Exploring Social Influence for Recommendation - A Probabilistic Generative Model Approach

In this paper, we propose a probabilistic generative model, called unified model, which naturally unifies the ideas of social influence, collaborative filtering and content-based methods for item recommendation. To address the issue of hidden social influence, we devise new algorithms to learn the model parameters of our proposal based on expectation maximization (EM). In addition to a single-machine version of our EM algorithm, we further devise a parallelized implementation on the Map-Reduce framework to process two large-scale datasets we collect. Moreover, we show that the social influence obtained from our generative models can be used for group recommendation. Finally, we conduct comprehensive experiments using the datasets crawled from last.fm and whrrl.com to validate our ideas. Experimental results show that the generative models with social influence significantly outperform those without incorporating social influence. The unified generative model proposed in this paper obtains the best performance. Moreover, our study on social influence finds that users in whrrl.com are more likely to get influenced by friends than those in last.fm. The experimental results also confirm that our social influence based group recommendation algorithm outperforms the state-of-the-art algorithms for group recommendation.

💡 Research Summary

The paper presents a unified probabilistic generative framework that explicitly models hidden social influence alongside traditional collaborative filtering (CF) and content‑based recommendation components. The authors argue that users are often swayed by friends when making consumption decisions, yet most existing recommender systems either ignore this factor or treat it with ad‑hoc heuristics that require pre‑computed trust scores. To address this gap, they introduce latent variables for both user interests (topics) and for the probability that a user is influenced by each of his friends. In the generative process, a user either selects an item directly from his own topic distribution or adopts a friend’s topic, which then generates the item. This yields a joint probability that combines three distributions: user‑topic, topic‑item, and friend‑to‑user influence.

Parameter learning is performed via Expectation‑Maximization (EM). In the E‑step, posterior probabilities over topics and influencing friends are computed for each observed user‑item interaction; in the M‑step, the model updates the topic priors, user‑topic, item‑topic, and friend‑influence parameters using the sufficient statistics gathered in the E‑step. Because EM requires multiple passes over the entire interaction log, the authors design a Map‑Reduce implementation to scale to millions of interactions. The mapper computes local posteriors and aggregates counts for topics, users, and friend influences; the reducer aggregates these counts across all mappers and normalizes them to obtain the new global parameters. This parallel EM converges in roughly 15 iterations and achieves near‑linear speed‑up on an 8‑node Hadoop cluster.

The framework is evaluated on two real‑world datasets: Last.fm (music listening logs) and Whrrl.com (location‑based check‑ins). Baselines include a pure probabilistic CF model, a CF+content model that incorporates item metadata, and several prior social‑influence methods that either linearly combine trust scores or add a regularization term. Across standard ranking metrics (Precision@10, Recall@10, NDCG@10), the unified model consistently outperforms all baselines. The performance gain is especially pronounced on Whrrl.com, where the learned friend‑influence parameters account for nearly half of the total improvement, indicating that users in that platform are more socially driven than Last.fm users.

Beyond individual recommendation, the authors extend the model to group recommendation. By aggregating each member’s inferred friend‑influence weights, they construct a group‑level topic distribution that reflects both individual preferences and intra‑group social dynamics. This approach outperforms traditional group‑recommendation strategies such as profile averaging or rank aggregation, delivering an average 12 % lift in NDCG over the best competing method.

Key contributions of the work are: (1) a principled probabilistic treatment of hidden social influence, (2) an EM‑based learning algorithm with a scalable Map‑Reduce implementation, and (3) a demonstration that the same influence parameters can improve both individual and group recommendation tasks. Limitations include sensitivity to the number of topics, potential instability when friend networks are extremely sparse, and the lack of an online updating mechanism for real‑time recommendation scenarios. Future research directions suggested include incremental EM, dynamic network modeling, and application to other domains such as e‑commerce or news recommendation.

In summary, the paper shows that integrating socially inferred influence into a generative recommender model yields substantial accuracy gains and provides a flexible foundation for extending recommendation services to group contexts, all while maintaining scalability through parallel processing.