Topic Modeling for Personalized Recommendation of Volatile Items

Abstract

One of the major strengths of probabilistic topic modeling is the ability to reveal hidden relations via the analysis of co-occurrence patterns on dyadic observations, such as document-term pairs. However, in many practical settings, the extreme sparsity and volatility of co-occurrence patterns within the data, when the majority of terms appear in a single document, limits the applicability of topic models. In this paper, we propose an efficient topic modeling framework in the presence of volatile dyadic observations when direct topic modeling is infeasible. We show both theoretically and empirically that often-available unstructured and semantically-rich meta-data can serve as a link between dyadic sets, and can allow accurate and efficient inference. Our approach is general and can work with most latent variable models, which rely on stable dyadic data, such as pLSI, LDA, and GaP. Using transactional data from a major e-commerce site, we demonstrate the effectiveness as well as the applicability of our method in a personalized recommendation system for volatile items. Our experiments show that the proposed learning method outperforms the traditional LDA by capturing more persistent relations between dyadic sets of wide and practical significance.