Enhancing Recommender Systems with Fuzzy Logic and Label Vectors

By Dr. Sampath LonkaReviewed by Susha Cheriyedath, M.Sc.Aug 30 2023

In recommender systems, predicting ratings is crucial for personalized recommendations. Improving the quality of recommended lists is a challenging task. In a recent paper published in the journal PLOS ONE, researchers enhanced item ranking in user lists by ensuring interpretability. This study introduces fuzzy membership functions for user attributes and a user similarity network.

Background

In recent times, the surge in information overload has spurred diverse recommender systems to aid users in managing data deluge. Fundamentally, these systems prioritize information based on user preferences. These systems have challenges mitigating the strain of information inundation and economizing time and energy.

Recommender systems draw upon user-generated data, interactions, and interactions between users and items. They are now ubiquitous across domains: YouTube, Netflix for entertainment, Amazon for competitive product recommendations, and Twitter and Spotify for social networking, enhancing user experiences.

Based on the working methodology, three primary recommendation algorithms are content-based, collaborative filtering (CF), and hybrid algorithms. Content-based algorithms focus on attributes; for instance, a movie site might suggest films based on viewing history. While CF utilizes a rating matrix and user neighborhoods, hybrid algorithms amalgamate content-based and CF strengths. Despite their benefits, recommender systems face challenges: refining list ranking, addressing cold start, sparsity, fairness, and privacy. Collaborative filtering's interpretability, ease, and research appeal drive its evolution. Diverse similarity measures enhance CF.

Matrix factorization (MF)-based recommendation techniques, decomposition of interaction matrices, and deep learning (DL) models leverage neural networks. Although both techniques offer improved recommendation systems, they are resource-intensive, slow, and lack explainability. The current study explores alternative methods to enhance recommender system performance.

Proposed techniques

Fuzzy logic and handling uncertain ratings enhance similarity calculations, and structural information aids prediction. Additionally, multi-dimensional data and advanced techniques optimize recommendations. The present study merges fuzzy logic and vector similarity, capturing uncertain preferences and intricate relationships.

The proposed algorithm consists of three main steps: rating fuzzification, item vectorization, and link prediction. Ratings are fuzzified into "like" and "dislike" using a fuzzy membership function, improving collaborative filtering. Items are transformed into label vectors, enhancing similarity and recommendation equations.

The paper introduces three prediction algorithms: fuzzy-preference label vector (FPLV) CF with all qualified neighbors (FPLV-ALL), FPLV CF with similarity selection (FPLV-SS), and FPLV CF with K nearest neighbors (FPLV-KNN). These algorithms are developed based on the similarity of fuzzy-preference label vectors. A user-user similarity network is established for prediction, and community characteristics are integrated. The algorithm's flow involves obtaining user fuzzy preferences, generating item label vectors, calculating similarities, and establishing user similarity networks.

Experiments and evaluations of proposed methods

Researchers employed the FSharp programming language with consistent caching and computation processes for algorithm implementation. Target data was read into memory and randomly divided into five groups for cross-validation. Parameters were tested for prediction, results were recorded, and analysis was performed using Python.

Researchers evaluated algorithms using real-world datasets: MovieLens-25M with 25 million movie ratings from 162,000 users, and the Netflix Dataset with 100,480,507 ratings from 480,189 users. For efficiency, a sample of 5,000 users and ratings was used. Baseline algorithms were compared, including similarity-network resource allocation (SRA), vector similarity collaborative filtering, entropy collaborative filtering, and the dual training error correction (DTEC)-SCoR algorithm.

Various metrics were included to measure the classification accuracy, namely mean absolute error (MAE), root mean square error (RMSE), F1-score, precision rate, and recall rate. Additionally, enhanced metrics such as half-life utility (HLU), sorting accuracy (SA), item degree diversity (IDD), and item genre diversity (IGD) are used for measuring recommendation list quality and diversity. These metrics provide comprehensive insight into the performance of the recommender system algorithms.

In prediction accuracy comparison, all three proposed algorithms achieve the lowest MAE and RMSE values compared to other algorithms. In the HLU and SA comparisons, the proposed algorithms and the base-line algorithm SRA performed better.

Experiment results show that FPLV-C’s similarity network improves ranking consistently across datasets. Finally, FPLV-C and FPLV-SS show identical performance, and FPLV-KNN performs slightly better. In classification accuracy comparison, FPLV-KNN performs remarkably; other FPLV algorithms show average performance. In diversity comparison, FPLV-C excels in diversifying user lists through community characteristics.

Algorithm scalability is assessed through time complexity analysis and empirical observations. FPLV-KNN demonstrates the lowest time complexity and strong performance, FPLV-SS and FPLV-C exhibit similar time complexity and FPLV-ALL and Entropy show varied time complexities. FPLV-KNN performs exceptionally well, but optimal neighbor count requires experimentation. These evaluations comprehensively understand the proposed algorithms' performance across various metrics and datasets.

Conclusion 

In summary, researchers introduced a novel approach using user fuzzy preferences and item label vectors to enhance predictive accuracy in recommender systems. They constructed user-to-item fuzzy preference label vectors, effectively capturing uncertain preferences. The proposed algorithms excel in ranking and rating prediction accuracy compared to classical methods. The FPLV series maintains commendable F1 classification accuracy and moderate recommendation list diversity. Future research should focus on expanding fuzzy preferences, weighted label attribution, and developing complex network systems.