Abstract
This paper addresses the challenge of jointly modeling user intent diversity and behavioral uncertainty in recommender systems. A unified representation learning framework is proposed, building a multi-intent representation module and an uncertainty modeling mechanism to extract multi-granularity interest structures from user behavior sequences. Behavioral ambiguity and preference fluctuation are captured using Bayesian distribution modeling. In the multi-intent modeling part, the model introduces multiple latent intent vectors, weighted and fused using an attention mechanism. In the uncertainty modeling part, the model learns the mean and covariance of behavior representations through Gaussian distributions. A learnable fusion strategy combines long-term intent and short-term behavior signals, improving both recommendation accuracy and robustness. Experimental results on standard public datasets show that it outperforms existing representative models across multiple metrics and demonstrates greater stability under cold-start and behavioral disturbance scenarios.
TL;DR
This research introduces a unified framework that jointly models user intent diversity (shifting desires) and behavioral uncertainty (stochastic noise in clicks). By combining multi-latent intent vectors with Bayesian Gaussian distributions, the model achieves superior accuracy and remains remarkably robust even when dealing with "cold-start" users or erratic behavior.
The Challenge: Intent Diversity and Behavioral Uncertainty
Traditional recommender systems often treat user behavior as a linear progression. However, real-world human behavior is inherently complex. The authors identify two major gaps in current models:
- Intent Diversity: A single user contains multitudes. Interests shift based on context (e.g., professional vs. personal), which most models struggle to capture simultaneously.
- Behavioral Uncertainty: Not every click is a strong signal of preference. Random browsing or curiosity-driven exploration introduces "noise." Standard models often mistake this noise for a hard preference, leading to poor recommendations.
Methodology: A Unified Architecture
The researchers propose a two-pronged architecture to create a more "human-like" representation of a user:
- Multi-Intent Representation Module: Instead of a single vector, the model uses multiple latent intent vectors ($z_k$). An attention mechanism weights these based on context to generate a semantically rich "long-term" preference.
- Uncertainty Modeling Mechanism: To handle noise, the model represents user behavior as a Gaussian distribution defined by a mean ($\mu$) and a covariance ($\Sigma$). The mean represents the predicted preference, while the covariance captures uncertainty. This is optimized using a Sequential Variational Autoencoder (SVAE) framework.
- Learnable Fusion: A learnable weight ($\gamma$) balances long-term intent signals with short-term, uncertain behavior signals to produce the final recommendation score.
Experimental Evaluation
The model was evaluated using the Amazon Electronics dataset, a standard benchmark known for high sparsity and a "long-tail" distribution. Raw interaction data were transformed into ordered click-path sequences to test the model's ability to handle temporal shifts and behavioral noise.
Key Results and Discussion
The proposed model significantly outperformed mainstream baselines like SASRec, BERT4Rec, and DenoiseRec across all metrics.
- Accuracy and Ranking: The model reached an HR@10 of 66.4% and an NDCG@10 of 47.2%, showing a superior ability to place relevant items at the top of the list.
- Diversity: With an IAS score of 0.712, the model confirmed its ability to capture a wider variety of user interests compared to traditional methods.
- Robustness in Cold-Start: One of the most striking insights is the model's performance with "cold-start" users (those with only 1–3 actions). By treating uncertainty as a feature rather than noise, the model avoids overfitting to initial random clicks. For users with only one action, it maintained an HR@10 of 42.3%.
Conclusion and Future Work
This unified modeling strategy effectively alleviates the bottlenecks faced by traditional methods in complex user behavior scenarios. Future work includes integrating multimodal data (images/text), exploring cross-domain intent transfer, and improving model interpretability to explain "why" specific intents are chosen.
Summary updated