framework, and Serverless Artillery on our SageMaker notebook instance. The cell that installs Node.js can take a long time to run, which is normal.

  1. Run the following cell to install Node.js and the Serverless Framework:

%conda install -c conda-forge nodejs !npm install -g serverless@1.80.0 serverless-artillery@0.4.9

Next, we deploy Serverless Artillery. The code first changes directories into the directory that contains the code for our load generating AWS Lambda function. Then it installs the function’s dependencies and uses the Serverless Artillery CLI to package and deploy the load generating function into our account via the Serverless Framework. For more information on what Serverless Artillery is doing under the hood, refer to the Jupyter notebook.

We set up Serverless Artillery to directly hit our SageMaker endpoints with manually signed requests using the AWS Signature Version 4 algorithm. The benefit of this approach is that we get to directly hit and measure the performance of exclusively the endpoints during the load test. If we front our endpoints with intermediary services like a Lambda-backed API Gateway, the load test results capture the performance characteristics of the all three services together rather than just the SageMaker resources.

  1. Deploy Serverless Artillery with the following code:

!cd serverless_artillery && npm install && slsart deploy –stage dev

After running these cells, you should have Node.js version 12.4.0 or higher, Serverless Framework version 1.80.0, and Serverless Artillery version 0.4.9.

The next task is to create the load test definition, which we do by running two cells. The first cell defines a custom magic command, and the second cell creates the load test definition and saves it into script.yaml.

The test definition has six phases, each of which runs 2 minutes in length. The first phase begins with an arrival rate of 20 users per second, meaning that approximately 10 requests are generated and sent to each endpoint every second for two minutes. The next three phases scale by an additional 20 users per second, and the last two phases scale up by 40. Each request contains 125 rows for inference. The Artillery documentation (the tool that Serverless Artillery is based on) is a good resource for learning about the structure and additional features of load test definitions.

  1. Create the load test definition with the following code:

from IPython.core.magic import register_line_cell_magic @register_line_cell_magic def writefilewithvariables(line, cell): with open(line, ‘w’) as f: f.write(cell.format(**globals())) # Get region that we’re currently in region = Session().boto_region_name %%writefilewithvariables script.yaml config: variables: unoptimizedEndpointName: {unoptimized_endpoint_name} # the xgboost model has 10000 trees optimizedEndpointName: {optimized_endpoint_name} # the xgboost model has 10000 trees numRowsInRequest: 125 # Each request to the endpoint contains 125 rows target: ‘https://runtime.sagemaker.{region}’ phases: – duration: 120 arrivalRate: 20 # 1200 total invocations per minute (600 per endpoint) – duration: 120 arrivalRate: 40 # 2400 total invocations per minute (1200 per endpoint) – duration: 120 arrivalRate: 60 # 3600 total invocations per minute (1800 per endpoint) – duration: 120 arrivalRate: 80 # 4800 invocations per minute (2400 per endpoint… this is the max of the unoptimized endpoint) – duration: 120 arrivalRate: 120 # only the neo endpoint can handle this load… – duration: 120 arrivalRate: 160 processor: ‘./processor.js’ scenarios: – flow: – post: url: ‘/endpoints/{{{{ unoptimizedEndpointName }}}}/invocations’ beforeRequest: ‘setRequest’ – flow: – post: url: ‘/endpoints/{{{{ optimizedEndpointName }}}}/invocations’ beforeRequest: ‘setRequest’

With the load test defined, we’re now ready to start it! Because there are six stages with each stage taking 2 minutes, the test runs for a total of 12 minutes. You can monitor the progression of the load test by clicking on the link generated by running the second cell. The link redirects you to the CloudWatch dashboard that you created earlier.

  1. Perform the load test with the following code:

!slsart invoke –stage dev –path script.yaml print(“Here’s the link to the dashboard again:”) print(f’{region}#dashboards:name={dashboard_name}’)

Review the CloudWatch metrics

After 12 minutes have passed, refresh the dashboard and look at the metrics that have been captured.

The plotted data should look similar to the following screenshot, which has several interesting observations to unpack.

First of all, even at the very beginning of the load test, when both endpoints were only handling about 10 requests per second (RPS), the model latency of the neo-optimized SageMaker endpoint was still almost three times lower than the unoptimized endpoint. This shows you the power of Neo—with one quick compilation job, we unlocked a performance improvement of nearly three times greater in our XGBoost model hosted on SageMaker!

Secondly, by the end of the load test, the ModelLatency metric of the unoptimized model spiked to almost 1.5 seconds per request. The unoptimized model’s CPUUtilization metric also reaches 181%, which is close to the endpoint’s theoretical maximum of 200% given that the ml.c5.large instance type has 2 vCPUs. On the other hand, the optimized endpoint’s ModelLatency metric never crosses 10,000 microseconds, and the CPUUtilization metric stays well below capacity at under 50%. This indicates that the Neo-optimized endpoint could definitely handle even more load if needed, much more than the load test’s maximum of 80 requests per second.

Looking at the following graph, we can also see that the unoptimized endpoint’s performance begins to drastically drop off around the 21:27 timestamp. To get a better idea of what’s going on, deselect the ModelLatency metric for the unoptimized endpoint (the green line) to get the graph of the subsequent image. Upon doing this, you can see that Invocations metrics confirm the story. Up till the 21:27 mark, both endpoints were handling almost the exact same number of requests from the load test (indicated by the blue and orange lines). Past the 21:27 mark when the number requests per second starts to go above 40, the unoptimized endpoint begins to struggle to keep up. This indicates that the maximum load that the unoptimized endpoint can sustain is around 40 RPS.

The load test report generated by Serverless Artillery is also available to us by navigating to CloudWatch in the console, choosing Log groups under Logs, and searching for the log group that has serverless-artillery in its name. If you choose the log group and then choose the most recent log stream, you can see that the last entries comprise of a report that looks similar to the following image. This report’s metrics are an aggregate of the performances of both SageMaker endpoints, so in this case it’s not very useful to us. The one interesting thing to point out is that under the heavier arrival rates, the unoptimized endpoint started to return 400 Status response codes—a sign of it being overwhelmed.

