Connect with us


Text classification for online conversations with machine learning on AWS

Online conversations are ubiquitous in modern life, spanning industries from video games to telecommunications. This has led to an exponential growth in the amount of online conversation data, which has helped in the development of state-of-the-art natural language processing (NLP) systems like chatbots and natural language generation (NLG) models. Over time, various NLP techniques for…



Online conversations are ubiquitous in modern life, spanning industries from video games to telecommunications. This has led to an exponential growth in the amount of online conversation data, which has helped in the development of state-of-the-art natural language processing (NLP) systems like chatbots and natural language generation (NLG) models. Over time, various NLP techniques for text analysis have also evolved. This necessitates the requirement for a fully managed service that can be integrated into applications using API calls without the need for extensive machine learning (ML) expertise. AWS offers pre-trained AWS AI services like Amazon Comprehend, which can effectively handle NLP use cases involving classification, text summarization, entity recognition, and more to gather insights from text.

Additionally, online conversations have led to a wide-spread phenomenon of non-traditional usage of language. Traditional NLP techniques often perform poorly on this text data due to the constantly evolving and domain-specific vocabularies that exist within different platforms, as well as the significant lexical deviations of words from proper English, either by accident or intentionally as a form of adversarial attack.

In this post, we describe multiple ML approaches for text classification of online conversations with tools and services available on AWS.


Before diving deep into this use case, please complete the following prerequisites:

  1. Set up an AWS account and create an IAM user.
  2. Set up the AWS CLI and AWS SDKs.
  3. (Optional) Set up your Cloud9 IDE environment.


For this post, we use the Jigsaw Unintended Bias in Toxicity Classification dataset, a benchmark for the specific problem of classification of toxicity in online conversations. The dataset provides toxicity labels as well as several subgroup attributes such as obscene, identity attack, insult, threat, and sexually explicit. Labels are provided as fractional values, which represent the proportion of human annotators who believed the attribute applied to a given piece of text, which are rarely unanimous. To generate binary labels (for example, toxic or non-toxic), a threshold of 0.5 is applied to the fractional values, and comments with values greater than the threshold are treated as the positive class for that label.

Subword embedding and RNNs

For our first modeling approach, we use a combination of subword embedding and recurrent neural networks (RNNs) to train text classification models. Subword embeddings were introduced by Bojanowski et al. in 2017 as an improvement upon previous word-level embedding methods. Traditional Word2Vec skip-gram models are trained to learn a static vector representation of a target word that optimally predicts that word’s context. Subword models, on the other hand, represent each target word as a bag of the character n-grams that make up the word, where an n-gram is composed of a set of n consecutive characters. This method allows for the embedding model to better represent the underlying morphology of related words in the corpus as well as the computation of embeddings for novel, out-of-vocabulary (OOV) words. This is particularly important in the context of online conversations, a problem space in which users often misspell words (sometimes intentionally to evade detection) and also use a unique, constantly evolving vocabulary that might not be captured by a general training corpus.

Amazon SageMaker makes it easy to train and optimize an unsupervised subword embedding model on your own corpus of domain-specific text data with the built-in BlazingText algorithm. We can also download existing general-purpose models trained on large datasets of online text, such as the following English language models available directly from fastText. From your SageMaker notebook instance, simply run the following to download a pretrained fastText model:

!wget -O

Whether you’ve trained your own embeddings with BlazingText or downloaded a pretrained model, the result is a zipped model binary that you can use with the gensim library to embed a given target word as a vector based on its constituent subwords:

# Imports import os from zipfile import ZipFile from gensim.models.fasttext import load_facebook_vectors # Unzip the model binary into ‘dir_path’ with ZipFile(‘’, ‘r’) as zipObj: zipObj.extractall(path=) # Load embedding model into memory embed_model = load_facebook_vectors(os.path.join(, ‘vectors.bin’)) # Compute embedding vector for ‘word’ word_embedding = embed_model[word]

After we preprocess a given segment of text, we can use this approach to generate a vector representation for each of the constituent words (as separated by spaces). We then use SageMaker and a deep learning framework such as PyTorch to train a customized RNN with a binary or multilabel classification objective to predict whether the text is toxic or not and the specific sub-type of toxicity based on labeled training examples.

To upload your preprocessed text to Amazon Simple Storage Service (Amazon S3), use the following code:

import boto3 s3 = boto3.client(‘s3’) bucket = prefix = s3.upload_file(‘train.pkl’, bucket, os.path.join(prefix, ‘train/train.pkl’)) s3.upload_file(‘valid.pkl’, bucket, os.path.join(prefix, ‘valid/valid.pkl’)) s3.upload_file(‘test.pkl’, bucket, os.path.join(prefix, ‘test/test.pkl’))

To initiate scalable, multi-GPU model training with SageMaker, enter the following code:

import sagemaker sess = sagemaker.Session() role = iam.get_role(RoleName= ‘AmazonSageMakerFullAccess’)[‘Role’][‘Arn’] from sagemaker.pytorch import PyTorch # hyperparameters, which are passed into the training job hyperparameters = { ‘epochs’: 20, # Maximum number of epochs to train model ‘train-batch-size’: 128, # Training batch size (No. sentences) ‘eval-batch-size’: 1024, # Evaluation batch size (No. sentences) ’embed-size’: 300, # Vector dimension of word embeddings (Must match embedding model) ‘lstm-hidden-size’: 200, # Number of neurons in LSTM hidden layer ‘lstm-num-layers’: 2, # Number of stacked LSTM layers ‘proj-size’: 100, # Number of neurons in intermediate projection layer ‘num-targets’: len(), # Number of targets for classification ‘class-weight’: ‘ ‘.join([str(c) for c in ]), # Weight to apply to each target during training ‘total-length’:, ‘metric-for-best-model’: ‘ap_score_weighted’, # Metric on which to select the best model } # create the Estimator pytorch_estimator = PyTorch( entry_point=’’, source_dir=, instance_type=, volume_size=200, instance_count=1, role=role, framework_version=’1.6.0’, py_version=’py36′, hyperparameters=hyperparameters, metric_definitions=[ {‘Name’: ‘validation:accuracy’, ‘Regex’: ‘eval_accuracy = (.*?);’}, {‘Name’: ‘validation:f1-micro’, ‘Regex’: ‘eval_f1_score_micro = (.*?);’}, {‘Name’: ‘validation:f1-macro’, ‘Regex’: ‘eval_f1_score_macro = (.*?);’}, {‘Name’: ‘validation:f1-weighted’, ‘Regex’: ‘eval_f1_score_weighted = (.*?);’}, {‘Name’: ‘validation:ap-micro’, ‘Regex’: ‘eval_ap_score_micro = (.*?);’}, {‘Name’: ‘validation:ap-macro’, ‘Regex’: ‘eval_ap_score_macro = (.*?);’}, {‘Name’: ‘validation:ap-weighted’, ‘Regex’: ‘eval_ap_score_weighted = (.*?);’}, {‘Name’: ‘validation:auc-micro’, ‘Regex’: ‘eval_auc_score_micro = (.*?);’}, {‘Name’: ‘validation:auc-macro’, ‘Regex’: ‘eval_auc_score_macro = (.*?);’}, {‘Name’: ‘validation:auc-weighted’, ‘Regex’: ‘eval_auc_score_weighted = (.*?);’} ] ) { ‘train’: ‘s3:////train’, ‘valid’: ‘s3:////valid’, ‘test’: ‘s3:////test’ } )

