Monthly Archives: August 2021

Amazon Personalize enables developers to build applications with the same machine learning (ML) technology used by Amazon.com for real-time personalized recommendations—no ML expertise required. Amazon Personalize makes it easy for developers to build applications capable of delivering a wide array of personalization experiences, including specific product recommendations, personalized product re-ranking, and customized direct marketing. Besides applications in retail and ecommerce, other common use cases for Amazon Personalize include recommending videos, blog posts, or newsfeeds based on users’ activity history.

What if you wanted to recommend users of common interest to connect with each other? As the pandemic pushes many of our normal activities virtual, connecting with people is a greater challenge than ever before. This post discusses how 6Connex turned this challenge into an opportunity by harnessing Amazon Personalize to elevate their “user matchmaking” feature.

6Connex and event AI

6Connex is an enterprise virtual venue and hybrid events system. Their cloud-based product portfolio includes virtual environments, learning management, and webinars. Attendee experience is one of the most important metrics for their success.

Attendees have better experiences when they are engaged not only with the event’s content, organizers, and sponsors, but also when making connections with other attendees. Engagement metrics are measured and reported for each attendee activity on the platform, as well as feedback from post-event surveys. The goal is to make the events system more attendee-centric by not only providing personalized content and activity recommendations, but also making matchmaking suggestions for attendees based on similar interests and activity history. By adding event AI features to their platform, 6Connex fosters more meaningful connections between attendees, and keeps their attendees more engaged with a personalized event journey.

Implementation and solution architecture

6Connex built their matchmaking solution using the related items recipe (SIMS) of Amazon Personalize. The SIMS algorithm uses collaborative filtering to recommend items that are similar to a given item. The novelty of 6Connex’s approach lies in the reverse mapping of users and items. In this solution, event attendees are items in Amazon Personalize terms, and content, meeting rooms, and so on are users’ in Amazon Personalize terms.

When a platform user joins a meeting room or views a piece of content, an interaction is created. To increase the accuracy of interaction types, also known as event_type, you can add logic to only count as an interaction when a user stays in a meeting room for at least a certain amount of time. This eliminates accidental clicks and cases when users join but quickly leave a room due to lack of interest.

As many users interact with the platform during a live event, interactions are streamed in real time from the platform via Amazon Kinesis Data Streams. AWS Lambda functions are used for data transformation before streaming data directly to Amazon Personalize through an event tracker. This mechanism also enables Amazon Personalize to adjust to changing user interest over time, allowing recommendations to adapt in real time.

After a model is trained in Amazon Personalize, a fully managed inference endpoint (campaign) is created to serve real-time recommendations for 6Connex’s platform. To answer the question “for each attendee, who are similar attendees?”, 6Connex’s client-side application queries the GetRecommendations API with a current user (represented as an itemId). The API response provides recommended connections because they have been identified as similar by the Amazon Personalize.

Due to its deep learning capabilities, Amazon Personalize requires at least 1,000 interaction data points before training the model. At the start of a live event, there aren’t enough interactions, therefore a rules engine is used at the beginning of an event to provide the initial recommendations prior to gathering 1000 data points. The following table shows the three main phases of an event where connection recommendations are generated.

For a high-level example architecture, see the following diagram.

The following are the steps involved in the solution architecture:

1. 6Connex web application calls the GetRecommendations API to retrieve recommended connections.
2. A matchmaking Lambda function retrieves recommendations.
3. Until the training threshold of 1,000 interaction data points is reached, the matchmaking function uses a simple rules engine to provide recommendations.
4. Recommendations are generated from Amazon Personalize and stored in Amazon ElastiCache. The reason for caching recommendations is to improve response performance while reducing the number of queries on the Amazon Personalize API. When new recommendations are requested, or when the cache expires (expiration is set to every 15 minutes), recommendations are pulled from Amazon Personalize.
5. New user interactions are ingested in real time via Kinesis Data Streams.
6. A Lambda function consumes data from the data stream, performs data transformation, persists the transformed data to Amazon Simple Storage Service (Amazon S3) and related metadata to Amazon DynamoDB, and sends the records to Amazon Personalize via the PutEvents API.
7. AWS Step Functions orchestrates the process for creating solutions, training, retraining, and several other workflows. More details on the Step Functions workflow are in the next section.
8. Amazon EventBridge schedules regular retraining events during the virtual events. We also use EventBridge to trigger batch recommendations after the virtual events are over and when the contents are served to end users on demand.
9. Recommendations are stored in DynamoDB for use during the on-demand period and also for future analysis.

