Source-linked AI summary
GHRS: Graph-based Hybrid Recommendation System with Application to Movie Recommendation
Zahra Zamanzadeh Darban, Mohammad Hadi Valipour
TL;DR
Recommender systems still need improved hybrid methods for the cold-start problem caused by missing user or item ratings. GHRS combines a graph of users’ rating similarities with demographic and location information, then uses Autoencoder-extracted features for user clustering. Experiments report competitive results and an improved best result on MovieLens 1M, while the Autoencoder structure and clustering method remain future research choices.
Problem
Collaborative filtering faces cold start when no ratings exist for a user or item, motivating hybrid models that use side information.
Method
GHRS combines users’ rating similarities in a graph with side information and uses Autoencoder feature extraction before clustering users.
Results
GHRS delivers good cold-start performance and improves the best result on the MovieLens 1M dataset.
Takeaways & Limitations
The method is applicable to user-item systems that provide side information for both users and items.
Takeaways & Limitations
The Autoencoder structure was predefined through heuristic and manual tuning, and alternative clustering methods were left for future investigation.
Abstract
from arXiv · showhide
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.
1. Introduction
Recommender systems address information overload through personalized recommendations, but collaborative filtering faces cold-start limitations. The paper introduces a hybrid Autoencoder framework that combines preferences, similarities, and side information.
- Recommendation systems provide personalized items across domains and can increase customer sales.They are used for movies, music, news, books, e-commerce, and tourism.
- Collaborative filtering estimates users’ unrated items from rating histories, whereas content-based filtering uses user-item side information.
- Cold-start occurs when no ratings exist for a user or item, motivating hybrid models that use side information to compensate for missing ratings.
- Deep learning improves recommendation by capturing nonlinear user-item relationships and complex patterns across data sources.
- The proposed hybrid Autoencoder learns nonlinear user-item representations and integrates side information to address cold start.It combines users’ preferences, similarities, and side information in one matrix, leading to improved collaborative-filtering results.
2. Related Works
Related work covers deep-learning recommendation models and Autoencoder-based methods for representation learning, rating prediction, and ranking prediction. Autoencoders can either learn lower-dimensional features or reconstruct missing interactions, while prior models differ in how they incorporate side information.
- Deep Learning based Recommendation Models: Deep-learning recommendation models are grouped into neural building blocks and deep hybrid models.MLPs model nonlinear interactions, while CNNs and RNNs handle heterogeneous or sequential information; hybrid models combine multiple neural components.
- Autoencoder based Recommendation Models: An Autoencoder reconstructs its input, using the bottleneck layer as a salient feature representation.The paper describes several variants, including denoising, sparse, contractive, and variational Autoencoders.
- Autoencoder based Recommendation Models: Autoencoders support recommendation by learning lower-dimensional bottleneck features or reconstructing missing entries in an interaction matrix.This paper employs feature extraction rather than direct interaction-matrix completion.
- Autoencoder based Collaborative Filtering Models: AutoRec reconstructs user- or item-partial rating vectors, with item-based and user-based variants.I-AutoRec reportedly performs better than U-AutoRec, while activation functions and hidden-unit size affect performance.
- Autoencoder based Collaborative Filtering Models: CFN extends AutoRec with denoising and side-information integration to improve robustness and mitigate sparsity and cold-start effects.Its variants use item- or user-partial vectors, and later extensions inject side information into every layer.
- Autoencoder based Collaborative Filtering Models: CDAE uses implicit-feedback preference vectors for ranking prediction, while Multi-VAE and Multi-DAE apply variational Autoencoders to implicit recommendation data.CDAE uses negative sampling to reduce the cost of training over feedbacks.
- Feature Representation Learning with Autoencoder: Feature-representation methods use Autoencoders to learn user-item content features, including CDL’s integration of stacked denoising Autoencoders with probabilistic matrix factorization.CDR instead targets top-n ranking through a pairwise framework and is reported to outperform CDL for ranking prediction.
3. Graph-based Hybrid Recommendation System
GHRS combines graph-derived user similarity features with Autoencoder-based feature extraction, clustering, and preference-based ranking to estimate user-item ratings and produce recommendations.
