[Paper Review] GHRS: Graph-based Hybrid Recommendation System with Application to Movie Recommendation

Zahra Zamanzadeh Darban, Mohammad Hadi Valipour|arXiv (Cornell University)|Nov 6, 2021

Recommender Systems and Techniques57 references145 citations

TL;DR

This paper proposes GHRS, a graph-based hybrid recommendation system that integrates user rating similarity, demographic data, and location information using autoencoder-based feature extraction. By constructing a user similarity graph and applying deep learning to extract low-dimensional, informative features, GHRS achieves state-of-the-art performance on MovieLens datasets, significantly improving recommendation accuracy and effectively addressing the cold-start problem with RMSE of 0.833 on MovieLens 1M.

ABSTRACT

Research about recommender systems emerges over the last decade and comprises valuable services to increase different companies' revenue. Several approaches exist in handling paper recommender systems. While most existing recommender systems rely either on a content-based approach or a collaborative approach, there are hybrid approaches that can improve recommendation accuracy using a combination of both approaches. Even though many algorithms are proposed using such methods, it is still necessary for further improvement. In this paper, we propose a recommender system method using a graph-based model associated with the similarity of users' ratings, in combination with users' demographic and location information. By utilizing the advantages of Autoencoder feature extraction, we extract new features based on all combined attributes. Using the new set of features for clustering users, our proposed approach (GHRS) has gained a significant improvement, which dominates other methods' performance in the cold-start problem. The experimental results on the MovieLens dataset show that the proposed algorithm outperforms many existing recommendation algorithms on recommendation accuracy.

Motivation & Objective

• To address the limitations of traditional collaborative filtering and content-based filtering in recommendation systems, particularly the cold-start problem.
• To improve recommendation accuracy by integrating user rating patterns, demographic data, and location information into a unified model.
• To leverage deep learning, specifically autoencoders, for effective feature extraction from heterogeneous user attributes.
• To develop a graph-based representation that captures user similarity and enhances clustering for better recommendation performance.
• To evaluate the model’s robustness in low-rating scenarios, simulating new users with no prior ratings.

Proposed method

• Construct a user similarity graph based on cosine similarity of user rating vectors to model relationships between users.
• Integrate user demographic and location data with rating patterns into a unified feature matrix for input to the autoencoder.
• Apply a deep autoencoder to learn low-dimensional, non-linear representations of users with reduced correlation and higher information content.
• Use the learned latent representations from the autoencoder for clustering users via K-Means and MiniBatchK-Means, with cluster count selected via the elbow and silhouette methods.
• Train and evaluate the final recommendation model using RMSE, precision, and recall on the MovieLens 100K and 1M datasets.
• Simulate cold-start scenarios by randomly removing user ratings from the rating matrix and evaluating performance on remaining side information.

Experimental results

Research questions

• RQ1 Can combining user rating similarity, demographic data, and location information through a graph-based model improve recommendation accuracy?
• RQ2 To what extent can autoencoder-based feature extraction enhance the performance of hybrid recommendation systems?
• RQ3 How well does the proposed model handle the cold-start problem when users have no prior rating history?
• RQ4 What is the optimal number of clusters for user groups derived from the graph and autoencoder features?
• RQ5 How does GHRS compare to state-of-the-art methods in terms of RMSE, precision, and recall on standard benchmark datasets?

Key findings

• GHRS achieved an RMSE of 0.833 on the MovieLens 1M dataset, outperforming most baseline models and matching the best result of AutoRec.
• On the MovieLens 100K dataset, GHRS achieved an RMSE of 0.887, which is competitive with state-of-the-art methods and the best result of AutoRec.
• The model demonstrated strong performance in cold-start scenarios, maintaining low RMSE even when up to 80% of user ratings were removed, indicating robustness to data sparsity.
• The use of autoencoders significantly improved feature representation, leading to more accurate clustering and faster training compared to raw feature inputs.
• The combination of graph-based user similarity and side information led to a 17.53 average silhouette score on MovieLens 1M, indicating well-separated and meaningful user clusters.
• The model outperformed several advanced methods, including SVD++, Bayesian PLSA, RBM-CF, and others, in both RMSE and recall metrics on the 1M dataset.