Clean up

With the load test completed and the results analyzed, all that’s left to do is to clean up the deployed resources by running the following two cells. The first cell deletes the two SageMaker endpoints (and their endpoint configurations) that were deployed, and the second cell destroys the Serverless Artillery resources.

# delete endpoints and endpoint configurations sm = boto3.client(‘sagemaker’) for name in [unoptimized_endpoint_name, optimized_endpoint_name]: sm.delete_endpoint(EndpointName=name) sm.delete_endpoint_config(EndpointConfigName=name) !slsart remove –stage dev

After you run the preceding cells, exit this notebook and stop or delete the notebook instance. To stop the notebook instance, on the SageMaker console, choose Notebook instances, select the NeoBlog notebook, and on the Actions menu, choose Stop.


Congratulations! You have successfully finished walking through this post. We were able to accomplish the following:

  • Optimize an XGBoost model artifact generated through a local training job with a Neo compilation job
  • Deploy both versions of the artifact to SageMaker endpoints
  • Deploy Serverless Artillery from our Jupyter notebook and configure the tool so that it directly invokes our SageMaker endpoints
  • Perform load tests against both endpoints with Serverless Artillery
  • Analyze our load test results and view how the Neo-optimized model outperforms the unoptimized model

The performance improvements gained through Neo can translate to significant cost savings. As a next step, you should look at your existing portfolio of models to evaluate them as potential candidates for optimization jobs. Creating Neo-optimized artifact versions allows you to achieve equivalent (if not better) performance metrics with less powerful resources, and it’s one of the easiest ways to save money on SageMaker endpoints.

Additionally, you can apply the load testing approach demonstrated in this post to any SageMaker endpoint. When used in tandem, Serverless Artillery and CloudWatch combine into a powerful framework for profiling the performance characteristics of your endpoints, which can then help you make data-driven decisions on what resource configurations best fit your needs. Simply deploy your models, update your load test definition, and start testing!

For more information about Neo, see Compile and Deploy Models with Neo. For other topics and services related to SageMaker, check out the AWS Machine Learning Blog.

About the Author

Adam Kozdrowicz is a Data and Machine Learning Engineer for AWS Professional Services. He specializes in bringing ML proof of concepts into production and automating the entire ML lifecycle. This includes data collection, data processing, model development and training, model deployments, and model monitoring. He also enjoys working with frameworks such as AWS Amplify, AWS SAM, and AWS CDK. During his free time, Adam likes to surf, travel, practice photography, and build machine learning models.


Continue Reading
Click to comment

Leave a Reply

Your email address will not be published.


Build a risk management machine learning workflow on Amazon SageMaker with no code

Since the global financial crisis, risk management has taken a major role in shaping decision-making for banks, including predicting loan status for potential customers. This is often a data-intensive exercise that requires machine learning (ML). However, not all organizations have the data science resources and expertise to build a risk management ML workflow. Amazon SageMaker…




Since the global financial crisis, risk management has taken a major role in shaping decision-making for banks, including predicting loan status for potential customers. This is often a data-intensive exercise that requires machine learning (ML). However, not all organizations have the data science resources and expertise to build a risk management ML workflow.

Amazon SageMaker is a fully managed ML platform that allows data engineers and business analysts to quickly and easily build, train, and deploy ML models. Data engineers and business analysts can collaborate using the no-code/low-code capabilities of SageMaker. Data engineers can use Amazon SageMaker Data Wrangler to quickly aggregate and prepare data for model building without writing code. Then business analysts can use the visual point-and-click interface of Amazon SageMaker Canvas to generate accurate ML predictions on their own.

In this post, we show how simple it is for data engineers and business analysts to collaborate to build an ML workflow involving data preparation, model building, and inference without writing code.

Solution overview

Although ML development is a complex and iterative process, you can generalize an ML workflow into the data preparation, model development, and model deployment stages.

Data Wrangler and Canvas abstract the complexities of data preparation and model development, so you can focus on delivering value to your business by drawing insights from your data without being an expert in code development. The following architecture diagram highlights the components in a no-code/low-code solution.

Amazon Simple Storage Service (Amazon S3) acts as our data repository for raw data, engineered data, and model artifacts. You can also choose to import data from Amazon Redshift, Amazon Athena, Databricks, and Snowflake.

As data scientists, we then use Data Wrangler for exploratory data analysis and feature engineering. Although Canvas can run feature engineering tasks, feature engineering usually requires some statistical and domain knowledge to enrich a dataset into the right form for model development. Therefore, we give this responsibility to data engineers so they can transform data without writing code with Data Wrangler.

After data preparation, we pass model building responsibilities to data analysts, who can use Canvas to train a model without having to write any code.

Finally, we make single and batch predictions directly within Canvas from the resulting model without having to deploy model endpoints ourselves.

Dataset overview

We use SageMaker features to predict the status of a loan using a modified version of Lending Club’s publicly available loan analysis dataset. The dataset contains loan data for loans issued through 2007–2011. The columns describing the loan and the borrower are our features. The column loan_status is the target variable, which is what we’re trying to predict.

To demonstrate in Data Wrangler, we split the dataset in two CSV files: part one and part two. We’ve removed some columns from Lending Club’s original dataset to simplify the demo. Our dataset contains over 37,000 rows and 21 feature columns, as described in the following table.

Column name Description
loan_status Current status of the loan (target variable).
loan_amount The listed amount of the loan applied for by the borrower. If the credit department reduces the loan amount, it’s reflected in this value.
funded_amount_by_investors The total amount committed by investors for that loan at that time.
term The number of payments on the loan. Values are in months and can be either 36 or 60.
interest_rate Interest rate on the loan.
installment The monthly payment owed by the borrower if the loan originates.
grade LC assigned loan grade.
sub_grade LC assigned loan subgrade.
employment_length Employment length in years. Possible values are between 0–10, where 0 means less than one year and 10 means ten or more years.
home_ownership The home ownership status provided by the borrower during registration. Our values are RENT, OWN, MORTGAGE, and OTHER.
annual_income The self-reported annual income provided by the borrower during registration.
verification_status Indicates if income was verified or not by the LC.
issued_amount The month at which the loan was funded.
purpose A category provided by the borrower for the loan request.
dti A ratio calculated using the borrower’s total monthly debt payments on the total debt obligations, excluding mortgage and the requested LC loan, divided by the borrower’s self-reported monthly income.
earliest_credit_line The month the borrower’s earliest reported credit line was opened.
inquiries_last_6_months The number of inquiries in the past 6 months (excluding auto and mortgage inquiries).
open_credit_lines The number of open credit lines in the borrower’s credit file.
derogatory_public_records The number of derogatory public records.
revolving_line_utilization_rate Revolving line utilization rate, or the amount of credit the borrower is using relative to all available revolving credit.
total_credit_lines The total number of credit lines currently in the borrower’s credit file.

