Recent advances in pre-trained language models (PLMs), such as GPT, BERT, and their larger variants, have significantly enhanced the ability to generate synthetic datasets for zero-shot learning, where models can generate data without the need for human annotations. However, existing methods often focus on generating domain-specific datasets, which restricts their generalization capabilities when applied to new, unseen domains. Furthermore, many of these methods rely on a single, static dataset that is either too broad or too narrowly tailored to a specific domain, making them less adaptable to diverse downstream domains. In this paper, we propose a dual prompt data generation framework, called Adaptive Domain-Specific Classification via Zero-shot Dataset Generation (AdaGen), which decouples universal representation learning from domain-specific adaptation. The framework consists of two components: domainagnostic training set generation via a universal prompt, followed by domain-specific validation set generation for weight learning. By combining these two components, our method achieves an optimal trade-off between cross-domain robustness and domainspecific accuracy, allowing a small task-aware model to effectively adapt across diverse domains while preserving generalization. This novel method overcomes the limitations of traditional methods by introducing dynamic feedback, enabling efficient zeroshot learning with improved adaptability. Extensive experiments in seven real-world datasets demonstrate the effectiveness of AdaGen, which outperforms the compared baseline models.

With the explosive growth of information, recommendation systems have emerged to alleviate the problem of information overload. In order to improve the performance of recommendation systems, many existing methods introduce Large Language Models to extract textual information from description text. However, Large Language Models are trained on large-scale generic textual data and may face a semantic gap for downstream recommendation tasks. To address the above issues, we propose Contrastive Learning for Adapting Language Model to Sequential Recommendation (CLA-Rec). In CLA-Rec, we first extract text embeddings from description text using Large Language Models and align the text embeddings learned by Large Language Models with the collaborative information through contrastive learning to obtain high-quality item representations. Through semantic alignment, we bridge the semantic gap between Large Language Models and the recommendation task. To map textual information and collaborative information into user representations, we utilize a Transformer model to learn user representations and capture user preferences by combining the semantically aligned item representations. Extensive experiments on three public datasets demonstrate that our method outperforms state-of-the-art approaches on multiple evaluation metrics, illustrating the effectiveness of the CLA-Rec model in adapting Large Language Models to recommendation tasks.

Next-basket recommendation (NBR) infers a set of items that a user will interact with in the next basket. Existing methods often struggle with the data sparsity problem, particularly when the number of baskets is significantly large due to diverse user behaviors. Cross-domain recommendation (CDR) can effectively alleviate this problem in NBR by transferring knowledge across different domains. Nevertheless, these methods often rely on the similarities of overlapping users, which leads to the negative transfer problem and ignores the overlapping items that are general in real-world scenarios like chain stores. In this paper, we provide a clear symbolic definition of cross-store recommendation (CSR) and distinguish it from CDR. We also propose a novel CSNBR model for cross-store next-basket recommendation task. To fully model the transferable collaborative information between two stores, we learn the embeddings of users, baskets, and items by two intra-store bipartite graphs, and use an inter-store unified bipartite graph to transfer the previously learned knowledge. Furthermore, to alleviate the negative transfer problem, we propose to reconstruct the inter-store unified bipartite graph by utilizing user embeddings obtained from the transfer layer and the disentanglement layer. We also employ two sequence encoders to model the historical sequential information at basket-level and item-level. Extensive experiments conducted on real-world datasets demonstrate the effectiveness of the CSNBR model.

With the rapid development of modern communication technology, telecom fraud has been increasing year by year. If fraudsters can be accurately identified before they carry out their scams, it can not only protect people from potential losses but also increase trust in telecom operators. Therefore, in recent years, telecom fraud detection has garnered widespread attention in both academia and industry. Although existing methods for telecom fraud detection have achieved good performance, there are still many unresolved issues for real-world telecom operators. First, existing methods only focus on a single telecom scenario, while real-world telecom scenarios are diverse. Utilizing the characteristics of these different telecom scenarios can improve the effectiveness of telecom fraud detection. Second, existing methods usually use Graph Neural Networks (GNNs) to aggregate neighbor information. However, real-world telecom operators can’t obtain information of users from other operators, resulting in the lacking destination node attributes, which degenerates the performance of GNNs. To address the above issues, in this paper, we propose a new model for Telecom Fraud Detection Based on Feature binning and Autoencoder (TFD-FA). In TFD-FA, a feature binning framework is designed to partition users into different telecom scenarios in order to reflect their unique characteristics. An autoencoder component is also designed to aggregate neighbor information. Furthermore, an imbalance classifier component is constructed to solve the problem of the significantly lower number of fraudsters compared to normal users. Extensive experiments in a real-world dataset demonstrate the effectiveness of TFD-FA, which outperforms the compared baseline models.