Knowledge graph-based recommendation systems utilize structured entity and relation representations to better model user preferences. However, many traditional approaches rely primarily on ID-based data and often overlook textual information associated with items and relations, leading to limited semantic understanding. While recent approaches have begun incorporating large language models (LLMs), most focus solely on enhancing relational embeddings and fail to fully exploit the semantic extraction capabilities of LLMs. To address these limitations, we propose LLMKnowRec, a novel LLM-enhanced, knowledge-aware recommendation framework designed to improve the semantic modeling of knowledge graphs. Our approach integrates the powerful language understanding abilities of LLMs with traditional ID-based recommendation by introducing an LLM-based embedding generator. This generator produces semantically rich embeddings using textual descriptions of user profiles and knowledge graph relations. Building on this, we further introduce a semantic user intent modeling module, which leverages LLMs to aggregate multiple intent signals into comprehensive, semantically enriched intent embed-dings. Additionally, we develop a relational intent-aware aggregation scheme that effectively combines higher-order representations, capturing both relational structures and user intent, thus enhancing the overall semantic understanding of users and items. Experimental conducted on three public benchmark datasets demonstrate that LLMKnowRec consistently outperforms stateof-the-art methods. Specifically, our model achieves improvements of up to 12.92%, 19.27%, and 8.23% in NDCG@10, and up to 13.41%, 15.62%, and 23.55% in Recall@10 across the three datasets, respectively. These results demonstrate the effectiveness and practical potential of our proposed method. The implementation code is publicly available at: https://github.com/ liujianfang2021/LLMKnowRec.

Large Reasoning Models (LRMs) achieve explicit chain-of-thought expansion by imitating deep thinking behaviors of humans, demonstrating excellent performance in complex task scenarios. However, the deep-thinking mode often leads to unnecessarily lengthy reasoning and resource inefficiency when handling simple tasks. This overthinking phenomenon may arise from the generation preference triggered by the reward function during post-training. Existing research attempts to mitigate overthinking from the perspective of prompt design or model training, but generally underestimates the importance of task difficulty awareness, which makes it difficult for LRMs to effectively allocate reasoning resources. In this paper, we propose Difficulty-aware Policy Optimization (DiPO), a reinforcement learning-based LRM training framework. DiPO encourages LRM to spontaneously model task complexity, and integrates them into reinforcement learning framework to adjust the generation preferences introduced by post-training. A difficulty modeling method based on model self-reasoning is proposed, which significantly reduces the dependence on manual annotation and formalize task complexity. We further develop a difficulty-signal-enhanced reward function that incorporates a penalty for lengthy reasoning while considering reasoning performance and output format. Experimental results indicate that DiPO enables the model to spontaneously adjust inference overhead, significantly reducing redundant tokens without losing performance due to thought compression.

Deep learning-based knowledge tracing (DLKT) models enhance personalized learning with strong predictions. However, their black-box nature hinders educational use, raising concerns about transparency and trust. To address this issue, this paper proposes a dual-strategy multiobjective evolutionary algorithm (DSMOEA) for counterfactual explanation, aimed at providing effective post hoc interpretations of DLKT predictions. Leveraging causal reasoning, DSMOEA avoids spurious correlations and generates actionable counterfactuals by constructing a tutoring-oriented framework that yields optimal, concise practice sequences to enhance future learning gains. We formulate the counterfactual generation task as a multi-objective optimization problem and solve it with a MOEA. Two strategies improve efficiency and quality: domain-informed mutation (DIM) guides solution generation, and secondary weighted selection (SWS) ranks Pareto candidates by educational relevance. Experiments demonstrate superior effectiveness and sparsity over baselines, delivering reliable counterfactuals and actionable guidance for learning path optimization.

