An experimental study on the performance of collaborative filtering based on user reviews for large-scale datasets

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Information Retrieval

Background:

• Collaborative filtering (CF) relies on explicit user ratings for recommendations.
• Explicit ratings are often sparse or unavailable in many domains.
• User reviews offer a valuable data source for implicit feedback.

Purpose of the Study:

• To improve CF performance by deriving implicit ratings from user reviews.
• To address limitations of existing methods that ignore sentiment degrees and review aspects.
• To propose novel methods for calculating implicit ratings that capture richer review information.

Main Methods:

• Developed four methods to calculate implicit ratings from user reviews.
• Methods incorporate sentiment word degrees and aspect-sentiment word pairs.
• Combined implicit ratings with explicit ratings for enhanced CF algorithms.

Main Results:

• Proposed implicit rating methods significantly improved CF rating prediction accuracy.
• Evaluated methods on large-scale Amazon and Yelp datasets.
• Outperformed traditional explicit ratings across three predictive accuracy metrics.

Conclusions:

• Leveraging sentiment analysis and aspect extraction from user reviews enhances CF.
• Novel implicit rating calculation methods offer a robust alternative to explicit ratings.
• This approach improves recommendation system accuracy and effectiveness.