Towards More Robust and Accurate Sequential Recommendation with Cascade-guided Adversarial Training

TL;DR

This paper tackles robustness gaps in deep sequential recommender systems by introducing Cascade-guided Adversarial Training, which uses cascade-aware perturbations of history-item embeddings to improve both ranking accuracy and resilience to adversarial item replacements. The method employs a two-level adversarial framework and a cascade-effect calculation to re-scale perturbations, optimizing a combined loss that includes two adversarial components in addition to the standard ranking loss. Empirical results on four public datasets using SASRec and GRU4Rec show consistent accuracy gains and increased end-of-sequence robustness, with larger benefits on sparser data. The approach is practical, requiring minimal extra parameters and training overhead, making it readily applicable to real-world sequential recommender systems.

Abstract

Sequential recommendation models, models that learn from chronological user-item interactions, outperform traditional recommendation models in many settings. Despite the success of sequential recommendation models, their robustness has recently come into question. Two properties unique to the nature of sequential recommendation models may impair their robustness - the cascade effects induced during training and the model's tendency to rely too heavily on temporal information. To address these vulnerabilities, we propose Cascade-guided Adversarial training, a new adversarial training procedure that is specifically designed for sequential recommendation models. Our approach harnesses the intrinsic cascade effects present in sequential modeling to produce strategic adversarial perturbations to item embeddings during training. Experiments on training state-of-the-art sequential models on four public datasets from different domains show that our training approach produces superior model ranking accuracy and superior model robustness to real item replacement perturbations when compared to both standard model training and generic adversarial training.

Figures (6)

• Figure 1: A toy example of applying adversarial training on sequential recommendation. (a) How adversarial perturbations are applied on the learned item and user embeddings. (b) Generic adversarial training applies adversarial perturbations of the same magnitude (green circles) to all item embeddings. (c) The Cascade-guided adversarial training method dynamically choose the magnitude of the perturbations (blue circles) according to the different cascade effects of each interaction in the user history.
• Figure 2: An example of calculating cascade effects. The item with blue bounding box has cascade effects on the $10$ items with green bounding boxes plus itself. Its cascade value is $11$.
• Figure 3: NDCG vs. the number of training epochs.
• Figure 4: Influence of $\epsilon$
• Figure 5: Drop of accuracy by attacking the the first, middle, and the last items in the user sequences respectively. The y-axis depicts negative values, with larger bars indicating larger decreases in accuracy of the recommendation models.