Symbolic Regression aims to automatically identify compact and interpretable mathematical expressions that model the functional relationship between input and output variables. Most existing search-based symbolic regression methods typically rely on the fitting error to inform the search process. However, in the vast expression space, numerous candidate expressions may exhibit similar error values while differing substantially in structure, leading to ambiguous search directions and hindering convergence to the underlying true function. To address this challenge, we propose a novel framework named EGRL-SR (Experience-driven Goal-conditioned Reinforcement Learning for Symbolic Regression). In contrast to traditional error-driven approaches, EGRL-SR introduces a new perspective: leveraging precise historical trajectories and optimizing the action-value network to proactively guide the search process, thereby achieving a more robust expression search. Specifically, we formulate symbolic regression as a goal-conditioned reinforcement learning problem and incorporate hindsight experience replay, allowing the action-value network to generalize common mapping patterns from diverse input-output pairs. Moreover, we design an all-point satisfaction binary reward function that encourages the action-value network to focus on structural patterns rather than low-error expressions, and concurrently propose a structure-guided heuristic exploration strategy to enhance search diversity and space coverage. Experiments on public benchmarks show that EGRL-SR consistently outperforms state-of-the-art methods in recovery rate and robustness, and can recover more complex expressions under the same search budget. Ablation results validate that the action-value network effectively guides the search, with both the reward function and the exploration strategy playing critical roles.

Situational Judgment Tests (SJTs) are a well-established class of assessment instruments that embed measurement within realistic scenarios, eliciting context-driven responses that are less susceptible to social desirability bias and more effective at capturing applied judgment than traditional self-report scales. Despite these advantages, the widespread adoption of SJTs in educational assessment remains constrained by the high cost and labor intensity of manual item development. To address this scalability bottleneck, we propose AES (Agents for Educational SJTs), a multi-agent framework powered by large language models (LLMs) that automatically transforms existing Likert-scale items into psychometrically informed, context-rich SJTs. AES comprises three cooperative modules, Generator, Simulator, and Calibrator, that emulate the expert driven development pipeline by combining qualitative review from simulated domain experts with quantitative analysis of synthetic student responses, enabling iterative item refinement without human intervention. We instantiate AES in the domain of AI literacy, a complex and affectively rich competency that typifies the constructs for which SJTs are most beneficial. Evaluation on both human participant and LLM-simulated response datasets shows that AES-generated SJT items provide improved psychometric evidence relative to their Likert counterparts, while learners report greater engagement and perceived authenticity. These findings provide initial evidence that AES can serve as a scalable and psychometrically informed approach for supporting the development of scenario-based AI literacy assessments.

Recent research in explainable recommendation commonly uses natural language explanations to improve transparency and user trust. However, reliably evaluating whether explanations are semantically faithful to users’ multi-dimensional preferences remains challenging. Existing evaluation methods capture different aspects of explanation quality but remain limited in modeling fine-grained aspect-level sentiment faithfulness. To address these limitations, we propose Feature Aspect-Level Sentiment Consistency (FASC), a framework that quantifies semantic consistency between generated explanations and user-authored reference explanations through aspect coverage and sentiment polarity. FASC uses Large Language Models (LLMs) to extract structured aspect-sentiment units from explanation texts, enabling reproducible metrics for aspect coverage, correctness, and sentiment alignment. To validate the framework, we re-annotated several widely used explainable recommendation datasets to construct benchmarks with fine-grained, aspect-level sentiment labels. We also conducted human studies using pairwise comparisons, showing that FASC captures a distinct dimension of human alignment not covered by existing metrics. Experimental results indicate that FASC can distinguish subtle differences in semantic faithfulness across models. Furthermore, FASC integrates complementarily with existing metrics, yielding substantial gains in human alignment.

The incorporation of knowledge graphs as auxiliary information can effectively enrich the feature representation of users and items in recommendation systems. Although Graph Neural Networks (GNNs) have developed into the dominant paradigm for graph-structured learning, generating node-level and graph-level embeddings through iterative message-passing mechanisms, current knowledge-aware GNN methods still exhibit two limitations: (1) Spatial adaptation deficiency arises when forcing integration of user-item bipartite graphs and knowledge graphs within Euclidean space, failing to reconcile their geometric heterogeneity. (2) Supervision-sparsity noise occurs as sparse user behavior data causes noise accumulation during multi-hop propagation, degrading representation quality. Addressing these challenges, this paper proposes the framework HANCL with Hierarchical Attention Networks with Multi-level Contrastive Learning for Knowledge-aware Recommendation. This approach employs a disentangled dual-space paradigm, representing user-item collaborative signals in Euclidean space while modeling knowledge graph hierarchies in hyperbolic space. It fuses dual-space features through a hierarchical attention network to extract highly discriminative vectors for refined user behavior modeling. Additionally, a multi-level contrastive optimization framework constructs cross-view contrastive loss and introduces structural supervision via negative sampling, suppressing irrelevant information interference. Comparative experiments on three public benchmark datasets demonstrate that HANCL outperforms existing state-of-the-art baselines in the vast majority of experimental scenarios.Extensive experiments on three public benchmark datasets demonstrate that HANCL significantly outperforms state-of-the-art baselines. Specifically, the framework achieves remarkable Recall@10 improvements of 6.22%, 13.35%, and 23.17% on the Last.FM, Book-Crossing, and MovieLens-1M datasets, respectively. Implementation code and datasets will be open-sourced at https://github.com/weiwang1992/HANCL.