Within, we define a PyTorch Dataset that is used by to prepare the text data for training and evaluation of the model:

def pad_matrix(m: torch.Tensor, max_len: int =100)-> tuple[int, torch.Tensor] : “””Pads an embedding matrix to a specified maximum length.””” if m.ndim == 1: m = m.reshape(1, -1) mask = np.ones_like(m) if m.shape[0] > max_len: m = m[:max_len, :] mask = mask[:max_len, :] else: m = np.pad(m, ((0, max_len – m.shape[0]), (0,0))) mask = np.pad(mask, ((0, max_len – mask.shape[0]), (0,0))) return m, mask class EmbeddingDataset(Dataset: “””PyTorch dataset representing pretrained sentence embeddings, masks, and labels.””” def __init__(self, text: str, labels: int, max_len: int=100): self.text = text self.labels = labels self.max_len = max_len def __len__(self) -> int: return len(self.labels) def __getitem__(self, idx: int) -> dict: e = embed_line(self.text[idx]) length = e.shape[0] m, mask = pad_matrix(e, max_len=self.max_len) item = {} item[’embeddings’] = torch.from_numpy(m) item[‘mask’] = torch.from_numpy(mask) item[‘labels’] = torch.tensor(self.labels[idx]) if length > self.max_len: item[‘lengths’] = torch.tensor(self.max_len) else: item[‘lengths’] = torch.tensor(length) return item

Note that this code anticipates that the file containing your fastText or BlazingText embeddings will be stored in.

Additionally, you can easily deploy pretrained fastText models on their own to live SageMaker endpoints to compute embedding vectors on the fly for use in relevant word-level tasks. See the following GitHub example for more details.

Transformers with Hugging Face

For our second modeling approach, we transition to the usage of Transformers, introduced in the paper Attention Is All You Need. Transformers are deep learning models designed to deliberately avoid the pitfalls of RNNs by relying on a self-attention mechanism to draw global dependencies between input and output. The Transformer model architecture allows for significantly better parallelization and can achieve high performance in relatively short training time.

Built on the success of Transformers, BERT, introduced in the paper BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, added bidirectional pre-training for language representation. Inspired by the Cloze task, BERT is pre-trained with masked language modeling (MLM), in which the model learns to recover the original words for randomly masked tokens. The BERT model is also pretrained on the next sentence prediction (NSP) task to predict if two sentences are in correct reading order. Since its advent in 2018, BERT and its variations have been widely used in text classification tasks.

Our solution uses a variant of BERT known as RoBERTa, which was introduced in the paper RoBERTa: A Robustly Optimized BERT Pretraining Approach. RoBERTa further improves BERT performance on a variety of natural language tasks by optimized model training, including training models longer on a 10 times larger bigger corpus, using optimized hyperparameters, dynamic random masking, removing the NSP task, and more.

Our RoBERTa-based models use the Hugging Face Transformers library, which is a popular open-source Python framework that provides high-quality implementations of all kinds of state-of-the-art Transformer models for a variety of NLP tasks. Hugging Face has partnered with AWS to enable you to easily train and deploy Transformer models on SageMaker. This functionality is available through Hugging Face AWS Deep Learning Container images, which include the Transformers, Tokenizers, and Datasets libraries, and optimized integration with SageMaker for model training and inference.

In our implementation, we inherit the RoBERTa architecture backbone from the Hugging Face Transformers framework and use SageMaker to train and deploy our own text classification model, which we call RoBERTox. RoBERTox uses byte pair encoding (BPE), introduced in Neural Machine Translation of Rare Words with Subword Units, to tokenize input text into subword representations. We can then train our models and tokenizers on the Jigsaw data or any large domain-specific corpus (such as the chat logs from a specific game) and use them for customized text classification. We define our custom classification model class in the following code:

class RoBERToxForSequenceClassification(CustomLossMixIn, RobertaPreTrainedModel): _keys_to_ignore_on_load_missing = [r”position_ids”] def __init__(self, config: PretrainedConfig, *inputs, **kwargs): “””Initialize the RoBERToxForSequenceClassification instance Parameters ———- config : PretrainedConfig num_labels : Optional[int] if not None, overwrite the default classification head in pretrained model. mode : Optional[str] ‘MULTI_CLASS’, ‘MULTI_LABEL’ or “REGRESSION”. Used to determine loss class_weight : Optional[List[float]] If not None, add class weight to BCEWithLogitsLoss or CrossEntropyLoss “”” super().__init__(config, *inputs, **kwargs) # Define model architecture self.roberta = RobertaModel(self.config, add_pooling_layer=False) self.classifier = RobertaClassificationHead(self.config) self.init_weights() @modeling_roberta.add_start_docstrings_to_model_forward( modeling_roberta.ROBERTA_INPUTS_DOCSTRING.format(“batch_size, sequence_length”) ) @modeling_roberta.add_code_sample_docstrings( tokenizer_class=modeling_roberta._TOKENIZER_FOR_DOC, checkpoint=modeling_roberta._CHECKPOINT_FOR_DOC, output_type=SequenceClassifierOutput, config_class=modeling_roberta._CONFIG_FOR_DOC, ) def forward( self, input_ids: torch.Tensor = None, attention_mask: torch.Tensor = None, token_type_ids: torch.Tensor = None, position_ids: torch.Tensor =None, head_mask: torch.Tensor =None, inputs_embeds: torch.Tensor =None, labels: torch.Tensor =None, output_attentions: torch.Tensor =None, output_hidden_states: torch.Tensor =None, return_dict: bool =None, sample_weights: torch.Tensor =None, ) -> : dict: “””Forward pass to return loss, logits, … Returns ——– output : SequenceClassifierOutput has those keys: loss, logits, hidden states, attentions “”” return_dict = return_dict or self.config.use_return_dict outputs = self.roberta( input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids, position_ids=position_ids, head_mask=head_mask, inputs_embeds=inputs_embeds, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) sequence_output = outputs[0] # [CLS] embedding logits = self.classifier(sequence_output) loss = self.compute_loss(logits, labels, sample_weights=sample_weights) if not return_dict: output = (logits,) + outputs[2:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def compute_loss(self, logits: torch.Tensor, labels: torch.Tensor, sample_weights: Optional[torch.Tensor] = None) -> torch.FloatTensor: return super().compute_loss(logits, labels, sample_weights)

Before training, we prepare our text data and labels using Hugging Face’s datasets library and upload the result to Amazon S3:

from datasets import Dataset import multiprocessing data_train = Dataset.from_pandas(df_train) … tokenizer = def preprocess_function(examples: examples) -> torch.Tensor: result = tokenizer(examples[“text”], padding=”max_length”, max_length=128, truncation=True) return result num_proc = multiprocessing.cpu_count() print(“Number of CPUs =”, num_proc) data_train = preprocess_function, batched=True, load_from_cache_file=False, num_proc=num_proc ) … import botocore from datasets.filesystems import S3FileSystem s3_session = botocore.session.Session() # create S3FileSystem instance with s3_session s3 = S3FileSystem(session=s3_session) # saves encoded_dataset to your s3 bucket data_train.save_to_disk(f’s3:////train’, fs=s3) …