Adoption of MLOps

It was crucial for 6Connex to quickly shift from a rules-based recommender engine to personalized recommendations using Amazon Personalize. To accelerate this shift and hydrate the interactions dataset, 6Connex infers interactions not only from content engagement, but also from other sources such as pre-events questionnaires. This is an important development that increased the speed to when users start receiving ML-based recommendations.

More importantly, the adoption of Amazon Personalize MLOps enabled 6Connex to automate and accelerate the transition from rule-based recommendations to personalized recommendations using Amazon Personalize. After the minimum threshold for data is met, Step Functions loads data into Amazon Personalize and manages the training process.

The following diagram shows the MLOps pipeline for the initial loading of data, training solutions, and deploying campaigns.

6Connex created their MLOps solution based on the Amazon Personalize MLOps reference solution to automate this process. There are several Step Functions workflows that offload long-running processes such loading batch recommendations in DynamoDB, retraining Amazon Personalize solutions, and cleaning up after an event is complete.

With Amazon Personalize and MLOps pipelines, 6Connex brought an AI solution to market in less than half the time it would have taken to develop and deploy their own ML infrastructure. Moreover, these solutions reduced the cost of acquiring data science and ML expertise. As a result, 6Connex realized a competitive advantage through AI-based personalized recommendations for each individual user.

Based on the success of this engagement, 6Connex plans to expand its usage of Amazon Personalize to provide content-based recommendations in the near future. 6Connex is looking forward to expanding the partnership not only in ML, but also in data analytics and business intelligence to serve the fast-growing hybrid event market.

Conclusion

With a well-designed MLOps pipeline and some creativity, 6Connex built a robust recommendation engine using Amazon Personalize in a short amount of time.

Do you have a use case for a recommendation engine but are short on time or ML expertise? You can get started with Amazon Personalize using the Developer Guide, as well as a myriad of hands-on resources such as the Amazon Personalize Samples GitHub repo.

If you have any questions on this matchmaking solution, please leave a comment!

About the Author

Shu Jackson is a Senior Solutions Architect with AWS. Shu works with startup customers helping them design and build solutions in the cloud, with a focus on AI/ML.

Luis Lopez Soria is a Sr AI/ML specialist solutions architect working with the Amazon Machine Learning team. He works with AWS customers to help them adopt machine learning on a large scale. He enjoys playing sports, traveling around the world, and exploring new foods and cultures.

Read More

This post was cowritten by Ziv Pollak, Machine Learning Team Lead, and Alex Thoreux, Web Analyst at Clearly.

A pioneer in online shopping, Clearly launched their first site in 2000. Since then, they’ve grown to become one of the biggest online eyewear retailers in the world, providing customers across Canada, the US, Australia and New Zealand with glasses, sunglasses, contact lenses, and other eye health products. Through their Mission to eliminate poor vision, Clearly strives to make eyewear affordable and accessible for everyone. Creating an optimized platform is a key part of this wider vision.

Predicting future sales is one of the biggest challenges every retail organization has – but it’s also one of the most important pieces of insight. Having a clear and reliable picture of predicted sales for the next day or week allows your company to adjust its strategy and increase the chances of meeting its sales and revenue goals.

We’ll talk about how Clearly built an automated and orchestrated forecasting pipeline using AWS Step Functions, and used Amazon Forecast APIs to train a machine learning (ML) model and predict sales on a daily basis for the upcoming weeks and months.

With a solution that also collects metrics and logs, provides auditing, and is invoked automatically, Clearly was able to create a serverless, well-architected solution in just a few weeks.

The challenge: Detailed sales forecasting

With a reliable sales forecast, we can improve our marketing strategy, decision-making process, and spend, to ensure successful operations of the business.

In addition, when a diversion between the predicted sales numbers and the actual sales number occurs, it’s a clear indicator that something is wrong, such as an issue with the website or promotions that may not be working properly. From there, we can problem solve the issues and address them in a timely manner.

For forecasting sales, our existing solution was based on senior members of the marketing team building manual predictions. Historical data was loaded into an Excel sheet and predictions were made using basic forecasting functionality and macros. These manual predictions were nearly 90% accurate and took 4–8 hours to complete, which was a good starting point, but still not accurate enough to confidently guide the marketing team’s next steps.

In addition, testing “what-if” future scenarios was difficult to implement because we could only perform the predictions for the following months, without further granularity such as weeks and days.

Having a detailed sales forecast allows us to identify situations where money is being lost due to outages or other technical or business issues. With a reliable and accurate forecast, when we see that actual sales aren’t meeting expected sales, we know there is an issue.

Another major challenge we faced was the lack of a tenured ML team – all members had been with the company less than a year when the project kicked off.

