Frequency-Aware Framework Improves Sequential Recommendation Accuracy Across Sparse Datasets

A sequential recommendation framework designated as DynFreqRec overall outperforms representative baseline models across multiple public datasets, according to a study published in Applied Intelligence. The experimental findings demonstrate that the system maintains favorable recommendation accuracy under sparse user interaction conditions and across short-sequence user groups. Sequential recommendation frameworks are designed to predict an individual’s next interaction, such as identifying the next book a user is likely to interact with, by modeling interest evolution over time from historical activity records.

Real-world interaction sequences present persistent challenges for recommendation algorithms, including interaction sparsity, feedback noise, and interest non-stationarity. Under these conditions, long-term trends, periodic preferences, and local variations become intertwined, complicating representation learning under limited supervision. Existing time-frequency architectures can jointly model multi-scale user preferences but focus predominantly on sequence encoding. Furthermore, conventional time-domain augmentations such as random cropping and masking can disrupt critical interaction patterns and semantic continuity, while single-pass predictions during test time remain vulnerable to input perturbations and model uncertainty.

To address these encoding and inference limitations, DynFreqRec coordinates training and inference stages around a frequency-aware architecture. The system builds upon an existing Base Time-Frequency Encoder that deploys adaptive frequency-domain filtering alongside the Haar Wavelet Transform. This foundational component characterizes global preferences, periodic interests, and local non-stationary variations within the input sequence.

During the training stage, the framework implements Frequency-Domain Phase Perturbation Contrastive Learning, known as PhaseCL. This contrastive learning mechanism generates controlled phase-perturbed augmented views while preserving magnitude spectrum stability. By maintaining magnitude stability during perturbation, PhaseCL mitigates representation degeneration that frequently arises under sparse supervision.

During the inference stage, the framework employs FreqTTA to aggregate prediction outputs derived from both the original view and frequency-domain phase-perturbed views under Monte Carlo Dropout. Aggregation weights across the views are based on prediction entropy.

Subject of Research: Frequency-aware sequential recommendation systems

Article Title: DynFreqRec: A frequency-aware training–inference synergistic enhancement framework for sequential recommendation

Article References: Huang, Z., Song, X., Xiao, S., & Wu, X. (2026). DynFreqRec: A frequency-aware training–inference synergistic enhancement framework for sequential recommendation. Applied Intelligence, 56(16), Article 532. https://doi.org/10.1007/s10489-026-07557-8

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07557-8

Keywords: sequential recommendation, DynFreqRec, frequency-domain filtering, Haar Wavelet Transform, contrastive learning, test-time augmentation, Monte Carlo Dropout, Applied Intelligence, frequency-aware, training, inference, synergistic