BI platform with AI and computer vision for a fashion retailer

Solution

Having delivered several successful retail BI solutions operating around the globe, our team used their expertise to build a centralized BI platform. The solution gathers and analyzes data in a near-real-time mode and provides accurate customer data for further website and mobile app personalization. The delivered solution processes clickstream data, mobile data, server events, and data on email campaign engagement.

Platform Load
10 TB of data processed	8 million trackable events on the website and in the mobile application	3.5 million emails in the system	30 thousand events per minute on the website (on average)

* Disclaimer: According to the Non-Disclosure Agreement, we cannot reveal the screenshots of the real system. Here we provide similar screenshots created to present an idea of the solution developed by Itransition

Solution architecture

Architecture Outlines
Event Tracking Layer	Tracking events from different sources (web, mobile, server, etc.)
Event Collecting Layer	Collecting both tracked events and operational data from the backend systems (ecommerce, CRM, etc.)
Event Processing Layer	Loading, normalizing, filtering, validating, and transforming collected data
Data Storage Layer	Storing data that is optimal for statistical analysis and machine learning
Data Consumption Layer	Building data marts with view and integration APIs to access the data
Integration Layer	Gathering data from third-party sources via connectors, adaptors, and ETL jobs

The platform supports two major user roles: the retailers’ marketing team and subscribed members.

The marketing team can:

• Get data on user activities in every channel (website, mobile application, emails, surveys, etc.)
• Review more than 100 types of custom reports on the website usage and views, product orders, etc.
• Create ad-hoc queries based on the collected data when needed

Subscribed members can:

• Enjoy their own personalized versions of the website, the web application, emails, and push notifications, complete with banners, site categories, and product types tailored to their preferences based on previous views and purchases

AI-powered recommendation engine

To help our customer provide personalized experience to their online visitors and improve the accuracy of recommendations, we developed a recommendation engine using collaborative filtering. The collaborative filtering algorithm operates on implicit user feedback such as purchases, views, clicks, and other metrics coming from the website, mobile app, or emails.

We chose this methodology as it scales easily to process terabytes of data, so we could run it on more than ten machines at a time.

Our team opted for the alternating least squares (ALS) algorithm, initially used during the Netflix Prize challenge, as it met the project’s scalability and performance criteria. We also used a random forest regressor to predict product scores, calculated as the combination of item clicks, add-to-cart button clicks, and purchases.

Taking into consideration more than 20 million platform users and 9 million SKUs, we selected Apache Spark as the main ETL platform and Spark MLLib, based on the ALS algorithm, to run ML pipelines in production.

Computer vision for product image recognition

Our team also applied computer vision for a number of internal tasks:

• Automated product attribute detection in product images, including color, clothing type, patterns, neckline, sleeve length, etc.
• Automated color detection based on pre-filtering to remove irrelevant image parts, such as background and skin. We used edge detection, adaptive thresholding, and clustering in the CIELAB color space to reduce the number of color tones used in basic colors.
• Multi-attribute image classification. We used a ready-made ResNet-50 convolutional neural network built with Keras and TensorFlow, which was pre-trained on ImageNet with a set of over 100,000 product images.
• Image similarity search developed with the same architecture as the multi-attribute classification. We added an additional embedding layer to the CNN to enable the nearest neighbor search (NNS).

Technologies & tools

One of the project’s key goals was to make the solution painlessly scalable while reducing the overall infrastructure costs. Initially, the solution was based on the Apache stack, including Kafka, Storm, and Hive Streaming. Together with the customer, we decided to host the solution on the Amazon Web Services (AWS) and build it with a serverless architecture.

This approach allowed the development team to make the solution easily scalable and fault-tolerant, as well as ensure the auto-scaling of the resources in place, minimize their idle time, and, as a result, reduce the infrastructure costs.

Itransition developed the data collection layer based on the Hortonworks Data Platform (HDP).

We also optimized costs associated with the Amazon DynamoDB management. With the application’s traffic alternating substantially during the day, it was very difficult to forecast and control it in order to effectively use the provisioned capacity mode. Therefore, we switched to the Amazon DynamoDB on-demand pricing with no planned capacity boundaries. It allowed the customer to avoid situations when the data storage capacity was under- or over-provisioned. The pay-per-request pricing helped the customer cut down the database management costs by almost 50 percent.

Another challenge that our team faced during this retail BI project was to reduce costs associated with training dockerized deep learning (DL) models. For example, the AWS ML platform SageMaker is expensive to train DL models on graphic processing unit (GPU) instances. It doesn’t support Amazon EC2 Spot Instances, which bring the benefit of spare computing capacities available in the AWS Cloud. With EC2 instances, it’s possible to rent virtual servers (instances) for the needed amount of time and then request an on-demand instance. To overcome this pitfall, our team developed a custom framework that allowed building TensorFlow-based models, dockerize them and deploy to EC2 Spot. As a result, we were able to save around 50 percent of costs.

Itransition also integrated the platform with third-party solutions and tools, including:

• Salesforce Marketing Cloud to manage email campaigns
• LiveIntent to support real-time advertising
• SurveyGizmo to create surveys and make informed decisions based on the results
• Evergage to collect and analyze user behavior in real time