We initiate training of the model in a similar fashion to the RNN:

import sagemaker sess = sagemaker.Session() role = sagemaker.get_execution_role() from sagemaker.huggingface import HuggingFace # hyperparameters, which are passed into the training job hyperparameters = { ‘model-name’: , ‘epochs’: 10, ‘train-batch-size’: 32, ‘eval-batch-size’: 64, ‘num-labels’: len(), ‘class-weight’: ‘ ‘.join([str(c) for c in ]), ‘metric-for-best-model’: ‘ap_score_weighted’, ‘save-total-limit’: 1, } # create the Estimator huggingface_estimator = HuggingFace( entry_point=’’, source_dir=, instance_type=, instance_count=1, role=role, transformers_version=’4.6.1′, pytorch_version=’1.7.1′, py_version=’py36′, hyperparameters=hyperparameters, metric_definitions=[ {‘Name’: ‘validation:accuracy’, ‘Regex’: ‘eval_accuracy = (.*?);’}, {‘Name’: ‘validation:f1-micro’, ‘Regex’: ‘eval_f1_score_micro = (.*?);’}, {‘Name’: ‘validation:f1-macro’, ‘Regex’: ‘eval_f1_score_macro = (.*?);’}, {‘Name’: ‘validation:f1-weighted’, ‘Regex’: ‘eval_f1_score_weighted = (.*?);’}, {‘Name’: ‘validation:ap-micro’, ‘Regex’: ‘eval_ap_score_micro = (.*?);’}, {‘Name’: ‘validation:ap-macro’, ‘Regex’: ‘eval_ap_score_macro = (.*?);’}, {‘Name’: ‘validation:ap-weighted’, ‘Regex’: ‘eval_ap_score_weighted = (.*?);’}, {‘Name’: ‘validation:auc-micro’, ‘Regex’: ‘eval_auc_score_micro = (.*?);’}, {‘Name’: ‘validation:auc-macro’, ‘Regex’: ‘eval_auc_score_macro = (.*?);’}, {‘Name’: ‘validation:auc-weighted’, ‘Regex’: ‘eval_auc_score_weighted = (.*?);’} ] ) { ‘train’: ‘s3:////train’, ‘valid’: ‘s3:////valid’, ‘test’: ‘s3:////test’ )

Finally, the following Python code snippet illustrates the process of serving RoBERTox via a live SageMaker endpoint for real-time text classification for a JSON request:

from sagemaker.huggingface import HuggingFaceModel from sagemaker import get_execution_role from sagemaker.predictor import Predictor from sagemaker.serializers import JSONSerializer from sagemaker.deserializers import JSONDeserializer class Classifier(Predictor): def __init__(self, endpoint_name, sagemaker_session): super().__init__(endpoint_name, sagemaker_session, serializer=JSONSerializer(), deserializer=JSONDeserializer()) hf_model = HuggingFaceModel( role=get_execution_role(), model_data=, entry_point=””, transformers_version=”4.6.1″, pytorch_version=”1.7.1″, py_version=”py36″, predictor_cls=Classifier ) predictor = hf_model.deploy(instance_type=, initial_instance_count=1)

Evaluation of model performance: Jigsaw unintended bias dataset

The following table contains performance metrics for models trained and evaluated on data from the Jigsaw Unintended Bias in Toxicity Detection Kaggle competition. We trained models for three different but interrelated tasks:

  • Binary case – The model was trained on the full training dataset to predict the toxicity label only
  • Fine-grained case – The subset of the training data for which toxicity>=0.5 was used to predict other toxicity sub-type labels (obscene, threat, insult, identity_attack, sexual_explicit)
  • Multitask case – The full training dataset was used to predict all six labels simultaneously

We trained RNN and RoBERTa models for each of these three tasks using the Jigsaw-provided fractional labels, which correspond to the proportion of annotators who thought the label was appropriate for the text, as well as with binary labels combined with class weights in the network loss function. In the binary labeling scheme, the proportions were thresholded at 0.5 for each available label (1 if label>=0.5, 0 otherwise), and the model loss functions were weighted based on the relative proportions of each binary label in the training dataset. In all cases, we found that using the fractional labels directly resulted in the best performance, indicating the added value of the information inherent in the degree of agreement between annotators.

We display two model metrics: the average precision (AP), which provides a summary of the precision-recall curve by computing the weighted mean of the precision values achieved at each classification threshold, and the area under the receiver operating characteristic curve (AUC), which aggregates model performance across classification thresholds with respect to the true positive rate and false positive rate. Note that the true class for a given text instance in the test set corresponds to whether the true proportion is greater than or equal to 0.5 (1 if label>=0.5, 0 otherwise).

. Subword Embedding + RNN RoBERTa
. Fractional labels Binary labels + Class weighting Fractional labels Binary labels + Class weighting
Binary AP=0.746, AUC=0.966 AP=0.730, AUC=0.963 AP=0.758, AUC=0.966 AP=0.747, AUC=0.963
Fine-grained AP=0.906, AUC=0.909 AP=0.850, AUC=0.851 AP=0.913, AUC=0.913 AP=0.911, AUC=0.912
Multitask AP=0.721, AUC=0.972 AP=0.535, AUC=0.907 AP=0.740, AUC=0.972 AP=0.711, AUC=0.961


In this post, we presented two text classification approaches for online conversations using AWS ML services. You can generalize these solutions across online communication platforms, with industries such as gaming particularly likely to benefit from improved ability to detect harmful content. In future posts, we plan to further discuss an end-to-end architecture for seamless deployment of models into your AWS account.

If you’d like help accelerating your use of ML in your products and processes, please contact the Amazon ML Solutions Lab.

About the Authors

Ryan Brand is a Data Scientist in the Amazon Machine Learning Solutions Lab. He has specific experience in applying machine learning to problems in healthcare and the life sciences, and in his free time he enjoys reading history and science fiction.

Sourav Bhabesh is a Data Scientist at the Amazon ML Solutions Lab. He develops AI/ML solutions for AWS customers across various industries. His specialty is Natural Language Processing (NLP) and is passionate about deep learning. Outside of work he enjoys reading books and traveling.

Liutong Zhou is an Applied Scientist at the Amazon ML Solutions Lab. He builds bespoke AI/ML solutions for AWS customers across various industries. He specializes in Natural Language Processing (NLP) and is passionate about multi-modal deep learning. He is a lyric tenor and enjoys singing operas outside of work.

Sia Gholami is a Senior Data Scientist at the Amazon ML Solutions Lab, where he builds AI/ML solutions for customers across various industries. He is passionate about natural language processing (NLP) and deep learning. Outside of work, Sia enjoys spending time in nature and playing tennis.

Daniel Horowitz is an Applied AI Science Manager. He leads a team of scientists on the Amazon ML Solutions Lab working to solve customer problems and drive cloud adoption with ML.


Continue Reading
Click to comment

Leave a Reply

Your email address will not be published.


Build a GNN-based real-time fraud detection solution using Amazon SageMaker, Amazon Neptune, and the Deep Graph Library

Fraudulent activities severely impact many industries, such as e-commerce, social media, and financial services. Frauds could cause a significant loss for businesses and consumers. American consumers reported losing more than $5.8 billion to frauds in 2021, up more than 70% over 2020. Many techniques have been used to detect fraudsters—rule-based filters, anomaly detection, and machine…




Fraudulent activities severely impact many industries, such as e-commerce, social media, and financial services. Frauds could cause a significant loss for businesses and consumers. American consumers reported losing more than $5.8 billion to frauds in 2021, up more than 70% over 2020. Many techniques have been used to detect fraudsters—rule-based filters, anomaly detection, and machine learning (ML) models, to name a few.

