Multi-agent LLM systems accumulate long, non-monotonic memories where claims are revised, retracted, and later resurfaced, making final aggregation a validity-adjudication problem rather than simple summarization. However, existing agent systems provide little support for final-stage validity adjudication, leaving the final aggregator to conflate conflicting versions and resurface outdated claims. We present EvoMem, a lightweight plug-in for post-retrieval evidence structuring that converts retrieved interaction traces into an evolutionary timeline: it pools backend evidence, selects salient timepoints under a fixed budget, and summarizes each timepoint to trace updates and suppress outdated answers. We also introduce EvoMemBench, a timeline-style benchmark with controlled revisions, retractions, and cross-topic noise. Across long-context agents, RAG base agents, and agentic memory frameworks, EvoMem consistently improves answer accuracy and reduces outdated errors with modest overhead.

In recommender systems, reranking is an important post-processing technique to reorder the items in a recommendation list. Recently, some LLM-based reranking approaches have been proposed to enjoy the semantic reasoning capability of a large language model. However, they are sensitive to the order of the input list and often incur large computational overheads. To address their limitations, we propose the APCR, an Anchor-based Pairwise Comparison for recommendation Reranking in this paper. It first leverages an LLM to conduct pairwise comparisons between those recommended items and an anchor and computes a kind of preference scores for producing a new LLM suggested list. We next propose a position-aware list reranking technique to reorder the items in the recommendation list by considering their positions in the LLM suggested list to output the final list. Experiments on real-world datasets show that our APCR outperforms the state-of-the-art LLM-based reranking techniques in terms of better list ranking performance.

Introducing a Knowledge Graph (KG) to facilitate a recommender system has become a tendency in recent years. Many existing methods leverage KGs to obtain side information of items to promote item representation learning for enhancing recommendation performance. However, they ignore that KGs also may contribute to better user representation learning. To solve this issue, we propose a novel algorithm, the KIGR ( K nowledge-aware I nteraction G raph for R ecommendation), to mine user–item interactions via KGs for assisting user representation learning. Specifically, a user–item interaction is encoded by attentively summing up the relation embedding about the item in the KG. Then, an unsupervised learning method is used to group the user–item interactions into different latent types. Further, a user–item interaction graph is divided into several subgraphs, which is referred to as a Knowledge-aware Interaction Graph, making each subgraph only contain one latent type of interaction. Finally, user representation is the fusion of user interest embedding, which is learned on the knowledge-aware interaction graph, whereas item representation is learned on the KG. Experimental results on MovieLens, LastFM and Amazon-Book validate that the proposed KIGR has a superior performance compared with the state-of-the-art algorithms.

With the development of recommendation algorithms, researchers are paying increasing attention to fairness issues such as user discrimination in recommendations. To address these issues, existing works often filter users’ sensitive information that may cause discrimination during the process of learning user representations. However, these approaches overlook the latent relationship between items’ content attributes and users’ sensitive information. In this article, we propose DALFRec, a fairness-aware recommendation algorithm based on user-side and item-side adversarial learning to mitigate the effects of sensitive information on both sides of the recommendation process. First, we conduct a statistical analysis to demonstrate the latent relationship between items’ information and users’ sensitive attributes. Then, we design a dual-side adversarial learning network that simultaneously filters out users’ sensitive information on the user and item side. Additionally, we propose a new evaluation strategy that leverages the latent relationship between items’ content attributes and users’ sensitive attributes to better assess the algorithm’s ability to reduce discrimination. Our experiments on three real datasets demonstrate the superiority of our proposed algorithm over state-of-the-art methods.

Learning representations from pairwise comparisons has achieved significant success in various fields, including computer vision and information retrieval. In recommendation systems, collaborative filtering algorithms based on pairwise learning are also rooted in this approach. However, a major challenge in collaborative filtering is the lack of labels for negative instances in implicit feedback data, leading to the inclusion of false negatives among randomly selected instances. This issue causes biased optimization objectives and results in biased parameter estimation. In this paper, we propose a novel method to address learning biases arising from implicit feedback data and introduce a modified loss function for pairwise learning, called debiased pairwise loss (DPL). The core idea of DPL is to correct the biased probability estimates caused by false negatives, thereby adjusting the gradients to more closely approximate those of fully supervised data. Implementing DPL requires only a small modification to the existing codebase. Experimental studies on public datasets demonstrate the effectiveness of the proposed method.

