TLDR: FreqRec is a new sequential recommendation model that uses a Frequency-Enhanced Dual-Path Network to improve predictions. It addresses limitations of current models by jointly capturing inter-session (across users) and intra-session (within a user’s history) behaviors using learnable frequency-domain MLPs. FreqRec also introduces a frequency-domain consistency loss to align predicted and actual spectral patterns, leading to more accurate and robust recommendations, especially in sparse or noisy data environments.
Sequential recommendation systems are everywhere, from suggesting your next binge-watch to curating your social media feed. These systems aim to predict what you’ll want next based on your past interactions. Over the years, they’ve evolved from simple Markov chains to sophisticated deep learning models, including those based on the powerful Transformer architecture.
However, even the most advanced Transformer-based recommenders face significant hurdles. One major issue is their vulnerability to noisy data, which can lead to models that perform well on training data but poorly on new, unseen data. Another challenge stems from the nature of self-attention mechanisms, which, while effective, can act like a ‘low-pass filter.’ This means they tend to smooth out sudden changes or periodic patterns in user behavior, losing the crucial high-frequency signals that are key to truly personalized recommendations.
To tackle these problems, recent research has explored incorporating frequency-domain modules into recommendation systems. These methods aim to recover those lost high-frequency signals. Yet, existing frequency-aware solutions often treat each user’s interaction session in isolation, ignoring valuable patterns that might span across multiple sessions. Furthermore, they typically optimize using only time-domain objectives, which don’t directly encourage the model to understand and align with the underlying frequency structure of user preferences.
Introducing FreqRec: A New Approach to Sequential Recommendation
A new research paper, titled “Exploiting Inter-Session Information with Frequency-enhanced Dual-Path Networks for Sequential Recommendation,” introduces a novel framework called FreqRec. Developed by Peng He, Yanglei Gan, Tingting Dai, Run Lin, Xuexin Li, Yao Liu, and Qiao Liu, FreqRec is designed to overcome the limitations of previous models by jointly capturing both inter-session (across multiple sessions) and intra-session (within a single session) behaviors. You can read the full paper here.
FreqRec employs a learnable Frequency-domain Multi-layer Perceptrons (FreqMLPs) within a dual-path network. This innovative design allows the model to intelligently amplify or suppress specific frequencies, rather than relying on predefined filters. The network processes user interaction history through two complementary frequency paths:
-
Global Spectral Aggregator (GSA): This path distills cohort-level rhythms, identifying common patterns and trends across many user sessions. This is particularly useful for addressing data sparsity, as population-wide trends can inform recommendations even when an individual user’s history is limited.
-
Local Spectral Refiner (LSR): This path focuses on sharpening user-specific nuances, capturing the fine-grained and often transient interests of individual users within their own interaction sequences.
These two spectral modules can be combined in either a parallel or serial fashion. Experiments showed that the parallel fusion strategy, which processes inter- and intra-session information independently before combining them, consistently achieved superior performance. This is because it preserves the integrity of both types of spectral information, allowing the model to leverage complementary insights without one overwriting the other.
A Dual-Objective Training Strategy
Beyond its unique architecture, FreqRec introduces a composite training objective. It combines the standard cross-entropy loss, which focuses on predicting the next item accurately, with a novel frequency-domain consistency loss. This consistency loss explicitly aligns the model’s predicted spectral signatures with the actual ground-truth spectral signatures. By doing so, FreqRec is directly incentivized to recover and utilize high-frequency interaction patterns that are often overlooked by models trained solely with time-domain objectives.
Also Read:
- HyMoERec: A Dynamic Approach to Personalized Recommendations
- Unlocking Adaptive Recommendations: How AdaRec Uses LLMs for Personalized Experiences
Robust Performance and Key Findings
Extensive experiments conducted on three real-world benchmark datasets (Amazon Beauty, Sports & Outdoors, and Toys & Games) demonstrate FreqRec’s effectiveness. The model consistently outperforms strong baselines, showing significant improvements in recommendation accuracy metrics like Hit Rate (HR@K) and Normalized Discounted Cumulative Gain (NDCG@K).
Ablation studies, where components of FreqRec were selectively removed, confirmed the critical role of both the Global Spectral Aggregator and the Local Spectral Refiner, as well as the frequency-domain consistency loss. Removing any of these components led to a noticeable drop in performance, highlighting their synergistic effect.
Furthermore, FreqRec proved to be remarkably robust under challenging conditions, including data sparsity (users with very short interaction histories) and noisy logs. Its ability to aggregate cohort-level patterns helps compensate for individual data scarcity, while its dual-path spectral processing and frequency-domain loss contribute to effective noise suppression and better generalization.
In conclusion, FreqRec represents a significant step forward in sequential recommendation. By intelligently exploiting inter-session information and explicitly aligning predicted and true spectral signatures, it offers a powerful and robust solution for delivering more personalized and accurate recommendations.
