GSPRec: Temporal-Aware Graph Spectral Filtering for Recommendation
Ahmad Bin Rabiah, Julian McAuley
TL;DR
GSPRec is a novel graph spectral recommendation model that incorporates temporal dynamics and frequency-aware filtering to better capture user preferences and sequential item transitions, leading to improved recommendation accuracy.
Contribution
It introduces a spectral filtering approach with dual filters and sequential graph construction to enhance recommendation performance over existing methods.
Findings
GSPRec outperforms baselines with an average 6.77% NDCG@10 improvement.
Dual-filtering captures both global trends and user-specific patterns.
Sequential graph augmentation improves the modeling of item transitions.
Abstract
Graph-based recommendation systems are effective at modeling collaborative patterns but often suffer from two limitations: overreliance on low-pass filtering, which suppresses user-specific signals, and omission of sequential dynamics in graph construction. We introduce GSPRec, a graph spectral model that integrates temporal transitions through sequentially-informed graph construction and applies frequency-aware filtering in the spectral domain. GSPRec encodes item transitions via multi-hop diffusion to enable the use of symmetric Laplacians for spectral processing. To capture user preferences, we design a dual-filtering mechanism: a Gaussian bandpass filter to extract mid-frequency, user-level patterns, and a low-pass filter to retain global trends. Extensive experiments on four public datasets show that GSPRec consistently outperforms baselines, with an average improvement of 6.77% in…
Peer Reviews
Decision·Submitted to ICLR 2026
1. By transforming raw click sequences into a symmetric graph before spectral filtering, GSPRec preserves temporal cues while maintaining the stability of eigendecomposition. 2. The entire pipeline requires only matrix decomposition and lightweight filters, demonstrating high efficiency as shown in the experiments. 3. The conducted experiments demonstrate the superior performance of the proposed method.
1. Throughout the paper the mid-frequency Gaussian bandpass is said to capture ‘user-specific sequential patterns’. Is there any case study to demonstrate it? 2. The experiments only report aggregate top-k accuracy, but emphasizing mid-frequency components may systematically boost long-tail items and thus change the exposure distribution. Could the authors include fairness-aware metrics in the experiments?
1. The paper introduces bandpass filtering to recommendation systems, explicitly targeting mid-frequency components—a distinct approach from prior GSP methods that focus on low-pass or high-pass filtering. The dual-filter architecture combining bandpass and lowpass filters represents a new design for balancing personalization and popularity in spectral collaborative filtering. 2. The paper offers a new lens for understanding collaborative signals through spectral analysis, proposing that differ
1. Unclear Utilization of Temporal Information The paper claims to be "Temporal-Aware" throughout, but the utilization of temporal order is not clearly demonstrated. Lines 196-198 symmetrize the sequential transition matrix (S⁰[i,j] = 1 if i→j OR j→i exists), and Equation (3) produces a symmetric matrix S̃ where S̃[a,b] = S̃[b,a]. While the authors note this symmetrization is required for spectral analysis (lines 183-185), it remains unclear how the method distinguishes "a→b" from "b→a" given
(S1) The paper proposed a new approach to modeling item–item relations by explicitly utilizing sequential information in a bidirectional manner. (S2) The newly proposed band-pass filter was theoretically well defined, enabling the model to exploit not only low-frequency components but also mid- and high-frequency components.
(W1) The method increases computational complexity, while the performance gain from the multi-hop diffusion component is relatively limited. (W2) The experimental results for hyperparameter sensitivity showed an inconsistent tendency that does not align with the implementation details. (W3) Most importantly, due to the fact that the method leverages additional information (e.g., sequential information) that is not typically available to conventional collaborative filtering baselines, the compa
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsRecommender Systems and Techniques · Advanced Graph Neural Networks · Machine Learning in Healthcare
MethodsDiffusion