Sparse Data Research Articles

Pairwise Intent Graph Embedding Learning for Context-Aware Recommendation with Knowledge Graph

Different from the data sparsity that traditional recommendations suffer from, context-aware recommender systems (CARS) face specific sparsity challenges related to contextual features, i.e., feature sparsity and interaction sparsity. How knowledge graphs address these challenges remains under-discussed. To bridge this gap, in this paper, we first propose a novel pairwise intent graph (PIG) containing nodes of users, items, entities, and enhanced intents to integrate knowledge graphs into CARS efficiently. Enhanced intent nodes are generated through the specific fusion of relational sub-intent and contextual sub-intent, and they are derived from semantic information and contextual information, respectively. We develop a pairwise intent graph embedding learning (PING) framework based on it. Specifically, our PING uses a pairwise intent joint graph convolution module to obtain refined embedding of all the features, where each enhanced intent node acts as a hub to effectively propagate information among different features and between all the features and knowledge graphs. Then, a recommendation module with refined embeddings is used to replace the randomly initialized embeddings of downstream recommendation models to improve model performance. Extensive experiments on three public datasets and some real-world scenarios verify the effectiveness and compatibility of our PING.

Optimizing tacrolimus dosing in Hispanic renal transplant patients: insights from real-world data

AimTacrolimus, an immunosuppressant used to prevent organ rejection in renal transplant patients, exhibits high inter-patient variability, necessitating therapeutic drug monitoring. Early post-transplant tacrolimus exposure in Hispanics is understudied. Although genotypic information is linked to pharmacokinetic differences, its clinical application remains limited. This study aimed to use a real-world data-driven, pharmacokinetic model-based approach for tacrolimus in Hispanics to determine a suitable initial dose and design an optimal dose titration strategy by simulations to achieve plasma trough concentration target levels of 10–12 ng/mL at the earliest.MethodsSparse concentration-time data of tacrolimus were obtained from electronic medical records for self-identified Hispanic subjects following renal transplant. Rich pharmacokinetic literature data was leveraged to estimate structural pharmacokinetic model parameters, which were then fixed in the current analysis. Only apparent clearance was estimated with the sparse tacrolimus data and potential covariates were identified. Simulations of various starting doses and different dose titration strategies were then evaluated.ResultsThe analysis included 121 renal transplant patients with 2,215 trough tacrolimus concentrations. A two-compartment transit absorption model with allometrically scaled body weight and time-varying hematocrit on apparent clearance adequately described the data. The estimated apparent clearance was 13.7 L/h for a typical patient weighing 70 kg and at 30% hematocrit, demonstrating a 40% decrease in clearance compared to other patient populations. Model based simulations indicated the best initial dose for the Hispanic population is 0.1 mg/kg/day. The proposed titration strategy, with three dose adjustments based on trough levels of tacrolimus, increased the proportion of patients within the target range (10–12 ng/mL) more than 2.5-fold and decreased the proportion of patients outside the therapeutic window by 50% after the first week of treatment.ConclusionHispanic renal transplant population showed an estimated 40% decrease of apparent clearance in the typical patient compared to other populations with similar characteristics. The proposed dose adjustment attained the target range rapidly and safely. This study advocates for tailored tacrolimus dosing regimens based on population pharmacokinetics to optimize therapy in Hispanic renal transplant recipients.

Enhancing Fake News Detection with Word Embedding: A Machine Learning and Deep Learning Approach

The widespread dissemination of fake news on social media has necessitated the development of more sophisticated detection methods to maintain information integrity. This research systematically investigates the effectiveness of different word embedding techniques—TF-IDF, Word2Vec, and FastText—when applied to a variety of machine learning (ML) and deep learning (DL) models for fake news detection. Leveraging the TruthSeeker dataset, which includes a diverse set of labeled news articles and social media posts spanning over a decade, we evaluated the performance of classifiers such as Support Vector Machines (SVMs), Multilayer Perceptrons (MLPs), and Convolutional Neural Networks (CNNs). Our analysis demonstrates that SVMs using TF-IDF embeddings and CNNs employing TF-IDF embeddings achieve the highest overall performance in terms of accuracy, precision, recall, and F1 score. These results suggest that TF-IDF, with its capacity to highlight discriminative features in text, enhances the performance of models like SVMs, which are adept at handling sparse data representations. Additionally, CNNs benefit from TF-IDF by effectively capturing localized features and patterns within the textual data. In contrast, while Word2Vec and FastText embeddings capture semantic and syntactic nuances, they introduce complexities that may not always benefit traditional ML models like MLPs or SVMs, which could explain their relatively lower performance in some cases. This study emphasizes the importance of selecting appropriate embedding techniques based on the model architecture to maximize fake news detection performance. Future research should consider integrating contextual embeddings and exploring hybrid model architectures to further enhance detection capabilities. These findings contribute to the ongoing development of advanced computational tools for combating misinformation.