In real-world data, entities often involve rich relationships with other entities. Such a graph structure can provide valuable information for anomaly detection. For example, in the following figure, users are connected via shared entities such as Wi-Fi IDs, physical locations, and phone numbers. Due to the large number of unique values of these entities, like phone numbers, it’s difficult to use them in the traditional feature-based models—for example, one-hot encoding all phone numbers wouldn’t be viable. But such relationships could help predict whether a user is a fraudster. If a user has shared several entities with a known fraudster, the user is more likely a fraudster.

Recently, graph neural network (GNN) has become a popular method for fraud detection. GNN models can combine both graph structure and attributes of nodes or edges, such as users or transactions, to learn meaningful representations to distinguish malicious users and events from legitimate ones. This capability is crucial for detecting frauds where fraudsters collude to hide their abnormal features but leave some traces of relations.

Current GNN solutions mainly rely on offline batch training and inference mode, which detect fraudsters after malicious events have happened and losses have occurred. However, catching fraudulent users and activities in real time is crucial for preventing losses. This is particularly true in business cases where there is only one chance to prevent fraudulent activities. For example, in some e-commerce platforms, account registration is wide open. Fraudsters can behave maliciously just once with an account and never use the same account again.

Predicting fraudsters in real time is important. Building such a solution, however, is challenging. Because GNNs are still new to the industry, there are limited online resources on converting GNN models from batch serving to real-time serving. Additionally, it’s challenging to construct a streaming data pipeline that can feed incoming events to a GNN real-time serving API. To the best of the authors’ knowledge, no reference architectures and examples are available for GNN-based real-time inference solutions as of this writing.

To help developers apply GNNs to real-time fraud detection, this post shows how to use Amazon Neptune, Amazon SageMaker, and the Deep Graph Library (DGL), among other AWS services, to construct an end-to-end solution for real-time fraud detection using GNN models.

We focus on four tasks:

  • Processing a tabular transaction dataset into a heterogeneous graph dataset
  • Training a GNN model using SageMaker
  • Deploying the trained GNN models as a SageMaker endpoint
  • Demonstrating real-time inference for incoming transactions

This post extends the previous work in Detecting fraud in heterogeneous networks using Amazon SageMaker and Deep Graph Library, which focuses on the first two tasks. You can refer to that post for more details on heterogeneous graphs, GNNs, and semi-supervised training of GNNs.

Businesses looking for a fully-managed AWS AI service for fraud detection can also use Amazon Fraud Detector, which makes it easy to identify potentially fraudulent online activities, such as the creation of fake accounts or online payment fraud.

Solution overview

This solution contains two major parts.

The first part is a pipeline that processes the data, trains GNN models, and deploys the trained models. It uses AWS Glue to process the transaction data, and saves the processed data to both Amazon Neptune and Amazon Simple Storage Service (Amazon S3). Then, a SageMaker training job is triggered to train a GNN model on the data saved in Amazon S3 to predict whether a transaction is fraudulent. The trained model along with other assets are saved back to Amazon S3 upon the completion of the training job. Finally, the saved model is deployed as a SageMaker endpoint. The pipeline is orchestrated by AWS Step Functions, as shown in the following figure.

The second part of the solution implements real-time fraudulent transaction detection. It starts from a RESTful API that queries the graph database in Neptune to extract the subgraph related to an incoming transaction. It also has a web portal that can simulate business activities, generating online transactions with both fraudulent and legitimate ones. The web portal provides a live visualization of the fraud detection. This part uses Amazon CloudFront, AWS Amplify, AWS AppSync, Amazon API Gateway, Step Functions, and Amazon DocumentDB to rapidly build the web application. The following diagram illustrates the real-time inference process and web portal.

The implementation of this solution, along with an AWS CloudFormation template that can launch the architecture in your AWS account, is publicly available through the following GitHub repo.

Data processing

In this section, we briefly describe how to process an example dataset and convert it from raw tables into a graph with relations identified among different columns.

This solution uses the same dataset, the IEEE-CIS fraud dataset, as the previous post Detecting fraud in heterogeneous networks using Amazon SageMaker and Deep Graph Library. Therefore, the basic principle of the data process is the same. In brief, the fraud dataset includes a transactions table and an identities table, having nearly 500,000 anonymized transaction records along with contextual information (for example, devices used in transactions). Some transactions have a binary label, indicating whether a transaction is fraudulent. Our task is to predict which unlabeled transactions are fraudulent and which are legitimate.

The following figure illustrates the general process of how to convert the IEEE tables into a heterogeneous graph. We first extract two columns from each table. One column is always the transaction ID column, where we set each unique TransactionID as one node. Another column is picked from the categorical columns, such as the ProductCD and id_03 columns, where each unique category was set as a node. If a TransactionID and a unique category appear in the same row, we connect them with one edge. This way, we convert two columns in a table into one bipartite. Then we combine those bipartites along with the TransactionID nodes, where the same TransactionID nodes are merged into one unique node. After this step, we have a heterogeneous graph built from bipartites.

For the rest of the columns that aren’t used to build the graph, we join them together as the feature of the TransactionID nodes. TransactionID values that have the isFraud values are used as the label for model training. Based on this heterogeneous graph, our task becomes a node classification task of the TransactionID nodes. For more details on preparing the graph data for training GNNs, refer to the Feature extraction and Constructing the graph sections of the previous blog post.

The code used in this solution is available in src/scripts/ You can also experiment with data processing through the Jupyter notebook src/sagemaker/01.FD_SL_Process_IEEE-CIS_Dataset.ipynb.

Instead of manually processing the data, as done in the previous post, this solution uses a fully automatic pipeline orchestrated by Step Functions and AWS Glue that supports processing huge datasets in parallel via Apache Spark. The Step Functions workflow is written in AWS Cloud Development Kit (AWS CDK). The following is a code snippet to create this workflow:

import { LambdaInvoke, GlueStartJobRun } from ‘aws-cdk-lib/aws-stepfunctions-tasks’; const parametersNormalizeTask = new LambdaInvoke(this, ‘Parameters normalize’, { lambdaFunction: parametersNormalizeFn, integrationPattern: IntegrationPattern.REQUEST_RESPONSE, }); … const dataProcessTask = new GlueStartJobRun(this, ‘Data Process’, { integrationPattern: IntegrationPattern.RUN_JOB, glueJobName: etlConstruct.jobName, timeout: Duration.hours(5), resultPath: ‘$.dataProcessOutput’, }); … const definition = parametersNormalizeTask .next(dataIngestTask) .next(dataCatalogCrawlerTask) .next(dataProcessTask) .next(hyperParaTask) .next(trainingJobTask) .next(runLoadGraphDataTask) .next(modelRepackagingTask) .next(createModelTask) .next(createEndpointConfigTask) .next(checkEndpointTask) .next(endpointChoice);

Besides constructing the graph data for GNN model training, this workflow also batch loads the graph data into Neptune to conduct real-time inference later on. This batch data loading process is demonstrated in the following code snippet:

from neptune_python_utils.endpoints import Endpoints from neptune_python_utils.bulkload import BulkLoad … bulkload = BulkLoad( source=targetDataPath, endpoints=endpoints, role=args.neptune_iam_role_arn, region=args.region, update_single_cardinality_properties=True, fail_on_error=True) load_status = bulkload.load_async() status, json = load_status.status(details=True, errors=True) load_status.wait()

GNN model training

After the graph data for model training is saved in Amazon S3, a SageMaker training job, which is only charged when the training job is running, is triggered to start the GNN model training process in the Bring Your Own Container (BYOC) mode. It allows you to pack your model training scripts and dependencies in a Docker image, which it uses to create SageMaker training instances. The BYOC method could save significant effort in setting up the training environment. In src/sagemaker/02.FD_SL_Build_Training_Container_Test_Local.ipynb, you can find details of the GNN model training.

