ABSTRACT. Deep Learning has recently caused very rapid advances in Artificial Intelligence, opening up unprecedented opportunities in machine learning. However, to date accuracy of deep learning systems can only be assessed statistically: they are essentially black boxes, meaning that their rationales for their decisions or predictions defy analysis. At the same time, they have been shown to display bias, and sometimes to hallucinate, i.e. to produce output that is not grounded in facts. In this presentation we will discuss recent results on the use of ontologies and deductive reasoning to understand the inner workings of trained deep learning systems.

ABSTRACT. Constraint Optimization Problems (COPs) ask for an assignment of values to variables in order to optimize an objective subject to constraints that restrict the value combinations in the assignment. They are usually solved by the classical Branch and Bound (B&B) search algorithm. Dominance breaking is an important technique in B&B to prune assignments that are subordinate to others concerning the objective value and/or the satisfiability of constraints. In practice, the addition of constraints for dominance breaking can drastically speed up the B&B search for solving many COPs. However, identification of suboptimal assignments in COPs and derivation of useful constraints for dominance breaking are usually problem-specific and require sophisticated human insights on the problem structure.In this talk, we present the first theoretical and practical framework for automatic generation of dominance breaking constraints for a class of COPs consisting of efficiently checkable objectives and constraints. Instead of synthesizing general constraints for dominance breaking, the framework focuses on generating nogoods representing incompatible value assignments and formulates nogood generation as solving auxiliary constraint satisfaction problems. The proposed method can generate nogoods of varying strengths for dominance breaking by controlling the number of involved variables. We would also discuss how to generalize the framework by exploiting functional constraints and their properties, such as monotonicity, commutativity, and associativity, to generate more effective dominance breaking nogoods for COPs with nested function calls.

ABSTRACT. In the early 1960s, the major modern technologies such as aerospace and military industry in the United States developed rapidly, and some major equipment came into being. Inevitably, the equipment will fail, causing huge losses and unimaginable consequences. In order to avoid such a disaster, the failure is killed in the cradle. Then artificial intelligence formed a new branch: fault diagnosis. In intelligent fault diagnosis, the discrete event system is usually modeled by finite state machine. In this paper, a new method is presented, that transforming discrete event system modeled by finite state machine into a Boolean expression. The Boolean function is represented by a binary decision diagram (BDD). Based on this method, the discrete system is further abstracted, and the time complexity and space complexity of the diagnosis operation are effectively reduced by using the algorithm based on the sequence of observed events.

ABSTRACT. The article is based on the ROS robot platform and introduces four SLAM algorithms, namely Gmapping, Hector, Cartographer, and Karto. It also introduces the global path planning algorithm and local path planning algorithm that cooperate with its navigation. The global path planning algorithm here uses Dijkstra and A *, while the local path planning algorithm uses DWA and TEB. After introducing the hardware equipment and experimental environment required for the implementation of this article's testing, the article utilized the above techniques to conduct multiple experimental tests in the actual application scenarios of this topic, accurately creating maps, completing navigation tests, and visualizing the test results on the RViz tool. By testing the robots using four different mapping methods and combining different global and local planning algorithms to obtain efficiency comparison results, the optimal combination for navigation can be obtained. Based on the experimental results, a basic service platform for SLAM and path planning based on ROS was developed, which provides further services for subsequent work.

ABSTRACT. Focus on the constructing diagnostic model and the problems of incomplete diagnosis caused by online undefined behavior, a clustering based diagnostic model construction method are proposed. Behaviors are obtained directly from the device data according to the clustering results, and then the behavior temporal sequence is constructed according to the time relationship and the diagnostic model is generated. At the same time, heuristics are added to the diagnostic model to speed up the diagnosis according to the information obtained when the clustering behavior is generated. In the online diagnosis process, once the undefined behavior is found, the model is reconstructed according to the classification and order relationship of the undefined behavior in the existing cluster, and the possible diagnostic interpretation is obtained.

