Source-linked AI summary
UniWalk: Explainable and Accurate Recommendation for Rating and Network Data
Haekyu Park, Hyunsik Jeon, Junghwan Kim, Beunguk Ahn, U Kang
TL;DR
Recommending appropriate items from ratings and social networks while explaining why they are recommended is challenging. UniWalk combines both data types in a unified graph and provides recommendations with reasons. The paper reports best explainability and state-of-the-art accuracy among competitors.
Problem
Using ratings and social-network links together is challenging because qualitative links do not directly provide rating values, yet recommendations and their reasons matter to users and suppliers.
Source: §1 · p. 1, §3.1 · p. 3Method
UniWalk builds a unified graph from rating and social data, learns user and item features from quantitative and qualitative links, and explains recommendations using similar users and items.
Source: §1 · p. 1, §3.2 · p. 4, §3.3 · p. 5Results
UniWalk reports the best explainability and state-of-the-art accuracy among competitors, with the lowest RMSE and MAE.
Source: §1 · p. 2, §4.3 · p. 7Takeaways & Limitations
Recommendations can be accompanied by reasons based on similar users' preferences and the target user's preferences for similar items.
Source: §3.3 · p. 5
Abstract
from arXiv · showhide
How can we leverage social network data and observed ratings to correctly recommend proper items and provide a persuasive explanation for the recommendations? Many online services provide social networks among users, and it is crucial to utilize social information since recommendation by a friend is more likely to grab attention than the one from a random user. Also, explaining why items are recommended is very important in encouraging the users' actions such as actual purchases. Exploiting both ratings and social graph for recommendation, however, is not trivial because of the heterogeneity of the data. In this paper, we propose UniWalk, an explainable and accurate recommender system that exploits both social network and rating data. UniWalk combines both data into a unified graph, learns latent features of users and items, and recommends items to each user through the features. Importantly, it explains why items are recommended together with the recommendation results. Extensive experiments show that UniWalk provides the best explainability and achieves the state-of-the-art-accuracy.
1 Introduction
UniWalk addresses the challenge of recommending appropriate items from ratings and social networks while explaining its recommendations. It combines both data types and reports the best explainability and state-of-the-art accuracy among competitors.
Motivation
- Using ratings and social links together is challenging because multi-step relations do not translate straightforwardly into ratings, and latent features are not human-interpretable.Existing approaches described here use social information only partially and do not provide explanations.Source: §1 · p. 1
Contributions
- UniWalk combines ratings and social-network data in a unified graph, samples quantitative and qualitative links with random walks, and learns latent features.Its objective uses a supervised term for quantitative links and unsupervised terms for qualitative links.Source: §1 · p. 1
- UniWalk reports the best explainability and accuracy among compared methods, and is the only method using all link types.Its accuracy is state of the art for recommendation with ratings and a social network in terms of RMSE and MAE.Source: Table 1 · p. 2, §1 · p. 2
- UniWalk explains recommendations through similar users who like the items and similar items preferred by the target user.It also gives meta-explanations for why supplementary users and items were judged similar.Source: §1 · p. 2
2 Preliminary
The preliminaries introduce matrix factorization for predicting ratings and network embedding for representing graph nodes as vectors. They provide the background for UniWalk's use of latent features.
2 Preliminary
- The preliminary section covers matrix factorization and network embedding as background for the proposed method.These are the two topics identified as the section's scope.Source: §2 · p. 2
2.1 Matrix Factorization with Bias Factors.
- Matrix factorization predicts unobserved ratings from global, user, and item biases plus the inner product of user and item vectors.Its objective fits observed ratings while regularization controls parameter magnitudes to prevent overfitting.Source: §2.1 · p. 2
2.2 Network embedding.
- Network embedding maps each graph node to a vector that encodes structural information.DeepWalk uses fixed-length random walks and Skipgram with negative sampling; negative sampling pushes randomly selected distant node pairs apart.Source: §2.2 · p. 3
3 Proposed Method
UniWalk addresses rating prediction and recommendation explanations using both observed ratings and social relationships. It learns from a unified graph and explains recommendations through similar users and items.