We use this dataset for our data preparation and model training.


Complete the following prerequisite steps:

  1. Upload both loan files to an S3 bucket of your choice.
  2. Make sure you have the necessary permissions. For more information, refer to Get Started with Data Wrangler.
  3. Set up a SageMaker domain configured to use Data Wrangler. For instructions, refer to Onboard to Amazon SageMaker Domain.

Import the data

Create a new Data Wrangler data flow from the Amazon SageMaker Studio UI.

Import data from Amazon S3 by selecting the CSV files from the S3 bucket where you placed your dataset. After you import both files, you can see two separate workflows in the Data flow view.

You can choose several sampling options when importing your data in a Data Wrangler flow. Sampling can help when you have a dataset that is too large to prepare interactively, or when you want to preserve the proportion of rare events in your sampled dataset. Because our dataset is small, we don’t use sampling.

Prepare the data

For our use case, we have two datasets with a common column: id. As a first step in data preparation, we want to combine these files by joining them. For instructions, refer to Transform Data.

We use the Join data transformation step and use the Inner join type on the id column.

As a result of our join transformation, Data Wrangler creates two additional columns: id_0 and id_1. However, these columns are unnecessary for our model building purposes. We drop these redundant columns using the Manage columns transform step.

We’ve imported our datasets, joined them, and removed unnecessary columns. We’re now ready to enrich our data through feature engineering and prepare for model building.

Perform feature engineering

We used Data Wrangler for preparing data. You can also use the Data Quality and Insights Report feature within Data Wrangler to verify your data quality and detect abnormalities in your data. Data scientists often need to use these data insights to efficiently apply the right domain knowledge to engineering features. For this post, we assume we’ve completed these quality assessments and can move on to feature engineering.

In this step, we apply a few transformations to numeric, categorical, and text columns.

We first normalize the interest rate to scale the values between 0–1. We do this using the Process numeric transform to scale the interest_rate column using a min-max scaler. The purpose for normalization (or standardization) is to eliminate bias from our model. Variables that are measured at different scales won’t contribute equally to the model learning process. Therefore, a transformation function like a min-max scaler transform helps normalize features.

To convert a categorial variable into a numeric value, we use one-hot encoding. We choose the Encode categorical transform, then choose One-hot encode. One-hot encoding improves an ML model’s predictive ability. This process converts a categorical value into a new feature by assigning a binary value of 1 or 0 to the feature. As a simple example, if you had one column that held either a value of yes or no, one-hot encoding would convert that column to two columns: a Yes column and a No column. A yes value would have 1 in the Yes column and a 0 in the No column. One-hot encoding makes our data more useful because numeric values can more easily determine a probability for our predictions.

Finally, we featurize the employer_title column to transform its string values into a numerical vector. We apply the Count Vectorizer and a standard tokenizer within the Vectorize transform. Tokenization breaks down a sentence or series of text into words, whereas a vectorizer converts text data into a machine-readable form. These words are represented as vectors.

With all feature engineering steps complete, we can export the data and output the results into our S3 bucket. Alternatively, you can export your flow as Python code, or a Jupyter notebook to create a pipeline with your view using Amazon SageMaker Pipelines. Consider this when you want to run your feature engineering steps at scale or as part of an ML pipeline.

We can now use the Data Wrangler output file as our input for Canvas. We reference this as a dataset in Canvas to build our ML model.

In our case, we exported our prepared dataset to the default Studio bucket with an output prefix. We reference this dataset location when loading the data into Canvas for model building next.

Build and train your ML model with Canvas

On the SageMaker console, launch the Canvas application. To build an ML model from the prepared data in the previous section, we perform the following steps:

  1. Import the prepared dataset to Canvas from the S3 bucket.

We reference the same S3 path where we exported the Data Wrangler results from the previous section.

  1. Create new model in Canvas and name it loan_prediction_model.
  2. Select the imported dataset and add it to the model object.

To have Canvas build a model, we must select the target column.

  1. Because our goal is to predict the probability of a lender’s ability to repay a loan, we choose the loan_status column.

Canvas automatically identifies the type of ML problem statement. At the time of writing, Canvas supports regression, classification, and time series forecasting problems. You can specify the type of problem or have Canvas automatically infer the problem from your data.

  1. Choose your option to start the model building process: Quick build or Standard build.

The Quick build option uses your dataset to train a model within 2–15 minutes. This is useful when you’re experimenting with a new dataset to determine if the dataset you have will be sufficient to make predictions. We use this option for this post.

The Standard build option choses accuracy over speed and uses approximately 250 model candidates to train the model. The process usually takes 1–2 hours.

After the model is built, you can review the results of the model. Canvas estimates that your model is able to predict the right outcome 82.9% of the time. Your own results may vary due to the variability in training models.

In addition, you can dive deep into details analysis of the model to learn more about the model.

Feature importance represents the estimated importance of each feature in predicting the target column. In this case, the credit line column has the most significant impact in predicting if a customer will pay back the loan amount, followed by interest rate and annual income.

The confusion matrix in the Advanced metrics section contains information for users that want a deeper understanding of their model performance.

Before you can deploy your model for production workloads, use Canvas to test the model. Canvas manages our model endpoint and allows us to make predictions directly in the Canvas user interface.

  1. Choose Predict and review the findings on either the Batch prediction or Single prediction tab.

In the following example, we make a single prediction by modifying values to predict our target variable loan_status in real time

We can also select a larger dataset and have Canvas generate batch predictions on our behalf.


End-to-end machine learning is complex and iterative, and often involves multiple personas, technologies, and processes. Data Wrangler and Canvas enable collaboration between teams without requiring these teams to write any code.