Overview of solution: Forecast

Amazon Forecast is a fully managed service that uses ML to deliver highly accurate forecasts. After we provided the data, Forecast automatically examined it, identified what was meaningful, and produced a forecasting model capable of making predictions on our different lines of products and geographical locations to deliver the most accurate daily forecasts. The following diagram illustrates our forecasting pipeline.

To operationalize the flow, we applied the following workflow:

1. Amazon EventBridge calls the orchestration pipeline daily to retrieve the predictions.
2. Step Functions help manage the orchestration pipeline.
3. An AWS Lambda function calls Amazon Athena APIs to retrieve and prepare the training data, stored on Amazon Simple Storage Service (Amazon S3).
4. An orchestrated pipeline of Lambda functions uses Forecast to create the datasets, train the predictors, and generate the forecasted revenue. The forecasted data is saved in an S3 bucket.
5. Amazon Simple Notification Service (Amazon SNS) notifies users when a problem occurs during the forecasting process or when the process completes successfully.
6. Business analysts build dashboards on Amazon QuickSight, which queries the forecast data from Amazon S3 using Athena.

We chose to work with Forecast for a few reasons:

• Forecast is based on the same technology used at Amazon.com, so we have a lot of confidence in the tool’s capabilities.
• The ease of use and implementation allowed us to quickly confirm we have the needed dataset to produce accurate results.
• Because the Clearly ML team was less than 1 year old, a fully managed service allowed us to deliver this project without needing deep technical ML skills and knowledge.

Data sources

Finding the data to use for this forecast, while making sure it was clear and reliable, was the most important element in our ability to generate accurate predictions. We ended up using the following datasets, training the model on 3 years of daily data:

• Web traffic.
• Number of orders.
• Average order value.
• Conversion rate.
• New customer revenue.
• Marketing spend.
• Marketing return on advertisement spend.
• Promotions.

To create the dataset, we went through many iterations, changing the number of data sources until the predictions reach our benchmark of at least 95% accuracy.

Dashboard and results

Writing the prediction results into our existing data lake allows us to use QuickSight to build metrics and dashboards for the senior-level managers. This enables them to understand and use these results when making decisions on the next steps needed to meet our monthly marketing targets.

We were able to present the forecast results on two levels, starting with overall business performance and then going deeper into performance per each line of business (contacts and glasses). For those three cases (overall, contacts, glasses) we presented the following information:

• Predicted revenue vs. target – This allows the marketing team to understand how we’re expected to perform this month, compared to our target, if they take no additional actions. For example, if we see that the projected sales don’t meet our marketing goals, we need to launch a new marketing campaign. The following screenshot shows an example analysis with a value of -17.47%, representing the expected total monthly revenue vs. the target.
  
• Revenue performance compared to predictions over the last month – This graph shows that the predicted revenue is within the forecasted range, which means that the predictions are accurate. The following example graph shows high bound, revenue, and low bound values.
  
• Month to date revenue compared to weekly and monthly forecasts – The following example screenshot shows text automatically generated by QuickSight that indicates revenue-related KPIs.
  

Thanks to Forecast, Clearly now has an automated pipeline that generates forecasts for daily and weekly scenarios, reaching or surpassing our benchmarks of 97%, which is an increase of 7.78% from a process that was done manually and was limited to longer periods.

Now, daily forecasts for weekly and monthly revenue take only 15 minutes in data gathering and preparation, with the forecasting process taking close to 15 minutes to complete on a daily basis. This is a huge improvement from 4-8 hours with the manual process, which could only perform predictions for the whole month.

With more granularity and better accuracy, our marketing team has better tools to act faster on discrepancies and create prediction scenarios on campaigns could achieve better revenue results.

Conclusion

Effective and accurate prediction of customer future behavior is one of the biggest challenges in ML in retail today, and having a good understanding of our customers and their behavior is vital for business success. Forecast provided a fully managed ML solution to easily create an accurate and reliable prediction with minimal overhead. The biggest benefit we get with these predictions is that we have accurate visibility of what the future will look like and can change it if it doesn’t meet our targets.

In addition, Forecast allows us to predict what-if scenarios and their impact on revenue. For example, we can project the overall revenue until the end of the month, and with some data manipulation we can also predict what will happen if we launch a BOGO (buy one, get one free) campaign next Tuesday.

“With leading ecommerce tools like Virtual Try On, combined with our unparalleled customer service, we strive to help everyone see clearly in an affordable and effortless manner—which means constantly looking for ways to innovate, improve, and streamline processes. Effective and accurate prediction of customer future behavior is one of the biggest challenges in machine learning in retail today. In just a few weeks, Amazon Forecast helped us accurately and reliably forecast sales for the upcoming week with over 97% accuracy, and with over 90% accuracy when predicting sales for the following month.”