Docker image

The first part of the Jupyter notebook file is the training Docker image generation (see the following code snippet):

*!* aws ecr get-login-password –region us-east-1 | docker login –username AWS –password-stdin image_name *=* ‘fraud-detection-with-gnn-on-dgl/training’ *!* docker build -t $image_name ./FD_SL_DGL/gnn_fraud_detection_dgl

We used a PyTorch-based image for the model training. The Deep Graph Library (DGL) and other dependencies are installed when building the Docker image. The GNN model code in the src/sagemaker/FD_SL_DGL/gnn_fraud_detection_dgl folder is copied to the image as well.

Because we process the transaction data into a heterogeneous graph, in this solution we choose the Relational Graph Convolutional Network (RGCN) model, which is specifically designed for heterogeneous graphs. Our RGCN model can train learnable embeddings for the nodes in heterogeneous graphs. Then, the learned embeddings are used as inputs of a fully connected layer for predicting the node labels.


To train the GNN, we need to define a few hyperparameters before the training process, such as the file names of the graph constructed, the number of layers of GNN models, the training epochs, the optimizer, the optimization parameters, and more. See the following code for a subset of the configurations:

edges *=* “,”*.*join(map(*lambda* x: x*.*split(“/”)[*-*1], [file *for* file *in* processed_files *if* “relation” *in* file])) params *=* {‘nodes’ : ‘features.csv’, ‘edges’: edges, ‘labels’: ‘tags.csv’, ’embedding-size’: 64, ‘n-layers’: 2, ‘n-epochs’: 10, ‘optimizer’: ‘adam’, ‘lr’: 1e-2}

For more information about all the hyperparameters and their default values, see in the src/sagemaker/FD_SL_DGL/gnn_fraud_detection_dgl folder.

Model training with SageMaker

After the customized container Docker image is built, we use the preprocessed data to train our GNN model with the hyperparameters we defined. The training job uses the DGL, with PyTorch as the backend deep learning framework, to construct and train the GNN. SageMaker makes it easy to train GNN models with the customized Docker image, which is an input argument of the SageMaker estimator. For more information about training GNNs with the DGL on SageMaker, see Train a Deep Graph Network.

The SageMaker Python SDK uses Estimator to encapsulate training on SageMaker, which runs SageMaker-compatible custom Docker containers, enabling you to run your own ML algorithms by using the SageMaker Python SDK. The following code snippet demonstrates training the model with SageMaker (either in a local environment or cloud instances):

from sagemaker.estimator import Estimator from time import strftime, gmtime from sagemaker.local import LocalSession localSageMakerSession = LocalSession(boto_session=boto3.session.Session(region_name=current_region)) estimator = Estimator(image_uri=image_name, role=sagemaker_exec_role, instance_count=1, instance_type=’local’, hyperparameters=params, output_path=output_path, sagemaker_session=localSageMakerSession) training_job_name = “{}-{}”.format(‘GNN-FD-SL-DGL-Train’, strftime(“%Y-%m-%d-%H-%M-%S”, gmtime())) print(training_job_name){‘train’: processed_data}, job_name=training_job_name)

After training, the GNN model’s performance on the test set is displayed like the following outputs. The RGCN model normally can achieve around 0.87 AUC and more than 95% accuracy. For a comparison of the RGCN model with other ML models, refer to the Results section of the previous blog post for more details.

Epoch 00099 | Time(s) 7.9413 | Loss 0.1023 | f1 0.3745 Metrics Confusion Matrix: labels positive labels negative predicted positive 4343 576 predicted negative 13494 454019 f1: 0.3817, precision: 0.8829, recall: 0.2435, acc: 0.9702, roc: 0.8704, pr: 0.4782, ap: 0.4782 Finished Model training

Upon the completion of model training, SageMaker packs the trained model along with other assets, including the trained node embeddings, into a ZIP file and then uploads it to a specified S3 location. Next, we discuss the deployment of the trained model for real-time fraudulent detection.

GNN model deployment

SageMaker makes the deployment of trained ML models simple. In this stage, we use the SageMaker PyTorchModel class to deploy the trained model, because our DGL model depends on PyTorch as the backend framework. You can find the deployment code in the src/sagemaker/03.FD_SL_Endpoint_Deployment.ipynb file.

Besides the trained model file and assets, SageMaker requires an entry point file for the deployment of a customized model. The entry point file is run and stored in the memory of an inference endpoint instance to respond to the inference request. In our case, the entry point file is the file in the src/sagemaker/FD_SL_DGL/code folder, which performs four major functions:

  • Receive requests and parse contents of requests to obtain the to-be-predicted nodes and their associated data
  • Convert the data to a DGL heterogeneous graph as input for the RGCN model
  • Perform the real-time inference via the trained RGCN model
  • Return the prediction results to the requester

Following SageMaker conventions, the first two functions are implemented in the input_fn method. See the following code (for simplicity, we delete some commentary code):

def input_fn(request_body, request_content_type=’application/json’): # ——————— receive request ———————————————— # input_data = json.loads(request_body) subgraph_dict = input_data[‘graph’] n_feats = input_data[‘n_feats’] target_id = input_data[‘target_id’] graph, new_n_feats, new_pred_target_id = recreate_graph_data(subgraph_dict, n_feats, target_id) return (graph, new_n_feats, new_pred_target_id)

The constructed DGL graph and features are then passed to the predict_fn method to fulfill the third function. predict_fn takes two input arguments: the outputs of input_fn and the trained model. See the following code:

def predict_fn(input_data, model): # ——————— Inference ———————————————— # graph, new_n_feats, new_pred_target_id = input_data with th.no_grad(): logits = model(graph, new_n_feats) res = logits[new_pred_target_id].cpu().detach().numpy() return res[1]

The model used in perdict_fn is created by the model_fn method when the endpoint is called the first time. The function model_fn loads the saved model file and associated assets from the model_dir argument and the SageMaker model folder. See the following code:

def model_fn(model_dir): # —————— Loading model ——————- ntype_dict, etypes, in_size, hidden_size, out_size, n_layers, embedding_size = initialize_arguments(os.path.join(BASE_PATH, ‘metadata.pkl’)) rgcn_model = HeteroRGCN(ntype_dict, etypes, in_size, hidden_size, out_size, n_layers, embedding_size) stat_dict = th.load(‘model.pth’) rgcn_model.load_state_dict(stat_dict) return rgcn_model

The output of the predict_fn method is a list of two numbers, indicating the logits for class 0 and class 1, where 0 means legitimate and 1 means fraudulent. SageMaker takes this list and passes it to an inner method called output_fn to complete the final function.