A data engineer can easily prepare data using Data Wrangler without writing any code and pass the prepared dataset to a business analyst. A business analyst can then easily build accurate ML models with just a few click using Canvas and get accurate predictions in real time or in batch.

Get started with Data Wrangler using these tools without having to manage any infrastructure. You can set up Canvas quickly and immediately start creating ML models to support your business needs.

About the Authors

Peter Chung is a Solutions Architect for AWS, and is passionate about helping customers uncover insights from their data. He has been building solutions to help organizations make data-driven decisions in both the public and private sectors. He holds all AWS certifications as well as two GCP certifications.

 Meenakshisundaram Thandavarayan is a Senior AI/ML specialist with AWS. He helps hi-tech strategic accounts on their AI and ML journey. He is very passionate about data-driven AI.

Dan Ferguson is a Solutions Architect at AWS, based in New York, USA. As a machine learning services expert, Dan works to support customers on their journey to integrating ML workflows efficiently, effectively, and sustainably.


Continue Reading


Optimize F1 aerodynamic geometries via Design of Experiments and machine learning

FORMULA 1 (F1) cars are the fastest regulated road-course racing vehicles in the world. Although these open-wheel automobiles are only 20–30 kilometers (or 12–18 miles) per-hour faster than top-of-the-line sports cars, they can speed around corners up to five times as fast due to the powerful aerodynamic downforce they create. Downforce is the vertical force…




FORMULA 1 (F1) cars are the fastest regulated road-course racing vehicles in the world. Although these open-wheel automobiles are only 20–30 kilometers (or 12–18 miles) per-hour faster than top-of-the-line sports cars, they can speed around corners up to five times as fast due to the powerful aerodynamic downforce they create. Downforce is the vertical force generated by the aerodynamic surfaces that presses the car towards the road, increasing the grip from the tires. F1 aerodynamicists must also monitor the air resistance or drag, which limits straight-line speed.

The F1 engineering team is in charge of designing the next generation of F1 cars and putting together the technical regulation for the sport. Over the last 3 years, they have been tasked with designing a car that maintains the current high levels of downforce and peak speeds, but is also not adversely affected by driving behind another car. This is important because the previous generation of cars can lose up to 50% of their downforce when racing closely behind another car due to the turbulent wake generated by wings and bodywork.

Instead of relying on time-consuming and costly track or wind tunnel tests, F1 uses Computational Fluid Dynamics (CFD), which provides a virtual environment to study the flow of fluids (in this case the air around the F1 car) without ever having to manufacture a single part. With CFD, F1 aerodynamicists test different geometry concepts, assess their aerodynamic impact, and iteratively optimize their designs. Over the past 3 years, the F1 engineering team has collaborated with AWS to set up a scalable and cost-efficient CFD workflow that has tripled the throughput of CFD runs and cut the turnaround time per run by half.

F1 is in the process of looking into AWS machine learning (ML) services such as Amazon SageMaker to help optimize the design and performance of the car by using the CFD simulation data to build models with additional insights. The aim is to uncover promising design directions and reduce the number of CFD simulations, thereby reducing the time taken to converge to optimal designs.

In this post, we explain how F1 collaborated with the AWS Professional Services team to develop a bespoke Design of Experiments (DoE) workflow powered by ML to advise F1 aerodynamicists on which design concepts to test in CFD to maximize learning and performance.

Problem statement

When exploring new aerodynamic concepts, F1 aerodynamicists sometimes employ a process called Design of Experiments (DoE). This process systematically studies the relationship between multiple factors. In the case of a rear wing, this might be wing chord, span, or camber, with respect to aerodynamic metrics such as downforce or drag. The goal of a DoE process is to efficiently sample the design space and minimize the number of candidates tested before converging to an optimal result. This is achieved by iteratively changing multiple design factors, measuring the aerodynamic response, studying the impact and relationship between factors, and then continuing testing in the most optimum or informative direction. In the following figure, we present an example rear wing geometry that F1 has kindly shared with us from their UNIFORM baseline. Four design parameters which F1 aerodynamicists could investigate in a DoE routine are labeled.

In this project, F1 worked with AWS Professional Services to investigate using ML to enhance DoE routines. Traditional DoE methods require a well-populated design space in order to understand the relationship between design parameters and therefore rely on a large number of upfront CFD simulations. ML regression models could use the results from previous CFD simulations to predict the aerodynamic response given the set of design parameters, as well as give you an indication of the relative importance of each design variable. You could use these insights to predict optimal designs and help designers converge to optimum solutions with fewer upfront CFD simulations. Secondly, you could use data science techniques to understand which regions in the design space haven’t been explored and could potentially hide optimal designs.

To illustrate the bespoke ML-powered DoE workflow, we walk through a real example of designing a front wing.

Designing a front wing

F1 cars rely on wings such as the front and rear wings to generate most of their downforce, which we refer to throughout this example by the coefficient Cz. Throughout this example, the downforce values have been normalized. In this example, F1 aerodynamicists used their domain expertise to parameterize the wing geometry as follows (refer to the following figure for a visual representation):

  • LE-Height – Leading edge height
  • Min-Z – Minimum ground clearance
  • Mid-LE-Angle – Leading edge angle of the third element
  • TE-Angle – Trailing edge angle
  • TE-Height – Trailing edge height

This front wing geometry was shared by F1 and is part of the UNIFORM baseline.

These parameters were selected because they are sufficient to describe the main aspects of the geometry efficiently and because in the past, aerodynamic performance has shown notable sensitivity with respect to these parameters. The goal of this DoE routine was to find the combination of the five design parameters that would maximize aerodynamic downforce (Cz). The design freedom is also limited by setting maximum and minimum values to the design parameters, as shown in the following table.

. Minimum Maximum
TE-Height 250.0 300.0
TE-Angle 145.0 165.0
Mid-LE-Angle 160.0 170.0
Min-Z 5.0 50.0
LE-Height 100.0 150.0

Having established the design parameters, the target output metric, and the bounds of our design space, we have all we need to get started with the DoE routine. A workflow diagram of our solution is presented in the following image. In the following section, we dive deep into the different stages.

