ABSTRACT. Picat is a logic-based multi-paradigm language that integrates logic programming, functional programming, constraint programming, and scripting. Picat takes many features from other languages, including logic variables, unification, backtracking, pattern-matching rules, functions, list/array comprehensions, loops, assignments, tabling for dynamic programming and planning, and constraint solving with CP (constraint programming), SAT (satisfiability), MIP (mixed integer programming), and SMT (Satisfiability Modulo Theories). In this talk, I will show how to declaratively model some combinatorial problems, including constraint satisfaction and planning problems, and solve them efficiently using Picat. I will also explain how high-level models are compiled to SAT encodings for fast solving.

ABSTRACT. An autonomous system is constructed by a manufacturer, operates in a society subject to norms and laws, and interacts with end users. All of these actors are stakeholders affected by the behavior of the autonomous system. We address the challenge of how the ethical views of such stakeholders can be integrated in the behavior of an autonomous system. We propose an ethical recommendation component called Jiminy which uses techniques from normative systems and formal argumentation to reach moral agreements among stakeholders. A Jiminy represents the ethical views of each stakeholder by using normative systems, and has three ways of resolving moral dilemmas that involve the opinions of the stakeholders. Jiminy represents the ethical views of each stakeholder by using normative systems, and has three ways of resolving moral dilemmas involving the opinions of the stakeholders. First, Jiminy considers how the arguments of the stakeholders relate to one another, which may already resolve the dilemma. Secondly, Jiminy combines the normative systems of the stakeholders such that the combined expertise of the stakeholders may resolve the dilemma. Thirdly, and only if these two other methods have failed, Jiminy uses context-sensitive rules to decide which of the stakeholders take preference. At the abstract level, these three methods are characterized by the addition of arguments, the addition of attacks among arguments, and the revision of attacks among arguments. We show how Jiminy can be used not only for ethical reasoning and collaborative decision making, but also for providing explanations about ethical behavior.

ABSTRACT. Graph (KGs) that cannot cover all the required content. The task of aligning Knowledge Graphs is to find nodes in source and target Knowledge Graphs that point to the same objective entities in the real world, which can achieve the goal of building a unified Knowledge Graph of rich content. In previous work, there is still potential for improvement in the utilization of structures, attributes, and relational information of Knowledge Graphs. Our work more accurately utilizes the important information contained in a large number of two-hop neighbors by sampling them based on CSLS simi-larity. It is more precise utilization of structural information contained in a large number of two-hop neighbors. At the same time, we use a simplified hyperplane generation method in learning embedding of relations of Knowledge Graph. Reduced computational costs without compromising accuracy, a Two-Hop neighbor sampling and Distinguishable Relation em-bedding (TH-DR) framework was proposed. Extensive experiments on DBP15K datasets demonstrate that our method significantly outperforms existing systems.

ABSTRACT. ct. Hierarchical time series refers to a collection of time series with a hierarchical structure, which is widely distributed in daily life and industrial fields. The differences among the various series in the hierarchical time series result in complex and diverse temporal features of the data. How to effectively extract the temporal features of each series is an important research question in prediction tasks. In addition, current research mainly focuses on ensuring the consistency of prediction results among different levels, and lacks consideration of the interaction of temporal patterns among different levels and the mutual influence among different time series. To alleviate these issues, this paper proposes a hierarchical time series prediction method that combines Temporal Convolutional Network (TCN) with attention mechanism. Firstly, the TCN model is constructed to extract the temporal features of different time series, and the time attention mechanism is introduced into the TCN to calculate the influence of each time periods on the target value, further improving the feature extraction ability of the model. Secondly,an optimal combination, bottom-up combined prediction method is proposed to realize the interaction between temporal patterns among different levels in the hierarchical structure and improve the utilization rate of hierarchical structure information. Finally, a hierarchical attention mechanism is proposed to address the issue of mutual influence between various time series in hierarchical time series, which is used to calculate the degree of mutual influence between each time series and improve the final prediction accuracy of the model. The effectiveness of the proposed model is verified by experiments on two typical hierarchical time series datasets, Dairy and Walmart, and the prediction performance of the proposed model is better than that of other b