How to sample high quality negative instances from unlabeled data, i.e., negative sampling, is important for training implicit collaborative filtering and contrastive learning models. Although previous studies have proposed some approaches to sample informative instances, discriminating false negative from true negative for unbiased negative sampling remains an unsolved problem. On the basis of our order relation analysis of negatives’ scores, we first derive the class conditional density of true negatives and that of false negatives. We next design a Bayesian classifier for negative classification, from which we define a model-agnostic posterior probability estimate of an instance being true negative as a quantitative negative signal measure. We also propose a Bayesian optimal sampling rule to sample high-quality negatives. The proposed Bayesian Negative Sampling (BNS) algorithm has a linear time complexity. Experimental studies validate the superiority of BNS over the peers in terms of better sampling quality and better recommendation performance.1

Discrimination against different user groups has received growing attention in the recommendation field. To address this problem, existing works typically remove sensitive attributes that may cause discrimination through adversary learning to achieve fair recommendations. However, these approaches leverage all available interactions for learning user representations and overlook the fact that different interactions have varying relevance to users’ sensitive attributes. Ignoring this issue may weaken the effectiveness of adversary learning in removing sensitive attributes. To tackle this challenge, we propose a novel model called GS-FairRec, which distinguishes between user interactions to achieve better removal of sensitive attributes. The model consists of three modules: graph sampling-based representation learning, pseudo-user representation learning, and adversarial learning. Firstly, the graph sampling-based representation learning module removes some irrelevant neighbors from a user-item bipartite graph and employs a graph convolutional network (GCN) to learn user/item representations. Next, items that are relevant to a user’s sensitive information but do not match their preferences are defined as the user’s pseudo-interest items, which are leveraged to learn the pseudo-user representation. In the adversarial learning module, the user’s two kinds of representations are fused for adversarial learning to remove sensitive information. Additionally, we design a new metric to measure the model’s ability to remove sensitive attributes based on how a generated recommendation list discloses the user’s sensitive attributes. Finally, we conduct experiments on two real-world datasets, and our results demonstrate the superiority of our proposed model in fairness tasks.

Some recent news recommendation (NR) methods introduce a Pre-trained Language Model (PLM) to encode news representation by following the vanilla pre-train and fine-tune paradigm with carefully-designed recommendation-specific neural networks and objective functions. Due to the inconsistent task objective with that of PLM, we argue that their modeling paradigm has not well exploited the abundant semantic information and linguistic knowledge embedded in the pre-training process. Recently, the pre-train, prompt, and predict paradigm, called prompt learning, has achieved many successes in natural language processing domain. In this paper, we make the first trial of this new paradigm to develop a Prompt Learning for News Recommendation (Prompt4NR) framework, which transforms the task of predicting whether a user would click a candidate news as a cloze-style mask-prediction task. Specifically, we design a series of prompt templates, including discrete, continuous, and hybrid templates, and construct their corresponding answer spaces to examine the proposed Prompt4NR framework. Furthermore, we use the prompt ensembling to integrate predictions from multiple prompt templates. Extensive experiments on the MIND dataset validate the effectiveness of our Prompt4NR with a set of new benchmark results.

Incorporating knowledge graph as side information has become a new trend in recommendation systems. Recent studies regard items as entities of a knowledge graph and leverage graph neural networks to assist item encoding, yet by considering each relation type independently. However, relation types are often too many and sometimes one relation type involves too few entities. We argue that there may exist some latent relevance among relations in KG. It may not necessary nor effective to consider all relation types for item encoding. In this paper, we propose a VRKG4Rec model (Virtual Relational Knowledge Graphs for Recommendation), which clusters relations with latent relevance to generates virtual relations. Specifically, we first construct virtual relational graphs (VRKGs) by an unsupervised learning scheme. We also design a local weighted smoothing (LWS) mechanism for node encoding on VRKGs, which iteratively updates a node embedding only depending on the node itself and its neighbors, but involve no additional training parameters. LWS mechanism is also employed on a user-item bipartite graph for user representation learning, which utilizes item encodings with virtual relational knowledge to help train user representations. Experiment results on two public datasets validate that our VRKG4Rec model outperforms the state-of-the-art methods. The implementations are available at https://github.com/lulu0913/VRKG4Rec.

Percolation theory shows that removing a small fraction of critical nodes can lead to the disintegration of a large network into many disconnected tiny subnetworks. The network dismantling task focuses on how to efficiently select the least such critical nodes. Most existing approaches focus on measuring nodes’ importance from either functional or topological viewpoint. Different from theirs, we argue that nodes’ importance can be measured from both of the two complementary aspects: The functional importance can be based on the nodes’ competence in relaying network information; While the topological importance can be measured from nodes’ regional structural patterns. In this paper, we propose an unsupervised learning framework for network dismantling, called DCRS, which encodes and fuses both node diffusion competence and role significance. Specifically, we propose a graph diffusion neural network which emulates information diffusion for competence encoding; We divide nodes with similar egonet structural patterns into a few roles, and construct a role graph on which to encode node role significance. The DCRS converts and fuses the two encodings to output a final ranking score for selecting critical nodes. Experiments on both real-world networks and synthetic networks demonstrate that our scheme significantly outperforms the state-of-the-art competitors for its mostly requiring much fewer nodes to dismantle a network.

