The track visual inspection system is a critical component in maintaining railway transportation safety. The scarcity of abnormal rail images presents a significant challenge for evaluating the performance of such inspection systems across diverse local railway divisions. Image harmonization emerges as a pivotal technique for generating evaluation resources for track vision inspection systems. However, the lack of suitable datasets has resulted in limited reporting on rail image harmonization techniques. This paper introduces the first Rail Image Harmonization Dataset (RHD). The dataset comprises 218 high-resolution rail images containing abnormal fasteners, captured by two inspection vehicles, and provides 14,712 pairs of inharmonious and harmonized rail image samples. Extensive experiments utilizing RHD were conducted, evaluating existing high-resolution harmonization methods alongside a specialized method termed Rail-DCCF—a simplification of the state-of-the-art DCCF method. Comprehensive analyses of the RHD and the harmonization techniques employed in these methods are presented. RHD is designed as a fair and comprehensive dataset to enable unified evaluation of track vision inspection systems across different railway divisions, providing fundamental data support for the research of rail image harmonization.

Large language models (LLM) have emerged as a pivotal technique for enhancing the performance of modern recommender systems. Despite enjoying many advantages, LLM-based recommender systems (LRS) exhibit more severe popularity bias than conventional recommender systems (CRS). Compounding this issue, current debiasing methods for LRS are limited to single-level and single-stage interventions, resulting in suboptimal effectiveness. In this article, we propose a novel LRS debiasing method that performs debiasing at both the token and item levels across the training and inference stages, so-called Hierarchical Dual Debiasing in LRS (HD \({}^{2}\) LRS) . Specifically, during the training stage, a token-wise inverse propensity score technique is embedded into the supervised fine-tuning phase, encouraging the LLM to focus more on unpopular tokens. Subsequently, a popularity-aware direct preference optimization method is developed to curb the LLM’s tendency to favor popular items. During the inference stage, we first train a highly biased CRS by amplifying the weights assigned to popular items. The prediction outputs of the biased CRS are then leveraged to constrain the LLM, preventing it from generating popular tokens and recommending popular items. Extensive experiments conducted on four real-world datasets confirm that our HD \({}^{2}\) LRS effectively alleviates popularity bias in LRS and yields significant improvements in recommendation accuracy.

Causal Inference (CI) plays a critical role in building unbiased recommender systems. However, most existing CI-based debiasing methods primarily pay attention to either popularity bias or conformity bias. In this paper, we propose a Disentangled Counterfactual Reasoning framework to alleviate dual biases in recommendation, called DCR. To be specific, in the training stage, we incorporate the impact of item popularity and user conformity to fit the biased recommendation process, and separate their indirect effects by disentangling user and item embeddings into biased and unbiased components. In the inference stage, we perform counterfactual reasoning to intervene in the item scoring process, which simultaneously mitigates the direct and indirect effects of bias factors. Extensive experiments on four widely used datasets demonstrate that DCR significantly outperforms existing debiasing methods with an average improvement of 9.2

Top-one recommendation with anonymous user behaviors, also known as session-based recommendation (SBR), faces challenges of top-one ranking and short anonymous sequences. To this end, we propose a novel objective that combines (1) a reciprocal rank loss to directly optimize the benchmark metric of top-one recommendation, with (2) a listwise contrastive loss to handle short sequences through listwise augmented consistency regularization. Empirical studies demonstrate that optimizing the proposed objective significantly improves the performance of existing SBR baselines.

Large language models (LLMs) have been widely applied in recommender systems, achieving remarkable success. However, LLM-based recommendation (LR) suffers from more severe popularity bias than conventional recommendation (CR), stemming from both training and inference stages. In this paper, we propose a novel debiasing method for LR, which performs debiasing in such two stages, so termed as Dual Debiasing in LR (D2LR). Concretely, in the training stage, we conduct token-wise inverse propensity score weighting to force the LLM to pay more attention on unpopular tokens. In the inference stage, we train a more biased CR model by increasing the weights of popular items, which adjusts the generation probability of corresponding tokens according to its scores for items, hoping to suppress the excessive generation of popular tokens. Experiments conducted on three real-world datasets validate the effectiveness of our D2LR in mitigating popularity bias in LR.

Recommender systems play an essential role in solving information overload but face openness challenges and are vulnerable to malicious attacks that can lead to data tampering, affecting recommendation results and jeopardizing the security of recommendation services. Existing robust recommendations that detect and block attacks directly from data may result in normal behavior being flagged as an attack being blocked. Statistical-based methods for manually extracting features from rating data are time-consuming and insufficient for sophisticated attacks. Embedded learning-based methods lack characterization of similar information and multi-order interactions. The imbalance between attack size and normal behavior also poses a challenge. To this end, we propose a robust recommendation-oriented malicious attack detection method. The attack detection component of the method utilizes graph learning to model the higher-order connectivity of user-item interaction graphs and Convolutional Neural Network (CNN) assisted feature extraction, and it is combined with an ensemble approach to deal with the imbalance problem to detect attacks. The recommendation component automatically determines the contribution of a user's behavior based on the probability that a click is recognized as normal. The experimental results show that the method performs well in malicious attack detection and is more robust to co-visitation injection attacks in recommendations.