To deploy our GNN model, we first wrap the GNN model into a SageMaker PyTorchModel class with the entry point file and other parameters (the path of the saved ZIP file, the PyTorch framework version, the Python version, and so on). Then we call its deploy method with instance settings. See the following code:

env = { ‘SAGEMAKER_MODEL_SERVER_WORKERS’: ‘1’ } print(f’Use model {repackged_model_path}’) sagemakerSession = sm.session.Session(boto3.session.Session(region_name=current_region)) fd_sl_model = PyTorchModel(model_data=repackged_model_path, role=sagemaker_exec_role, entry_point=’./FD_SL_DGL/code/’, framework_version=’1.6.0′, py_version=’py3′, predictor_cls=JSONPredictor, env=env, sagemaker_session=sagemakerSession) fd_sl_predictor *=* fd_sl_model*.*deploy(instance_type*=*’ml.c5.4xlarge’, initial_instance_count*=*1,)

The preceding procedures and code snippets demonstrate how to deploy your GNN model as an online inference endpoint from a Jupyter notebook. However, for production, we recommend using the previously mentioned MLOps pipeline orchestrated by Step Functions for the entire workflow, including processing data, training the model, and deploying an inference endpoint. The entire pipeline is implemented by an AWS CDK application, which can be easily replicated in different Regions and accounts.

Real-time inference

When a new transaction arrives, to perform real-time prediction, we need to complete four steps:

  1. Node and edge insertion – Extract the transaction’s information such as the TransactionID and ProductCD as nodes and edges, and insert the new nodes into the existing graph data stored at the Neptune database.
  2. Subgraph extraction – Set the to-be-predicted transaction node as the center node, and extract a n-hop subgraph according to the GNN model’s input requirements.
  3. Feature extraction – For the nodes and edges in the subgraph, extract their associated features.
  4. Call the inference endpoint – Pack the subgraph and features into the contents of a request, then send the request to the inference endpoint.

In this solution, we implement a RESTful API to achieve real-time fraudulent predication described in the preceding steps. See the following pseudo-code for real-time predictions. The full implementation is in the complete source code file.

For prediction in real time, the first three steps require lower latency. Therefore, a graph database is an optimal choice for these tasks, particularly for the subgraph extraction, which could be achieved efficiently with graph database queries. The underline functions that support the pseudo-code are based on Neptune’s gremlin queries.

def handler(event, context): graph_input = GraphModelClient(endpoints) # Step 1: node and edge insertion trans_dict, identity_dict, target_id, transaction_value_cols, union_li_cols = load_data_from_event(event, transactions_id_cols, transactions_cat_cols, dummied_col) graph_input.insert_new_transaction_vertex_and_edge(trans_dict, identity_dict , target_id, vertex_type = ‘Transaction’) # Setp 2: subgraph extraction subgraph_dict, transaction_embed_value_dict = graph_input.query_target_subgraph(target_id, trans_dict, transaction_value_cols, union_li_cols, dummied_col) # Step 3 & 4: feature extraction & call the inference endpoint transaction_id = int(target_id[(target_id.find(‘-‘)+1):]) pred_prob = invoke_endpoint_with_idx(endpointname = ENDPOINT_NAME, target_id = transaction_id, subgraph_dict = subgraph_dict, n_feats = transaction_embed_value_dict) function_res = { ‘id’: event[‘transaction_data’][0][‘TransactionID’], ‘flag’: pred_prob > MODEL_BTW, ‘pred_prob’: pred_prob } return function_res

One caveat about real-time fraud detection using GNNs is the GNN inference mode. To fulfill real-time inference, we need to convert the GNN model inference from transductive mode to inductive mode. GNN models in transductive inference mode can’t make predictions for newly appeared nodes and edges, whereas in inductive mode, GNN models can handle new nodes and edges. A demonstration of the difference between transductive and inductive mode is shown in the following figure.

In transductive mode, predicted nodes and edges coexist with labeled nodes and edges during training. Models identify them before inference, and they could be inferred in training. Models in inductive mode are trained on the training graph but need to predict unseen nodes (those in red dotted circles on the right) with their associated neighbors, which might be new nodes, like the gray triangle node on the right.

Our RGCN model is trained and tested in transductive mode. It has access to all nodes in training, and also trained an embedding for each featureless node, such as IP address and card types. In the testing stage, the RGCN model uses these embeddings as node features to predict nodes in the test set. When we do real-time inference, however, some of the newly added featureless nodes have no such embeddings because they’re not in the training graph. One way to tackle this issue is to assign the mean of all embeddings in the same node type to the new nodes. In this solution, we adopt this method.

In addition, this solution provides a web portal (as seen in the following screenshot) to demonstrate real-time fraudulent predictions from business operators’ perspectives. It can generate the simulated online transactions, and provide a live visualization of detected fraudulent transaction information.

Clean up

When you’re finished exploring the solution, you can clean the resources to avoid incurring charges.


In this post, we showed how to build a GNN-based real-time fraud detection solution using SageMaker, Neptune, and the DGL. This solution has three major advantages:

  • It has good performance in terms of prediction accuracy and AUC metrics
  • It can perform real-time inference via a streaming MLOps pipeline and SageMaker endpoints
  • It automates the total deployment process with the provided CloudFormation template so that interested developers can easily test this solution with custom data in their account

For more details about the solution, see the GitHub repo.

After you deploy this solution, we recommend customizing the data processing code to fit your own data format and modify the real-time inference mechanism while keeping the GNN model unchanged. Note that we split the real-time inference into four steps without further optimization of the latency. These four steps take a few seconds to get a prediction on the demo dataset. We believe that optimizing the Neptune graph data schema design and queries for subgraph and feature extraction can significantly reduce the inference latency.

About the authors

Jian Zhang is an applied scientist who has been using machine learning techniques to help customers solve various problems, such as fraud detection, decoration image generation, and more. He has successfully developed graph-based machine learning, particularly graph neural network, solutions for customers in China, USA, and Singapore. As an enlightener of AWS’s graph capabilities, Zhang has given many public presentations about the GNN, the Deep Graph Library (DGL), Amazon Neptune, and other AWS services.

Mengxin Zhu is a manager of Solutions Architects at AWS, with a focus on designing and developing reusable AWS solutions. He has been engaged in software development for many years and has been responsible for several startup teams of various sizes. He also is an advocate of open-source software and was an Eclipse Committer.

Haozhu Wang is a research scientist at Amazon ML Solutions Lab, where he co-leads the Reinforcement Learning Vertical. He helps customers build advanced machine learning solutions with the latest research on graph learning, natural language processing, reinforcement learning, and AutoML. Haozhu received his PhD in Electrical and Computer Engineering from the University of Michigan.


Continue Reading


New – AWS Private 5G – Build Your Own Private Mobile Network

Back in the mid-1990’s, I had a young family and 5 or 6 PCs in the basement. One day my son Stephen and I bought a single box that contained a bunch of 3COM network cards, a hub, some drivers, and some cables, and spent a pleasant weekend setting up our first home LAN. Introducing…




Back in the mid-1990’s, I had a young family and 5 or 6 PCs in the basement. One day my son Stephen and I bought a single box that contained a bunch of 3COM network cards, a hub, some drivers, and some cables, and spent a pleasant weekend setting up our first home LAN.

Introducing AWS Private 5G
Today I would like to introduce you to AWS Private 5G, the modern, corporate version of that very powerful box of hardware and software. This cool new service lets you design and deploy your own private mobile network in a matter of days. It is easy to install, operate, and scale, and does not require any specialized expertise. You can use the network to communicate with the sensors & actuators in your smart factory, or to provide better connectivity for handheld devices, scanners, and tablets for process automation.