ABSTRACT. Deep neural networks have been increasingly used in safety-critical systems as an important component. Most methods of verifying neural networks are difficult to apply to large neural networks because of the complexity of the neural network. Abstraction has proven to be an effective way to improve scalability. Through abstraction, we can obtain a smaller neural network of an original neural network, which mitigates the state-explosion problem of neural network verification. A challenging problem is that the output precision is negatively correlated with the degree of abstraction. Therefore, we propose an abstract approach that selects neuron pairs to merge dynamically with a heuristic strategy, which over-approximates the output. Over-approximation guarantees that if the property holds for the abstract network, then it also holds for the original one. Experimental results show that our method outperforms other state-of-the-art methods in terms of output precision.

ABSTRACT. We focus on the direct limits in the category with EMV-algebras as objects and EMV-homomorphisms as morphisms. We show that the direct limits exist in the category of EMV-algebras.

ABSTRACT. Multi-view discriminant analysis method (MvDA) is an effective supervised multi-view learning method. However, in practical multi-view learning application scenarios, there may be the problem that the sample label information from some viewpoints is unreliable. In the MvDA method, the information of all views is considered to have equal reliability. The data information in views containing unreliable label information may make it difficult for MvDA to learn accurate common features among data from different views. Therefore, in this paper, a weighted consistent multi-view discriminant analysis (WMvDA-VC) is proposed for the multi-view learning task in unreliable labeling environment. By considering the data characteristics under different views, the information from different views is assigned different reliability weights. And the weighted multi-view intra-class scatter matrix and the weighted multi-view inter-class scatter matrix are constructed. Using these two weighted scatter matrices, multiple projection matrices are learned. Then the data from different views of the same object are projected into the same common projection space. In the common subspace, dissimilar samples in different views are dispersed as much as possible and similar samples are aggregated as much as possible. The impact of data from views containing unreliable labels on the multi-view learning performance is reduced by adjusting the weights of the views with unreliable labels. The method is experimented on four widely used datasets. The experimental results show that the proposed WMvDA-VC exhibits excellent performance in the classification recognition task with unreliably labeled environment.

ABSTRACT. Arguments cannot be evaluated in isolation; their reliability hinges on the evidence that supports them. However, in some extensions to Dung's seminal work, the crucial role that evidence plays has not been adequately captured. To address this issue, this paper proposes an extension of Dung's abstract argumentation framework that includes a set of evidence and the concept of evidential support. Under this framework, arguments derive support from evidence and can then defeat their opponents. Moreover, the incremental semantics of the framework explicate the fallibility and cumulative nature of evidence.

ABSTRACT. Events occurrences examined through the spatio-temporal attributes of events are considered in many applications like digital forensics, storm occurrences, and forest fires. Constructing periodical timeslots or spatial hierarchies like country, state, region etc approaches are used to explore the knowledge about the event occurrences for the public security or safety purposes. Granular Computing approach is used with Formal Concept Analysis to discover the knowledge in such applications at different hierarchical levels of data. Existing approaches use spatiotemporal aspects of data about the events occurrences and re-occurrences through Granular Computing and binary or classical Formal Concept Analysis. However, Intuitionistic Fuzzy Sets three parameters µ, γ and π can be used to identify the events aspects and analysed through the Granular Computing approach and Formal Concept Analysis technique. Additionally, formal concepts formed through the binary or classical Formal Concept Analysis and Intuitionistic Fuzzy formal concepts formed through the Formal Concept Analysis operations on Intuitionistic Fuzzy Sets data are analysed through different Granular Computing measures. In this research article we discuss these Granular Computing measures used to analyse and interpret binary or classical formal concepts and Intionistic Fuzzy formal concepts of the formal contexts related to the events occurrences, non-occurrences and event indeterminacy of occurrences or non-occurrences. Moreover, these Granular Computing measures are explained through the examples of different events occurrences data.