Knowledge Graph Embedding (KGE) methods have achieved great success in predicting missing links in knowledge graphs, a task also known as Knowledge Graph Completion (KGC). Under this task, the Reciprocal Rank (RR) of ground-truth items serve as a key indicator for evaluating the method’s performance. However, most existing studies have overlooked the inconsistency between the ranking metric, RR, and the optimization objective functions, resulting in sub-optimal KGC performance. To address this issue, we propose a KGC framework called KGCRR by designing a novel upper bound function named CRR. By introducing the parameter-pressure ρ to shift the sigmoid function, CRR achieves a better approximation to RR compared with existing objective functions. We theoretically proved that by adjusting ρ, CRR can achieve a more effective approximation to RR. By narrowing the discrepancy with RR and alleviating the gradient vanishing issue associated with the direct optimization of RR loss, CRR demonstrates an advantage in optimizing RR. CRR serves as a plug-and-play objective, capable of seamless integration into various KGE methods. Through extensive experiments conducted on FB15k-237 and WN18RR datasets, we have obtained promising results, with an average improvement of 19.06% in MRR, indicating that CRR significantly enhances the performance of existing methods.

Sequence augmentation based contrastive learning (SACL) plays a critical role in user behavior modeling towards sequential recommendation tasks. However, SACL cannot work well in the scenario of session-based recommendation (SBR), where the anonymous user behavior sequences (known as sessions) are very short (e.g., with no more than 5 interactions), making most augmentation techniques ineffective. In this paper, we propose a novel method named LCAA (Listwise C ontrastive learning with A ssociation-rule-based sequence A ugmentation), which lengthens the current session with association rules to create an augmented session, and then leverages a corresponding listwise contrastive loss to maximize the agreement of two recommendation lists generated from the original session and its augmentation. Remarkably, LCAA is a model-agnostic method that can be easily plugged into a wide range of existing SBR models towards better accuracy. To evaluate the effectiveness of LCAA, we implement it with five SBR models utilizing various deep learning techniques (NARM, STAMP, SRGNN, CORE, and ATTMIX) and then compare the performance of each SBR baseline with its LCAAmodified version. Extensive experiments on three datasets (Diginetica, Nowplaying, and Tmall) demonstrate that LCAA yields the average improvement of around 5% on the complete testing sets and around 3% on the short session testing sets in terms of HR and MRR metrics. The code is publicly available.1

Causal Inference (CI) plays a crucial role in building unbiased recommender systems. However, most current CI-based debiasing methods only pay attention on either popularity bias or conformity bias. This paper presents a Disentangled Counterfactual Reasoning framework to alleviate dual biases in recommendation, so called DCR. Concretely, we consider the impact of both item popularity and user conformity during training, and separate their indirect effects by disentangling user and item embeddings into biased and unbiased components. In the inference stage, we perform counterfactual reasoning to simultaneously mitigate the indirect and direct effects of bias factors. Experimental results demonstrate the effectiveness of our DCR.

Deep Neural Network (DNN) stands out as one widely adopted and effective technique for Click-Through Rate (CTR) prediction in live recommender systems. However, the prevalent DNN-based CTR methods exhibit two main drawbacks. On one hand, they fail to align their optimization objectives with the benchmark metric, such as the Area Under the ROC Curve (AUC), designed for ranking tasks. On the other hand, current DNNbased CTR solutions indiscriminately treat all positive-negative item pairs, ignoring the fact that each item pair differently contributes to AUC optimization. To this end, we propose Rank Gap Sensitive Deep AUC maximization method for accurate CTR prediction, namely RgsAUC. Specifically, we target AUC as the learning objective by relaxing the Heaviside function via sigmoid function to render it differentiable and thus can be optimized directly using gradient-descent methods, which is the de facto choice for solving DNN-based CTR tasks. Furthermore, we incorporate a rank gap sensitive weight in estimating gradients for items, aiming to assign greater significance to item pairs with substantial rank gaps during the learning process. In particular, we reduce the computational complexity from quadratic to linear through reformulation, enabling efficient deployment. Consequently, these designs sharply minimize the number of erroneously-ranked item pairs, which is beneficial to AUC optimization. Notably, RgsAUC is model-agnostic and we implement it in five classic DNN models for the CTR prediction task. Extensive experiments on six real-world datasets clearly demonstrate the effectiveness of our proposed method.