- Architecture: GHRS constructs a user-similarity graph, extracts graph features, and combines them with user and item information for recommendation.Users are connected when their ratings are sufficiently similar; extracted features include centrality and neighborhood measures.
- Graph-based Features: Graph features represent user connectivity and structural position through PageRank, degree, closeness, betweenness, load, and average-neighbor-degree measures.PageRank measures transitive influence; degree centrality measures relationships; closeness reflects distances; betweenness identifies shortest-path bridges; load measures shortest-path flow.
- Architecture: The framework encodes combined features with an Autoencoder, clusters users using the encoded features, and ranks movies for a target user.The upper stage creates clustered user representations, while the lower stage retrieves predicted movie rankings.
- Prediction: For new users, GHRS assigns a cluster from encoded features and estimates item ratings using cluster-average ratings, with similarity-based item information when available.The resulting estimated ratings matrix is used to compute the recommendation list for the target user.
- Clustering: K-means assigns users to clusters by minimizing within-cluster variation over extracted Autoencoder features.The method uses the Elbow and Average Silhouette approaches to select the number of clusters.
4. Empirical Experiments and Performance Evaluation
The experiments evaluate GHRS on MovieLens 100K and 1M using graph, demographic, and rating-derived features, cross-validation, and several recommendation metrics. Results examine graph construction, Autoencoder training, clustering, cold-start performance, and comparisons with existing methods.
- Datasets and features: GHRS is evaluated on MovieLens 100K and MovieLens 1M, containing 100,000 and 1,000,209 ratings, respectively.MovieLens 100K has 943 users and 1,682 movies; MovieLens 1M has 6,040 users and approximately 3,900 movies.
- Datasets and features: The raw feature set combines one-hot demographic features with graph-based features extracted from users’ similarity graph.The combined features are formed before dimensionality reduction with an Autoencoder.
- Similarity graph: The graph connects users whose ratings overlap beyond threshold α, which controls graph sparsity from very sparse to nearly full-mesh structures.The experiments visualize α values of 0.005, 0.01, 0.02, and 0.03 for MovieLens 100K.
- Evaluation protocol: Cross-validation uses 10 folds for MovieLens 1M and 5 folds for MovieLens 100K, with RMSE, Precision, and Recall used for evaluation.The training data includes ratings, user demographics, and item side information.
- Results and comparison: GHRS achieves lower testing RMSE, performs well with a high percentage of new users, and improves the best MovieLens 1M RMSE while matching AutoRec after Autoencoder COFILS.The reported cross-validation errors are almost similar across iterations, and the comparison includes primary and state-of-the-art methods.
- Parameter analysis: The optimum graph threshold is near α = 0.01 for MovieLens 100K and α = 0.005 for MovieLens 1M.Very small α values connect most users, whereas larger values make the graph more sparse, so the minimum RMSE occurs at an intermediate setting.
5. Conclusion and Future Works
GHRS combines user-similarity graphs, side information, Autoencoder-derived features, and clustering for user-item recommendation. The authors report competitive results and identify feature extraction, Autoencoder design, clustering, and dataset availability as future-work considerations.
- Proposed method: GHRS models relations between users through a similarity graph and combines those relations with user side information to address cold-start recommendation.The method is intended for user-item systems providing side information for both users and items.
- Proposed method: Autoencoder feature extraction produces low-dimensional, less-correlated features that improve clustering accuracy and time consumption.
- Results: Experiments and comparisons showed competitive results, including an improvement over the best result on the MovieLens 1M dataset.
- Future works: Future work could extract additional features from the similarity graph and examine how different Autoencoder structures affect feature extraction, training duration, and final performance.The current study used a predefined Autoencoder structure selected through heuristic methods and manual tuning.
- Future works: Future studies should investigate and measure alternative user-clustering methods for the resulting feature-space distribution.
- Future works: Because few datasets contain side information for both users and items, evaluating the method on additional datasets with this information is desirable.Demographic data for users is given as an example of such side information.