Creating a web page recommendation system for Haystack

Abstract

The driving goal of this thesis was to create a web page recommendation system for Haystack, capable of tracking a user's browsing behavior and suggesting new, interesting web pages to read based on the past behavior. However, during the course of this thesis, 3 salient subgoals were met. First, Haystack's learning framework was unified so that, for example, different types of binary classifiers could be used with black box access under a single interface, regardless of whether they were text learning algorithms or image classifiers. Second, a tree learning module, capable of using hierarchical descriptions of objects and their labels to classify new objects, was designed and implemented. Third, Haystack's learning framework and existing user history faculties were leveraged to create a web page recommendation system that uses the history of a user's visits to web pages to produce recommendations of unvisited links from user-specified web pages. Testing of the recommendation system suggests that using tree learners with both the URL and tabular location of a web page's link as taxonomic descriptions yields a recommender that significantly outperforms traditional, text-based systems.