Information extraction is a critical technology for intelligent analysis and risk assessment in the dairy financial domain. However, real-world applications face three major challenges: the complexity and diversity of entity-relation types, significant data imbalance, and ambiguity in textual expressions. Traditional methods often fail to capture rare patterns, struggle with vague mentions, and exhibit poor generalization in low-resource settings. To address these issues, we propose a novel framework that integrates large language models (LLMs) with targeted data augmentation and agent-based retrieval-augmented generation (RAG). Our approach builds on the BaiChuan2 model, which is first adapted to the dairy finance domain via secondary pretraining. We introduce a two-stage data augmentation strategy: the first stage uses ChatGPT to generate pseudo-samples for rare types, and the second stage refines model weaknesses based on prediction-guided feedback. These augmented datasets are used to fine-tune the model through prompt-based supervised learning with LoRA. To further enhance robustness, we incorporate an agent-based RAG module for completing vague or underspecified entities by retrieving external contextual knowledge. Extensive experiments demonstrate that our framework achieves state-of-the-art performance, with the improved metric, i.e., F1+ scores, of 0.876 and 0.824 for entity recognition and relation extraction, respectively. The RAG component boosts entity completion accuracy to 0.802 while reducing retrieval latency by over 6x, showcasing both the effectiveness and practicality of our method in real-world dairy financial applications.

Temporal Knowledge Graph Completion (TKGC) involves predicting and filling in missing facts within time series data, a crucial task with wide-ranging applications across various domains. The dynamic evolution of Temporal Knowledge Graphs (TKGs) adds complexity to this task, making it inherently challenging. Existing research predominantly relies on historical data to complete the missing facts. However, these approaches often overlook the potential of future information and the significance of node weights.To address these challenges, we propose Neo-TKGC, a novel temporal knowledge graph completion model that integrates a graph structure encoding module and a temporal encoding module. The graph structure encoding module introduces node weights to enhance the capabilities of graph neural networks (GNNs) for entity and relation representation learning, implemented using CompGCN. This module can be easily extended to any GNN models utilizing node and edge aggregation. The temporal encoding module leverages both future and historical information to capture relevant contexts and temporal dependencies among entities and relations.By combining node weights and future information, Neo-TKGC achieves more accurate entity and relation representations, thereby improving the model's ability to infer unknown entities. Extensive experiments on three real-world TKGC datasets demonstrate the superior performance of our model compared to existing approaches, achieving at least a 1.7% relative improvement in Hits@1 across most metrics.

Session-based Social Recommendation (SSR) harnesses social relationships within online social networks to improve Session-based Recommendation (SR) performance. However, existing SSR algorithms often face the challenge of "friend data sparsity". Additionally, significant discrepancies may exist between the purchase preferences of social network friends and those of the target user, thereby diminishing the influence of friends relative to the target user’s preferences. To tackle these challenges, this paper introduces the concept of "Like-minded Peers" (LMP), representing users whose preferences are aligned with the target user’s current session in their history sessions. To the best of our knowledge, this is the first work to use LMP as an enhancement for modeling the social influence in SSR. It not only alleviates the problem of friend data sparseness but also effectively incorporates users with similar preferences. Furthermore, we propose a novel model named Transformer Encoder with Graph Attention Aggregator Recommendation. Experimental results on four real-world datasets demonstrates the efficacy and superiority of our proposed model.

Session-based recommendation (SBR) is a challenging task that aims to make item recommendations based on anonymized user session data. Mainstream SBR efforts focus on modeling information within a session and do not use information from other sessions. Although some works try to use other session information, there are still many limitations, and how to model other session information is still a challenging task. To overcome these limitations, we propose a new method for learning similar intentions between sessions, aiming to better model the recommendation information contained in other sessions. Specifically, we contribute a new model named ISIM-GNN that learns and integrates three levels of information simultaneously: (i) In the intra-session representation learning layer, we represent the session as a session graph and model it using a gated graph neural network. (ii) In the global item embedding learning layer, we use the graph attention mechanism to propagate and aggregate relevant item information from other sessions on the global graph. (iii) In the inter-session similar intent learning layer, we employ both “hard similarity” and “soft similarity” to select similar sessions, and use the attention mechanism to conduct session-level aggregation on the selected similar sessions to make better use of the inter-session collaboration information. Experiments on three real-world datasets show a significant performance improvement of our approach compared to state-of-the-art work.

