Exploring and Tailoring the Test-Time Augmentation for Sequential Recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Data augmentation is crucial for addressing data sparsity in sequential recommendation (SR).
• Existing training-time augmentation methods require costly retraining or architectural modifications for well-trained models.
• Test-time augmentation (TTA) offers a promising alternative by augmenting data during inference, avoiding training overhead.

Purpose of the Study:

• To explore the effectiveness of test-time augmentation (TTA) for sequential recommendation (SR).
• To identify limitations of existing augmentation operators for TTA and propose novel solutions.
• To develop an efficient and generalizable TTA method that improves SR performance without extensive retraining.

Main Methods:

• Experimentally evaluated existing augmentation operators (Substitute, Mask) for TTA in SR.
• Introduced TNoise (uniform noise injection) and TMask (mask token handling) to address limitations of existing operators.
• Implemented strategies for improved data diversity (uniform ratio sampling) and sequence length adaptation (smoothing/lengthening for short, thresholding for long sequences).

Main Results:

• Substitute and Mask operators showed potential for TTA, retaining sequential patterns with appropriate perturbations.
• TNoise and TMask demonstrated effectiveness, efficiency, and generalizability across various SR scenarios.
• Proposed methods improved data diversity and mitigated performance degradation on both short and long sequences.

Conclusions:

• Test-time augmentation (TTA) is a viable and efficient approach for enhancing sequential recommendation models.
• The proposed TNoise and TMask methods, along with sequence length adaptation strategies, significantly improve SR performance.
• This work provides a practical solution for leveraging data augmentation in SR without the need for model retraining.