ABSTRACT. With the rapid increase of the data, not only the scale of data is very large, but also the dimensionality of data is very high. It is increasingly difficult to perform feature extraction on high-dimensional datasets, called curse of dimensionality. Traditional manifold learning algorithms propose to mine the local or global geometric structure of high-dimensional datasets for feature extraction. However, such algorithms can only mine the local linear structure of the datasets, and there is no way to mine the higher-order geometric structure. In this paper, we propose a new nonlinear manifold learning algorithm to mine the curvature information of dataset, which can quantify the spatial similarity between data points more accurately and improve the accuracy of feature extraction. We perform clustering experiments on several types of image datasets, and the related experimental results show that our algorithm outperforms the traditional manifold learning algorithms.

ABSTRACT. Link prediction is one of the key problems in network science and can be applied to many scenarios such as recommender system, knowledge graph completion. Traditional link prediction methods are either based on heuristics or shallow network embedding methods. The heuristic methods tend to have strong assumptions on nodes similarities, which may hardly generalize well. Relatively, the shallow network embedding methods solve the link prediction problem in a two-stage architecture, which cannot optimize node representations and link probabilities simultaneously. Recently, some link prediction methods based on graph neural networks have been proposed to alleviate above problems. Although the graph neural network-based methods achieve promising performances, they do not preserve the original pairwise proximity at each hidden layer since non-linear operations may hinder it. In this paper, we propose a novel end-to-end link prediction method named Pairwise Proximity Preserving Graph neural network (PPPG), which can preserve the pairwise proximity at each layer. Specifically, inspired by graph signal denoising techniques, we preserve the structural pairwise proximity by solving an optimization problem after the non-linear operations at each layer. We evaluate our model on various standard link prediction benchmarks, and obtain competitive results.

ABSTRACT. Korean is the native and official language of Chinese-Korean people. Although the Chinese-Korean and the Korean in South Korea are the same language, there are differences in vocabulary, grammar, and other aspects due to cultural influences. Reasonable analysis and utilization of texts are of great significance for promoting the development of minority languages and enhancing ethnic unity. In this paper, we propose a few-shot Chinese-Korean text classification method(CKKP) based on pre-trained language models(PLMS) combined with knowledge base(KB) and prompt learning . In this work, Chinese-Korean news data are scraped from Yanbian Daily , and we paper clean and label the data to obtain a Chinese-Korean datasets. We focus on collecting external knowledge and integrating it into the utterance to form a prompt of knowledge to improve and stabilize the prompt. Specifically,we use the clustering algorithm to construct the label wordspace of the external knowledge base extended language.Then, before using the extended label word space for prediction, we use PLM itself to refine the extended label word space.Finally, a projection is constructed between the label space for prompt learning and the label word space. Extensive experiments on a Chinese-Korean news datasets collected from Yanbian Daily, Naver Korean, and Twitter Korean datasets which demonstrate the effectiveness of the knowledge prompt learning.

ABSTRACT. The advent of intelligent education systems and widespread distance learning have revolutionized the educational landscape. Extracting meaningful insights from this wealth of information is crucial for improving student learning outcomes. Knowledge tracking play a pivotal role in monitoring and optimizing students’ comprehension by simulating their understanding of knowledge concepts. Previous studies on knowledge tracing models have largely overlooked the influence of forgetting law and lack effective representation methods for forgetting. To address this problem, we proposes a novel knowledge tracking model based on GRU-attention with integrated forgetting law, abbreviated as FGAKT. FGAKT is composed of four essential modules: an embedding module that incorporates forgetting laws to optimize memory retention, an attention extraction module that employs an encoder-decoder structure for effective feature selection, a GRU module to capture long-term timing information, and a prediction module. To enhance the interpretability of the model, we introduce the weight of the attention extraction module’s output into the loss function, which encourages the model to prioritize features. Additionally, we use regulation parameters to balance the predictive performance and interpretability of the model. Experimental results demonstrate that FGAKT outperforms the baselines in terms of ACC and AUC. Furthermore, ablation experiments show the feasibility of each module and the impact of layers within the encoder-decoder structure. The student answering process is instantiated to provide an intuitive depiction of how the model simulates a student’s response.