ABSTRACT. Fuzzy c-means (i.e., FCM) is a representative clustering method that is widely used in machine learning and pattern recognition. It achieves clustering by constructing a membership matrix of objects to clusters and ends up with a specific but possibly wrong clustering result, i.e., an object is assigned to a single cluster which the object does not belong to. Motivated by such issues, in this paper, we propose a multi-view fuzzy clustering method based on FCM and a fuzzy assessment is introduced to assess the fuzziness of clustering results. Specifically, we constructed a multi-view FCM to obtain a reasonable membership matrix by fully integrating multi-view information. Based on this, we propose a fuzzy assessment method to evaluate the fuzziness of the clustering results and mark the objects where the clustering results are imprecise (i.e., the object is in the overlapping part of two clusters). The experimental result shows that the proposed multi-view fuzzy clustering method achieves better performance than several related clustering methods, including CGD, Co-FW-MVFCM, GMC, LMSC, TBGL-MVC and MCSSC.

ABSTRACT. With the frequent occurrence of security issues in open-source software, the public is gradually paying attention to the trustworthiness of open-source software. As the cradle of open-source software, the open-source community affects the quality and security of open-source software. Current researches about trustworthiness in open source communities mainly focus on single attribute, rather than conducting a comprehensive analysis from a macro perspective. This paper proposes an open-source community trustworthiness measurement model that combines multi-level fuzzy comprehensive evaluation method with trustworthiness attribute hierarchical model. The model includes five attributes: support, activeness, standardization, maturity, and community team. This study is the first to propose a comprehensive trustworthiness model for open-source community, reflecting the multi-dimensional attributes of open-source community. At the same time, We construct membership functions to optimize experts' evaluations of open-source community. Finally, this paper takes OpenEuler community as an empirical object and uses the fuzzy comprehensive evaluation model proposed in this paper to measure and verify the practicality of the method.

ABSTRACT. Fuzzy reasoning is based on fuzzy sets, and aims to describe things with fuzzy and uncertain concepts and convert them into information that computers can process. In the history of fuzzy inference development, Zadeh first gave the most basic inference rule in fuzzy inference theory, i.e., the fuzzy separation rule FMP, which was then algorithmized by Zadeh and Mamdani and others to produce various fuzzy inference methods mainly based on the inference synthesis rule CRI. Later, Wang thought that the inference algorithm based on CRI has certain defective nature and proposed the triple I algorithm with fully implication operator. Following that, the α-universal triple I algorithm was proposed. As an important part of the fuzzy inference algorithm, the implication operator directly affects the robustness of the inference algorithm, that is, affects the quality of the inference algorithm. Besides, the quality of fuzzy reasoning algorithm depends on the robustness of reasoning algorithm to a large extent. In this paper, the ordinary implication operator is extended to a more general interval implication operator. And the simple perturbation of the α-universal triple I algorithm is studied for some specific interval implication operators under the interval-valued environment. It is found that the α-universal triple I algorithm possesses good simple perturbation performance for single rule or multi-rules, in which specific interval implications are used.

ABSTRACT. In the field of intelligent education, knowledge tracing (KT), which can estimate and trace students' grasp of knowledge concepts to provide high-quality education, is drawing increasing attention. Recent studies use flexible deep neural network models for knowledge tracing task and have gain good results. However, most of the existing studies ignore the influence of difficulty factor and discrimination factor on knowledge tracing task. In fact, the different difficulty of the same knowledge concept and the difference in the degree of discrimination of the question also indicate the degree of students' mastery of knowledge. In this paper, we first incorporate question difficulty and question differentiation into the student answer record, and then propose a novel knowledge tracing model(called FDDKT) based on the question difficulty and discrimination embedding, so as to make more use of the rich information. We calculate the question difficulty and discrimination respectively through the existing datasets, and embed them into the representation, and then we integrate them through a linear layer to process the features. In addition, we use a monotonic attention mechanism to connect the learner's responses to evaluation questions at different time periods. In other words, the attention weight of the current problem on past problems depends not only on the similarity between query and key, but also on their relative temporal distance. Experiments on several real world benchmark datasets and show that FDDKT has good predictive performance, which demonstrates the difficulty of knowledge concept and discrimination of question are not ignored factors.