The retention of newly learned second language (L2) words is influenced by pre-existing word representations through interactions driven by similarity. When learners encounter new L2 words, similar previously learned words influence retention through cross-lexical activation. This study integrates established memory models with advanced machine learning techniques, providing a comprehensive approach. It analyzes a large-scale Duolingo dataset and quantifies the impact of previously learned words on current word memory. By improving the Half-Life Regression model with interlexical features, we constructed an L2 interlexical influence matrix and examined how similarity dimensions affect vocabulary retention through complementary analytical approaches. The results indicate that semantic similarity is the primary facilitator with significant nonlinear impacts, whereas orthographic similarity causes interference when considered in isolation. Notably, high semantic similarity mitigates orthographic interference, enabling successful vocabulary retention even when visual form similarities would otherwise impede learning. Two-dimensional and three-dimensional interactions significantly enhanced model fit, while four-dimensional interactions showed no significant improvement, suggesting an optimal complexity threshold for similarity-based memory facilitation. These findings reveal the complex mechanisms through which multiple similarity dimensions work synergistically in L2 vocabulary learning and offer practical implications for optimizing vocabulary learning sequences in language education.

Memory behavior modeling seeks to predict individual recall performance and understand its underlying cognitive mechanisms. However, the dynamic and heterogeneous nature of memory data poses significant challenges to the generalization ability of models under unseen conditions. To address this challenge, we propose an invariant representation learning framework I-Mem that integrates self-supervised contrastive learning with decorrelation constraints, enabling the adaptive identification and suppression of environment-related factors in sequential behavioral data, thereby mitigating the influence of spurious features and enhancing the modeling of stable cognitive structures. Importantly, the method does not rely on explicit environment partitioning or predefined environment labels, while our theoretical analysis demonstrates that it can effectively resist environmental perturbations and facilitate the extraction of invariant structural representations, thereby ensuring adaptability and generalization. Empirical evaluations on both synthetic and real-world datasets further confirm its superiority over mainstream methods in terms of generalization performance and stable feature identification. Feature attribution analysis reveals that I-Mem extracts invariant features aligned with classical cognitive effects, and reflects short-term behavioral patterns that may indicate latent cognitive mechanisms beyond existing theories, highlighting both interpretability and discovery potential.

Symbolic Regression aims to discover interpretable mathematical expressions from data. Equation Learner (EQL) is a gradient-based method with strong fitting capability and expressive potential, yet it often activates redundant operators as model complexity grows, leading to over-complex expressions and unstable equation recovery. We analyze a gradient residual issue induced by operators that do not vanish at zero, which can prevent the ideal sparse expression from being a local optimum and bias training toward unnecessarily complex structures, making exact recovery nearly unattainable in practice. To address this, we propose EQL-Z, a structurally controllable symbolic regression framework. EQL-Z enforces zero-point constraints via zero-point consistent operator transformations to eliminate residual gradients on silent paths, and performs an incremental small-to-large structure search that grows depth/width from a compact seed under a complexity-penalized validation score. After selecting a compact structure, we optionally apply BFGS fine-tuning to refine coefficients. Experiments on synthetic and real-world datasets show that EQL-Z substantially improves exact equation recovery and in-/out-of-distribution generalization over vanilla EQL, achieving performance close to the best existing symbolic regression baselines. The code is available at https://anonymous.4open.science/r/EQL-Z-BE6C/.