Initial sampling of the design space

The first step of the DoE workflow is to run in CFD an initial set of candidates that efficiently sample the design space and allow us to build the first set of ML regression models to study the influence of each feature. First, we generate a pool of N samples using Latin Hypercube Sampling (LHS) or a regular grid method. Then, we select k candidates to test in CFD by means of a greedy inputs algorithm, which aims to maximize the exploration of the design space. Starting with a baseline candidate (the current design), we iteratively select candidates furthest away from all the previously tested candidates. Suppose that we already tested k designs; for the remaining design candidates, we find the minimum distance d with respect to the tested k designs:

The greedy inputs algorithm selects the candidate that maximizes the distance in the feature space to the previously tested candidates:

In this DoE, we selected three greedy inputs candidates and ran those in CFD to assess their aerodynamic downforce (Cz). The greedy inputs candidates explore the bounds of the design space and at this stage, none of them proved superior to the baseline candidate in terms of aerodynamic downforce (Cz). The results of this initial round of CFD testing together with the design parameters are displayed in the following table.

. TE-Height TE-Angle Mid-LE-Angle Min-Z LE-Height Normalized Cz
Baseline 292.25 154.86 166 5 130 0.975
GI 0 250 165 160 50 100 0.795
GI 1 300 145 170 50 100 0.909
GI 2 250 145 170 5 100 0.847

Initial ML regression models

The goal of the regression model is to predict Cz for any combination of the five design parameters. With such a small dataset, we prioritized simple models, applied model regularization to avoid overfitting, and combined the predictions of different models where possible. The following ML models were constructed:

  • Ordinary Least Squares (OLS)
  • Support Vector Regression (SVM) with an RBF kernel
  • Gaussian Process Regression (GP) with a Matérn kernel
  • XGBoost

In addition, a two-level stacked model was built, where the predictions of the GP, SVM, and XGBoost models are assimilated by a Lasso algorithm to produce the final response. This model is referred to throughout this post as the stacked model. To rank the predictive capabilities of the five models we described, a repeated k-fold cross validation routine was implemented.

Generating the next design candidate to test in CFD

Selecting which candidate to test next requires careful consideration. The F1 aerodynamicist must balance the benefit of exploiting options predicted by the ML model to provide high downforce with the cost of failing to explore uncharted regions of the design space, which may provide even higher downforce. For that reason, in this DoE routine, we propose three candidates: one performance-driven and two exploration-driven. The purpose of the exploration-driven candidates is also to provide additional data points to the ML algorithm in regions of the design space where the uncertainty around the prediction is highest. This in turn leads to more accurate predictions in the next round of design iteration.

Genetic algorithm optimization to maximize downforce

To obtain the candidate with the highest expected aerodynamic downforce, we could run a prediction over all possible design candidates. However, this wouldn’t be efficient. For this optimization problem, we use a genetic algorithm (GA). The goal is to efficiently search through a huge solution space (obtained via the ML prediction of Cz) and return the most optimal candidate. GAs are advantageous when the solution space is complex and non-convex, so that classical optimization methods such as gradient descent are an ineffective means to find a global solution. GA is a subset of evolutionary algorithms and inspired by concepts from natural selection, genetic crossover, and mutation to solve the search problem. Over a series of iterations (known as generations), the best candidates of an initially randomly selected set of design candidates are combined (much like reproduction). Eventually, this mechanism allows you to find the most optimal candidates in an efficient manner. For more information about GAs, refer to Using genetic algorithms on AWS for optimization problems.

Generating exploration-driven candidates

In generating what we term exploration-driven candidates, a good sampling strategy must be able to adapt to a situation of effect sparsity, where only a subset of the parameters significantly affects the solution. Therefore, the sampling strategy should spread out the candidates across the input design space but also avoid unnecessary CFD runs, changing variables that have little effect on performance. The sampling strategy must take into account the response surface predicted by the ML regressor. Two sampling strategies were employed to obtain exploration-driven candidates.

In the case of Gaussian Process Regressors (GP), the standard deviation of the predicted response surface can be used as an indication of the uncertainty of the model. The sampling strategy consists of selecting out of the pool of N samples , the candidate that maximizes . By doing so, we’re sampling in the region of the design space where the regressor is least confident about its prediction. In mathematical terms, we select the candidate that satisfies the following equation:

Alternatively, we employ a greedy inputs and outputs sampling strategy, which maximizes both the distances in the feature space and in the response space between the proposed candidate and the already tested designs. This tackles the effect sparsity situation because candidates that modify a design parameter of little relevance have a similar response, and therefore the distances in the response surface are minimal. In mathematical terms, we select the candidate that satisfies the following equation, where the function f is the ML regression model:

Candidate selection, CFD testing, and optimization loop

At this stage, the user is presented with both performance-driven and exploration-driven candidates. The next step consists of selecting a subset of the proposed candidates, running CFD simulations with those design parameters, and recording the aerodynamic downforce response.

After this, the DoE workflow retrains the ML regression models, runs the genetic algorithm optimization, and proposes a new set of performance-driven and exploration-driven candidates. The user runs a subset of the proposed candidates and continues iterating in this fashion until the stopping criteria is met. The stopping criteria is generally met when a candidate deemed optimum is obtained.


In the following figure, we record the normalized aerodynamic downforce (Cz) from the CFD simulation (blue) and the one predicted beforehand using the ML regression model of choice (pink) for each iteration of the DoE workflow. The goal was to maximize aerodynamic downforce (Cz). The first four runs (to the left of the red line) were the baseline and the three greedy inputs candidates outlined previously. From there on, a combination of performance-driven and exploration-driven candidates were tested. In particular, the candidates at iterations 6 and 8 were exploratory candidates, both showing lower levels of downforce than the baseline candidate (iteration 1). As expected, as we recorded more candidates, the ML prediction became increasingly accurate, as denoted by the decreasing distance between the predicted and actual Cz. At iteration 9, the DoE workflow managed to find a candidate with a similar performance to the baseline, and at iteration 12, the DoE workflow was concluded when the performance-driven candidate surpassed the baseline.

The final design parameters together with the resultant normalized downforce value is presented in the following table. The normalized downforce level for the baseline candidate was 0.975, whereas the optimum candidate for the DoE workflow recorded a normalized downforce level of 1.000. This is an important 2.5% relative increase.

For context, a traditional DoE approach with five variables would require 25 upfront CFD simulations before achieving a good enough fit to predict an optimum. On the other hand, this active learning approach converged to an optimum in 12 iterations.

