Deciphering user purchase preferences, their likes and dislikes is a very tricky task even for humans, making its automation a very complex job. This research work augments heuristic-driven user interest profiling with reviewer credibility analysis and fine-grained feature sentiment analysis to devise a robust recommendation methodology. The proposed credibility, interest and sentiment enhanced recommendation (CISER) model has five modules namely candidate feature extraction, reviewer credibility analysis, user interest mining, candidate feature sentiment assignment and recommendation module. Review corpus is given as an input to the CISER model. Candidate feature extraction module uses context and sentiment confidence to extract features of importance. To make our model robust to fake and unworthy reviews and reviewers, reviewer credibility analysis proffers an approach of associating expertise, trust and influence scores with reviewers to weigh their opinion according to their credibility. The user interest mining module uses aesthetics of review writing as heuristics for interest-pattern mining. The candidate feature sentiment assignment module scores candidate features present in review based on their fastText sentiment polarity. Finally, the recommendation module uses credibility weighted sentiment scoring of user preferred features for purchase recommendations. The proposed recommendation methodology harnesses not only numeric ratings, but also sentiment expressions associated with features, customer preference profile and reviewer credibility for quantitative analysis of various alternative products. The mean average precision (MAP@1) for CISER is 93% and MAP@3 is 49%, which is better than current state-of-the-art systems.

No abstract

The scarcity of dependable product descriptions and limited emotion unmasking capabilities of user-ratings compromise the accuracy of content-based filtering (CBF) systems. This work puts forward a sentiment-enhanced content-based recommender system (SEC-Rec). The model has four modules, namely key feature extraction module, feature sentiment analysis module, recommendation module, and rating prediction module. Key feature extraction module uses hybrid of RAKE and TextRank to uncover key product features. The authors propose a hybridized model HSVADER (Hybrid SVM and VADER) for feature sentiment evaluation. The recommendation module combines sentiment and similarity for robust product ranking strategy. The practical benefits of SEC-Rec are demonstrated using Amazon Camera dataset, and the results are compared to the state of the art. The rating prediction module uses key feature sentiment score to estimate the overall user-rating resolving the multi-criteria decision-making issue. The RMSE value obtained ascertains the effectiveness of the approach compared to recent models.