Toward Personalized Product Search for eCommerce Sites: A Case Study in Yahoo! Taiwan
Yi-An Chen, Yahoo! Taiwan
Yahoo Taiwan Shopping is the B2C eCommerce site. It hosts more than 500k products. Based on our statistics, there are nearly 50k queries have more than 500 items re- turned. However, users click only few of those items and also use paging buttons heavily to search products further. Therefore surfacing up products relevant to user to high rank is crucial for Yahoo Taiwan Shopping. Conventional strategy of optimizing search relevance focuses on improving text relevance; however in an eCommerce site, there are often numerous products with high text relevance for a query while only a few match the user’s intention or requirement since each user has his/her preferences - preferred brand, beloved color, and acceptable price range, etc., Therefore, relevance could be improved by personalizing the search result. We proposed a personalized search optimization method for Yahoo Taiwan Shopping. We suggest that besides text-based relevance, there are still two important factors influencing whether a product is relevant to the user. The first one is if the product fits in the user’s preferences. We proposed an algorithm to model a user’s preferences and calculate the affinities between a product and the user. The second one is the popularity of the product itself. Therefore, web analytic measurements, such as historical conversion and click through rate, greatly correlate with a product’s click through rate or probability of conversion, are integrated with affinities together as features in our training model. GBDT is employed to build the ranking model to maximize clicks. We take text relevance based approach as baseline and conduct both offline and online experiments. The personalized ranking model outperforms text relevance based approach in both experiments.