– Dr. Ziv Pollak, Machine Learning Team Leader.

For more information about how to get started building your own MLOps pipelines with Forecast, see Building AI-powered forecasting automation with Amazon Forecast by applying MLOps, and for other use cases, visit the AWS Machine Leaning Blog.

The content and opinions in this post are those of the third-party author and AWS is not responsible for the content or accuracy of this post.

About the Author

Dr Ziv Pollak is an experienced technical leader who transforms the way organizations use machine learning to increase revenue, reduce costs, improve customer service, and ensure business success. He is currently leading the Machine Learning team at Clearly.

Alex Thoreux is a Jr Web Analyst at Clearly who built the forecasting pipeline, as well as other ML applications for Clearly.

Fernando Rocha is a Specialist SA. As Clearly’s Solutions Architect, he helps them build analytics and machine learning solutions on AWS.

Read More

Amazon Comprehend is a natural language processing (NLP) service that provides APIs to extract key phrases, contextual entities, events, sentiment from unstructured text, and more. Entities refer to things in your document such as people, places, organizations, credit card numbers, and so on. But what if you want to add entity types unique to your business, like proprietary part codes or industry-specific terms? Custom entity recognition (CER) in Amazon Comprehend enables you to train models with entities that are unique to your business in just a few easy steps. You can identify almost any kind of entity, simply by providing a sufficient number of details to train your model effectively.

Training an entity recognizer from the ground up requires extensive knowledge of machine learning (ML) and a complex process for model optimization. Amazon Comprehend makes this easy for you using a technique called transfer learning to help build your custom model. Internally, Amazon Comprehend uses base models that have been trained on data collected by Amazon Comprehend and optimized for the purposes of entity recognition. With this in place, all you need to supply is the data. ML model accuracy is typically dependent on both the volume and quality of data. Getting good quality annotation data is a laborious process.

Until today, you could train an Amazon Comprehend custom entity recognizer with only 1,000 documents and 200 annotations per entity. Today, we’re announcing that we have improved underlying models for the Amazon Comprehend custom entity API by reducing the minimum requirements to train the model. Now, with as few as 250 documents and 100 annotations per entity (also referred to as shots), you can train Amazon Comprehend CER models to predict entities with greater accuracy. To take advantage of the updated performance offered by the new CER model framework, you can simply retrain and deploy improved models.

To illustrate the model improvements, we compare the result of previous models with that of the new release. We selected a diverse set of entity recognition datasets across different domains and languages from the open-source domain to showcase the model improvements. In this post, we walk you through the results from our training and inference process between the previous CER model version and the new CER model.

Datasets

When you train an Amazon Comprehend CER model, you provide the entities that you want the custom model to recognize, and the documents with text containing these entities. You can train Amazon Comprehend CER models using entity lists or annotations. Entity lists are CSV files that contain the text (a word or words) of an entity example from the training document along with a label, which is the entity type that the text is categorized as. With annotations, you can provide the positional offset of entities in a sentence along with the entity type being represented. When you use the entire sentence, you’re providing the contextual reference for the entities, which increases the accuracy of the model you’re training.

We selected the annotations option for labeling our entities because the datasets we selected already contained the annotations for each of the entity types represented. In this section, we discuss the datasets we selected and what they describe.

CoNLL

The Conference on Computational Natural Language Learning (CoNLL) provides datasets for language-independent (doesn’t use language-specific resources for performing the task) named entity recognition with entities provided in English, Spanish, and German. Four types of named entities are provided in the dataset: persons, locations, organizations, and names of miscellaneous entities that don’t belong to the previous three types.

We used the CoNLL-2003 dataset for English, and the CoNLL-2002 dataset for Spanish languages for our entity recognition training. We ran some basic transformations to convert the annotations data to a format that is required by Amazon Comprehend CER. We converted the entity types from their semantic notation to actual words they represent, such as person, organization, location, and miscellaneous.

SNIPS

The SNIPS dataset was created in 2017 as part of benchmarking tests for natural language understanding (NLU) by Snips. The results from these tests are available in the 2018 paper “Snips Voice Platform: an embedded Spoken Language Understanding system for private-by-design voice interfaces” by Coucke, et al. We used the GetWeather and the AddToPlaylist datasets for our experiments. The entities for the GetWeather dataset we considered are timerange, city, state, condition_description, country, and condition_temperature. For AddToPlaylist, we considered the entities artist, playlist_owner, playlist, music_item, and entity_name.

