Spurious correlations cause deep learning models to rely on predictive shortcuts that hold in the training data but break under distribution shifts, leading to large performance drops for minority groups. Existing strategies often rely on costly group annotations or employ unstable adversarial training. In this paper, we propose Prototype-guided debiasing using Robust Invariant Feature Transformations (PRIFT), a novel framework that mitigates spurious correlations by manipulating latent space geometry. Specifically, we introduce a prototype-guided modeling approach that leverages natural language prompts to represent confounders, transforming abstract biases into interpretable geometric anchors without auxiliary classifiers. Based on these anchors, we introduce a centered projection operator that adaptively purifies representations by removing confounding deviations specific to instances while preserving essential semantic structure. Furthermore, PRIFT can handle confounding factor information at different levels, ranging from true labels to unsupervised latent inference. Experiments on four text classification benchmarks demonstrate the superiority of our method; notably, PRIFT outperforms state-of-the-art baselines and improves worst-group accuracy by over 20% on the CivilComments dataset compared to standard empirical risk minimization.

Knowledge Tracing (KT) evaluates students’ dynamic knowledge states from historical interaction sequences. Prior work has largely focused on predictive accuracy, leaving model interpretability underexplored and limiting practical educational applications. In real learning processes, a student’s knowledge state is often shaped by a few critical interactions, termed high-influence intervention signals. Identifying and leveraging these signals is crucial for enhancing both model interpretability and predictive accuracy. To address this challenge, we propose a high-influence intervention-based dual-stream framework for knowledge tracing, termed HIDKT . This method actively identifies high-influence intervention signals through a contrastive learning mechanism. Specifically, we design a dynamic importance scoring mechanism to identify high-influence intervention signals in historical interactions. By intervening on these interactions, we construct a counterfactual interaction stream that runs parallel to the factual interaction stream. Both streams are fed into a shared encoder, where a counterfactual-guided intervention contrastive loss guides the model. This loss explicitly encourages the model to amplify differences induced by high-influence intervention signal interventions in the prediction space, while constraining low-influence perturbations in the latent state space to preserve the stability of the student’s underlying knowledge structure. Extensive experiments on five real-world datasets demonstrate that HIDKT significantly outperforms current state-of-the-art baselines. It improves AUC and ACC by an average of 11.78% and 7.69%, respectively. Further analysis confirms that by explicitly guiding the model to focus on high-influence intervention signals, HIDKT transforms KT from a black-box predictor into a more transparent knowledge tracing tool. This provides a reliable basis for precise educational interventions.

Large-scale studies indicate that the distinct approach to opinion fusion employed by extreme agents exerts a more potent influence on overall opinion evolution when compared to regular agents. The presence of extreme agents within the network tends to undermine the development of opinion neutrality, which is harmful to the guidance of online public opinion. Notably, prior research often overlooks the existence of opinion extreme agents in social networks. However, existing researches seldom consider the time sunk cost in the evolution of opinions. Building upon this foundation, we introduce a temporal dimension to the opinion evolution, integrating the time sunk cost with the opinion evolution process. Furthermore, we devise an agent partitioning method that categorizes agents into four states based on their opinion values: watch state, subjective state, firm state, and extreme state, with extreme state agents generally expressing radical opinions. We constructed an agent network based on the phenomenon of time sunk costs and proposed a model for the evolution of extreme opinions in this network. Our study found that the information sharing among extreme agents significantly influences the extremization of opinions in various networks. After restricting the exchange of opinions on extreme agents, the number of extreme agents in the network decreased by 40% to 50% compared to the initial situation. Additionally, we also discovered that imposing restrictions on extreme agents in the early stages can help increase the possibility of network opinions moving towards neutral positions. When restriction of extreme agents(REA) was performed at the beginning of the experiment compared to REA in the midway of the experiment, the final number of extreme state agents decreased by 15.57%. The results show that extreme agents have a great influence on the spread and evolution of extreme opinions on platforms.

Emojis in texts provide lots of additional information in sentiment analysis. Previous implicit sentiment analysis models have primarily treated emojis as unique tokens or deleted them directly, and thus have ignored the explicit sentiment information inside emojis. Considering the different relationships between emoji descriptions and texts, we propose a pre-training Bidirectional Encoder Representations from Transformers (BERT) with emojis (BEMOJI) for Chinese and English sentiment analysis. At the pre-training stage, we pre-train BEMOJI by predicting the emoji descriptions from the corresponding texts via prompt learning. At the fine-tuning stage, we propose a fusion layer to fuse text representations and emoji descriptions into fused representations. These representations are used to predict text sentiment orientations. Experimental results show that BEMOJI gets the highest accuracy (91.41% and 93.36%), Macro-precision (91.30% and 92.85%), Macro-recall (90.66% and 93.65%) and Macro-F1-measure (90.95% and 93.15%) on the Chinese and English datasets. The performance of BEMOJI is 29.92% and 24.60% higher than emoji-based methods on average on Chinese and English datasets, respectively. Meanwhile, the performance of BEMOJI is 3.76% and 5.81% higher than transformer-based methods on average on Chinese and English datasets, respectively. The ablation study verifies that the emoji descriptions and fusion layer play a crucial role in BEMOJI. Besides, the robustness study illustrates that BEMOJI achieves comparable results with BERT on four sentiment analysis tasks without emojis, which means BEMOJI is a very robust model. Finally, the case study shows that BEMOJI can output more reasonable emojis than BERT.

