Abstract
Collaborative Filtering (CF) often suffers from data sparsity and cold-start issues. While social recommendation leverages trust networks to mitigate these problems, existing methods frequently conflate social trust with user preference or treat social influence as a static weight. This paper introduces TrustEV, a multi-task learning framework that utilizes a probabilistic Feature Evolution Unit and a Bayesian meta-controller to bridge social trust and recommendations. TrustEV allows for dynamic information exchange between distinct feature spaces, significantly improving recommendation accuracy and robustness in sparse data environments.
Introduction
Modern recommender systems rely heavily on Collaborative Filtering, yet they struggle when interaction data is scarce. Social recommendation seeks to address this by incorporating auxiliary data from a user's social trust network, based on the premise that friends often share similar tastes. However, traditional approaches often fall into the "Common Space Fallacy," assuming that friendship and preference similarity are identical and forcing both into a single feature space. Furthermore, they often treat social influence as a fixed parameter, failing to capture its dynamic nature.
TrustEV addresses these limitations by treating recommendation and social network analysis as separate but related tasks. By maintaining independent vectors for preference and trust while allowing for stochastic feature exchange, the model respects the distinct nature of social ties and consumer tastes.
The TrustEV Framework
Multi-Task Architecture
The TrustEV architecture consists of two primary components: a Collaborative Filtering model for predicting user-item ratings and a Network Embedding model for predicting social links between users. Each user is represented by two independent vectors: a preference factor and a trust factor.
Feature Evolution Unit
The core innovation of TrustEV is the Feature Evolution Unit (FEU). Inspired by biological assortment, the FEU enables the preference and trust vectors to "exchange" features element-wise based on a Bernoulli distribution.
Key characteristics of the FEU include:
- Probabilistic Sharing: Features are swapped stochastically, introducing beneficial noise that helps the model escape local optima during training.
- Residual Connections: To mitigate gradient vanishing, the unit employs a weighted summation of original and exchanged features, ensuring stable learning across layers.
Meta-Controller Optimization
To avoid the manual tuning of exchange probabilities, TrustEV incorporates a meta-controller governed by Bayesian Optimization (specifically a Gaussian Process). This "outer loop" automatically identifies optimal settings for the evolution units by evaluating model performance on a validation set, aligning with modern AutoML trends.
Experimental Evaluation
Datasets and Baselines
TrustEV was evaluated on three benchmark datasets: Epinions, Ciao, and FilmTrust. The model was compared against nine state-of-the-art baselines, including PMF, TrustSVD, and NCF.
Performance Analysis
TrustEV consistently outperformed all baselines across all datasets. On the Epinions dataset, it achieved a 1.63% improvement in Mean Absolute Error (MAE) and a 0.47% improvement in Root Mean Square Error (RMSE) over the strongest competitors.
A crucial finding was the model's superior performance in cold-start scenarios. For users with fewer than five ratings, TrustEV showed significant gains, demonstrating that the feature evolution mechanism effectively leverages social trust to compensate for missing interaction data. Ablation studies further confirmed that both the meta-controller and the evolution unit are vital for optimal performance; removing them (as in TrustEV-FIX and TrustEV-COMM variants) led to a marked decline in accuracy.
Conclusion
TrustEV represents a significant step forward in social recommendation by decoupling social trust from user preference while facilitating intelligent information exchange. The use of stochastic feature evolution and automated meta-control provides a robust framework for multi-task learning. Future research may explore the application of this architecture to dynamic social networks and other multi-task scenarios where information sharing between distinct domains is required.
Summary updated