Personalized Web Search Ranking CS 229 : Project Report

In this project, we investigate new approaches to personalize web search result pages for users using anonymized search logs with scant data per user. We investigate modelling users into groups and personalize the user’s SERP based on the response of the users’ ‘peers’ towards the documents in the SERP. For evaluation, we compare our personalization strategy against the popular ranking strategy of sorting of documents in the SERP based on overall number of clicks. We quickly realize that personalizing every SERP makes our algorithm perform worse. When then try to learn when to personalize a SERP using a SVM on our chosen features. Although we do not achieve high tranining/test accuracy in learning when to personalize, the accuracy we achieve is sufficient to beat the click sorted ranking scheme.