Memory behavior modeling is a fundamental issue in the fields of cognitive psychology and education. Classical theoretical models of memory are characterized by insufficient accuracy and ongoing controversies, while data-driven memory modeling methods often require large amount of training data and lack interpretability, highlighting the need for new approaches to memory behavior modeling. This paper integrates classic psychological theories of memory to explore the feasibility of knowledge-driven neural networks in memory behavior modeling. It proposes the EPsyINN model, which combines temporal neural networks with sparse differential regression in a unified framework, enabling the joint optimization of neural networks and classical symbolic models. More specifically, to address the controversies in classical psychological theories and the ambiguity of descriptors, it proposes a descriptor evolution method based on differential operators to achieve precise descriptor characterization and advance the evolution of classical symbolic models. Additionally, it introduces a caching mechanism for regression coefficient matrices and an alternating iterative optimization method for multiple modules, effectively alleviating local optima in model optimization. On five large-scale real-world memory behavior datasets, the proposed method surpasses state-of-the-art memory modeling approaches in predictive accuracy, while the evolved classical symbolic models also achieve performance improvements. Ablation experiments validate the effectiveness of the proposed improvements, and application experiments demonstrate its potential to inspire psychological research. The code for the experiments is available at: https://github.com/hellowads/PsyINN.

Cross-domain recommendation (CDR) offers an efficient and effective solution to mitigate data sparsity in recommender systems. Existing research primarily focuses on exploring knowledge transfer based on overlapping entities or auxiliary contents between domains. However, there is little research on the real non-overlapping cross-domain recommendation (NCDR) problems, even though it poses a more general and applicable prospect. The core challenge of NCDR lies in the difficulty of finding the correct and useful knowledge transfer bridge between domains without relying on the explicit overlapping identities. Utilizing the inherent similarity and fuzzy characteristics of users and items in the latent feature space, this paper investigates a Fuzzy Prototype Transfer (FPT) learning method for the NCDR problem. FPT jointly optimizes prototypes and individual features for both users and items in target domain under the guidance of source features. An end-to-end learnable fuzzy clustering module based on maximum entropy regularization is proposed to learn both user and item fuzzy clustering assignments and fuzzy fusion prototypes. Lastly, by constructing an asymmetric dual-prototype fuzzy transfer module, similar user and item features across domains are found and aligned effectively. Extensive experiments demonstrate FPT's superior performance over the state-of-the-art methods while maintaining lower inference and memory costs than those of the baselines.

Knowledge tracing is one of the key technologies for achieving personalized learning, aiming to predict students’ future performance by analyzing their historical responses during the learning process. In recent years, the development of deep learning has led to the emergence of deep knowledge tracing, which employs neural network models to better capture the complex dynamic patterns in students’ learning processes. However, most existing deep learning-based knowledge tracing models primarily learn a student’s cognitive state through sequential modeling of single-scale exercise-answer interactions, neglecting the multi-scale characteristics of both short-term and long-term learning processes. This approach often struggles to fit the complex cognitive states of students. Moreover, these methods typically model students’ mastery of exercises but overlook their problem-solving abilities, a critical factor in determining whether students can correctly answer the exercises to be predicted.To address the shortcomings of existing knowledge tracing models, we propose a Dual-view Multi-scale Cognitive Representation-based Knowledge Tracing model (DMKT). From the exercise mastery view, the model designs a multi-scale representation method that simultaneously considers students’ long-term static cognitive states and short-term dynamic cognitive states to better fit their complex cognitive conditions. From the problem solving view, the model adopts a Transformer architecture to explicitly learn joint representations of exercises answered by students at each time step and the most similar historical exercises, integrating response information to model their performance on both new and old exercises. Furthermore, to effectively learn the similarity relationships between exercises, we extend multiple exercise-centered heterogeneous relations and propose a dual-channel heterogeneous graph-based online feature distillation method to fully model exercise representations. Extensive experimental results demonstrate that our model exhibits significant superiority, while also providing a certain level of interpretability for the model’s predictions from both the exercise mastery and problem solving view.

Knowledge Tracing (KT) is the task of predicting students' future performance based on their past interactions with educational resources. A key aspect of KT is representation learning, which aims to capture meaningful features from students' learning behaviors to improve prediction performance. Recently, contrastive learning methods have shown great promise in representation learning. As a result, KT models based on contrastive learning have been introduced to enhance representation learning for KT. However, these models have posed several challenges. Firstly, most of these models adopt the contrastive learning approach used in other fields, which involves data augmentation followed by contrastive learning, yet effectively applying data augmentation in KT remains an open challenge. Secondly, these models typically apply contrastive learning to only one of the fundamental components of KT: questions, interactions, or knowledge states, thereby limiting their overall performance. To address these issues, this article proposes a Multi-level Contrastive learning model for Knowledge Tracing (MCKT). MCKT (The code can be found at https://github.com/lilstrawberry/MCKT.) does not rely on data augmentation strategies; instead, it deeply integrates domain knowledge and performs contrastive learning at three levels: questions, interactions, and knowledge states. Experimental results on four publicly available datasets, compared against a total of 20 state-of-the-art KT models, demonstrate that MCKT consistently outperforms other models. Subsequent experiments further validate the effectiveness of the multi-level contrastive learning approach.

