TL;DR
This paper introduces Social Poisson Factorization (SPF), a probabilistic model that enhances item recommendations by incorporating social network influence alongside traditional user preferences. SPF effectively bridges the gap between preference-based and social-based recommendation systems, leading to more accurate and explainable suggestions.
What Problem Does This Paper Address?
Traditional recommendation systems excel at suggesting items based on a user's past behavior and the behavior of similar users. However, they often fail to capture a crucial aspect of how we discover new things: recommendations from friends. A trusted friend might suggest an item that doesn't align with our usual tastes but is still highly relevant. Existing factorization methods, while powerful, cannot leverage this social signal. This paper aims to address this gap by developing a model that explicitly accounts for the influence of a user's social network on their consumption patterns, leading to more personalized and potentially serendipitous recommendations.
Methods
The core of this paper is the Social Poisson Factorization (SPF) model. SPF builds upon Poisson Factorization (PF), a probabilistic model well-suited for implicit feedback data (like clicks or favorites). SPF extends PF by introducing two key components:
- Latent User Preferences ($\theta_u$) and Item Attributes ($\beta_i$): Similar to standard factorization, SPF learns underlying preferences for users and attributes for items.
- Latent Social Influence ($\tau_{uv}$): For each user $u$ and their friend $v$, SPF learns an "influence" value ($\tau_{uv}$) that quantifies how much user $u$ is influenced by user $v$'s clicks.
The model specifies that a user's click on an item is influenced by both their own latent preferences and the latent influences of their friends who have clicked on that item. To handle the computational complexity of fitting such a model, the authors employ variational inference, specifically using a mean-field variational family. This approach approximates the posterior distribution of the model's hidden variables, allowing for scalable learning. The paper also introduces Social Factorization (SF) as a component of SPF, which models influence without factorization, serving as a useful baseline.
Datasets
The authors evaluated SPF on six diverse real-world datasets:
- Ciao: A consumer review website with an underlying social network and trust values.
- Epinions: Another consumer review website with user ratings and trust information.
- Flixster: A social movie review website with binarized ratings.
- Douban: A Chinese social service for music, movies, and books with user ratings.
- Etsy: A marketplace for handmade and vintage items with a "following" network and item favorites.
- Social Reader: Data from a popular online social network's reader application, including friendship networks and article clicks.
These datasets vary in interaction type (ratings, clicks, favorites), network type (directed, undirected), and size, providing a robust evaluation environment.
Key Insights and Ideas
- Social Influence is a Distinct Signal: SPF demonstrates that social influence provides a unique and valuable signal for recommendations, distinct from a user's general preferences. This allows for recommending items that might be outside a user's typical taste but are popular among their friends.
- Explainability through Attribution: SPF's probabilistic nature allows for attributing recommendations to either general preferences or social influence, enhancing the explainability of suggestions to users.
- Scalability with Variational Inference: The use of variational inference makes SPF computationally feasible for large-scale datasets, a critical factor for real-world recommendation systems.
- Performance Gains Across Diverse Datasets: SPF consistently outperforms competing methods on most datasets, highlighting its generalizability.
- Etsy Anomaly: The fact that SF (social influence only) outperformed SPF on Etsy suggests that for certain platforms, social influence might be a dominant factor, or that the factorization component in SPF might not have fully captured the nuances of that specific dataset's item attributes.
Key Results
- Superior Performance: SPF achieved state-of-the-art performance (measured by Normalized Cumulative Reciprocal Rank - NCRR) on five out of six datasets, outperforming traditional factorization methods and other social recommendation approaches.
- Competitive Runtime: SPF's training and testing runtime was found to be average compared to other methods, indicating good scalability.
- Impact of Social Influence: The paper shows that SPF's ability to leverage social influence is crucial, especially for users with lower degrees (fewer friends), where it significantly outperforms models that only consider general preferences.
- Etsy Outperformance: On the Etsy dataset, the Social Factorization (SF) component of SPF achieved the top performance, suggesting the strong role of social connections in that specific marketplace.
Limitations and Future Work
- Time Dependency: The current SPF model does not account for the temporal aspect of interactions. For instance, it doesn't explicitly model whether a friend's click preceded the user's. Future work could incorporate time to better understand the dynamics of influence.
- Hierarchical and Topical Influence: The authors suggest exploring hierarchical influence (e.g., influence from friends of friends) and topical influence (influence related to specific item categories) as potential extensions.
- Cold-Start Users: While SPF performs well in cold-start scenarios (users with no click history but a social network), further research could refine its handling of such users.
- Hyperparameter Sensitivity: While the authors found their model to be not overly sensitive to hyperparameter settings in their pilot study, further investigation into optimal hyperparameter tuning for various datasets could be beneficial.
Summary updated