Semi-supervised Review-Aware Rating Regression (Student Abstract)

Xiangkui Lu, Jun WuAAAI 2023 · DOI

Introduction

Review-aware Rating Regression (RaRR) has emerged as a powerful technique in recommendation systems, leveraging textual reviews to provide context and explanation for user ratings. However, RaRR models frequently suffer from extreme data sparsity, as most users only rate a small fraction of items and provide even fewer detailed reviews. While Semi-Supervised Learning (SSL) techniques like co-training attempt to utilize unlabeled user-item pairs through pseudo-labeling, they often fall into a "noise trap." If a model generates an incorrect pseudo-label, it subsequently trains on its own error, leading to a degradation in predictive accuracy.

Methodology: The $CoT^2$ Framework

To address these challenges, we propose $CoT^2$ (Co-Training-Teaching), a framework that integrates the data-augmentation strengths of co-training with the noise-filtering robustness of co-teaching. The system maintains two distinct predictors based on Factorization Machines that serve as mutual "labelers" and "validators."

The $CoT^2$ process follows a four-step cycle:

  1. Labeling: Predictor A generates a pseudo-rating for an unlabeled data point.
  2. Validating: Predictor B predicts a rating for the same point and calculates the difference ($\Delta(x)$) between its prediction and Predictor A's label.
  3. Filtering: Data points with a small difference are deemed reliable and used for parameter updates; those with large discrepancies are discarded as noise.
  4. Role Swap: The predictors switch roles to ensure mutual learning.

To prevent the models from converging on the same errors, we enforce diversity through three mechanisms: Feature Diversity (using different review subsets), Data Diversity (using different bootstrap samples for initialization), and Parameter Diversity (using different random initializations for the underlying Factorization Machines).

Experimental Setup

We evaluated $CoT^2$ against several baselines, including review-only models (HFT, AHN) and existing semi-supervised models (CSEL, CoFM). Testing was conducted on three Amazon product datasets: Musical Instruments, Office Products, and Video Games. To test resilience to sparsity, we utilized three training-to-testing ratios: 8:2, 6:4, and 4:6.

Results and Analysis

Experimental results demonstrate that $CoT^2$ consistently outperforms all baseline models across every dataset and training split.

Table 1: Performance comparison (MSE) on three Amazon datasets across different training-to-testing ratios.

As shown in Table 1, the performance advantage of $CoT^2$ is most pronounced in high-sparsity scenarios (the 4:6 split). For instance, on the Musical Instruments dataset at 4:6, $CoT^2$ achieved a Mean Squared Error (MSE) of 0.775, a significant improvement over the next best semi-supervised model, CoFM (0.798). This suggests that the "labeler-validator" loop effectively filters out the noise that typically degrades performance in standard co-training paradigms.

Conclusion and Future Work

The $CoT^2$ framework provides a robust solution for leveraging unlabeled data in recommendation systems without the risk of error propagation. By treating disagreement between diverse models as a proxy for uncertainty, the system safely exploits massive amounts of unlabeled data. Future research will focus on reducing the computational overhead of maintaining dual predictors and exploring the framework's generalizability to other semi-supervised tasks beyond rating regression.

Summary updated

BibTeX
@inproceedings{lu2023semi,
  author = {Xiangkui Lu and Jun Wu},
  title = {{Semi-supervised Review-Aware Rating Regression (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i13.26996},
}
Figure