Deep-learning-based knowledge tracing (DLKT) models have achieved high predictive accuracy, but their opaque "black box" nature limits practical value: educators cannot trace why predictions are made and learners cannot obtain transparent feedback. Existing explainability techniques, which aim to explain why a model makes a particular prediction and provide the evidence for it, predominantly rely on correlational analyses, often yielding unfaithful or suboptimal explanations. To address this, we propose a post hoc reinforcement-learning-based causal deep knowledge tracing explainer (RCKTE). RCKTE operates through a structured workflow: First, it formulates the explanation task as a globally optimal subsequence screening problem, aiming to identify the most causally influential historical interactions from a student's learning sequence. Then, a reinforcement learning agent, guided by a causal attribution reward and a dual-optimizer scheme, iteratively constructs the optimal subsequences by assessing the causal impact of each interaction. This process results in more faithful and concise explainable subsequences than those produced by correlation-based methods, achieving this within about 1 s to support real-time use. Finally, these explainable subsequences directly support actionable educational applications, including identifying a learner's weak knowledge for targeted review, constructing personalized knowledge structure graphs for intervention tracking, and deriving group-level knowledge structures to guide curriculum design. Extensive experiments across multiple DLKT models and datasets confirm that RCKTE consistently outperforms existing post hoc methods in both the faithfulness and readability of explanations. By integrating causal attribution with reinforcement learning, RCKTE provides accurate, efficient, and educationally meaningful explanations that enhance the usability of DLKT in real learning environments.

Knowledge Tracing (KT) aims to predict students' future performance on answering questions based on their historical exercise sequences. To alleviate the problem of data sparsity in KT, recent works have introduced auxiliary information to mine question similarity, resulting in the enhancement of question embeddings. Nonetheless, there remains a gap in developing an approach that effectively incorporates various forms of auxiliary information, including relational information (e.g., question-student, question-skill relation), relationship attributes (e.g., correctness indicating a student's performance on a question), and node attributes (e.g., student ability). To tackle this challenge, the Similarity-enhanced Question Embedding (SimQE) method for KT is proposed, with its central feature being the utilization of weighted and attributed meta-paths for extracting question similarity. To capture multi-dimensional question similarity semantics by integrating multiple relations, various meta-paths are constructed for learning question embeddings separately. These embeddings, each encoding different similarity semantics, are then fused to serve the task of KT. To capture finer-grained similarity by leveraging the relationship attributes and node attributes on the meta-paths, the biased random walk algorithm is designed. In addition, the auxiliary node generation method is proposed to capture high-order question similarity. Finally, extensive experiments conducted on six datasets demonstrate that SimQE performs the best among 10 representative question embedding methods. Furthermore, SimQE proves to be more effective in alleviating the problem of data sparsity.

Math problems are an important knowledge carrier and evaluation means in personalized teaching. Their high cost of manual compilation promotes the research of math problem generation. Many previous studies have focused on the generation of math word problems, which are difficult to meet the real teaching needs due to the single task-objective orientation and small differences in generation results. By fusing external knowledge through retrieval-augmented generation (RAG), large language model (LLM) can generate a variety of math problems, but the generated results still have limitations such as poor knowledge consistency, uncontrollability, and high computational cost. In this paper, we propose the task of multi-objective math problem generation (MMPG). This task introduces the triple objectives of generation including "question type, knowledge point and difficulty"in respond to teaching needs in real scene. To the best of our knowledge, this is the first study considering multiple objectives on the process of math problem generation. Based on this, we further design an adaptive multi-level retrieval augmentation framework (AMRAF) for LLM to generate multi-objective math problems. This plug-and-play framework can effectively improve the generation performance without parameter tuning of the target model due to the fine-grained information retrieval and fusion. To verify the effectiveness of the proposed framework and provide a benchmark for subsequent research, we construct an MMPG dataset containing 9,000 samples. Experimental results demonstrate the superiority and effectiveness of our framework.

