In music recommendation systems, multimodal interest learning is pivotal, which allows the model to capture nuanced preferences, including textual elements such as lyrics and various musical attributes such as different instruments and melodies. Recently, methods that incorporate multimodal content features through semantic IDs have achieved promising results. However, existing methods suffer from two critical limitations: 1) intra-modal semantic degradation, where residual-based quantization processes gradually decouple discrete IDs from original content semantics, leading to semantic drift; and 2) inter-modal modeling gaps, where traditional fusion strategies either overlook modal-specific details or fail to capture cross-modal correlations, hindering comprehensive user interest modeling. To address these challenges, we propose a novel multimodal recommendation framework with two stages. In the first stage, our Progressive Semantic Residual Quantization (PSRQ) method generates modal-specific and modal-joint semantic IDs by explicitly preserving the prefix semantic feature. In the second stage, to model multimodal interest of users, a Multi-Codebook Cross-Attention (MCCA) network is designed to enable the model to simultaneously capture modal-specific interests and perceive cross-modal correlations. Extensive experiments on multiple real-world datasets demonstrate that our framework outperforms state-of-the-art baselines. This framework has been deployed on one of China's largest music streaming platforms, and online A/B tests confirm significant improvements in commercial metrics, underscoring its practical value for industrial-scale recommendation systems.

The escalating challenges in urban traffic highlight the importance of accurate traffic forecasting for enhancing the Intelligent Transportation Systems (ITS), given its pivotal role in optimizing traffic management strategies. Due to the impressive performance of Graph Neural Networks in spatio-temporal data analysis, Spatio-Temporal Graph Neural Networks (STGNNs) have emerged as a mainstream approach for traffic forecasting. However, many recent STGNN-based works often utilize redundant and intricate architectures, leading to increased computational costs with only minor improvements in performance. To solve this issue, we propose a concise but effective Lightweight Spatio-Temporal Neural Network architecture, LSTNN. Specifically, we design two sampling strategies, downsampling and patched sampling, to decompose the time series and separately fit the seasonal component and trend components. Based on sampling strategies, we introduce a spatio-temporal localization module based on the raw traffic data and learnable spatio-temporal indexing. In particular, this module models spatio-temporal correlations by appending the spatio-temporal indexing to the traffic sequences. We visualized the spatio-temporal indexing, confirming the effectiveness of the spatio-temporal localization module design. Extensive experiments are conducted on four widely utilized real-world traffic datasets, showing the significant improvement of LSTNN compared to baselines, including several state-of-the-art models. Furthermore, we conducted extensive experiments on four long-term time series datasets to validate the effectiveness of LSTNN in the Long-term Time Series Forecasting (LTSF) task. Our findings may inspire and encourage the design of more concise neural prediction architectures for real-world applications. We release the datasets and baseline implementations at: github.com/HduDBSI/LSTNN