Group-aware User Behavior Simulation (G-UBS)

• G-UBS is a framework that incorporates group-level context to robustly simulate user behaviors, addressing challenges of sparse and noisy implicit feedback.
• It employs a dual-agent approach with a User Group Manager for dynamic clustering and a User Feedback Modeler using reinforcement learning for accurate behavior interpretation.
• Empirical results on the IF-VR benchmark demonstrate improved play ratios and reasoning accuracy, validating its effectiveness over traditional individual-based models.

Group-aware User Behavior Simulation (G-UBS) refers to computational frameworks that emulate and analyze user behaviors in digital environments, explicitly leveraging group-level context in both the modeling and interpretation of user actions. The core premise is that certain user behaviors—especially those involving complex, sparse, or noisy feedback—can be understood with greater robustness and depth by incorporating shared dynamics, preferences, and characteristics from relevant user groups. This paradigm is increasingly fundamental in domains such as recommendation systems, behavioral analytics, and synthetic data generation, where understanding or simulating the interplay of individual and group signals is essential to accurate system evaluation and robust decision-making.

1. Motivation and Conceptual Underpinnings

G-UBS arises chiefly from the observation that explicit feedback (likes, ratings) is uncommon in many real-world applications, while implicit feedback (such as fast-negation actions or skip behaviors) is abundant yet highly noisy. For example, a rapidly skipped video might result from external perturbations (e.g., accidental taps or distractions) rather than genuine disinterest. Traditional models that interpret user actions in isolation risk misattributing these behaviors, undermining recommendation accuracy and user satisfaction. The G-UBS paradigm aims to resolve this by contextualizing individual feedback within group-derived profiles, thus calibrating interpretation through shared characteristics and historical patterns seen in relevant user cohorts (Chen et al., 7 Aug 2025).

2. System Architecture: User Group Management and Reinforcement Modeling

G-UBS operates through a dual-agent workflow:

• User Group Manager (UGM): Responsible for forming user clusters and generating canonical group profiles. Employing a "summarize-cluster-reflect" workflow, UGM utilizes LLMs to process a user population—each represented as [ID, occupation, age, gender, interest tags]—to (i) summarize and categorize into $k$ groups; (ii) cluster users by similarity to group representatives, using dynamic thresholds; (iii) reflect and refine assignment through inspection of behavioral histories (e.g., play rates, click sequences) ensuring the resulting group profiles $P_g$ are behaviorally consistent and demographically meaningful.
• User Feedback Modeler (UFM): Employs a group-aware reinforcement learning (RL) process for simulating and interpreting implicit user feedback on content recommendations. UFM is first pre-trained via supervised fine-tuning with explicit feedback and associated chain-of-thought (CoT) rationales. During RL, UFM samples three profile types per instance: the focal user $u_T$, their group profile $P_g$, and a similar user $u_S$. For each, the model generates predicted feedback (e.g., skip signals, reasoning chains). Rewards are structured and weighted, including format correctness, skip prediction accuracy, and reasoning alignment, and are aggregated as:

$R(o) = r_{format} + r_{skip} + r_{choice}$

with normalized quality $A_R$ computed for variance reduction. The optimization objective includes a KL-divergence regularization to encourage policy stability:

$$\max_{\pi_\theta} \mathbb{E}_{o \sim \pi_{\theta_{old}}}\left[\sum_{o \in O} \frac{\pi_\theta(o)}{\pi_{\theta_{old}}(o)} A_R - \beta D_{KL}(\pi_\theta \Vert \pi_{ref}) \right]$$

This process ensures that UFM's decision logic is robustly aligned with group-informed context while respecting the reliability of supervised pre-training.

3. Clustering and Profiling: Summarize–Cluster–Reflect Workflow

The UGM workflow consists of three tightly integrated modules:

• Summarize: An LLM generates an initial categorization of the userbase into $k$ clusters, selecting representative users $U_g$ and assigning users to tentative groups $\mathcal{C}_g$ using similarity metrics and dynamic thresholds $\tau_g$.
• Cluster: Users are assigned into groups where $Sim(u, u_g) \geq \tau_g$. This enables nuanced, context-sensitive clusters that capture both demographic and behavioral affinities.
• Reflect: The clustering is refined by reviewing each user’s interaction sequence (e.g., completions, skips, clicks, titles, durations); only users whose behavioral history matches the group profile $\hat{P}_g$ (as validated by a Match function against expected behaviors) are retained for the final group profile. This ensures coherence of the shared group context that will be distilled to guide UFM.

