# STUDY: Socially Aware Temporally Casual Decoder Recommender Systems

**Authors:** Eltayeb Ahmed, Diana Mincu, Lauren Harrell, Katherine Heller, Subhrajit Roy  
**arXiv:** [2306.07946](https://arxiv.org/abs/2306.07946)  
**Date:** 2023-06-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.07946v1/

## Key points

- Recommender systems have become essential in the era of abundant data
- Existing methods that use social network information for recommendations are complex and difficult to train
- The proposed approach, called STUDY, performs joint inference over groups of adjacent users using a modified transformer decoder network
- STUDY outperforms both social and sequential methods in a school-based educational content setting
- The success of the model relies on leveraging a social network structure that captures similarities in user behavior
- Click-through rate prediction is a popular approach for recommender systems, with various modeling approaches available when social connectivity information is present
- Unlike previous work, the proposed approach utilizes a single transformer decoder with a specially designed attention mask for joint inference over groups of users
- The problem addressed is making personalized recommendations based on historical user-item interactions and social connectivity.

## Summary

In the era of abundant data, recommender systems have become essential for helping users find personalized recommendations. However, existing methods that utilize social network information to improve recommendations often suffer from complex architectures and training procedures. Additionally, many of these methods rely on graph neural networks, which are notoriously difficult to train. To address these challenges, we propose a novel approach called Socially-aware Temporally caUsal Decoder recommender sYstems (STUDY). STUDY performs joint inference over groups of users who are adjacent in the social network graph using a modified transformer decoder network. This allows us to make better recommendations while maintaining the simplicity of a single homogeneous network that models all interactions in the data. We evaluate our method in a school-based educational content setting, using classroom structure to define social networks. Our results show that STUDY outperforms both social and sequential methods. To provide more context and insights into our model, we compare it against modern and classical recommendation baselines. We also conduct ablation studies to understand the factors driving our performance gains. Our findings reveal that our model's success relies on leveraging a social network structure that effectively captures similarities in user behavior. In related work, click-through rate prediction is a popular approach for recommender systems where the goal is to model the probability of a user clicking on an item presented to them. Model architectures used in this problem range from standard models like Transformers to more task-specific architectures. When social connectivity information is available for users, there are various modeling approaches that leverage this information for recommendation systems. These include matrix factorization approaches like TrustMF and Sorec as well as deep-learning based methods utilizing graph neural networks or attention mechanisms. However, unlike previous work that uses separate components or custom architectures to handle cross-user interactions, our approach utilizes a single transformer decoder with a specially designed attention mask for joint inference over groups of users. This allows us to handle both within-user and cross-user interactions consistently. The problem we address is making personalized recommendations of new items that are likely to be interesting given historical observational data of user-item interactions and the users' social connectivity.