. TE-Height TE-Angle Mid-LE-Angle Min-Z LE-Height Normalized Cz
Baseline 292.25 154.86 166 5 130 0.975
Optimal 299.97 156.79 166.27 5.01 135.26 1.000

Feature importance

Understanding the relative feature importance for a predictive model can provide a useful insight into the data. It can help feature selection with less important variables being removed, thereby reducing the dimensionality of the problem and potentially improving the predictive powers of the regression model, particularly in the small data regime. In this design problem, it provides F1 aerodynamicists an insight into which variables are the most sensitive and therefore require more careful tuning.

In this routine, we implemented a model-agnostic technique called permutation importance. The relative importance of each variable is measured by calculating the increase in the model’s prediction error after randomly shuffling the values for that variable alone. If a feature is important for the model, the prediction error increases greatly, and vice versa for lesser important features. In the following figure, we present the permutation importance for a Gaussian Process Regressor (GP) predicting aerodynamic downforce (Cz). The trailing edge height (TE-Height) was deemed the most important.


In this post, we explained how F1 aerodynamicists are using ML regression models in DoE workflows when designing novel aerodynamic geometries. The ML-powered DoE workflow developed by AWS Professional Services provides insights into which design parameters will maximize performance or explore uncharted regions in the design space. As opposed to iteratively testing candidates in CFD in a grid search fashion, the ML-powered DoE workflow is able to converge to optimal design parameters in fewer iterations. This saves both time and resources because fewer CFD simulations are required.

Whether you’re a pharmaceutical company looking to speed up chemical composition optimization or a manufacturing company looking to find the design dimensions for the most robust designs, DoE workflows can help reach optimal candidates more efficiently. AWS Professional Services is ready to supplement your team with specialized ML skills and experience to develop the tools to streamline DoE workflows and help you achieve better business outcomes. For more information, see AWS Professional Services, or reach out through your account manager to get in touch.

About the Authors

Pablo Hermoso Moreno is a Data Scientist in the AWS Professional Services Team. He works with clients across industries using Machine Learning to tell stories with data and reach more informed engineering decisions faster. Pablo’s background is in Aerospace Engineering and having worked in the motorsport industry he has an interest in bridging physics and domain expertise with ML. In his spare time, he enjoys rowing and playing guitar.


Continue Reading


Detect social media fake news using graph machine learning with Amazon Neptune ML

In recent years, social media has become a common means for sharing and consuming news. However, the spread of misinformation and fake news on these platforms has posed a major challenge to the well-being of individuals and societies. Therefore, it is imperative that we develop robust and automated solutions for early detection of fake news…




In recent years, social media has become a common means for sharing and consuming news. However, the spread of misinformation and fake news on these platforms has posed a major challenge to the well-being of individuals and societies. Therefore, it is imperative that we develop robust and automated solutions for early detection of fake news on social media. Traditional approaches rely purely on the news content (using natural language processing) to mark information as real or fake. However, the social context in which the news is published and shared can provide additional insights into the nature of fake news on social media and improve the predictive capabilities of fake news detection tools. In this post, we demonstrate how to use Amazon Neptune ML to detect fake news based on the content and social context of the news on social media.

Neptune ML is a new capability of Amazon Neptune that uses graph neural networks (GNNs), a machine learning (ML) technique purpose-built for graphs, to make easy, fast, and accurate predictions using graph data. Making accurate predictions on graphs with billions of relationships requires expertise. Existing ML approaches such as XGBoost can’t operate effectively on graphs because they’re designed for tabular data. As a result, using these methods on graphs can take time, require specialized skills, and produce suboptimal predictions.

Neptune ML uses the Deep Graph Library (DGL), an open-source library to which AWS contributes, and Amazon SageMaker to build and train GNNs, including Relational Graph Convolutional Networks (R-GCNs) for tasks such as node classification, node regression, link prediction, or edge classification.

The DGL makes it easy to apply deep learning to graph data, and Neptune ML automates the heavy lifting of selecting and training the best ML model for graph data. It provides fast and memory-efficient message passing primitives for training GNNs. Neptune ML uses the DGL to automatically choose and train the best ML model for your workload. This enables you to make ML-based predictions on graph data in hours instead of weeks. For more information, see Amazon Neptune ML for machine learning on graphs.

Amazon SageMaker is a fully managed service that provides every developer and data scientist with the ability to prepare, build, train, and deploy ML models quickly.

Overview of GNNs

GNNs are neural networks that take graphs as input. These models operate on the relational information in data to produce insights not possible in other neural network architectures and algorithms. A graph (sometimes called a network) is a data structure that highlights the relationships between components in the data. It consists of nodes (or vertices) and edges (or links) that act as connections between the nodes. Such a data structure has an advantage when dealing with entities that have multiple relationships. Graph data structures have been around for centuries, with a wide variety of modern use cases.

GNNs are emerging as an important class of deep learning (DL) models. GNNs learn embeddings on nodes, edges, and graphs. GNNs have been around for about 20 years, but interest in them has dramatically increased in the last 5 years. In this time, we’ve seen new architectures emerge, novel applications realized, and new platforms and libraries enter the scene. There are several potential research and industry use cases for GNNs, including the following:

  • Computer vision – Generating scene graphs
  • Forecasting – Predicting traffic volume
  • Node classification – Implementing targeted campaigns, detecting fake news
  • Graph classification – Predicting the properties of a chemical compound
  • Link prediction – Building recommendation systems
  • Other – Predicting adversarial attacks


For this post, we use the BuzzFeed dataset from the 2018 version of FakeNewsNet. The BuzzFeed dataset consists of a sample of news articles shared on Facebook from nine news agencies over 1 week leading up to the 2016 US election. Every post and the corresponding news article have been fact-checked by BuzzFeed journalists. The following table summarizes key statistics about the BuzzFeed dataset from FakeNewsNet.

Category Amount
Users 15,257
Authors 126
Publishers 28
Social Links 634,750
Engagements 25,240
News Articles 182
Fake News 91
Real News 91

To get the raw data, you can complete the following steps:

  1. Clone the FakeNewsNet repository from GitHub.
  2. Check out the old version branch.
  3. Change the directory to Data/BuzzFeed.