ABSTRACT. Over the past decade, multivariate time series (MTS) forecasting is becoming a research hotspot in the fields of data mining. Despite the emergence of several deep learning-based models achieve superior predictive performance, the lack of credible explanations limits their usability in many fields. To bridge this gap, on the premise of ensuring the accuracy of model prediction, we propose an interpretable forecasting model called XCTF (eXplainable CNN for multivariate Time series Forecasting), which mainly combines a Dual-Attention Module and a Saliency Detection Module. XCTF architecture enables a good generalization ability on multivariate time series forecasting. The Dual-Attention Module is leveraged to extract variable features related to prediction and parallelly extract temporal features from the input data, which leads to more discriminative features by incorporating all the relevant information. The Saliency Detection Module has two main parts: ConvDown block and ConvUp block. The ConvDown block can filter out the small, redundant features but keep the important information. The ConvUp block can enlarge the extracted feature map to obtain more accurate prediction data. Then, by using a method based on Grad-CAM, XCTF can generate an attribution map of feature importance. The attribution map can precisely demonstrate the important information of the input data both on observed variables and time. Experimental results on three benchmark datasets (ETTh1, ETTm1, and Exchange) demonstrate that XCTF has high performance when compares with current mainstream models. Moreover, XCTF offers strong interpretability and practical relevance by providing an attribution map of observed variables with time.

ABSTRACT. Multi-label feature selection is an essential component of data mining that can improve classification accuracy. By reducing redundant label information, the feature selection method can select a better feature subset. However, current methods remove too much redundant information and fail to preserve the original label information effectively. To address this problem, this paper proposes a new multi-label feature selection method called granular label dynamic feature selection (GLDFS). First, a label granularity algorithm is proposed to remove redundant label information by granulating highly correlated labels into the same information granule and representing the information granule with the label closest to the center. Second, an automatic granulation quantity algorithm is used to control the granularity size and reduce the loss of useful information. Third, dynamic relevance is introduced into multi-label feature selection from single-label feature selection to select a better feature subset. Finally, compared with six other methods, GLDFS outperforms other methods on three evaluation indicators of thirteen data sets in most cases.

ABSTRACT. Low-resolution face image recognition is a research focus in computer vision and pattern recognition. In this paper, the potential relationship between matched high-resolution face and low-resolution face images is considered to propose the regularized loose coupled deep non-negative basis matrix factor-ization (RLCDNBMF) architecture, which uses regularized deep non-negative basis matrix factorization for high-resolution images and couples its coefficient matrix with the coefficient matrix obtained from low-resolution non-negative factorization, so that the discriminative information of high-resolution face images can be fully utilized to guide the non-negative matrix factorization of low-resolution face images. Comparative experiments are conducted on three mainstream face databases, and the experimental results show that the method can improve the accuracy of low-resolution face recognition.

ABSTRACT. Consistent Discriminant Correlation Analysis is an effective dual-view learning algorithm. However, in practical applications, different sample and feature information may have different importance levels, and CDCA does not fully utilize the information from different features. Therefore, this paper extends CDCA to multi-views and proposes Consistent Discriminant Correlation Analysis with weight correction. During the training process, it utilizes CDCA for multi-views to solve the feature subspace mapping of training samples, and performs weight correction on the feature and sample dimensions of the data as new samples for the second training. During testing, the KNN classifier is used for similarity measurement. Experiments on three public datasets for face recognition show that the Consistent Discriminant Correlation Analysis with weight correction can effectively improve the accuracy of low-resolution face image recognition.