Recommendation systems are an effective solution to deal with information overload, particularly in the e-commerce sector, in which sequential recommendation is extensively utilized. Sequential recommendations aim to acquire users’ interests and provide accurate recommendations by analyzing users’ historical interaction sequences. To improve recommendation performance, it is vital to take into account the long- and short-term interests of users. Despite significant advancements in this domain, some issues need to be addressed. Conventional sequential recommendation models typically express each item with a uniform embedding, ignoring evolutionary patterns among item attributes, such as category, brand, and price. Moreover, these models often model users’ long- and short-term interests independently, failing to adequately address the issues of interest drift and short-term interest evolution. This study proposes a new model, the Feature-aware Long-Short Interest Evolution Network (FLSIE), to address the above-mentioned issues. Specifically, the model uses explicit feature embedding to represent item attribute information and employs a two-dimensional (2D) attention mechanism to distinguish the significance of individual features in a specific item and the relevance of each item in the interaction sequence. Furthermore, to avoid the issue of interest drift, the model employs a long-term interest guidance mechanism to enhance the representation of short-term interest and adopts a gated recurrent unit with attentional update gate to model the dynamic evolution of users’ short-term interest. Experimental results indicate that our presented model outperforms existing methods on three real-world datasets.

Abstract Membrane computing is a type of parallel computing system (generally called P system) abstracted from information exchange mechanisms in biological cells, tissues, or neurons, which can process data in a distributed and interpretable manner. LSTM-SNP, the first model of long short-term memory networks based on parameterized nonlinear Spiking neural P systems, was proposed recently. However, a systematic understanding and leveraging of the LSTM-SNP model to address named entity recognition (NER) and other natural language processing (NLP) tasks are still lacking. The bottleneck of the NER task lies in the scarcity of data and the vague definition of entity edges. Most approaches center on dataset handling, and there have been few attempts to address the issue in Spiking neural P (SNP) systems. This paper proposes a model named CLSTM-SNP based on the LSTM-SNP, aiming to tackle the NER problem in the field of SNP systems for the first time. First, this study employs a CNN layer to obtain character-level characteristics. Second, GloVe word vectors are utilized as word representations. Third, the research employs the LSTM-SNP to analyze textual features. We subsequently studied CLSTM-SNP’s effectiveness in addressing NER problems on CoNLL-2003 and OntoNotes 5.0 datasets and compared it to the results of five other baseline methods. Our model CLSTM-SNP achieved a macro F1-score of 89.2 $$\%$$ % on CoNLL-2003 and 75.5 $$\%$$ % on OntoNotes 5.0, respectively. The performance of CLSTM-SNP and LSTM-SNP indicates a great potential for handling named entity recognition or other sequential tasks in NLP.

Target-oriented opinion word extraction (TOWE) is critical in aspect-based sentiment analysis. It aims at extracting opinion words that are related to aspect terms. Existing TOWE approaches primarily focused on explicit or implicit target aspects. However, few methods dealt with them simultaneously. For compensating this limitation, this study proposes a dual-cell recurrent network (DCRN) that combines aspect term extraction (ATE) and target-oriented opinion word extraction. The DCRN model is trained and evaluated on global fields, including explicit and implicit target aspects. Empirical results demonstrate that the proposed DCRN model outperforms existing methods by an average of 4.90% on the SemEva114–16 datasets. Furthermore, the DCRN model achieves higher Macro-F1 values than the IOG model on the Restaurant 14–16 datasets by 8.97%, 7.90 %, and 8.70 %, respectively. These results indicate that the DCRN model significantly improves the performance of TOWE and exhibits robust generalization capabilities.

Influence Maximization (IM), an NP-hard central issue for social network research, aims to recognize the influential nodes in a network so that the message can spread faster and more effectively. A large number of existing studies mainly focus on the heuristic methods, which generally lead to sub-optimal solutions and suffer time-consuming and inapplicability for large-scale networks. Furthermore, the present community-aware random walk to analyze IM using network representation learning considers only the node’s influence or network community structures. No research has been found that surveyed both of them. Hence, the present study is designed to solve the IM problem by introducing a novel influence network embedding (NINE) approach and a novel influence maximization algorithm, namely NineIM, based on network representation learning. First, a mechanism that can capture the diffusion behavior proximity between network nodes is constructed. Second, we consider a more realistic social behavior assumption. The probability of information dissemination between network nodes (users) is different from other random walk based network representation learning. Third, the node influence is used to define the rules of random walk and then get the embedding representation of a social network. Experiments on four real-world networks indicate that our proposed NINE method outperforms four state-of-the-art network embedding baselines. Finally, the superiority of the proposed NineIM algorithm is reported by comparing four traditional IM algorithms. The code is available at https://github.com/baiyazi/NineIM.

Misinformation has brought great challenges to the government and network media in social networks. To clarify the influences of behaviors of the network media, government, and netizen on misinformation propagation, a large number of influence parameters are proposed for the three participants. Then, a tripartite evolutionary game model for misinformation propagation is constructed. According to the proposed game model, the expected payoffs of three participants are analyzed when they adopt different strategies. The evolutionary stabilities of the game model are also analyzed theoretically. Finally, the impacts of different parameters on expected payoffs of three participants are analyzed experimentally. Meanwhile, coping strategies of three participants under different conditions are given. The experimental results show that the proposed tripartite evolutionary game model can properly describe the influence of network media, government, and netizen on misinformation propagation.