The private mobile network makes use of CBRS spectrum. It supports 4G LTE (Long Term Evolution) today, and will support 5G in the future, both of which give you a consistent, predictable level of throughput with ultra low latency. You get long range coverage, indoors and out, and fine-grained access control.

AWS Private 5G runs on AWS-managed infrastructure. It is self-service and API-driven, and can scale with respect to geographic coverage, device count, and overall throughput. It also works nicely with other parts of AWS, and lets you use AWS Identity and Access Management (IAM) to control access to both devices and applications.

Getting Started with AWS Private 5G
To get started, I visit the AWS Private 5G Console and click Create network:

I assign a name to my network (JeffCell) and to my site (JeffSite) and click Create network:

The network and the site are created right away. Now I click Create order:

I fill in the shipping address, agree to the pricing (more on that later), and click Create order:

Then I await delivery, and click Acknowledge order to proceed:

The package includes a radio unit and ten SIM cards. The radio unit requires AC power and wired access to the public Internet, along with basic networking (IPv4 and DHCP).

When the order arrives, I click Acknowledge order and confirm that I have received the desired radio unit and SIMs. Then I engage a Certified Professional Installer (CPI) to set it up. As part of the installation process, the installer will enter the latitude, longitude, and elevation of my site.

Things to Know
Here are a couple of important things to know about AWS Private 5G:

Partners – Planning and deploying a private wireless network can be complex and not every enterprise will have the tools to do this work on their own. In addition, CBRS spectrum in the United States requires Certified Professional Installation (CPI) of radios. To address these needs, we are building an ecosystem of partners that can provide customers with radio planning, installation, CPI certification, and implementation of customer use cases. You can access these partners from the AWS Private 5G Console and work with them through the AWS Marketplace.

Deployment Options – In the demo above, I showed you the cloud–based deployment option, which is designed for testing and evaluation purposes, for time-limited deployments, and for deployments that do not use the network in latency-sensitive ways. With this option, the AWS Private 5G Mobile Core runs within a specific AWS Region. We are also working to enable on-premises hosting of the Mobile Core on a Private 5G compute appliance.

CLI and API Access – I can also use the create-network, create-network-site, and acknowledge-order-receipt commands to set up my AWS Private 5G network from the command line. I still need to use the console to place my equipment order.

Scaling and Expansion – Each network supports one radio unit that can provide up to 150 Mbps of throughput spread across up to 100 SIMs. We are working to add support for multiple radio units and greater number of SIM cards per network.

Regions and Locations – We are launching AWS Private 5G in the US East (Ohio), US East (N. Virginia), and US West (Oregon) Regions, and are working to make the service available outside of the United States in the near future.

Pricing – Each radio unit is billed at $10 per hour, with a 60 day minimum.

To learn more, read about AWS Private 5G.



Continue Reading


Build an air quality anomaly detector using Amazon Lookout for Metrics

Today, air pollution is a familiar environmental issue that creates severe respiratory and heart conditions, which pose serious health threats. Acid rain, depletion of the ozone layer, and global warming are also adverse consequences of air pollution. There is a need for intelligent monitoring and automation in order to prevent severe health issues and in…




Today, air pollution is a familiar environmental issue that creates severe respiratory and heart conditions, which pose serious health threats. Acid rain, depletion of the ozone layer, and global warming are also adverse consequences of air pollution. There is a need for intelligent monitoring and automation in order to prevent severe health issues and in extreme cases life-threatening situations. Air quality is measured using the concentration of pollutants in the air. Identifying symptoms early and controlling the pollutant level before it’s dangerous is crucial. The process of identifying the air quality and the anomaly in the weight of pollutants, and quickly diagnosing the root cause, is difficult, costly, and error-prone.

The process of applying AI and machine learning (ML)-based solutions to find data anomalies involves a lot of complexity in ingesting, curating, and preparing data in the right format and then optimizing and maintaining the effectiveness of these ML models over long periods of time. This has been one of the barriers to quickly implementing and scaling the adoption of ML capabilities.

This post shows you how to use an integrated solution with Amazon Lookout for Metrics and Amazon Kinesis Data Firehose to break these barriers by quickly and easily ingesting streaming data, and subsequently detecting anomalies in the key performance indicators of your interest.

Lookout for Metrics automatically detects and diagnoses anomalies (outliers from the norm) in business and operational data. It’s a fully managed ML service that uses specialized ML models to detect anomalies based on the characteristics of your data. For example, trends and seasonality are two characteristics of time series metrics in which threshold-based anomaly detection doesn’t work. Trends are continuous variations (increases or decreases) in a metric’s value. On the other hand, seasonality is periodic patterns that occur in a system, usually rising above a baseline and then decreasing again. You don’t need ML experience to use Lookout for Metrics.

We demonstrate a common air quality monitoring scenario, in which we detect anomalies in the pollutant concentration in the air. By the end of this post, you’ll learn how to use these managed services from AWS to help prevent health issues and global warming. You can apply this solution to other use cases for better environment management, such as detecting anomalies in water quality, land quality, and power consumption patterns, to name a few.

Solution overview

The architecture consists of three functional blocks:

  • Wireless sensors placed at strategic locations to sense the concentration level of carbon monoxide (CO), sulfur dioxide (SO2), and nitrogen dioxide(NO2) in the air
  • Streaming data ingestion and storage
  • Anomaly detection and notification

The solution provides a fully automated data path from the sensors all the way to a notification being raised to the user. You can also interact with the solution using the Lookout for Metrics UI in order to analyze the identified anomalies.

The following diagram illustrates our solution architecture.


You need the following prerequisites before you can proceed with solution. For this post, we use the us-east-1 Region.

  1. Download the Python script ( and data file from the GitHub repo.
  2. Open the live_data.csv file in your preferred editor and replace the dates to be today’s and tomorrow’s date. For example, if today’s date is July 8, 2022, then replace 2022-03-25 with 2022-07-08. Keep the format the same. This is required to simulate sensor data for the current date using the IoT simulator script.
  3. Create an Amazon Simple Storage Service (Amazon S3) bucket and a folder named air-quality. Create a subfolder inside air-quality named historical. For instructions, see Creating a folder.
  4. Upload the live_data.csv file in the root S3 bucket and historical_data.json in the historical folder.
  5. Create an AWS Cloud9 development environment, which we use to run the Python simulator program to create sensor data for this solution.

Ingest and transform data using AWS IoT Core and Kinesis Data Firehose

We use a Kinesis Data Firehose delivery stream to ingest the streaming data from AWS IoT Core and deliver it to Amazon S3. Complete the following steps:

  1. On the Kinesis Data Firehose console, choose Create delivery stream.
  2. For Source, choose Direct PUT.
  3. For Destination, choose Amazon S3.
  4. For Delivery stream name, enter a name for your delivery stream.
  5. For S3 bucket, enter the bucket you created as a prerequisite.
  6. Enter values for S3 bucket prefix and S3 bucket error output prefix.One of the key points to note is the configuration of the custom prefix that is configured for the Amazon S3 destination. This prefix pattern makes sure that the data is created in the S3 bucket as per the prefix hierarchy expected by Lookout for Metrics. (More on this later in this post.) For more information about custom prefixes, see Custom Prefixes for Amazon S3 Objects.
  7. For Buffer interval, enter 60.
  8. Choose Create or update IAM role.
  9. Choose Create delivery stream.

