Fashion Discovery Website

Challenge

We started developing this project, from its very beginning, in 2015. Over the years, the website has undergone various implementatios and upgrades.

The biggest challenge was to improve the internal workflow for the update/inclusion of products in the marketplace.

As StylishCircle is an ever growing platform, and from early stages of the upgrade process the Infobest team has used predictive analytics methods combined with data science techniques, the next step in the development of the project is turning to Machine Learning and Artificial Intelligence’s way.

Solution

We are integrating machine learning in our processes to reduce the administrative effort required to manage tens of thousands of clothing items. More exactly, we use machine learning to classify products into predefined sub-categories. To do that, we built a multi-modal model for commercial product classification, that combines features extracted by multiple pre-trained models (VGG19, Distilbert Base Multilingual Cased), using simple fusion techniques. The input comprises of images / texts / genders and the output is represented by a probability distribution. At this moment, the multi-modal model can identify 74 sub-categories. Regarding the training process, this can be done both locally and in AWS SageMaker.

Going further, model is served with the help of AWS. The end user has the possibility of uploading csv files in Amazon S3 and download the results from the same place. Taking into account the fact that sometimes the inference process might take long, we decided to use AWS Fargate. AWS Fargate is a serverless, pay-as-you-go compute engine. In addition, we also use AWS Lambda, Amazon Elastic Container Registry (ECR).

In the future, we aim to improve the accuracy of the multi-modal model by expanding the current dataset and finding the proper hyperparameters.