PRODUCT RECOMMENDATION SYSTEM ON E-COMMERCE USING MULTI-TASK TWO TOWER RETRIEVAL MODEL

• Main Author: Ahmad Zuhri, Hamdi
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/55523
• Institution: Institut Teknologi Bandung

As the trend of e-commerce is positive time by time, a product recommendation system should be improved to recommend better product for users while maintaining its scalability. In an e-commerce platform, there are various user or product features to enrich the context of learning model and there are various interactions to utilize multi-task learning for recommendation model. Lately, Deep Neural Network (DNN) seems to work well for recommendation task and it is possible to use as much as feature since DNN is good for high complexity pattern. In particular, the two tower neural network has been demonstrated as an effective structure to build a scalable recommendation system. In the other side, multi-task neural-based learning has been also successfully used in many largescale real problem applications. Therefore, this paper proposed a product recommendation system on e-commerce using a multi-task two tower retrieval model. The objective of this experiment is to build a two tower based retrieval model which is able to recommend a personalized list of products that are more likely to be bought by users. The experiment shows that with proper features and proper supporting tasks, two tower retrieval model can give better performance in recommendation. In this research, brand or category feature and add-to-cart event are the components that made two tower model got better performance. Performance improvement from baseline for recall@100 metric reached 0.9% for brand feature addition, reached 1.3% for utilizing add-to-cart event, and reached 1.6% for using both.