Can we employ one neural model to efficiently dismantle many complex yet unique networks? This article provides an affirmative answer. Diverse real-world systems can be abstracted as complex networks each consisting of many functional nodes and edges. Percolation theory has indicated that removing only a few vital nodes can cause the collapse of whole network. However, finding the least number of such vital nodes is a rather challenging task for large networks due to its NP-hardness. Previous studies have proposed many centrality measures and heuristic algorithms to tackle this network dismantling (ND) problem. Different from theirs, this article tries to approach the ND task by designing a neural model which can be trained from tiny synthetic networks but will be applied for various real-world networks. It seems a discouraging mission at first sight, as network sizes and topologies are quite different across distinct real-world networks. Nonetheless, this article initiates insightful efforts of designing and training a neural influence ranking model (NIRM). Experiments on fifteen real-world networks validate its effectiveness for its mostly requiring fewer vital nodes to dismantle a network, compared with the state-of-the-art competitors. The key to its success lies in that our NIRM can efficiently encode both local structural and global topological signals for ranking nodes, in addition to our innovative labelling method in training dataset construction.

Session-based recommendation (SBR) is to predict the next item for an anonymous item sequence. Although many neural models have proven effectiveness in the SBR task, how to learn better items’ embeddings for neural models still remains a key challenge due to the anonymity of sessions and sparsity of users’ behaviors. This paper proposes a graph-based neural model, called Graph N eighborhood Routing and Random Walk (GNRRW), which learns two kinds of item embeddings for the SBR task. We first construct an item graph based on items’ co-occurrences in all sessions, on which we learn a local embedding and a global embedding for each item. For local embedding learning, we propose a novel neighborhood routing (NR) algorithm to exploit the compositive relations between an item and its neighbors. The NR algorithm has an excellent feature in that no additional parameters are needed in the training process. For global embedding learning, we propose a random walk-based approach to explore a kind of global relations between an item and representative items. Furthermore, we propose a switch-based shared gated recurrent unit (GRU) network to alternatively learn session local representation to make a local prediction, and learn session global representation to make a global prediction. Finally, we design a decision fusion mechanism to adaptively fuse both local and global predictions to output final items’ preference scores. Experiments on the public Yoochoose and Diginetica dataset validate the superiority of our GNRRW model over the state-of-the-art neural models.

Social emotion classification is to predict the distribution of different emotions evoked by an article among its readers. Prior studies have shown that document semantic and topical features can help improve classification performance. However, how to effectively extract and jointly exploit such features have not been well researched. In this paper, we propose an end-to-end topic-enhanced self-attention network (TESAN) that jointly encodes document semantics and extracts document topics. In particular, TESAN first constructs a neural topic model to learn topical information and generates a topic embedding for a document. We then propose a topic-enhanced self-attention mechanism to encode semantic and topical information into a document vector. Finally, a fusion gate is used to compose the document representation for emotion classification by integrating the document vector and the topic embedding. The entire TESAN is trained in an end-to-end manner. Experimental results on three public datasets reveal that TESAN outperforms the state-of-the-art schemes in terms of higher classification accuracy and higher average Pearson correlation coefficient. Furthermore, the TESAN is computation efficient and can generate more coherent topics.

Social emotion classification is to estimate the distribution of readers' emotion evoked by an article. In this paper, we design a new neural network model by encoding sentence syntactic dependency and document topical information into the document representation. We first use a dependency embedded recursive neural network to learn syntactic features for each sentence, and then use a gated recurrent unit to transform the sentences' vectors into a document vector. We also use a multi-layer perceptron to encode the topical information of a document into a topic vector. Finally, a gate layer is used to compose the document representation from the gated summation of the document vector and the topic vector. Experiment results on two public datasets indicate that our proposed model outperforms the state-of-the-art methods in terms of better average Pearson correlation coefficient and MicroF1 performance.

How to decompose a large global matrix into many small local matrices has been recently researched a lot for matrix approximation. However, the distance computation in matrix decomposition is a challenging issue, as no prior knowledge about the most appropriate feature vectors and distance measures are available. In this paper, we propose a novel scheme for local matrix construction without involving distance computation. The basic idea is based on the application of convergence probabilities of graph random walk. At first, a user-item bipartite graph is constructed from the global matrix. After performing random walk on the bipartite graph, we select some user-item pairs as anchors. Then another random walk with restart is applied to construct the local matrix for each anchor. Finally, the global matrix approximation is obtained by averaging the prediction results of local matrices. Our experiments on the four real-world datasets show that the proposed solution outperforms the state-of-the-art schemes in terms of lower prediction errors and higher coverage ratios.

Event recommendation has become an important issue in event-based social networks (EBSN). In this paper, we study how to exploit diverse relations in an EBSN as well as individual history preferences to recommend preferred events. We first construct a hybrid graph consisting of different types of nodes to represent available entities in an EBSN. The graph uses explicit relations as edges to connect nodes of different types; while transferring implicit relations of event attributes to interconnect the event nodes. After executing the graph random walking, we obtain the candidate events with high convergency probabilities. We next extract a user preference from his attended events to further compute his interest similarities to his candidate events. The recommended event list is then obtained by combining the two similarity scores. Data sets from a real EBSN are used to examine the proposed scheme, and experiment results validate its superiority over peer schemes.