With the explosive growth of users and items, Recommender Systems are facing unprecedented challenges in terms of retrieval efficiency and storage overhead. Learning to Hash techniques have emerged as a promising solution to these issues by encoding high-dimensional data into compact hash codes. As a result, hashing-based recommendation methods (HashRec) have garnered growing attention for enabling large-scale and efficient recommendation services. This survey provides a comprehensive overview of state-of-the-art HashRec algorithms. Specifically, we begin by introducing the common two-tower architecture used in the recall stage and by detailing two predominant hash search strategies. Then, we categorize existing works into a three-tier taxonomy based on: (i) learning objectives, (ii) optimization strategies, and (iii) recommendation scenarios. Additionally, we summarize widely adopted evaluation metrics for assessing both the effectiveness and efficiency of HashRec algorithms. Finally, we discuss current limitations in the field and outline promising directions for future research. We index these HashRec methods at the repository \href{https://github.com/Luo-Fangyuan/HashRec}{https://github.com/Luo-Fangyuan/HashRec}.

To perform online inference efficiently, hashing techniques, devoted to encoding model parameters as binary codes, play a key role in reducing the computational cost of content-aware recommendation (CAR), particularly on devices with limited computation resource. However, current hashing methods for CAR fail to align their learning objectives (e.g., squared loss) with the ranking-based metrics (e.g., Normalized Discounted Cumulative Gain (NDCG)), resulting in suboptimal recommendation accuracy. In this article, we propose a novel ranking-based CAR hashing method based on Factorization Machine (FM), called Discrete Listwise FM (DLFM), for fast and accurate recommendation. Concretely, our DLFM is to optimize NDCG in the Hamming space for preserving the listwise user-item relationships. We devise an efficient algorithm to resolve the challenging DLFM problem, which can directly learn binary parameters in a relaxed continuous solution space, without additional quantization. Particularly, our theoretical analysis shows that the optimal solution to the relaxed continuous optimization problem is approximately the same as that of the original discrete optimization problem. Through extensive experiments on two real-world datasets, we show that DLFM consistently outperforms state-of-the-art hashing-based recommendation techniques.

Hashing-based recommendation (HR) methods, whose core idea is mapping users and items into hamming space, are common practice to improve item retrieval efficiency. However, existing HR fails to align optimization objective (i.e., Bayesian Personalized Ranking) and evaluation metric (i.e., Recall), leading to suboptimal performance. In this paper, we propose a smooth recall loss (termed as SRLoss), which targets Recall as the optimization objective. Due to the existence of discrete constraints, the optimization problem is NP-hard. To this end, we propose an approximation-adjustable gradient estimator to solve our problem. Experimental Results demonstrate the effectiveness of our proposed method.

Social media platforms are inundated with an extensive volume of unverified information, most of which originates from heterogeneous data from a variety of diverse sources, spreading rapidly and widely, thereby posing a significant threat to both individuals and society. An existing challenge in multimodal fake news detection is its limitation to acquiring textual and visual data exclusively from a single source, which leads to a high level of subjectivity in news reporting, incomplete data coverage, and difficulties in adapting to the various forms and sources of fake news. In this paper, we propose a fake news detection model (MHDF) for multi-source heterogeneous data progressive fusion. Our approach begins with gathering, filtering, and cleaning data from multiple sources to create a reliable multi-source multimodal dataset, which involved obtaining reports from diverse perspectives on each event. Subsequently, progressive fusion is achieved by combining features from diverse sources. This is achieved by inputting the features obtained from the textual feature extractor and visual feature extractor into the news textual and visual feature fusion module. We also integrated sentiment features from the text into the model, allowing for multi-level feature extraction. Experimental results and analysis indicate that our approach outperforms other methods.

Review-aware Rating Regression (RaRR) suffers the severe challenge of extreme data sparsity as the multi-modality interactions of ratings accompanied by reviews are costly to obtain. Although some studies of semi-supervised rating regression are proposed to mitigate the impact of sparse data, they bear the risk of learning from noisy pseudo-labeled data. In this article, we propose a simple yet effective paradigm, called co-training-teaching ( CoT 2 ), for integrating the merits of both co-training and co-teaching toward robust semi-supervised RaRR. CoT 2 employs two predictors trained with different feature sets of textual reviews, each of which functions as both “labeler” and “validator.” Specifically, one predictor (labeler) first labels unlabeled data for its peer predictor (validator); after that, the validator samples reliable instances from the noisy pseudo-labeled data it received and sends them back to the labeler for updating. By exchanging and validating pseudo-labeled instances, the two predictors are reinforced by each other in an iterative learning process. The final prediction is made by averaging the outputs of both the refined predictors. Extensive experiments show that our CoT 2 considerably outperforms the state-of-the-art recommendation techniques in the RaRR task, especially when the training data is severely insufficient.