The 4D Camera: An 87 kHz Direct Electron Detector for Scanning/Transmission Electron Microscopy

Abstract We describe the development, operation, and application of the 4D Camera—a 576 by 576 pixel active pixel sensor for scanning/transmission electron microscopy which operates at 87,000 Hz. The detector generates data at ∼480 Gbit/s which is captured by dedicated receiver computers with a parallelized software infrastructure that has been implemented to process the resulting 10–700 Gigabyte-sized raw datasets. The back illuminated detector provides the ability to detect single electron events at accelerating voltages from 30 to 300 kV. Through electron counting, the resulting sparse data sets are reduced in size by 10--300× compared to the raw data, and open-source sparsity-based processing algorithms offer rapid data analysis. The high frame rate allows for large and complex scanning diffraction experiments to be accomplished with typical scanning transmission electron microscopy scanning parameters.

Using Exact Tests from Algebraic Statistics in Sparse Multi-Way Analyses: An Application to Analyzing Differential Item Functioning

Asymptotic goodness-of-fit methods in contingency table analysis can struggle with sparse data, especially in multi-way tables where it can be infeasible to meet sample size requirements for a robust application of distributional assumptions. However, algebraic statistics provides exact alternatives to these classical asymptotic methods that remain viable even with sparse data. We apply these methods to a context in psychometrics and education research that leads naturally to multi-way contingency tables: the analysis of differential item functioning (DIF). We explain concretely how to apply the exact methods of algebraic statistics to DIF analysis using the R package algstat, and we compare their performance to that of classical asymptotic methods.

Autonomous Vehicles Traversability Mapping Fusing Semantic–Geometric in Off-Road Navigation

This paper proposes an evaluating and mapping methodology of terrain traversability for off-road navigation of autonomous vehicles in unstructured environments. Terrain features are extracted from RGB images and 3D point clouds to create a traversal cost map. The cost map is then employed to plan safe trajectories. Bayesian generalized kernel inference is employed to assess unknown grid attributes due to the sparse raw point cloud data. A Kalman filter also creates density local elevation maps in real time by fusing multiframe information. Consequently, the terrain semantic mapping procedure considers the uncertainty of semantic segmentation and the impact of sensor noise. A Bayesian filter is used to update the surface semantic information in a probabilistic manner. Ultimately, the elevation map is utilized to extract geometric characteristics, which are then integrated with the probabilistic semantic map. This combined map is then used in conjunction with the extended motion primitive planner to plan the most effective trajectory. The experimental results demonstrate that the autonomous vehicles obtain a success rate enhancement ranging from 4.4% to 13.6% and a decrease in trajectory roughness ranging from 5.1% to 35.8% when compared with the most developed outdoor navigation algorithms. Additionally, the autonomous vehicles maintain a terrain surface selection accuracy of over 85% during the navigation process.

Active learning inspired multi-fidelity probabilistic modelling of geomaterial property

The identification of geomaterial properties is costly but pivotal for engineering design. A wide range of approaches perform well with sufficiently measured data but their performance is problematic for sparse data. To address this issue, this study proposes an active learning based multi-fidelity residual Gaussian process (AL-MR-GP) modelling framework. A low-fidelity (LF) prediction model is first trained using extensive LF data collected from worldwide sites to generate preliminary estimations. Subsequent training employs active learning to prioritize high-fidelity data from the specific site of interest with larger information gain for calibrating the LF model to make ultimate predictions. The compression index of clays is selected as an example to examine the capability of the proposed framework. The results indicate that using the same number of site-specific datasets, the compression index of clays can be well captured by AL-MR-GP, exhibiting superior accuracy and reliability than models without incorporating multi-fidelity data or active learning. Based on unified LF data, the proposed framework becomes data-efficient for the model development of three sites and is significantly competitive in extrapolation, compared with site-specific models even with active learning. These promising characteristics indicate substantial potential to be extended to broader applications in geotechnical engineering.

