Modeling Social Influence for Personalized Recommendations

This direction is split into other directions: social graph integration. Its last version stays readable and citable here.

Version 1 of 1246 papers263 conceptsfrozen 2026-10-04

Traditional recommendation systems often treat users as isolated islands. By focusing primarily on a user’s historical clicks or ratings, these models assume that preferences are static and self-contained. However, human behavior is fundamentally social. Whether we are choosing a restaurant, clicking an ad, or joining a forum, our decisions are shaped by the visible actions and perceived preferences of our peers.

A growing body of research argues that to move beyond simple pattern matching, we must model social influence—the mechanism by which the social graph modulates individual behavior.

The Limits of the Isolated User

Most predictive models optimize for individual signals, such as click-through rates or sequential item consumption. As noted in Social Influence Does Matter: User Action Prediction for In-Feed Advertising, this "IID" (independent and identically distributed) assumption ignores the social signals generated when users interact with content. If a user sees that several friends have liked an advertisement, their likelihood of engagement increases—a phenomenon often called the social influence effect.

When models ignore these signals, they suffer from several persistent issues:

  • Data Sparsity: New users or "cold-start" items lack the history needed for accurate prediction.
  • Interest Drift: As explored by Weiqing Wang and colleagues in Geo-SAGE, user preferences shift when they travel to unfamiliar regions, making historical data less relevant than the real-time "crowd sentiment" or local social cues.
  • Static Intent: Models like DIGMN suggest that failing to account for how a user's connections drive their future behavior prevents systems from adapting to evolving professional or personal intents.

Encoding the "Who" and the "How"

To integrate social influence, researchers have moved toward hybrid architectures that combine structural graph data with temporal sequence modeling.

In Social Influence Does Matter, Hongyang Wang and Ju Fan propose a framework that uses structural graph encoding to identify who is likely to be influenced, paired with Graph Attention Networks (GATs) to determine how that influence manifests. This allows the model to weigh a friend's interaction more heavily than a stranger's, capturing the nuance of interpersonal trust.

Similarly, the DAN-SNR framework developed by Yanbo Liu and Keqing He 0002 replaces traditional Recurrent Neural Networks with a dual-channel self-attention mechanism. One channel tracks the user's personal movement history, while the other captures long-term social influences. Their work in DAN-SNR: A Deep Attentive Network for Social-Aware Next Point-of-Interest Recommendation demonstrates that social pressure and shared preferences are critical components of human mobility that complement individual habits.

The Reciprocal Nature of Networks

Social influence is not a one-way street; it is a mutually reinforcing system. As highlighted by Shoujin Wang and Liang Hu in The Era of Intelligent Recommendation, a user’s social connections inform their consumption patterns, and their consumption patterns, in turn, can lead to the formation of new social ties.

This reciprocity suggests that "preference" is a multi-dimensional construct. The SREPS model by Chuan Zhou and Chun-Yi Liu addresses this by projecting ratings, social ties, and item consumption into an "Essential Preference Space." This acknowledges that the reasons we "trust" a friend in a social network may be distinct from the reasons we "like" a specific movie, yet both dimensions are necessary to predict future behavior accurately.

From Content to Context

The shift toward social awareness also changes how we interpret content. In the realm of micro-videos, the RADAR framework (Xiao Wang et al., Micro-video Tagging via Jointly Modeling Social Influence and Tag Relation) treats tagging not just as a computer vision problem, but as a social imitation problem. By modeling how creators imitate their peers, the system uses social influence as an extrinsic signal to improve label prediction.

Even in sensitive areas like cyberbullying detection, researchers are moving away from isolated text analysis. Lu Cheng and Huan Liu demonstrate in TGBully that modeling the "flow" of social sessions and the established relationships between users provides a much clearer picture of online behavior than analyzing comments in a vacuum.

Future Directions: Explicit vs. Implicit Graphs

While the field has made strides in using implicit signals (like co-location or content similarity), a recurring challenge is the integration of explicit social structures. Many current models, such as those discussed by Yozen Liu in the context of Action Graphs, still struggle to fully bridge the gap between microscopic in-app actions and the broader ego-networks of users.

The intuition for the next generation of recommenders is clear: a user is defined not just by what they have done, but by who they are connected to. By treating the social graph as a primary feature rather than a secondary metadata source, recommendation systems can finally capture the collaborative nature of human decision-making.

Papers in this version

Showing 60 of 246 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_modeling_social_influence_for_personaliz_f8bb8a69,
  title = {Modeling Social Influence for Personalized Recommendations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/modeling-social-influence-for-personalized-recommendations/rdr_f8bb8a69ae074a448fee18e8c1117aa0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_modeling_social_influence_for_personaliz_f8bb8a69_v1,
  title = {Modeling Social Influence for Personalized Recommendations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/modeling-social-influence-for-personalized-recommendations/rdr_f8bb8a69ae074a448fee18e8c1117aa0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}