Authors: Mohammad Ali, Jin Xie, Wu Wenhuan
Abstract: The machine behind e-commerce recommendation systems is often based on Collaborative Filtering (CF) algorithms which are highly dependent on numerical star ratings ignoring the fact that user review text is a rich source of semantic information. This constraint increases the issue of data sparsity and lowers the quality of the recommendations. We suggest SA-BTM-CF, a Sentiment-Aware Collaborative Filtering system, which encompasses the Biterm Topic Model (BTM) with a conventional CF. The suggested methodology is a latent topic extractor of review literature and a fine-grained sentiment score generator of latent topic models which are employed to build an enriched user-item rating matrices. Personalized product suggestions are then created by computing user similarity on these sentiment-augmented ratings. Large scale shampoo review data on the JD.com site is experimented with and experimentally validates that SA-BTM-CF can be used to effectively reduce sparsity in data, and it can be used to obtain better recommendations than current baseline algorithms.
DOI: https://doi.org/10.5281/zenodo.23310790