Leveraging gene correlations in single cell transcriptomic data

BackgroundMany approaches have been developed to overcome technical noise in single cell RNA-sequencing (scRNAseq). As researchers dig deeper into data—looking for rare cell types, subtleties of cell states, and details of gene regulatory networks—there is a growing need for algorithms with controllable accuracy and fewer ad hoc parameters and thresholds. Impeding this goal is the fact that an appropriate null distribution for scRNAseq cannot simply be extracted from data in which ground truth about biological variation is unknown (i.e., usually).ResultsWe approach this problem analytically, assuming that scRNAseq data reflect only cell heterogeneity (what we seek to characterize), transcriptional noise (temporal fluctuations randomly distributed across cells), and sampling error (i.e., Poisson noise). We analyze scRNAseq data without normalization—a step that skews distributions, particularly for sparse data—and calculate p values associated with key statistics. We develop an improved method for selecting features for cell clustering and identifying gene–gene correlations, both positive and negative. Using simulated data, we show that this method, which we call BigSur (Basic Informatics and Gene Statistics from Unnormalized Reads), captures even weak yet significant correlation structures in scRNAseq data. Applying BigSur to data from a clonal human melanoma cell line, we identify thousands of correlations that, when clustered without supervision into gene communities, align with known cellular components and biological processes, and highlight potentially novel cell biological relationships.ConclusionsNew insights into functionally relevant gene regulatory networks can be obtained using a statistically grounded approach to the identification of gene–gene correlations.

Forecasting Urban Traffic States with Sparse Data Using Hankel Temporal Matrix Factorization

Forecasting urban traffic states is crucial to transportation network monitoring and management, playing an important role in the decision-making process. Despite the substantial progress that has been made in developing accurate, efficient, and reliable algorithms for traffic forecasting, most existing approaches fail to handle sparsity, high-dimensionality, and nonstationarity in traffic time series and seldom consider the temporal dependence between traffic states. To address these issues, this work presents a Hankel temporal matrix factorization (HTMF) model using the Hankel matrix in the lower dimensional spaces under a matrix factorization framework. In particular, we consider an alternating minimization scheme to optimize the factor matrices in matrix factorization and the Hankel matrix in the lower dimensional spaces simultaneously. To perform traffic state forecasting, we introduce two efficient estimation processes on real-time incremental data, including an online imputation (i.e., reconstruct missing values) and an online forecasting (i.e., estimate future data points). Through extensive experiments on the real-world Uber movement speed data set in Seattle, Washington, we empirically demonstrate the superior forecasting performance of HTMF over several baseline models and highlight the advantages of HTMF for addressing sparsity, nonstationarity, and short time series. History: Accepted by Ram Ramesh, Area Editor for Data Science & Machine Learning. Funding: This research was supported by the Institute for Data Valorisation, the Interuniversity Research Centre on Enterprise Networks, Logistics and Transportation, the National Natural Science Foundation of China [Grants 12371456, 72101049, 72232001], the Sichuan Science and Technology Program [Grant 2024NSFJQ0038], and the Fundamental Research Funds for the Central Universities [Grant DUT23RC(3)045].

A roadmap for generating annual bycatch estimates from sparse at-sea observer data

Abstract To support ecosystem-based fisheries management, monitoring data from at-sea observer (ASO) programs should be leveraged to understand the impact of fisheries on discarded species (bycatch). Available techniques to estimate fishery-scale quantities from observations range from simple mean estimators to more complex spatiotemporal models, each making assumptions with differing degrees of support. However, the resulting implementation and analytical trade-offs are rarely discussed when applying these techniques in practice. Using blue shark (Prionace glauca) bycatch in the Canadian pelagic longline fishery as a case study, we evaluated the performance of seven contrasting approaches to estimating total annual discard amounts and assessed their trade-offs in application. Results demonstrated that simple approaches such as mean estimator and nearest neighbors are feasible to implement and can be as efficient for prediction as complex models such as random forest and mixed-effects models. The traditionally used catch-ratio estimator consistently underperformed among all tested models, likely due to misspecified correlative relationships between target and bycatch species. Overall, efforts in model-based approaches were rewarded with very small gains in predictive ability, suggesting that such models relying on environmental, biological, spatial, and/or temporal patterns to improve prediction of bycatch may lack sufficient foundation in data-limited contexts.