    Now we configure AWS IoT Core and run the air quality simulator program.
  10. On the AWS IoT Core console, create an AWS IoT policy called admin.
  11. In the navigation pane under Message Routing, choose Rules.
  12. Choose Create rule.
  13. Create a rule with the Kinesis Data Firehose(firehose) action.
    This sends data from an MQTT message to a Kinesis Data Firehose delivery stream.
  14. Choose Create.
  15. Create an AWS IoT thing with name Test-Thing and attach the policy you created.
  16. Download the certificate, public key, private key, device certificate, and root CA for AWS IoT Core.
  17. Save each of the downloaded files to the certificates subdirectory that you created earlier.
  18. Upload to the iot-test-publish folder.
  19. On the AWS IoT Core console, in the navigation pane, choose Settings.
  20. Under Custom endpoint, copy the endpoint.
    This AWS IoT Core custom endpoint URL is personal to your AWS account and Region.
  21. Replace customEndpointUrl with your AWS IoT Core custom endpoint URL, certificates with the name of certificate, and Your_S3_Bucket_Name with your S3 bucket name.
    Next, you install pip and the AWS IoT SDK for Python.
  22. Log in to AWS Cloud9 and create a working directory in your development environment. For example: aq-iot-publish.
  23. Create a subdirectory for certificates in your new working directory. For example: certificates.
  24. Install the AWS IoT SDK for Python v2 by running the following from the command line.
  25. To test the data pipeline, run the following command:

You can see the payload in the following screenshot.

Finally, the data is delivered to the specified S3 bucket in the prefix structure.

The data of the files is as follows:

  • {“TIMESTAMP”:”2022-03-20 00:00″,”LOCATION_ID”:”B-101″,”CO”:2.6,”SO2″:62,”NO2″:57}
  • {“TIMESTAMP”:”2022-03-20 00:05″,”LOCATION_ID”:”B-101″,”CO”:3.9,”SO2″:60,”NO2″:73}

The timestamps show that each file contains data for 5-minute intervals.

With minimal code, we have now ingested the sensor data, created an input stream from the ingested data, and stored the data in an S3 bucket based on the requirements for Lookout for Metrics.

In the following sections, we take a deeper look at the constructs within Lookout for Metrics, and how easy it is to configure these concepts using the Lookout for Metrics console.

Create a detector

A detector is a Lookout for Metrics resource that monitors a dataset and identifies anomalies at a predefined frequency. Detectors use ML to find patterns in data and distinguish between expected variations in data and legitimate anomalies. To improve its performance, a detector learns more about your data over time.

In our use case, the detector analyzes data from the sensor every 5 minutes.

To create the detector, navigate to the Lookout for Metrics console and choose Create detector. Provide the name and description (optional) for the detector, along with the interval of 5 minutes.

Your data is encrypted by default with a key that AWS owns and manages for you. You can also configure if you want to use a different encryption key from the one that is used by default.

Now let’s point this detector to the data that you want it to run anomaly detection on.

Create a dataset

A dataset tells the detector where to find your data and which metrics to analyze for anomalies. To create a dataset, complete the following steps:

  1. On the Amazon Lookout for Metrics console, navigate to your detector.
  2. Choose Add a dataset.
  3. For Name, enter a name (for example, air-quality-dataset).
  4. For Datasource, choose your data source (for this post, Amazon S3).
  5. For Detector mode, select your mode (for this post, Continuous).

With Amazon S3, you can create a detector in two modes:

    • Backtest – This mode is used to find anomalies in historical data. It needs all records to be consolidated in a single file.
    • Continuous – This mode is used to detect anomalies in live data. We use this mode with our use case because we want to detect anomalies as we receive air pollutant data from the air monitoring sensor.
  1. Enter the S3 path for the live S3 folder and path pattern.
  2. For Datasource interval, choose 5 minute intervals.If you have historical data from which the detector can learn patterns, you can provide it during this configuration. The data is expected to be in the same format that you use to perform a backtest. Providing historical data speeds up the ML model training process. If this isn’t available, the continuous detector waits for sufficient data to be available before making inferences.
  3. For this post, we already have historical data, so select Use historical data.
  4. Enter the S3 path of historical_data.json.
  5. For File format, select JSON lines.

At this point, Lookout for Metrics accesses the data source and validates whether it can parse the data. If the parsing is successful, it gives you a “Validation successful” message and takes you to the next page, where you configure measures, dimensions, and timestamps.

Configure measures, dimensions, and timestamps

Measures define KPIs that you want to track anomalies for. You can add up to five measures per detector. The fields that are used to create KPIs from your source data must be of numeric format. The KPIs can be currently defined by aggregating records within the time interval by doing a SUM or AVERAGE.

Dimensions give you the ability to slice and dice your data by defining categories or segments. This allows you to track anomalies for a subset of the whole set of data for which a particular measure is applicable.

In our use case, we add three measures, which calculate the AVG of the objects seen in the 5-minute interval, and have only one dimension, for which pollutants concentration is measured.

Every record in the dataset must have a timestamp. The following configuration allows you to choose the field that represents the timestamp value and also the format of the timestamp.

The next page allows you to review all the details you added and then save and activate the detector.

The detector then begins learning the data streaming into the data source. At this stage, the status of the detector changes to Initializing.

It’s important to note the minimum amount of data that is required before Lookout for Metrics can start detecting anomalies. For more information about requirements and limits, see Lookout for Metrics quotas.

With minimal configuration, you have created your detector, pointed it at a dataset, and defined the metrics that you want Lookout for Metrics to find anomalies in.

Visualize anomalies

Lookout for Metrics provides a rich UI experience for users who want to use the AWS Management Console to analyze the anomalies being detected. It also provides the capability to query the anomalies via APIs.

Let’s look at an example anomaly detected from our air quality data use case. The following screenshot shows an anomaly detected in CO concentration in the air at the designated time and date with a severity score of 93. It also shows the percentage contribution of the dimension towards the anomaly. In this case, 100% contribution comes from the location ID B-101 dimension.

Create alerts

Lookout for Metrics allows you to send alerts using a variety of channels. You can configure the anomaly severity score threshold at which the alerts must be triggered.

In our use case, we configure alerts to be sent to an Amazon Simple Notification Service (Amazon SNS) channel, which in turn sends an SMS. The following screenshots show the configuration details.

You can also use an alert to trigger automations using AWS Lambda functions in order to drive API-driven operations on AWS IoT Core.


In this post, we showed you how easy to use Lookout for Metrics and Kinesis Data Firehose to remove the undifferentiated heavy lifting involved in managing the end-to-end lifecycle of building ML-powered anomaly detection applications. This solution can help you accelerate your ability to find anomalies in key business metrics and allow you focus your efforts on growing and improving your business.

We encourage you to learn more by visiting the Amazon Lookout for Metrics Developer Guide and try out the end-to-end solution enabled by these services with a dataset relevant to your business KPIs.

About the author

Dhiraj Thakur is a Solutions Architect with Amazon Web Services. He works with AWS customers and partners to provide guidance on enterprise cloud adoption, migration, and strategy. He is passionate about technology and enjoys building and experimenting in the analytics and AI/ML space.


Continue Reading


Copyright © 2021 Today's Digital.