Technical Summary: FEDIN – Frequency-Enhanced Deep Interest Network for Click-Through Rate Prediction

Problem Statement

Sequential recommendation models for Click-Through Rate (CTR) prediction face significant challenges in capturing latent periodic patterns within user interests. Real-world user behavior sequences are often a mixture of long-term periodic habits and short-term stochastic noise. Purely time-domain approaches, including those utilizing Recurrent Neural Networks (RNNs), Transformers, and State Space Models (SSMs), tend to be sensitive to point-wise noise. These models often struggle to decouple stable periodic patterns from random clicks, particularly in long sequences.

While frequency-domain analysis offers a global perspective suitable for denoising and pattern recognition, existing frequency-based approaches (e.g., FMLP-Rec, FEARec, DIFF) typically treat user sequences in isolation. They apply frequency analysis to the raw behavior sequence without considering the context of the target item. The authors argue this "target-agnostic" filtering is suboptimal because a user's behavior sequence is polymorphic; without the specific target item as context, the model cannot effectively distinguish between frequency components representing true signal and those representing background noise.

Methodology: FEDIN Architecture

The proposed Frequency-Enhanced Deep Interest Network (FEDIN) addresses these limitations through a dual-branch architecture that fuses time-domain and frequency-domain perspectives. The framework takes a user behavior sequence $X \in \mathbb{R}^{L \times D}$ and a target item $X_{tar} \in \mathbb{R}^D$ as inputs to predict a user interest representation $U$.

1. Preprocessing

The model employs Reversible Instance Normalization (RevIN) as a preprocessing step to mitigate distribution shifts and the non-stationary nature of user behaviors before feeding data into the branches.

2. Time-Domain Branch (Local Evolution)

This branch focuses on capturing the local sequential evolution of interests:

• Coarse Filtering via Target Attention: A target-aware attention mechanism computes weighted scores to filter out pointwise noise while preserving temporal resolution, producing a weighted sequence $X_{attn}$.
• Fine-grained Evolution via Patching Transformer: To handle long sequences efficiently and capture local temporal semantics, $X_{attn}$ is segmented into non-overlapping patches. A Transformer encoder models dependencies between these patches, yielding the time-domain interest representation $X_{time}$.

3. Frequency-Domain Branch (Global Periodicity)

This branch is the core innovation, designed to isolate structured harmonic patterns using a target-aware spectral filtering mechanism:

• Target-Conditioned Spectrum: Unlike prior works that transform raw sequences, FEDIN first calculates target attention scores $X_{score}$ to capture the relevance of each historical behavior to the target item. These scores are transformed into the frequency domain via Fast Fourier Transform (FFT) to create a target-conditioned spectrum $\hat{X}_{score}$.
• Target-Aware Spectrum Filtering: The spectrum is modulated by the attention scores ($\hat{X}_{attn}$) to highlight frequency components where the user shows strong resonance with the target.
• Learnable Spectral Filter: A Complex-Valued MLP acts as a learnable filter on the complex spectrum. It adaptively amplifies dominant harmonic frequencies while suppressing high-entropy noise, producing a cleaned frequency representation $\hat{X}_{freq}$.
• Adaptive Resonance Scaling: An adaptive gating mechanism scales the frequency branch output based on the spectral clarity (entropy) of the target-conditioned spectrum. This ensures the model relies on the frequency view only when clear resonance patterns are detected, falling back to the time domain otherwise.

4. User Interest Aggregator

The outputs from both branches ($X_{time}$ and $X_{freq}$) are summed to form a mixed representation. A Top-k Target Attention mechanism is then applied to derive the final user interest $U$. This mechanism retains only the $k$ items with the highest similarity to the target, using a masking strategy to ensure differentiability and prevent the dilution of strong signals by weak or conflicting ones.

Key Contributions

1. Empirical Observation: The authors identify that user attention scores exhibit distinct spectral entropy distributions when conditioned on positive versus negative target items. True user interests manifest as highly concentrated spectral patterns with lower entropy, whereas irrelevant behaviors appear as high-entropy noise. This provides a theoretical grounding for target-aware spectral recommendation.
2. FEDIN Framework: The proposal of a hybrid framework that integrates a target-aware frequency branch with traditional time-domain modeling. This design effectively combines the global noise robustness of frequency analysis with the local sequential precision of time-domain modeling.
3. Performance and Robustness: Extensive experiments demonstrate that FEDIN consistently outperforms state-of-the-art sequential baselines and exhibits superior robustness against noise.

Experimental Results

The model was evaluated on three public datasets: Tmall, Taobao, and Alipay.

• Baselines: Comparisons were made against six state-of-the-art models (DIN, DIEN, SASRec, BERT4Rec, GRU4Rec, BST, DIFF) and a Sum Pooling baseline.
• Performance: FEDIN achieved the best performance across all datasets in both GAUC and AUC metrics. For instance, on the Tmall dataset, FEDIN achieved a GAUC of 0.9658 and AUC of 0.9666, significantly outperforming the best baseline (DIN: 0.9547/0.9518).
• Ablation Studies: Removing either the Time-Domain or Frequency-Domain branch resulted in performance degradation, confirming the necessity of both. Removing the "Target-Aware" aspect of the frequency filter (replacing it with a static filter) also caused a drop, validating the hypothesis that target context is crucial for effective spectral filtering.
• Noise Resistance: Synthetic noise experiments (Drop Noise and Replace Noise) showed that the Frequency-Domain Branch maintains significantly higher stability as noise rates increase compared to the Time-Domain Branch. The spectral filtering effectively attenuates high-entropy noise components.

Significance and Claims

The paper claims that FEDIN offers a novel perspective on sequential recommendation by leveraging the target-conditioned nature of spectral patterns. The authors assert that simply analyzing the raw behavior sequence in the frequency domain is insufficient; the target item acts as a necessary condition to activate specific resonance patterns.

By integrating frequency-domain analysis as a complement rather than a replacement for time-domain modeling, FEDIN achieves a balance between capturing global periodicities and preserving local sequential details. The work demonstrates that this dual-branch approach provides superior robustness against the inherent noise and sparsity of real-world user behavior logs, leading to more stable and accurate CTR predictions across diverse datasets. The authors note that the computational cost of the frequency branch is $O(L \log L)$ due to FFT operations, offering an efficiency advantage over standard self-attention mechanisms ($O(L^2)$) for long sequences.