Accurately modeling the continuous evolution of user preferences remains a fundamental challenge in sequential recommendation. While existing models have made significant strides in capturing sequential dependencies, they predominantly encode historical behaviors via discrete aggregation, often failing to explicitly model the continuous evolution trends and intrinsic momentum underlying preference shifts. To address this limitation, we propose Kinematic Evolution for Sequential Recommendation (KERec), which models user preference evolution from a continuous kinematic perspective. KERec adopts a dual-branch encoding architecture: the global context branch captures stable long-term preference structures, while the kinematic evolution branch explicitly models the first-order velocity and second-order acceleration of preference evolution based on Taylor series expansion and multi-scale difference operators. To effectively integrate these heterogeneous signals, we introduce an adaptive feature fusion mechanism that balances steady-state consistency with dynamic sensitivity. Furthermore, we design a momentum-guided contrastive learning paradigm that constrains augmentation views via evolutionary momentum, suppressing random noise while reinforcing the perception of genuine evolutionary trajectories. Extensive experiments on multiple public benchmark datasets demonstrate that KERec significantly outperforms state-of-the-art methods in recommendation accuracy and exhibits superior robustness under varying sequence lengths and noisy levels. The code of our method is available at https://github.com/yyt-1219/KERec.

Abstract Knowledge graph-based recommendation systems have strong capabilities in deep association mining and structured reasoning, which effectively alleviate data sparsity and cold-start problems in recommendation. However, traditional knowledge graph-based recommendation considers each graph relation in isolation, failing to capture the potential semantic correlations between different relation types, which leads to incomplete semantic representations. In addition, existing methods generally ignore negative feedback signals in users’ historical interactions, resulting in biased preference modeling. To overcome the above challenges, we introduce S ign-aware R ecommendation based on V irtual S emantic K nowledge G raph(SRVSKG), which enhances recommendation performance by combining virtual semantic collaborative representations with sign-aware learning techniques. Firstly, we propose a virtual semantic subgraph collaborative representation module to learn user and item embeddings. In this module, we build virtual semantic subgraphs through relation clustering based on latent semantic similarity measurement. Then a hierarchical feature extraction mechanism based on graph attention network is applied to virtual semantic subgraphs, which captures semantic associations across relations and enriches item embedding. At the same time, we emphasize user preference for attributes to enrich user embedding. Secondly, we design a sign-aware learning module, which constructs a user-item signed graph, applies Laplacian matrix factorization to simultaneously model the topological features of positive and negative feedback, and introduces the transformer architecture to dynamically fuse signed information, effectively utilizing negative feedback information to eliminate bias in user preference modeling. Finally, we establish a multi-feature fusion mechanism that deeply combines features from the virtual semantic subgraph collaborative representation module and the sign-aware learning module to enable recommendation. Experiment results on three public datasets demonstrate that SRVSKG outperforms state-of-the-art recommendation baselines.

Code summarization generation aims to automatically produce summaries from source code based on its analysis, providing information such as the design goals of code segments and relevant parameters. Recently, fine-tuning generalized code models, which are based on pre-trained large-scale code datasets, has garnered considerable attention. Several studies have proposed data augmentation methods to mitigate the limitation of dataset sizes on the effectiveness of fine-tuned models. However, these approaches continue to face challenges, such as the generation of low-quality augmented data and the limited ability of models to capture code representations. To address these issues, we propose CATS, a fine-tuning method for code summarization generation that incorporates keyword augmentation and contrastive learning. CATS employs a two-stage training strategy. In the first stage, a data augmentation method that retains keyword information is designed to construct similar code pairs, followed by a contrastive learning method to effectively capture code representations from these pairs. In the second stage, a generic autoregressive task for code summarization is used to further refine the high-quality code representations obtained in the first stage. The effectiveness of CATS is evaluated using three different pre-trained models: CodeBERT, GraphCodeBERT, and UniXcoder. Experimental results on two generalized datasets demonstrate that CATS, when applied to UniXcoder, significantly improves performance across all three evaluation metrics compared to the baseline methods. Moreover, a series of ablation experiments validate the contributions of CATS in terms of data augmentation, contrastive learning, and the two-stage training strategy.

Conversational recommender systems model user dynamic preferences and recommend items based on multi-turn interactions. Though the conversational recommender system has achieved good performance, it has two limitations. On the one hand, researchers usually random select an anchor item from user's historical interactions to simulate the interaction with the real user, but some items in the historical interactions do not fit the user realistic preferences (item noise). On the other hand, it pays too much attention to user dynamic preferences, but nurses some static preferences that are difficult to change over a short period. In fact, when there is no explicit attribute preference in user's conversation, the user static preferences can also be used to make recommendations. To address the aforementioned issues, a novel method that combines graph path reasoning with multi-turn conversation is proposed, called Graph Path reasoning for conversational Recommendation (GPR). In GPR, a soft-clustering is designed to classify items and then set operations are utilized to filter the noise in the user's historical interactions. To capture user dynamic preferences and take account of the user inherent static preferences, GPR asks questions about attributes in the attribute-level reasoning and asks whether the items fit user static preferences in the item-level reasoning on a heterogeneous graph. In the multi-turn of two-level graph path reasoning, a reinforcement learning is used to obtain the optimal path and accurately recommend items to users. Extensive experiments conducted on two benchmark datasets verify that GPR can significantly improve recommendation performance and reduce the turn of path reasoning.