The container has become the mainstream technology in the industry due to its fast delivery speed, low resource consumption, and good portability. Accurate and effective container workload prediction models are crucial for proactive autoscaling. Most current workload prediction models either focused on a single application or just predicted the demand for a specific resource type, such as CPU. They can hardly support the comprehensive prediction of multiple resource demands for hybrid applications in container cloud scenarios. In this paper, we proposed a novel framework called "MAN4Container" to bridge the existing gap. This framework is comprised of two essential components: a Multi-Resource Correlation Weight (MRCW) algorithm and a Multivariate Workload Aware Correlation (MWAC) model. The MRCW algorithm classifies multivariate workloads into different types, and the MWAC model predicts each type of workload by combining long-term and short-term predictions. The experimental results on the Alibaba-Cluster-Trace-V2018 and Google cluster datasets have shown that our framework MAN4Container has higher prediction accuracy than other baseline models.

Spam has plagued Internet users for a long time, and it is of great significance to design an efficient spam detection method.In recent years, spam detection methods based on fine-tuning pre-trained language models (PLM) have achieved great success.The approach is to fine-tune a pre-trained language model on a large dataset to adapt it to the downstream spam detection task.However, the objective of the initial training phase of PLM is inconsistent with the objective of downstream tasks, which results in the downstream tasks cannot fully utilize the latent knowledge in PLM.In this paper, we use Prompt Tuning and PLM to identify Chinese spam by constructing additional prompt templates, converting the email classification task into a fill-in-the-blank task, and then getting the email classification results according to the filling content on the prompt templates.This process is very similar to the process of initial training of PLM, which can more fully utilize the rich knowledge in PLM.We use prompt tuning to train the model on public datasets.Through experiments, we found that the accuracy score of the proposed model on trec06 datasets can reach 0.996, and the F1 score can reach 0.994, which is better than the comparison model.In terms of model convergence speed, the proposed model only needs less than 200 training steps to converge, which is faster than the comparison model.

Dynamic spectrum sharing will facilitate the interference coordination in device-to-device (D2D) communications. In the absence of network level coordination, the timing synchronization among D2D users will be unavailable, leading to inaccurate channel state estimation and device detection, especially in time-varying fading environments. In this paper, we design an asynchronous device detection/discovery framework for cognitive-D2D applications, which acquires timing drifts and dynamical fading channels when directly detecting the existence of a proximity D2D device (e.g. or primary user). To model and analyze this, a new dynamical system model is established, where the unknown timing deviation follows a random process, while the fading channel is governed by a discrete state Markov chain. To cope with the mixed estimation and detection problem, a novel sequential estimation scheme is proposed, using the conceptions of statistic Bayesian inference and random finite set. By tracking the unknown states (i.e. varying time deviations and fading gains) and suppressing the link uncertainty, the proposed scheme can effectively enhance the detection performance. The general framework, as a complimentary to a network-aided case with the coordinated signaling, provides the foundation for development of flexible D2D communications along with proximity-based spectrum sharing.

Due to the potential detection error caused by information loss in sampling process, collaborative scheme is especially important for compressed sensing-based spectrum detection to improve the detection accuracy. In this paper, a novel energy-efficient low-complexity collaborative scheme is proposed for cognitive radio networks. In the proposed scheme, based on the prediction results of signal sparsity level by Lempel-Ziv-based prediction algorithm, the number of detection devices for spectrum detection is evaluated with the aim of minimizing the objective function, which takes into account both the detection accuracy and energy consumption. Finally, extensive simulation results are presented to show the effectiveness of our proposed collaborative scheme by comparing with the existing ones.

Spectrum sensing is one of the key issues for spectrum sharing in cognitive radio networks to deal with the more and more serious problem on exhaustive spectrum resource. In this paper, we study the problem of wideband spectrum sensing in cognitive radio networks with multiple primary networks. By dividing the total wideband spectrum into many groups with relatively small number of narrowbands, a novel wideband spectrum sensing scheme is proposed. In the proposed scheme, the number and corresponding probability of free bands in each group are predicted by Lempel-Ziv based prediction algorithm. Then the optimal number of secondary users that will sense the narrowbands in each group is found to maximize the defined reward. Not only the sensing speed and accuracy but also the system reward are improved in this paper. Finally, extensive simulation results are provided to show the effectiveness of our proposed wideband spectrum sensing scheme by comparing with the existing ones.

Dynamic spectrum access is one of the most important premises of spectrum reuse based on cognitive radio technologies, which are considered to be the best way to alleviate the controversy on spectrum scarcity and low efficiency. However, most of previous work focuses on the increasing of system throughput, ignoring the QoS requirement of secondary users. In this paper, dynamic spectrum access with QoS provisioning is studied. We adopt discrete-time Markov chain to analyze and model the spectrum usage in time-slotted cognitive radio networks. Furthermore, three admission control schemes are proposed to minimize the forced termination probability of secondary users. Simulation results show that our proposed schemes can significantly improve the forced termination probability of secondary users, though slightly increase the blocking probability.