An answerer recommendation approach on Stack Overflow using Historical Posts with User Profile Data

Abstract

The exponential growth of community-driven knowledge platforms such as Stack Overflow has created a pressing need for effective recommendation systems that can personalize content to users expertise and interests. This thesis presents a hybrid recommendation approach that integrates semantic topic embeddings, temporal ac tiveness features, and reputation-based contribution metrics into a unified two-tower neural recommendation framework. To address the challenge of high-dimensional and composite tag vocabularies, a pre processing strategywasdevelopedtoadaptpretrainedtopicembeddingstoover10,000 StackOverflow tags, including hyphenated and dot-separated forms. User represen tations were constructed by aggregating embeddings of interacted tags, augmented with features derived from activity patterns and reputation dynamics. Negative sam pling was employed using cosine similarity thresholds to differentiate relevant from irrelevant tags, while candidate sampling ensured balanced evaluation across users. Experimental results across multiple configurations demonstrate that topic embed dings provide a strong semantic baseline, but recommendation quality does not im prove with the addition of activeness and contribution features. The contribution of this paper is an empirical insight that semantic topic signals dominate, and adding activeness and reputational features naively reducesperformance. Thefindingiscon tradictory to the assumptions of other papers, such as IEA [22], on the fact that using different types of features improves the result