Exploring protein structural ensembles: Integration of sparse experimental data from electron paramagnetic resonance spectroscopy with molecular modeling methods.

Under physiological conditions, proteins continuously undergo structural fluctuations on different timescales. Some conformations are only sparsely populated, but still play a key role in protein function. Thus, meaningful structure-function frameworks must include structural ensembles rather than only the most populated protein conformations. To detail protein plasticity, modern structural biology combines complementary experimental and computational approaches. In this review, we survey available computational approaches that integrate sparse experimental data from electron paramagnetic resonance spectroscopy with molecular modeling techniques to derive all-atom structural models of rare protein conformations. We also propose strategies to increase the reliability and improve efficiency using deep learning approaches, thus advancing the field of integrative structural biology.

Cascading graph contrastive learning for multi-behavior recommendation

Traditional recommendation techniques often prioritize target behavior in practical recommendation scenarios(e.g., follow, play and buy). However, these approaches suffer from data sparsity issues and may not fully capture user’s personal preferences. To address this deficiency, multi-behavior recommendation technology has emerged, leveraging users’ multi-behavioral interactions for recommendation. Nevertheless, certain multi-behavior recommendation methods learning behavioral information from each behavior separately and then aggregate them before making recommendation, which inadvertently neglects the intrinsic connections between different behaviors. In some scenarios, user behavior often occurs in a fixed order, such as view -> cart -> buy in e-commerce platforms. In this work, we propose a novel Cascading Graph Constrastive Learning (CGCL) framework for Multi-Behavior recommendation. Specifically, we devise a graph contrastive learning block to learn distinctive user behavioral representations for each type of interaction. Leveraging the recommendation task, we aim to capture user preferences, while the contrastive learning provides supplementary supervisory signals to refine the user and item representation. By acknowledging the sequential order of behaviors, we utilize the cascading structure within our model to iteratively propagate and purify the personalized preferences of users. Extensive experimental results and ablation studies on three real-world datasets have shown that our CGCL framework outperforms various state-of-the-art recommendation methods and validated the effectiveness of our approach.

Prior-image-based low-dose CT reconstruction for adaptive radiation therapy.

The study aims to reduce the imaging radiation dose in Adaptive Radiotherapy (ART) while maintaining high-quality CT images, critical for effective treatment planning and monitoring. We developed the Prior-aware Learned Primal-Dual Network (pLPD-UNet), which uses prior CT images to enhance reconstructions from low-dose scans. The network was separately trained on thorax and abdomen datasets to accommodate the unique imaging requirements of each anatomical region. The pLPD-UNet demonstrated improved reconstruction accuracy and robustness in handling sparse data compared to traditional methods. It effectively maintained image quality essential for precise organ delineation and dose calculation, while achieving a significant reduction in radiation exposure. This method offers a significant advancement in the practice of ART by integrating prior imaging data, potentially setting a new standard for balancing radiation safety with the need for high-resolution imaging in cancer treatment planning.

Multi-Granularity and Multi-Interest Contrast-Enhanced Hypergraph Convolutional Networks for Session Recommendation

Aiming at the problems of the incomplete recommendation and sparsity of session data in session recommendation, a new multi-granularity and multi-interest contrast-enhanced hypergraph convolutional network (MGMI-CEHCN) model for session recommendation is proposed. The session data are modeled as a heterogeneous hypergraph, the information is embedded by the hypergraph using two granularities of item price and category, and then the information between different granularities is fused, while the final item embedding is obtained through multi-layer convolution. Finally, an interest perceptron is used to detect multiple potential interests for each item, and a decentralized interest extraction network based on a gated recurrent unit (GRU) is used to integrate the user’s final interests and obtain a global session representation through a soft attention mechanism; a local session representation is generated with the help of a weighted line graph convolutional network. A further joint contrast enhancement strategy is used to maximize the mutual information between the global session representation and local session representation, to improve the recommendation performance. Experiments on several real datasets showed that the recommendation performance of the MGMI-CEHCN model outperformed the current mainstream models. On Cosmetics, the P@20 reached up to 55.25% and M@20 reached up to 38.26%, improvements of 3.06% and 3.09%, respectively; on Diginetica-buy, the P@20 reached up to 65.60% and M@20 reached up to 27.47%, improvements of 2.47% and 6.64%, respectively, which proved the validity of the model.