Sampling configuration

The following table represents the dataset configuration for our tests. Each row represents an Amazon Comprehend CER model that was trained, deployed, and used for entity prediction with our test dataset.

For more details on how to format data to create annotations and entity lists for Amazon Comprehend CER, see Training Custom Entity Recognizers. We created a benchmarking approach based on the sampling configuration for our tests, and we discuss the results in the following sections.

Benchmarking process

As shown in the sampling configuration in the preceding section, we trained a total of 12 models, with four models each for CoNLL English and Spanish datasets with varying document and annotation configurations, three models for the SNIPS-GetWeather dataset, again with varying document and annotation configurations, and one model with the SNIPS-AddToPlaylist dataset, primarily to test the new minimums of 250 documents and 100 annotations per entity.

Two inputs are required to train an Amazon CER model: entity representations and the documents containing these entities. For an example of how to train your own CER model, refer to Setting up human review of your NLP-based entity recognition models with Amazon SageMaker Ground Truth, Amazon Comprehend, and Amazon A2I. We measure the accuracy of our models using metrics such as F1 score, precision, and recall for the test set at training and the blind test set at inference. We run subsequent inference on these models using a blind test dataset of documents that we set aside from our original datasets.

Precision indicates how many times the model makes a correct entity identification compared to the number of attempted identifications. Recall indicates how many times the model makes a correct entity identification compared to the number of instances of that the entity is actually present, as defined by the total number of correct identifications (true positives) and missed identifications (false negatives). F1 score indicates a combination of the precision and recall metrics, which measures the overall accuracy of the model for custom entity recognition. To learn more about these metrics, refer to Custom Entity Recognizer Metrics.

Amazon Comprehend CER provides support for both real-time endpoints and batch inference requirements. We used the asynchronous batch inference API for our experiments. Finally, we calculated the F1 score, precision, and recall for the inference by comparing what the model predicted with what was originally annotated for the test documents. The metrics are calculated by doing a strict match for the span offsets, and a partial match isn’t considered nor given partial credit.

Results

The following tables document the results from our experiments we ran using the sampling configuration and the benchmarking process we explained previously.

Previous limits vs. new limits

The limits have reduced from 1,000 documents and 200 annotations per entity for CER training in the previous model to 250 documents and 100 annotations per entity in the improved model.

The following table shows the absolute improvement in F1 scores measured at training, between the old and new models. The new model improves the accuracy of your entity recognition models even when you have a lower count of training documents.

Next, we report the evaluation on a blind test set that was split before the training process from the dataset.

Overall, we observe an improvement in F1 scores with the new model even with half the number of annotations provided, as seen in the preceding table.

Continued improvement with more data

In addition to the improved F1 scores at lower limits, we noticed a trend where the new model’s accuracy measured with the blind test dataset continued to improve as we trained with increased annotations. For this test, we considered the SNIPS GetWeather and AddToPlaylist datasets.

The following graph shows a distribution of absolute blind test F1 scores for models trained with different datasets and annotation counts.

We generated the following metrics during training and inference for the SNIPS-AddToPlaylist model trained with 250 documents in the new Amazon Comprehend CER model.

SNIPS-AddToPlaylist metrics at training time

SNIPS-AddToPlaylist inference metrics with blind test dataset

Conclusion

In our experiments with the model improvements in Amazon Comprehend CER, we observe accuracy improvements with fewer annotations and lower document volumes. Now, we consistently see increased accuracy across multiple datasets even with half the number of data samples. We continue to see improvements to the F1 score as we trained models with different dataset sampling configurations, including multi-lingual models. With this updated model, Amazon Comprehend makes it easy to train custom entity recognition models. Limits have been lowered to 100 annotations per entity and 250 documents for training while offering improved accuracy with your models. You can start training custom entity models on the Amazon Comprehend console or through the API.

About the Authors

Prem Ranga is an Enterprise Solutions Architect based out of Houston, Texas. He is part of the Machine Learning Technical Field Community and loves working with customers on their ML and AI journey. Prem is passionate about robotics, is an Autonomous Vehicles researcher, and also built the Alexa-controlled Beer Pours in Houston and other locations.

Chethan Krishna is a Senior Partner Solutions Architect in India. He works with Strategic AWS Partners for establishing a robust cloud competency, adopting AWS best practices and solving customer challenges. He is a builder and enjoys experimenting with AI/ML, IoT and Analytics.

Mona Mona is an AI/ML Specialist Solutions Architect based out of Arlington, VA. She helps customers adopt machine learning on a large scale. She is passionate about NLP and ML Explainability areas in AI/ML.

Read More