Each row in the Users.txt file provides a UUID for the corresponding user.

Each row in the News.txt file provides a name and ID for the corresponding news in the dataset.

In the BuzzFeedNewsUser.txt file, the news_id in the first column is posted or shared by the user_id in the second column n times, where n is the value in the third column.

In the BuzzFeedUserUser.txt file, the user_id in the first column follows the user_id in the second column.

User features such as age, gender, and historical social media activities (109,626 features for each user) are made available in UserFeature.mat file. Sample news content files, shown in the following screenshot, contain information such as news title, news text, author name, and publisher web address.

We processed the raw data from the FakeNewsNet repository and converted it into CSV format for vertices and edges in a heterogeneous property graph that can be readily loaded into a Neptune database with Apache TinkerPop Gremlin. The constructed property graph is composed of four vertex types and five edge types, as demonstrated in the following schematic, which together describe the social context in which each news item is published and shared. The News vertices have two properties: news_title and news_type (Fake or Real). The edges connecting News and User vertices have a weight property describing how many times the user has shared the news. The User vertices have a 100-dimension property representing user features such as age, gender, and historical social media activities (reduced from 109,626 to 100 using principal coordinate analysis).

The following screenshot shows the first 10 rows of the processed nodes.csv file.

The following screenshot shows the first 10 rows of the processed edges.csv file.

To follow along with this post, start by using the following AWS CloudFormation quick-start template to quickly spin up an associated Neptune cluster and AWS graph notebook, and set up all the configurations needed to work with Neptune ML in a graph notebook. You then need to download and save the sample dataset in the default Amazon Simple Storage Service (Amazon S3) bucket associated with your SageMaker session, or in an S3 bucket of your choice. For rapid experimentation and initial data exploration, you can save a copy of the dataset under the home directory of the local volume attached to your SageMaker notebook instance, and follow the create_graph_dataset.ipynb Jupyter notebook. After you generate the processed nodes and edges files, you can run the following commands to upload the transformed graph data to Amazon S3:

bucket = ‘‘ prefix = ‘fake-news-detection/data’ s3_client = boto3.client(‘s3’) resp = s3_client.upload_file(‘./Data/upload/nodes.csv’, bucket, f”{prefix}/nodes.csv”) resp = s3_client.upload_file(‘./Data/upload/edges.csv’, bucket, f”{prefix}/edges.csv”)

You can use the %load magic command, which is available as part of the AWS graph notebook, to bulk load data to Neptune:

%load -s {s3_uri} -f csv -p OVERSUBSCRIBE –run

You can use the %graph_notebook_config magic command to see information about the Neptune cluster associated with your graph notebook. You can also use the %status magic command to see the status of your Neptune cluster, as shown in the following screenshot.

Solution overview

Neptune ML uses graph neural network technology to automatically create, train, and deploy ML models on your graph data. Neptune ML supports common graph prediction tasks, such as node classification and regression, edge classification and regression, and link prediction. In our solution, we use node classification to classify news nodes according to the news_type property.

The following diagram illustrates the high-level process flow to develop the best model for fake news detection.

Graph ML with Neptune ML involves five main steps:

  1. Export and configure the data – The data export step uses the Neptune-Export service to export data from Neptune into Amazon S3 in CSV format. A configuration file named training-data-configuration.json is automatically generated, which specifies how the exported data can be loaded into a trainable graph.
  2. Preprocess the data – The exported dataset is preprocessed using standard techniques to prepare it for model training. Feature normalization can be performed for numeric data, and text features can be encoded using word2vec. At the end of this step, a DGL graph is generated from the exported dataset for the model training step. This step is implemented using a SageMaker processing job, and the resulting data is stored in an Amazon S3 location that you have specified.
  3. Train the model – This step trains the ML model that will be used for predictions. Model training is done in two stages:
    1. The first stage uses a SageMaker processing job to generate a model training strategy configuration set that specifies what type of model and model hyperparameter ranges are used for the model training.
    2. The second stage uses a SageMaker model tuning job to try different hyperparameter configurations and select the training job that produced the best-performing model. The tuning job runs a pre-specified number of model training job trials on the processed data. At the end of this stage, the trained model parameters of the best training job are used to generate model artifacts for inference.
  4. Create an inference endpoint in SageMaker – The inference endpoint is a SageMaker endpoint instance that is launched with the model artifacts produced by the best training job. The endpoint is able to accept incoming requests from the graph database and return the model predictions for inputs in the requests.
  5. Query the ML model using Gremlin – You can use extensions to the Gremlin query language to query predictions from the inference endpoint.

Before we proceed with the first step of machine learning, let’s verify that the graph dataset is loaded in the Neptune cluster. Run the following Gremlin traversal to see the count of nodes by label:

%%gremlin g.V().groupCount().by(label).unfold().order().by(keys)

If nodes are loaded correctly, the output is as follows:

  • 126 author nodes
  • 182 news nodes
  • 28 publisher nodes
  • 15,257 user nodes

Use the following code to see the count edges by label:

%%gremlin g.E().groupCount().by(label).unfold().order().by(keys)

If edges are loaded correctly, the output is as follows:

  • 634,750 follows edges
  • 174 published edges
  • 250 wrote edges
  • 250 wrote_for edges

Now let’s go through the ML development process in detail.

Export and configure the data

The export process is triggered by calling to the Neptune-Export service endpoint. This call contains a configuration object that specifies the type of ML model to build, in our case node classification, as well as any feature configurations required.

The configuration options provided to the Neptune-Export service are broken into two main sections: selecting the target and configuring features. Here we want to classify news nodes according to the news_type property.