3.1 Challenges and Solutions.
- UniWalk faces three challenges: selecting useful cross-domain qualitative links, learning from them alongside ratings, and using them systematically in explanations.Random-walk sampling selects similar or dissimilar pairs; a joint objective combines supervised rating learning with unsupervised qualitative-link learning, and explanations report similarity values. Qualitative links also supply taste and item-property information that mitigates sparsity.Source: §3.1 · pp. 3–4
3.2 UniWalk for Rating Prediction.
- A weighted undirected graph represents ratings as user-item edges and social ties as user-user edges weighted by c; embeddings fit ratings, pull similar entities together, and separate dissimilar ones before predicting ratings.The embedding learns entity biases and latent vectors, with rating prediction based on those learned representations.Source: §3.2 · p. 4
- Positive, negative, and unweighted walks sample distinct relation patterns; sliding-window pairs populate rating-supervised, similarity-positive, and dissimilarity-negative sets for learning.Positive walks favor high-weight edges, negative walks invert score-link weights while retaining social-link weights, and unweighted walks favor higher-degree entities.Source: §3.2 · pp. 4–5
3.3 UniWalk for Explainable Recommendation.
- Recommendations are explained by similar users’ preferences and a user’s preference for similar items; meta-explanations trace those similarities to shared friends or rating behavior.Similar users are selected from those who rated a recommended item, while similar-item explanations draw on items rated by the target user.Source: §3.3 · pp. 5–6
4 Experiment
Experiments assess UniWalk’s explanations, rating accuracy, and learning behavior using social-rating datasets and comparisons with collaborative-filtering and social-recommendation baselines.
4 Experiment
- The experiment is organized around explanation quality, rating-prediction accuracy, and how accuracy changes during learning and with hyperparameters.Source: §4 · p. 6
4.1 Experimental settings.
- The evaluation uses FilmTrust, Epinions, and Flixster social-rating data, compares explainable UCF and ICF with MF and TrustSVD, and evaluates RMSE and MAE using five-fold cross-validation.Directed social edges are converted to undirected edges; each observed rating appears in a test set once.Source: §4.1 · p. 6, §4.3 · p. 7
4.2 Explainability of UniWalk.
- Recommendations are explained through similar users who prefer the recommended items and similar items liked by the target user, with meta-explanations for those similarities.The examples trace user similarity to common liked or disliked items and shared friends, and item similarity to common raters, including friends.Source: §4.2 · pp. 6–7
4.3 Accuracy.
- UniWalk achieves the lowest RMSE and MAE among competitors; TrustSVD ranks second, while comparison with MF indicates the unsupervised terms enhance accuracy.Tables 3 and 4 report RMSE and MAE, respectively.Source: §4.3 · p. 7, Table 3 · p. 7, Table 4 · p. 7
4.4 Learning Details.
- On FilmTrust, UniWalk’s minimum test error occurs after three iterations; with momentum, learning is 1.45× faster and requires 2.5× fewer iterations than without momentum in the reported experiment.UniWalk is less sensitive to social-link weight c than to α and β.Source: §4.4 · p. 7
5 Related Works
The related-work review covers recommendation with auxiliary data, explainable recommendation, and network embedding.
5 Related Works
- The review groups closely related work into auxiliary-data recommendation, explainable recommendation, and network embedding.Source: §5 · p. 7
5.1 Recommendation with Additional Data.
- Auxiliary information is used to address rating sparsity and improve recommendation performance, including social networks, geography, and entity attributes.Prior work also uses heterogeneous-data categories and coupled tensor factorization.Source: §5.1 · p. 8
5.2 Explainable
- Review-based recommenders explain results with preference keywords or product features, while friend-based approaches use social links and friends’ favored items or visited locations.Source: §5.2 · p. 8
- Network embedding methods represent graph proximity; DeepWalk uses random-walk node sequences, while other approaches model proximity directly or customize walks.Source: §5.2 · p. 8
6 Conclusion
UniWalk combines rating data and a social network in a unified graph, applies network embedding to learn user and item features, and predicts ratings from those features.
6 Conclusion
- UniWalk constructs a unified user-item graph whose weights reflect association, then uses network embedding to extract latent features and predict ratings; the paper reports best explainability and accuracy.Future work includes extending the method to distributed systems for scalable learning.Source: §6 · p. 9