Toward Improving Similar Item Recommendation for a C2C Marketplace

Abstract

This study stems from actual challenges that we face in building a recommender system at Bukalapak. Our existing recommendation system, which is an item-to-item collaborative filtering based on co-view method has helped users during their shopping journey by offering them similar items recommendation in the product detail page. However, our baseline system has not utilized the information on how the user perceives the recommendation offered in our recommendation widget. This paper discusses how we improve our recommendation system with point-wise Learning-to-Rank (LTR) method. We also present specific features that support the LTR performance. Finally, we evaluate the LTR model by offline evaluation metrics and online A/B test, where we compare our baseline by using the model to rerank the candidates item before presenting them as recommendation sets. By calculating the Click Through Rate (CTR), it shows that the LTR method can outperform the baseline by 1.63%.