Recommender systems based on knowledge graphs enhance the explainability of recommendations by incorporating external knowledge. Nevertheless, the accuracy of recommendations heavily depends on dense interaction data and high-quality knowledge graphs, both of which commonly suffer from data sparsity. Introducing graph contrastive learning to enhance representation quality can effectively improve recommendation performance. Existing graph contrastive learning methods that use graph augmentation can alleviate the data sparsity problem. However, they often neglect the semantic modeling of relation embeddings and lack sufficient contrastive information, leading to insufficient utilization of the embedding space for relations and nodes. To address this, we propose a semantic relation-aware graph attention network with a noise augmented layer-wise contrastive learning model for recommendation, named SRGAN. Specifically, we design a semantic relation-aware graph attention network that updates the semantics of relations during multi-layer iterations to better capture user preferences. Additionally, we construct a noise-augmented layer- wise contrastive learning model, employing simple yet effective noise perturbations to generate contrastive views for entities and relations. By maximizing the consistency of the representations in each layer, the model achieves alignment with the lower-level features of the intermediate layers. Extensive experiments on three public benchmark datasets demonstrate that our proposed method significantly outperforms current approaches. To ensure reproducibility, we make the code and data from our experiments publicly available on https://github.com/liujianfang2021/SRGAN.

The rapid advancement of multimedia technologies and their increasing integration in education have underscored the importance of multimedia learning. Knowledge Tracing (KT) plays a crucial role in enabling adaptive multimedia learning by continuously monitoring students' progress and forecasting their performance throughout the learning process. Question lies at the heart of the KT process, making its representation crucial for building efficient KT models. However, the sparsity and complexity of question data pose significant challenges for existing methods to capture the underlying features of questions, thereby affecting the accuracy of knowledge state predictions. To address this issue, this paper attempts to introduce the diffusion model to the KT field, proposing a novel knowledge tracing model, DiffuQKT. The model presents a diffusion-based generative approach for question representation and enhances the stability of knowledge states through contrastive learning. Specifically, DiffuQKT first constructs question representations based on their concepts, difficulty, and variations, and then, during the forward phase, progressively adds noise to the question representations, disrupting them into a Gaussian distribution. In the reverse phase, DiffuQKT gradually recovers the representations from noise, generating higher-quality question representations for knowledge tracing. Furthermore, to guide more meaningful question generation, we incorporate question concepts and difficulty as conditions during the denoising process. In addition, to improve the robustness of knowledge states against subtle variations in question representations, we employ contrastive learning to stabilize knowledge states across both original and denoised question representations. We conduct extensive experiments on four public datasets, comparing DiffuQKT with 15 baseline methods. The results demonstrate that DiffuQKT significantly outperforms existing models. Moreover, we find that the diffusion-based generative approach for question representation proposed in this paper has the ability to significantly improve the performance of baseline models. The code can be found at https://github.com/lilstrawberry/DiffuQKT.

Modern recommender systems place great inclination towards facilitating user experience, as more applications enabling users to critique and then refine recommendations immediately. Considering the real-time requirements, critique-able recommender systems typically straight modify the model parameters and update the recommend list through analyzing the user critiquing keyphrases in the inference phase. Current critiquing methods require first constructing a specially designated model which establish direct correlations between users and keyphrases during the training phase allowing for innovative recommendations upon the critiquing,restricting the applicable scenarios. Additionally, all these approaches ignore the catastrophic forgetting problem, where the cumulative changes in parameters during continuous multi-step critiquing may lead to a collapse in model performance. Thus, We conceptualize a proxy bridging users and keyphrases, proposing a streamlined yet potent Items Proxy Generic Critiquing Framework (IPGC) framework, which can serve as a universal plugin for most knowledge graph recommender models based on collaborative filtering (CF) strategies. IPGC provides a new paradigm for frictionless integration of critique mechanisms to enable iterative recommendation refinement in mainstream recommendation scenarios. IPGC describes the items proxy mechanism for transforming the critiquing optimization objective of user-keyphrase pairs into user-item pairs, adapting it for general CF recommender models without the necessity of specifically designed user-keyphrase correlation module. Furthermore, an anti-forgetting regularizer is introduced in order to efficiently mitigate the catastrophic forgetting problem of the model as a prior for critiquing optimization.