Time-Frequency Fusion and Feature Cross-Fusion Sequence Recommendation Algorithm

为了有效融合项目ID嵌入和文本嵌入，提出一种时频融合和特征交叉融合的序列推荐算法(Time-frequency fusion and feature cross-fusion sequential recommendation algorithm, TFFCRec)。使用对项目文本进行编码，多样化混合专家调制方法获得的是易于区分的文本表示，将项目ID嵌入和文本嵌入通过结合快速傅里叶变换(FFT)和短时傅里叶变换(STFT)，提取用户的全局频域特征和局部时频特征。这样的方法使得算法既能考虑用户的长期兴趣偏好，又能捕捉用户的短期兴趣变化。此外，我们引入了特征交叉融合，并通过优化的Mamba-like的线性注意力(OMLLA)来捕获特征之间更深层次的非线性关系，提取更深层次的特征表示。我们设计了一个融合网络，自适应地学习不同嵌入表示的权重，将FFT、STFT和OMLLA得到的特征向量进行加权融合，通过SASRec来进行序列推荐。在Instant video、Beauty、Digital Music、Tools and Home improvement数据集上进行实验，本文方法较基准方法在Recall@10指标上分别提升了6.3%、13.2%、3.7%、6.5%。To effectively integrate project ID embeddings and text embeddings, we propose a sequence recommendation algorithm called Time-frequency Fusion and Feature Cross-fusion Sequential Recommendation Algorithm (TFFCRec). RoBERTa is used to encode the project text, and a diversity mixture expert modulation method is applied to obtain distinguishable text representations. Project ID embeddings and text embeddings are combined using Fast Fourier Transform (FFT) and Short-Time Fourier Transform (STFT), extracting the user’s global frequency-domain features and local time-frequency features. This approach enables the algorithm to capture both the user’s long-term interest preferences and short-term interest variations. In addition, we introduce feature cross-fusion and use the optimized Mamba-like Linear Attention (OMLLA) to capture deeper non-linear relationships between features and extract more profound feature representations. We design a fusion network that adaptively learns the weights of different embedding representations and performs weighted fusion of the feature vectors obtained from FFT, STFT, and OMLLA. These fused features are then passed into SASRec for sequence recommendation. Experiments are carried out on the Instant Video, Beauty, Digital Music, and Tools and Home Improvement datasets. Compared with the benchmark methods, the proposed method in this paper has improved the Recall@10 metric by 6.3%, 13.2%, 3.7%, and 6.5% respectively.