Objectives Herbal prescription recommendation (HPR) is a hot topic and challenging issue in field of clinical decision support of traditional Chinese medicine (TCM). However, almost all previous HPR methods have not adhered to the clinical principles of syndrome differentiation and treatment planning of TCM, which has resulted in suboptimal performance and difficulties in application to real-world clinical scenarios.Materials and Methods We emphasize the synergy among diagnosis and treatment procedure in real-world TCM clinical settings to propose the PresRecST model, which effectively combines the key components of symptom collection, syndrome differentiation, treatment method determination, and herb recommendation. This model integrates a self-curated TCM knowledge graph to learn the high-quality representations of TCM biomedical entities and performs 3 stages of clinical predictions to meet the principle of systematic sequential procedure of TCM decision making.Results To address the limitations of previous datasets, we constructed the TCM-Lung dataset, which is suitable for the simultaneous training of the syndrome differentiation, treatment method determination, and herb recommendation. Overall experimental results on 2 datasets demonstrate that the proposed PresRecST outperforms the state-of-the-art algorithm by significant improvements (eg, improvements of P@5 by 4.70%, P@10 by 5.37%, P@20 by 3.08% compared with the best baseline).Discussion The workflow of PresRecST effectively integrates the embedding vectors of the knowledge graph for progressive recommendation tasks, and it closely aligns with the actual diagnostic and treatment procedures followed by TCM doctors. A series of ablation experiments and case study show the availability and interpretability of PresRecST, indicating the proposed PresRecST can be beneficial for assisting the diagnosis and treatment in real-world TCM clinical settings.Conclusion Our technology can be applied in a progressive recommendation scenario, providing recommendations for related items in a progressive manner, which can assist in providing more reliable diagnoses and herbal therapies for TCM clinical task.

In recommender systems (RSs), inverse propensity score (IPS) has been a key technique to mitigate popularity bias by decreasing the contribution of popular items in modeling user-item interactions. However, conventional IPS treats all users equally, which tends to over-debias the popularity-insensitive (PI) users and under-debias the popularity-sensitive (PS) users. Furthermore, in such a treatment, IPS only performs slightly well on the debiased test while does not work on the normal biased test. To this end, we propose a user-dependent IPS (UDIPS in short) method, which adaptively conducts propensity estimation for each user-item pair based on the user's sensitivity to item popularity. Like IPS, our theoretical analysis validates the unbiasedness of UDIPS. Remarkably, our solution is model-agnostic and can be easily used to upgrade current unbiased recommenders. We implemented it in four state-of-the-art models for unbiased recommendation, and experimental results on two benchmark datasets demonstrate the effectiveness of our method in both unbiased and normal biased test.

Machine vision-based automatic anti-bird thorn failure inspection, instead of manual identification, remains a great challenge. In this paper, we proposed a novel Object Position Embedding Network (OPENnet), which can improve the precision of anti-bird thorn localization. OPENnet can simultaneously predict the location boxes of the support device and anti-bird thorn by using the proposed double-head network. And then, OPENnet is optimized using the proposed symbiotic loss function (SymLoss), which embeds the object position into the network. The comprehensive experiments are conducted on the real railway video dataset. OPENnet yields competitive performance on anti-bird thorn localization. Specifically, the localization performance gains +3.65 AP, +2.10 AP50, and +1.22 AP75.

In the era of information explosion, people are constantly exposed to a wealth of unreliable information. Nevertheless, societal stability and public trust may be seriously threatened by such unverified or even fabricated false information. In order to discriminate the fake news, several fake news detection approaches have been proposed. However, most of the current multimodal fake news detection methods only rely on news text and images from a single data source. Owing to the scarcity of multimodal information and the limited volume of data, the effectiveness of the majority of these methods is low. As a result, we embrace the approach of amalgamating multi-source heterogeneous data and propose a new network structure FusionNet. In this network structure, we creatively use web crawler technology to expand the data. We achieve double amplification of text and image as opposed to the prior single-mode amplification of text, and by data screening and cleaning, we create multi-source and multimodal datasets for various viewpoints on the same event. We also propose a novel modal fusion approach that leverages the inherent information in news content, integrates an attention mechanism, and explores the shared relevant features between original news and multi-source news. Through empirical results and analysis, our approach is superior to other methods.

Semi-supervised learning is a promising solution to mitigate data sparsity in review-aware rating regression (RaRR), but it bears the risk of learning with noisy pseudo-labelled data. In this paper, we propose a paradigm called co-training-teaching ( CoT 2 ), which integrates the merits of both co-training and co-teaching towards the robust semi-supervised RaRR. Concretely, CoT 2 employs two predictors and each of them alternately plays the roles of "labeler" and "validator" to generate and validate pseudo-labelled instances. Extensive experiments show that CoT 2 considerably outperforms state-of-the-art RaRR techniques, especially when training data is severely insufficient.