The second section of the configuration, configuring features, is where we specify details about the types of data stored in our graph and how the ML model should interpret that data. When data is exported from Neptune, all properties of all nodes are included. Each property is treated as a separate feature for the ML model. Neptune ML does its best to infer the correct type of feature for a property, but in many cases, the accuracy of the model can be improved by specifying information about the property used for a feature. We use word2vec to encode the news_title property of news nodes, and the numerical type for user_features property of user nodes. See the following code:

export_params={ “command”: “export-pg”, “params”: { “endpoint”: neptune_ml.get_host(), “profile”: “neptune_ml”, “useIamAuth”: neptune_ml.get_iam(), “cloneCluster”: False }, “outputS3Path”: f”{s3_uri}/neptune-export”, “additionalParams”: { “neptune_ml”: { “version”: “v2.0”, “targets”: [ { “node”: “news”, “property”: “news_type”, “type”: “classification” } ], “features”: [ { “node”: “news”, “property”: “news_title”, “type”: “text_word2vec” }, { “node”: “user”, “property”: “user_features”, “type”: “numerical” } ] } }, “jobSize”: “medium”}

Start the export process by running the following command:

%%neptune_ml export start –export-url {neptune_ml.get_export_service_host()} –export-iam –wait –store-to export_results ${export_params}

Preprocess the data

When the export job is complete, we’re ready to train our ML model. There are three machine learning steps in Neptune ML. The first step (data processing) processes the exported graph dataset using standard feature preprocessing techniques to prepare it for use by the DGL. This step performs functions such as feature normalization for numeric data and encoding text features using word2vec. At the conclusion of this step, the dataset is formatted for model training. This step is implemented using a SageMaker processing job, and data artifacts are stored in a pre-specified Amazon S3 location when the job is complete. Run the following code to create the data processing configuration and begin the processing job:

# The training_job_name can be set to a unique value below, otherwise one will be auto generated training_job_name=neptune_ml.get_training_job_name(‘fake-news-detection’) processing_params = f””” –config-file-name training-data-configuration.json –job-id {training_job_name} –s3-input-uri {export_results[‘outputS3Uri’]} –s3-processed-uri {str(s3_uri)}/preloading “””

Train the model

Now that you have the data processed in the desired format, this step trains the ML model that is used for predictions. The model training is done in two stages. The first stage uses a SageMaker processing job to generate a model training strategy. A model training strategy is a configuration set that specifies what type of model and model hyperparameter ranges are used for the model training. After the first stage is complete, the SageMaker processing job launches a SageMaker hyperparameter tuning job. The hyperparameter tuning job runs a pre-specified number of model training job trials on the processed data, and stores the model artifacts generated by the training in the output Amazon S3 location. When all the training jobs are complete, the hyperparameter tuning job also notes the training job that produced the best performing model.

We use the following training parameters:

training_params=f””” –job-id {training_job_name} –data-processing-id {training_job_name} –instance-type ml.c5.18xlarge –s3-output-uri {str(s3_uri)}/training –max-hpo-number 20 –max-hpo-parallel 4 “””

The hyperparameter tuning finds the best version of a model by running many training jobs on the dataset. You can summarize hyperparameters of the five best training jobs and their respective model performance as follows:

tuning_job_name = training_results[‘hpoJob’][‘name’] tuner = sagemaker.HyperparameterTuningJobAnalytics(tuning_job_name) full_df = tuner.dataframe() if len(full_df) > 0: df = full_df[full_df[“FinalObjectiveValue”] > -float(“inf”)] if len(df) > 0: df = df.sort_values(“FinalObjectiveValue”, ascending=False) print(“Number of training jobs with valid objective: %d” % len(df)) print({“lowest”: min(df[“FinalObjectiveValue”]), “highest”: max(df[“FinalObjectiveValue”])}) pd.set_option(“display.max_colwidth”, None) # Don’t truncate TrainingJobName else: print(“No training jobs have reported valid results yet.”)

We can see that the best performing training job achieved an accuracy of approximately 94%. This training job will be automatically selected by Neptune ML for creating an endpoint in the next step.

Create an endpoint

The final step of machine learning is to create an inference endpoint, which is a SageMaker endpoint instance that is launched with the model artifacts produced by the best training job. We use this endpoint in our graph queries to return the model predictions for the inputs in the request. After the endpoint is created, it stays active until it’s manually deleted. Create the endpoint with the following code:

endpoint_params=f””” –id {training_job_name} –model-training-job-id {training_job_name} “”” #Create endpoint %neptune_ml endpoint create –wait –store-to endpoint_results {endpoint_params}

Our new endpoint is now up and running.

Query the ML model

Now let’s query your trained graph to see how the model predicts news_type for one unseen news node:

# Random fake news: test node: Actual %%gremlin g.V().has(‘news_title’, ‘BREAKING: Steps to FORCE FBI Director Comey to Resign In Process – Hearing Decides His Fate Sept 28’).properties(“news_type”).value() # Random fake news: test node: Predicted %%gremlin g.with(“Neptune#ml.endpoint”, “${endpoint}”). V().has(‘news_title’, “BREAKING: Steps to FORCE FBI Director Comey to Resign In Process – Hearing Decides His Fate Sept 28”).properties(“news_type”).with(“Neptune#ml.classification”).value()

If your graph is continuously changing, you may need to update ML predictions frequently using the newest data. Although you can do this simply by rerunning the earlier steps (from data export and configuration to creating your inference endpoint), Neptune ML supports simpler ways to update your ML predictions using new data. See Workflows for handling evolving graph data for more details.


In this post, we showed how Neptune ML and GNNs can help detect social media fake news using node classification on graph data by combining information from the complex interaction patterns in the graph. For instructions on implementing this solution, see the GitHub repo. You can also clone and extend this solution with additional data sources for model retraining and tuning. We encourage you to reach out and discuss your use cases with the authors via your AWS account manager.

Additional references

For more information related to Neptune ML and detecting fake news in social media, see the following resources:

About the Authors

Hasan Shojaei is a Data Scientist with AWS Professional Services, where he helps customers across different industries such as sports, insurance, and financial services solve their business challenges through the use of big data, machine learning, and cloud technologies. Prior to this role, Hasan led multiple initiatives to develop novel physics-based and data-driven modeling techniques for top energy companies. Outside of work, Hasan is passionate about books, hiking, photography, and ancient history.

Sarita Joshi is a Senior Data Science Manager with the AWS Professional Services Intelligence team. Together with her team, Sarita plays a strategic role for our customers and partners by helping them achieve their business outcomes through machine learning and artificial intelligence solutions at scale. She has several years of experience as a consultant advising clients across many industries and technical domains, including AI, ML, analytics, and SAP. She holds a master’s degree in Computer Science, Specialty Data Science from Northeastern University.


Continue Reading


Copyright © 2021 Today's Digital.