This workflow is crucial for distilling group context that is not only demographically grounded but also aligned with shared consumption patterns and latent interests.

4. Group-aware Reinforcement Learning for Feedback Interpretation

Within the UFM, the reinforcement learning strategy is group-aware in its policy optimization. For a given user, the model generates outputs for their profile $u_T$, group profile $P_g$, and similar peer $u_S$. These outputs—denoted $o_T$, $o_G$, and $o_S$—each receive a reward reflecting output structure, skip/click prediction accuracy, and explanation validity.

The RL procedure employs a relative scoring $A_R$ (normalized across samples) to counteract reward scaling variability due to group or user differences. Optimization is formally regularized to balance innovation (policy improvement) and conservative update, controlled by a KL penalty. When group-level context is not available, the model falls back to user-specific context, ensuring robustness at both cohort and individual resolutions.

5. Benchmarking: IF-VR Dataset and Empirical Results

The G-UBS approach is validated on the IF-VR (Implicit Feedback Video Recommendation) benchmark, which contains:

• 15,000 user profiles (with age, gender, occupation, and interest tags),
• 25,000 video items (with titles and visual data),
• Approximately 933,000 implicit interaction records (fast-skip, completion, explicit dislikes, click events).

Key empirical findings include:

• A 4.0% higher rate of videos achieving $>$30% play ratio compared to mainstream LLM and MLLM baselines.
• 14.9% higher reasoning (explanation) accuracy on IF-VR, indicating substantial improvement in interpreting implicit feedback origins.
• Consistently superior scores across metrics such as Person Play Rate, Judge F1, and Reason F1—suggesting that contextualizing behavior through group-level profiles yields more precise behavioral simulation and prediction in real-world, noisy feedback scenarios.

System	Play Rate (>30%)	Reasoning Accuracy	Judge F1
G-UBS (proposed)	Higher by 4.0%	Higher by 14.9%	Superior
Mainstream LLM/MLLM	Baseline	Baseline	Baseline

These gains are attributed to the ability of G-UBS to dilute individual noise through aggregation and context provided by group profiles, thereby preventing overfitting to anomalies or spurious signals in implicit data.

6. Broader Implications and Future Directions

The success of G-UBS demonstrates the effectiveness of incorporating group-level context into user behavior simulation and feedback interpretation. The paradigm achieves:

• Enhanced robustness to feedback noise, closing the gap between explicit and implicit signal interpretation.
• Stronger reasoning and explanatory capacity, facilitating interpretability and transparency in recommendation.
• Higher engagement and accuracy, supporting user retention and system trustworthiness.

Potential future avenues include integrating additional modalities (audio content, fine-grained temporal interaction sequences), developing adaptive or dynamic clustering for evolving user interests, and extending the G-UBS methodology to broader content domains (such as news, shopping, or multi-modal entertainment), as well as accommodating cross-domain behavioral patterns.

A plausible implication is that G-UBS-like frameworks may become the default paradigm for synthetic user simulation and implicit feedback modeling in next-generation recommender systems, especially in scenarios where explicit analysis is infeasible or prohibitively expensive.

7. Summary

Group-aware User Behavior Simulation systematically integrates contextual group-level profiles into user modeling and feedback interpretation, with a focus on robustly handling sparse, noisy, and implicit data scenarios. Through a synergistic dual-agent architecture—comprising a group manager for cluster-based profile distillation and a reinforcement-learning-based modeler for behavior simulation—G-UBS achieves notable improvements in feedback understanding and recommendation outcomes compared to baseline LLM methods. Its empirical validation on the IF-VR benchmark substantiates its advantage and positions G-UBS as a robust foundation for resilient, interpretable, and high-fidelity simulation of user behaviors in complex digital ecosystems (Chen et al., 7 Aug 2025).