TV-Rec: Time-Variant Convolutional Filter for Sequential Recommendation

TL;DR

This work addresses the inefficiencies and limitations of fixed convolutional filters and self-attention in sequential recommendation by introducing TV-Rec, a time-variant convolutional filter framework inspired by graph signal processing. TV-Rec models temporal dynamics with per-position filter taps on a directed cyclic graph, removing the need for positional embeddings and heavy attention while enabling fast inference through linear spectral operators. The architecture comprises an embedding layer, a time-variant encoder with filter and feed-forward blocks, and a prediction layer trained with cross-entropy and an orthogonal regularizer on the filter bases. Across six public benchmarks, TV-Rec consistently outperforms state-of-the-art baselines by an average of $7.49\%$, including strong gains on long-range sequences, and exhibits favorable parameter efficiency and runtime performance, illustrating practical impact for scalable recommender systems.

Abstract

Recently, convolutional filters have been increasingly adopted in sequential recommendation for their ability to capture local sequential patterns. However, most of these models complement convolutional filters with self-attention. This is because convolutional filters alone, generally fixed filters, struggle to capture global interactions necessary for accurate recommendation. We propose Time-Variant Convolutional Filters for Sequential Recommendation (TV-Rec), a model inspired by graph signal processing, where time-variant graph filters capture position-dependent temporal variations in user sequences. By replacing both fixed kernels and self-attention with time-variant filters, TV-Rec achieves higher expressive power and better captures complex interaction patterns in user behavior. This design not only eliminates the need for self-attention but also reduces computation while accelerating inference. Extensive experiments on six public benchmarks show that TV-Rec outperforms state-of-the-art baselines by an average of 7.49%.

Figures (8)

• Figure 1: Comparison of a fixed filter in (a) and a time-variant convolutional filter in (b) under our line graph expression of a sequence of signals $x_i$, with $K=2$ and $N=3$. The output $y_j$, i.e., the filtered signal at index $j$, is produced by summing the filtered results. Arrow colors show each filter’s contribution to the output $y_j$, while different $h_i$ box colors represent different filters. In the fixed filter case in (a), the same filter $h_i$ is applied to every node, while the time-variant convolutional filter in (b) allows each node to have its own filter.
• Figure 2: Architecture of our proposed TV-Rec.
• Figure 3: Sensitivity to the number of basis vectors $m$.
• Figure 5: Visualization of learned graph filters on LastFM. The x-axis denotes the number of shifts in graph convolution, while the y-axis represents individual nodes, with higher numbers indicating more recent time points.
• Figure 7: Comparison of model inference time and NDCG@20 on Beauty. The size of each circle corresponds to the number of parameters.

Theorems & Definitions (1)

• proof