Source-linked AI summary
Collaborative Topic Regression with Social Matrix Factorization for Recommendation Systems
Sanjay Purushotham, Yan Liu, C. -C. Jay Kuo
TL;DR
Recommendation systems need to determine whether social networks add predictive value beyond item content, particularly under sparse rating data. The paper develops a hierarchical Bayesian model that combines LDA-based content modeling with social matrix factorization, and experiments show improved performance while identifying social-information leakage as a limitation.
Problem
The paper asks whether and how social networks can improve recommendation systems beyond the information provided by item content, amid sparse collaborative-filtering data and limited joint modeling of content and social information.
Method
The proposed hierarchical Bayesian model jointly integrates item content through LDA and social-network structure through matrix factorization using a shared user latent feature space.
Results
The model consistently outperforms state-of-the-art approaches on two large social-media datasets, with social-network fusion improving recommendation accuracy by 2.5 ∼3%.
Takeaways & Limitations
Social-network information can improve recommendation predictions, and its influence may differ between usefulness-of-information decisions and personal taste.
Takeaways & Limitations
Using a final static social network may leak future information into prediction, and the paper identifies modeling evolving social networks as future work.
Abstract
from arXiv · showhide
Social network websites, such as Facebook, YouTube, Lastfm etc, have become a popular platform for users to connect with each other and share content or opinions. They provide rich information for us to study the influence of user's social circle in their decision process. In this paper, we are interested in examining the effectiveness of social network information to predict the user's ratings of items. We propose a novel hierarchical Bayesian model which jointly incorporates topic modeling and probabilistic matrix factorization of social networks. A major advantage of our model is to automatically infer useful latent topics and social information as well as their importance to collaborative filtering from the training data. Empirical experiments on two large-scale datasets show that our algorithm provides a more effective recommendation system than the state-of-the art approaches. Our results reveal interesting insight that the social circles have more influence on people's decisions about the usefulness of information (e.g., bookmarking preference on Delicious) than personal taste (e.g., music preference on Lastfm). We also examine and discuss solutions on potential information leak in many recommendation systems that utilize social information.
1. Introduction
Recommendation systems can benefit from social-network information, but existing work has rarely combined it with item content. This paper proposes a hierarchical Bayesian model that jointly integrates both sources through shared latent features and evaluates their recommendation value.
- The paper asks whether social networks provide additional recommendation gains when rich item content is already available.The question is framed as relevant to product recommendation, advertisement targeting, and scientific-paper suggestion.
- Collaborative filtering predicts user interests from similar users’ or items’ ratings but is unreliable with sparse and imbalanced data, especially for new or infrequent users.Existing approaches address these weaknesses by incorporating item content or user social networks.
- Most prior recommendation methods use either item content or social-network information, while few consider both jointly.
- The proposed hierarchical Bayesian model combines social-network matrix factorization with LDA-based item content modeling through a shared user latent feature space.Social matrix factorization learns user features, while topic modeling represents items.
- Experiments on Lastfm and Delicious show that the proposed model outperforms state-of-the-art collaborative-filtering methods including CTR and PMF.The model also examines how social information affects prediction performance and introduces social information leak as a recommendation-system issue.
2. Related Work
Prior recommendation research includes memory-based and model-based collaborative filtering, content-aware topic models, and social-network matrix factorization. The paper’s related work motivates combining these information sources while addressing limitations involving new users and unseen items.
- Memory-based collaborative filtering uses user-based or item-based approaches, whereas model-based methods learn predictive models from the data.Memory-based approaches are easy to implement but do not guarantee good prediction results.
- Some collaborative-filtering approaches ignore item content and users’ social connections during recommendation.
- CTR combines traditional collaborative filtering with probabilistic topic modeling by representing users with topic interests and generating items through a topic model.CTR also includes a latent item-specific offset in its formulation.
- CTR’s generative process draws user latent vectors, item topic proportions, item offsets, topic assignments, words, and user-item ratings.
- Earlier social recommendation methods use matrix factorization with users’ social networks and rating records, while CTR does not reliably learn latent spaces for new or inactive users.
- The proposed model addresses recommendation when user content, ratings, and social-network information are known, including predictions for new items and new or inactive users.
3. Proposed Approach
The proposed hierarchical Bayesian model jointly learns user, item, and social latent spaces by combining LDA-based item content modeling with matrix factorization of the social network. It uses these learned representations for both in-matrix and out-matrix recommendation.
- The model combines LDA for item content with social matrix factorization, linking both sources through a shared user latent feature space.LDA represents items in a topic space, while social matrix factorization derives user features from the social graph.
- The social network is represented as a user-relation matrix whose factorization yields l-dimensional user and social-factor representations.Relation confidence parameters encode the strength of connections between users.
- The model places Gaussian priors on user and factor vectors and uses Bayesian inference to learn the latent variables and topic proportions.The posterior includes user, item, social, and topic-related variables controlled by regularization parameters.
- An EM-style optimization procedure alternates gradient-based updates for collaborative-filtering variables, social factors, topic proportions, and topic-word parameters.Topic proportions are optimized with projection-gradient methods because they cannot be optimized analytically.
- After learning the parameters, the model supports in-matrix prediction for partially observed items and out-matrix prediction for entirely new items.For new items, the expected item-specific offset is set to zero when predicting ratings.
4. Experimental Analysis
Experiments on sparse Lastfm and Delicious datasets evaluate recall against collaborative filtering baselines while varying model parameters and recommendation-list size. The proposed CTR with SMF model consistently improves recall, with parameter settings and social-network effects differing by dataset.
- 4.1. Description of Datasets: The experiments use Lastfm music and Delicious bookmarking datasets, whose user-item matrices are 99.7% and 99.91% sparse, respectively.
- 4.2. Experimental Setup: The evaluation uses 90% training and 10% testing splits, with user-oriented recall as the primary metric for in-matrix prediction.
- 4.4. Comparisons: CTR with SMF consistently outperforms CTR by 2.5 ∼3% on both Lastfm and Delicious when varying λv with M = 250.
- 4.4. Comparisons: Across M = 50, 100, ..., 250, the proposed model improves recall as M increases and outperforms both CTR and CF at each tested value.
- 4.5. Impact of parameters λv, λq: For Lastfm, higher λv and λq values perform best near balanced values around 150, whereas Delicious performs best at λq = 0.05 and λv = 0.01.
- 4.6. Complexity Analysis: LDA makes the model more computationally expensive than traditional matrix factorization, while smaller latent dimension K converges 45x faster with reduced accuracy.
- 4.7. Discussion on Social Network Structure: Using the full static social network improves prediction accuracy by 3-5% over timestamped social-network information, indicating potential information leakage.
5. Conclusions and Future Work
The paper presents a hierarchical Bayesian recommender that combines users’ social-network information with item content information, and reports consistent gains over existing approaches. Future work targets scalability and modeling evolving social networks to study information leakage.
- The proposed model exploits users’ social networks and item content information for recommendation.
- Experiments on two large social-media datasets show consistent outperformance of CTR and matrix-factorization approaches.
- The paper identifies social information leakage as an underexplored issue in social recommendation systems.
- Future work will examine parallel implementations for scalability to large-scale datasets.
- Future work will model evolving social networks and analyze their effect on prediction accuracy.