A fast laser ultrasonic SAFT method with sparse scanning data and F-K based interpolation

ABSTRACT Laser ultrasonic SAFT can be applied for noncontact testing and imaging of defects in structures. However, due to the large demand for signal data by laser scanning, the time consumption of this method would be relatively long. To address this problem, this paper proposes a fast LU-SAFT method with sparse scanning data and F-K based interpolation. It aims to reduce the inspection time but without decreasing the imaging resolution. First, the method determines the minimum number of scanning sparse points using simulation. Then the F-K interpolation method is utilised to recover the full-field unscanned point signal from the sparse scan data. Finally, the reconstructed signal is used for SAFT image reconstruction of the defect. The experimental results show that, for the same scanning area, the proposed method in this paper reduces the number of scanning points by 87.5% and shortens the scanning time by 86.0% compared to conventional LU-SAFT point-by-point scanning. Additionally, the signal-to-noise ratio of the defective SAFT imaging decreases by only about 7.9%.

Cross-domain recommender system with embedding- and mapping-based knowledge correlation

A knowledge transfer-based cross-domain recommender system is currently a research hotspot. Existing research has reached a high level of maturity in mining potential knowledge and establishing transfer mechanisms. However, most of them ignore the impact of the dissimilarity of potential knowledge on the transfer performance. Herein, a cross-domain recommender system based on knowledge correlation-induced the embedding and mapping approach is proposed, denoted by KCEM-CDRS. First, we propose a knowledge correlation measure, which captures the consistency of knowledge between the target and source domains to build the bridge for knowledge transfer. Second, we construct a joint matrix triple factorization model to solve the data sparsity in the target domain while introducing graph regularization to solve the problem of negative knowledge transfer. Results of extensive experiments on real Amazon metadata indicate that compared with three existing cross-domain recommendation methods, KCEM-CDRS shows performance improvements of 0.05–9.55 % and 0.02–2.63 % on mean absolute error and root mean square error, respectively. Additionally, the results of the ablation experiments indicate that consideration of the knowledge correlation between domains is beneficial for knowledge transfer when the density of the source domain is rich.

Clarifying urban flood response characteristics and improving interpretable flood prediction with sparse data considering the coupling effect of rainfall and drainage pipeline siltation

With climate warming and accelerated urbanisation, severe urban flooding has become a common problem worldwide. Frequent extreme rainfall events and the siltation of drainage pipes further increase the burden on urban drainage networks. However, existing studies have not fully considered the effects of rainfall and pipeline siltation on the response characteristics of flooding when constructing numerical models of urban flooding simulations. To solve this problem, a surface-subsurface coupling model was constructed by combining the Saint-Venant equation, Manning equation, a one-dimensional pipeline model (SWMM), and a two-dimensional surface overflow model (LISFLOOD-FP). Then, the SWMM model considering pipeline siltation and the two-dimensional surface overflow model (LISFLOOD-FP) were coupled with the flow exchange governing equation, and the urban flooding response characteristics considering the coupling effect of “rainfall and drainage pipeline siltation” were analysed. To enhance the solvability of waterlogging prediction, an intelligent prediction model of urban flooding based on Bayes-CNN-BLSTM was established by combining a convolutional neural network (CNN), bidirectional long short-term memory neural network (BLSTM), Bayesian optimisation (Bayes), and an interpretable loss function error correction method. The actual rainfall events and flooding processes recorded by the monitoring equipment at Huizhou University were used to calibrate and verify the model. The results show that in the Rainfall 1 and Rainfall 2 scenarios, the overload rates of the pipelines in the current siltation scenario were 60.06 % and 68.37 %, respectively, and the proportions of overflow nodes were 24.87 % and 25.89 %, respectively. When the drainage network was initially put into operation, the overload rates of the pipeline were 36.67 % and 41.16 %, and the overflow nodes accounted for 3.05 % and 4.06 %, respectively. The inundated area and volume of urban flooding increased when the combined siltation coefficient (CSC) was 0.2; therefore, two desilting schemes were determined. Under Rainfall 1, Rainfall 2, and the four rainfall recurrence periods, the Bayes-CNN-BLSTM model had clear advantages in terms of accuracy, reliability, and robustness.

The Burden of Psychiatric Manifestations in Inflammatory Bowel Diseases: A Systematic Review With Meta-analysis.