ABSTRACT. For the feature selection algorithm based on information gain and Pearson correlation coefficient FSIP of high-dimensional and small sample biological data, which is inconvenient for feature subset selection, we propose a fully adaptive 2D feature selection algorithm based on discernibility matrix DFSIP(Discernibility matrix based Fully Adaptive 2D FSIP).However, the set-ting of the threshold is not suitable for each dimension of the data set, which makes the algorithm perform poorly in some specific complex situations. There-fore, a feature selection algorithm CDFSIP (Clustering based Fully adaptive 2D DFSIP) based on the adaptive density peak clustering algorithm ADA-DPC is proposed. By introducing the ADA-DPC clustering algorithm (An Adaptive Clustering Algorithm by Finding Density Peaks), the defect that the neighbor-hood parameters of the DFSIP algorithm are globally consistent is improved, so that the appropriate neighborhood parameters can be adaptively selected when calculating the neighborhood information of each feature. The CDFSIP algorithm evaluates the distribution of each feature by Gini coefficient, judges to use the traditional domain division method, or performs neighborhood division by clustering the category information, so as to achieve adaptive selection of appropriate neighborhood parameters in each feature dimension as much as possible, and thus obtain a feature subset with stronger classification ability under complex conditions. The classification performance of K-ELM classifier is used as the evaluation index of the selected feature subset. Experimental tests on classical data sets show that the algorithm can select feature subsets with strong recognition ability, and the classifier based on this feature subset also has strong classification performance, compared with DFSIP, FSIP, mRMR, LLE score, DRJMIM, AVC and AMID algorithms.

ABSTRACT. The existing low-resolution (LR) image recognition methods based on coupled relationship learning of high- and low-resolution images usually take similar feature extraction steps for both high- and low-resolution images. However, the high- and low-resolution features thus obtained may differ in terms of information content and information scale. This leads to a robust coupled relationship between the obtained high- and low-resolution images feature is difficult to be learned. Therefore, a coupled non-negative matrix factorization with a deep part-based feature learning method is proposed for the LR image recognition task. This method is divided into two main steps. Firstly, the depth non-negative matrix factorization method by graph convolution is used to obtain a robust depth underlying basis images of HR images for feature representation. Then the coupled relationship between the depth basis images feature of the HR images and the LR basis images feature is learned. Learning More Robust LR Basis images features using depth underlying features of HR images to assist LR images. Improve the expression of LR basis images of features. Experimental results show that the method can provide higher recognition performance. It outperforms existing image recognition methods in image recognition.

ABSTRACT. Based on improvement of the existing person re-identification network model and investigating person re-identification related problems on open datasets, a pedestrian re-identification method that incorporates shared feature branching as well as a fused attention mechanism is proposed. First, the benchmark network framework of pedestrian re-identification method is formed by image pre-processing method, generalized average pooling and loss function together, and BNNeck structure is added to optimize the problem of inconsistent multi-loss objectives before obtaining classification features. Second, the global feature branching structure is proposed to be assisted by shared feature branching, which can effectively assist global feature branching to obtain more distinguishable pedestrian features and the attention network is combined with person re-recognition using a network structure that fuses non-local modules with channel attention to fully utilize the feature information of pedestrian images. Finally, the performance improvement of the benchmark network with shared feature branching and the fused attention mechanism approach is demonstrated by experiments on the re-identification dataset.

ABSTRACT. To tackle the fine-grained classification problem encountered in mushroom identification, a classification model IAMR-Net(Integrating Attention Mechanism and ResNet-Net) which combined bilinear residual network and attention mechanism was proposed. The model combined a modified residual block to extract features from input images. Extracted features will be embedded to fit the Multi-Headed-Self-Attention blocks for spatial dimensional modeling to achieve the purpose of extracting fine-grained relationships in feature maps. Our model trained by mixing loss functions, on the benchmarks, such as Oxford 102 Flowers, CUB-200-2011, Stanford Cars, Stanford Dogs, and a Mushroom-96 datasets. The model obtained an accuracy score of over 91% on both the fine-grained benchmarks and Mushroom-96 datasets, demonstrating its efficacy in classifying finegrained mushroom images.

ABSTRACT. to be confirmed