استغلال معلومات التغذية الراجعة الصريحة والضمنية لتحسين نظام التوصية

Exploiting Explicit and Implicit Feedback Information for Recommendation System Improvement

Recommendation systems (RSs) are intelligent information filtering systems that deal with the information overload problem. Recently, social trust information has become a significant additional factor in obtaining high-quality recommendations. In addition, textual review information plays a core role in many RS methods that can improve the accuracy of recommendations.
RS techniques suffer from problems such as sparsity, cold-start, and class imbalance, which reduce their recommendation accuracy. In this work, two approaches have been proposed for RS improvement. They are based on implicit feedback inferences and explicit feedback ratings. The first one is the trust-based prediction model (TPM) and the other is offered with two tracks of review-based recommendation models (RRMs).
In TPM, explicit and implicit trust relations are obtained depending on the trust propagation attribute. They are incorporated to benefit from more ratings of trustworthy neighbors to alleviate the rating sparsity and cold-start problems. Accordingly, the weighted voting technique of the ensemble classifier is applied for the election of the most appropriate trusted neighbors’ ratings. The K Nearest Neighbor (KNN) method is employed with a linear combination of original and trust-elected ratings to obtain the best coverage and accuracy of the prediction.
In RRMs, on the other hand, two recommendation models are developed based on the presence and absence of textual reviews in the test set. In particular, five essential steps have shaped these models: data preprocessing, text classification, topic modeling, text similarity, and the step of deducing adjustment weights to mitigate the rating sparsity and class imbalance problems. The Naïve Bayes model is used to integrate the adjustment weights as implicit feedback for recommending the preferable items to the target user.
Extensive experiments have been conducted using five real-world datasets: FilmTrust, Epinions, Musical Instruments, Automotive, and Amazon Instant Video. The experimental results show that the proposed TPM and RRMs significantly outperform all methods compared in both rating prediction and Top-N recommendation tasks. Specifically, in TPM, the improvement ratios ranged approximately from 4% as a minimum to 20% and 10% as a maximum on FilmTrust and Epinions respectively, in terms of the F-measure of the rating prediction task. While, in RRMs, the recommendation accuracy improved by about 10% as a minimum to 61%, 26%, and 22% as a maximum on Musical Instruments, Automotive, and Amazon Instant Video respectively, in terms of the F1-measure of the Top-N recommendation task.