Psychiatric disorders in patients with inflammatory bowel disease (IBD) represent a significant but uncertain facet of the disease, with unsolved questions regarding their overall magnitude, their impact on intestinal disease, and the whole burden of psychiatric manifestations. This systematic review summarizes the evidence on the prevalence and impact of psychiatric disorders, including depression, anxiety, bipolar disorder (BD), and schizophrenia, among patients with IBD. A systematic search across PubMed/MEDLINE, Embase, and Scopus databases from January 2010 to January 2023 was performed to identify relevant studies. The focus was on studies exploring the prevalence of specific psychiatric disorders in IBD patients compared to the general population and that reported specific outcome measures. A subsequent meta-analysis (MA) assessed the strength of the association between IBD and these psychiatric disorders, with data reliability ensured through rigorous extraction and quality assessment. Out of 3,209 articles, 193 met the inclusion criteria and only 26 provided complete data for comprehensive analysis. These studies showed a significantly higher overall prevalence of psychiatric comorbidities in IBD patients compared to the general population. The MA showed a significant association between IBD and depression (pooled OR 1.42, 95% CI = 1.33-1.52, P < .0001) and anxiety (pooled OR 1.3, 95% CI = 1.22-1.44, P < .0001). The association between IBD and BD was significant (pooled OR 1.64, 95% CI = 1.20-2.24, P < .0001) but showed considerable heterogeneity (I2 = 94.01%). Only 3 studies examined the association between schizophrenia and IBD, providing widely heterogeneous results, with an inconclusive OR, estimated at 0.93 (95% CI = 0.62-1.39, P = .73). This MA highlights the high prevalence of psychiatric disorders, particularly depression and anxiety, in IBD patients, which exceeds rates in the general population. BD in IBD is proving to be an important but under-researched area. The sparse and contradictory data on schizophrenia requires further investigation. These findings highlight the need for better understanding, early detection, and tailored mental health interventions in the management of IBD to significantly improve patients' quality of life.

GIAE-DTI: Predicting Drug-Target Interactions Based on Heterogeneous Network and GIN-based Graph Autoencoder.

Accurate prediction of drug-target interactions (DTIs) is essential for advancing drug discovery and repurposing. However, the sparsity of DTI data limits the effectiveness of existing computational methods, which primarily focus on sparse DTI networks and have poor performance in aggregating information from neighboring nodes and representing isolated nodes within the network. In this study, we propose a novel deep learning framework, named GIAE-DTI, which considers cross-modal similarity of drugs and targets and constructs a heterogeneous network for DTI prediction. Firstly, the model calculates the cross-modal similarity of drugs and proteins from the relationships among drugs, proteins, diseases, and side effects, and performs similarity integration by taking the average. Then, a drug-target heterogeneous network is constructed, including drug-drug interactions, protein-protein interactions, and drug-target interactions processed by weighted K nearest known neighbors. In the heterogeneous network, a graph autoencoder based on a graph isomorphism network is employed for feature extraction, while a dual decoder is utilized to achieve better self-supervised learning, resulting in latent feature representations for drugs and targets. Finally, a deep neural network is employed to predict DTIs. The experimental results indicate that on the benchmark dataset, GIAE-DTI achieves AUC and AUPR scores of 0.9533 and 0.9619, respectively, in DTI prediction, outperforming the current state-of-the-art methods. Additionally, case studies on four 5-hydroxytryptamine receptor-related targets and five drugs related to mental diseases show the great potential of the proposed method in practical applications.

Multimodal Recommender Systems: A Survey

The recommender system (RS) has been an integral toolkit of online services. They are equipped with various deep learning techniques to model user preference based on identifier and attribute information. With the emergence of multimedia services, such as short videos, news and etc. , understanding these contents while recommending becomes critical. Besides, multimodal features are also helpful in alleviating the problem of data sparsity in RS. Thus, M ultimodal R ecommender S ystem (MRS) has attracted much attention from both academia and industry recently. In this paper, we will give a comprehensive survey of the MRS models, mainly from technical views. First, we conclude the general procedures and major challenges for MRS. Then, we introduce the existing MRS models according to four categories, i.e., Modality Encoder , Feature Interaction , Feature Enhancement and Model Optimization . Besides, to make it convenient for those who want to research this field, we also summarize the dataset and code resources. Finally, we discuss some promising future directions of MRS and conclude this paper. To access more details of the surveyed papers, such as implementation code, we open source a repository.