نظام التوصية المبني على تنقيب الآراء باستخدام تقنيات التعلم الآلي

Recommendation System Based On Opinion Mining using Machine Learning Techniques

Abstract

The Internet and the Web have made it possible for a vast amount of information to be shared and accessed by large numbers of people. This has led to a problem called information overload (the challenge of making decisions when faced with too much information). This problem necessitated the creation of recommendation systems, which address the information overload challenge by suggesting products or services that may be useful to users and their interests.
Recommendation systems may face several problems, including cold-start and sparsity. These problems lead to a decline in the performance of the recommender system.
In this work, a Recommender System based on Textual Reviews and using the Deep Learning method (RS-TRDL) was proposed to perform two tasks namely alleviate the user cold-start and alleviate sparsity problems, this leads to improving the performance of the proposed method. Textual reviews were used as additional information alongside the users’ numerical ratings. Important aspects were extracted from these reviews, in addition to the polarity of sentiment by using one of the deep learning algorithms, which is Long Short-Term Memory (LSTM) algorithm, to then benefit from these aspects in the recommendation process.
The RS-TRDL model employed a comprehensive pre-processing stage for the dataset. This stage encompassed various steps, including Handling Missing Values and Data Labeling. Additionally, it incorporated text-specific pre-processing operations such as Text Cleaning and general Text Preprocessing, it then proceeded with aspects extraction. This step employed spaCy for Noun Extraction via Part-of-Speech (POS) tagging. Additionally, Topic Modeling was performed using the BERTopic algorithm. Finally, Sentiment Analysis was conducted utilizing the Long Short-Term Memory (LSTM) algorithm.
After extracting the aspects, the cold-start users and non-cold-start users were treated separately. For cold-start users, the Rating prediction process was done using K-Nearest Neighbours (KNN) algorithm based on the ratings of the non-cold-start users who share the same aspects of the same items and have high helpfulness value.
For non-cold start users, firstly grouping process was done based on aspects extracted from users’ reviews, then a similarity matrix was created for each group using Cosine similarity measure. Finally, the rating prediction process was performed using KNN based on the ratings of the nearest users belonging to the same group and having a high helpfulness value.
Extensive Experiments were conducted by the proposed system on two Amazon datasets: Amazon Electronics and Amazon Fine Food. The experimental results show that the proposed RS-TRDL model exceeded all literature-reviewed comparison methods in the rating prediction process for both tasks it was built to perform. It is worth highlighting the model’s consistent performance across both tasks, as evidenced by the improvement range of 0.24% to 34.32% for alleviating the user cold start problem task and 3.21% to 58.7% for alleviating the sparsity problem and enhancing the recommender model task.