Causally-Informed Reinforcement Learning for Adaptive Emotion-Aware Social Media Recommendation

TL;DR

This work tackles the problem of emotional well-being in social media recommendations by proposing ESMR, a causally informed, emotion-aware hybrid policy that switches between LightGBM-based engagement and reinforcement learning during prolonged negative states. Emotion states are predicted with a Transformer-based TabTransformer, and causal discovery via DirectLiNGAM identifies top drivers used to shape rewards. The approach demonstrates that incorporating emotion dynamics and causal factors into the reward function can reduce negative emotional trajectories and volatility while maintaining engagement—achieving a more emotionally intelligent recommender. The findings highlight the feasibility and significance of dynamically balancing engagement with emotional stability, offering a path toward healthier and more adaptive content curation in practice.

Abstract

Social media recommendation systems play a central role in shaping users' emotional experiences. However, most systems are optimized solely for engagement metrics, such as click rate, viewing time, or scrolling, without accounting for users' emotional states. Repeated exposure to emotionally charged content has been shown to negatively affect users' emotional well-being over time. We propose an Emotion-aware Social Media Recommendation (ESMR) framework that personalizes content based on users' evolving emotional trajectories. ESMR integrates a Transformer-based emotion predictor with a hybrid recommendation policy: a LightGBM model for engagement during stable periods and a reinforcement learning agent with causally informed rewards when negative emotional states persist. Through behaviorally grounded evaluation over 30-day interaction traces, ESMR demonstrates improved emotional recovery, reduced volatility, and strong engagement retention. ESMR offers a path toward emotionally aware recommendations without compromising engagement performance.

Figures (8)

• Figure 1: Overview of our methods. Phases I–II simulate users and predict emotional states; Phase III implements ESMR, a hybrid LightGBM–RL model with causal reward shaping based on emotion-influencing features.
• Figure 2: Architecture of TabTransformer.
• Figure 3: DAG of Causal Parents for the Happy and Stressed State, highlighting emotional, behavioral, and transitional features identified through causal discovery. Color-coded nodes reveal key drivers of happiness and stress based on user engagement dynamics.
• Figure 4: Symbolic execution flow of the ESMR hybrid framework. At each day $t$, state $s_t$ and emotion history determine the active policy $\pi_t$. Actions $a_t$ are selected using LightGBM (during stable periods) or RL (during vulnerable states). Rewards $r_t$ and emotional states $\hat{e}_t$ are updated accordingly and logged for learning and evaluation.
• Figure 5: DAG of Causal Parents for Next Day Happy and Next Day Stressed State, highlighting emotional, behavioral, and transitional features identified through causal discovery. Color-coded nodes reveal key drivers of happiness and stress based on user engagement dynamics.