Data Augmentation as Free Lunch: Exploring the Test-Time Augmentation for Sequential Recommendation

Abstract: Data augmentation has become a promising method of mitigating data sparsity in sequential recommendation. Existing methods generate new yet effective data during model training to improve performance. However, deploying them requires retraining, architecture modification, or introducing additional learnable parameters. The above steps are time-consuming and costly for well-trained models, especially when the model scale becomes large. In this work, we explore the test-time augmentation (TTA) for sequential recommendation, which augments the inputs during the model inference and then aggregates the model's predictions for augmented data to improve final accuracy. It avoids significant time and cost overhead from loss calculation and backward propagation. We first experimentally disclose the potential of existing augmentation operators for TTA and find that the Mask and Substitute consistently achieve better performance. Further analysis reveals that these two operators are effective because they retain the original sequential pattern while adding appropriate perturbations. Meanwhile, we argue that these two operators still face time-consuming item selection or interference information from mask tokens. Based on the analysis and limitations, we present TNoise and TMask. The former injects uniform noise into the original representation, avoiding the computational overhead of item selection. The latter blocks mask token from participating in model calculations or directly removes interactions that should have been replaced with mask tokens. Comprehensive experiments demonstrate the effectiveness, efficiency, and generalizability of our method. We provide an anonymous implementation at https://github.com/KingGugu/TTA4SR.

• The paper demonstrates that test-time augmentation using Mask and Substitute effectively enhances sequential recommendation by preserving key interaction patterns.
• It introduces novel TNoise and TMask operators that inject noise or apply masking to improve efficiency and maintain sequential dynamics during inference.
• Experiments on diverse datasets and models confirm that TTA methods, especially TMask-R, outperform traditional training-time augmentation approaches.

This paper investigates the application of Test-Time Augmentation (TTA) to sequential recommendation (SR) systems. The core idea is to improve recommendation accuracy by augmenting input sequences during inference and aggregating the model's predictions on these augmented versions, avoiding the need for retraining or modifying the model architecture.

The paper begins by highlighting the challenges of data sparsity in SR and the limitations of existing training-time data augmentation methods, which often require retraining or architectural changes, incurring significant time and cost overhead. TTA is presented as a cost-effective alternative.

The research is structured around three key questions:

1. Q1: Can existing sequence data augmentation operators be used for TTA? The paper performs an empirical study using several representative augmentation operators (Crop, Reorder, Sliding Windows, Mask, Substitute, and Insert) with GRU4Rec and SASRec on Amazon Beauty and Sports datasets. The results show that Mask and Substitute operators consistently outperform others when used for TTA.
2. Q2: What factors make Mask and Substitute effective for TTA, and what limits the performance of other operators? The paper analyzes the performance based on data similarity between augmented and original data and the impact on sequential patterns. It finds that Mask and Substitute introduce "appropriate perturbations" while largely preserving the original sequential patterns. Other operators, like Crop, Reorder, and Insert, tend to disrupt these patterns or lose crucial recent interaction information. The paper also explores the impact of randomly selecting items for augmentation versus selecting "key interactions" identified by a LLM. The study found that randomly selecting items for augmentation is a satisfactory scheme, possibly because key interactions are naturally less frequently selected in random sampling.
3. Q3: Based on the analysis, can more effective operators for TTA be proposed? Addressing limitations of Substitute (high computational cost) and Mask (interference from unlearned mask token embeddings), the authors propose two new TTA operators:
   • TNoise: Injects uniform noise directly into the item representation, avoiding the item selection overhead of Substitute.

• TMask: Two variants are proposed:
  • TMask-B: Blocks mask tokens by setting their embeddings to zero during model calculation.
  • TMask-R: Removes items that would have been masked.

The paper presents extensive experiments on four datasets (Amazon Beauty, Sports, Home, and Yelp) using GRU4Rec and SASRec as backbone models. The results demonstrate the effectiveness, efficiency, and generalizability of TNoise and TMask, especially TMask-R. TMask-R consistently achieves strong performance and often outperforms existing TTA methods. The authors also benchmark against training-time augmentation (CLS4Rec) to highlight the benefits of TTA in terms of deployment simplicity and computational efficiency. Further experiments with other SR models (NARM, NextItNet, LightSANs, and FMLP-Rec) demonstrate the general applicability of the proposed TTA methods. Finally, hyperparameter analysis of the proposed method is performed.

In summary, the paper makes the following key contributions:

• Identifies the potential of existing sequence data augmentation operators for TTA.
• Analyzes the factors that contribute to the effectiveness of Mask and Substitute.
• Proposes TNoise and TMask, novel and efficient TTA operators.
• Provides comprehensive experimental validation of the proposed methods, demonstrating their effectiveness, efficiency, and generalizability.