Enhancing Topic Extraction in Recommender Systems with Entropy Regularization
This work addresses the issue of poor explainability in recommender systems for users and developers, though it is incremental as it builds on existing probabilistic matrix factorization models.
The paper tackled the problem of low keyword coherence in topic extraction for recommender systems by introducing entropy regularization, resulting in a significant improvement in topic coherence measured by cosine similarity on word embeddings while maintaining competitive performance on the primary task.
In recent years, many recommender systems have utilized textual data for topic extraction to enhance interpretability. However, our findings reveal a noticeable deficiency in the coherence of keywords within topics, resulting in low explainability of the model. This paper introduces a novel approach called entropy regularization to address the issue, leading to more interpretable topics extracted from recommender systems, while ensuring that the performance of the primary task stays competitively strong. The effectiveness of the strategy is validated through experiments on a variation of the probabilistic matrix factorization model that utilizes textual data to extract item embeddings. The experiment results show a significant improvement in topic coherence, which is quantified by cosine similarity on word embeddings.