ABSTRACT. Heterogeneous multicore system integrated with FPGA has been shown to yield superior performance when executing real-time applications. In this paper, we propose a heterogeneous multicore co-scheduling algorithm based on multi-characteristic fuzzy cluster that utilizes fuzzy cluster to optimize task scheduling. Experimental results demonstrate that our algorithm achieves better scheduling results which is 8.85\% faster than ESHCS, 6.03\% faster than MARCO, and 9.56\% faster than HSPA. Our algorithm is suitable for applications with strong real-time requirements, such as environmental perception tasks in the field of autonomous driving.

ABSTRACT. Formal concept analysis (FCA), as a conceptual clustering method, is well suited to explain the meaning of clusters. However, most of the existing FCA-based concept clustering algorithms cannot handle fuzzy linguistic information. To solve this problem, in this work, we propose a conceptual clustering based on linguistic-valued layered concept lattice. is generated by the fuzzy linguistic coupling relationship between attributes. To obtain bi-clustering results for linguistic concepts, attribute-induced and object-induced matrices are first proposed. Second, the problem of overlapping clustering results is handled by calculating the stability of linguistic-valued layered concepts. In addition, the conceptual clustering results can be obtained at different granularities by varying the fuzzy linguistic-valued trust degrees to meet the individual needs of different users. Finally, illustration example is used to demonstrate the effectiveness of our proposed method.

ABSTRACT. The automated deduction mechanism based on contradiction separation is a novel deduction mechanism, which has some outstanding features, such as dynamicity, multi-clauses involving, and synergy. In order to better utilize the synergy between clauses of this state-of-art automated deduction mechanism, this paper proposes two types of strategies for clause and literal selection during the deduction process. The first one is clause selection strategy based on complementary pairs in unit clauses, which consider mainly the synergized effect between non-unit clause and unit clause. The second type consists of several literal and clause selection strategies based on distribution of complementary pairs and deduction distance, which consider mainly the synergized effect of non-unit clauses. We integrate the proposed strategies into CSE1.5, which is the latest version of contradiction separation based automated theorem prover, and denote the new prover as CSE1.5_CD. We then test CSE1.5_CD and CSE1.5 on the theorems of CADE ATP system competition (CASC-28) FOF division. CSE1.5_CD with the proposed strategies can solve 11 more theorems than that solved by CSE1.5. Experimental results show that the integration of the proposed strategies can enhance the capability of CSE1.5 to a certain extent.

ABSTRACT. In the process of stock price forecasting, there are the following problems: how to find the more effective factors for stock price forecasting, and how to calculate the weight of the constructed stock correlation factor sets. To solve the above problems, this paper proposes a method of factor construction in the field of stock price prediction based on genetic programming. The method can automatically construct the factor by reading the original data set of the stock, and calculate the weight of each factor. In addition, this paper also proposes a new crossover operator, which can dynamically adjust the selection of crossover nodes by using the information in the execution process of genetic programming algorithm, so as to improve the quality of the constructed factor set. A lot of experiments have been carried out with this method. The results show that the factors constructed by this method can improve the accuracy of the stock price prediction algorithm in most cases.

ABSTRACT. Fraud detection based on graph neural networks has received wide attention in recent years. In fraud detection, there exists a class imbalance problem because the number of fraudulent nodes is often much smaller than the number of normal nodes. Besides, fraudulent nodes usually camouflage themselves feature-wise and structure-wise, making the fraudulent information hidden from the normal information. This also introduces inconsistency to graphs. There must be edges connecting nodes with different labels, which violates the homophily assumption of vanilla GNNs. To address the above problems, we propose a Dual-Channel Graph Neural Network (DCGNN) for fraud detection. Firstly, we use the Class-balanced Node Sample Module for sampling so that the model can better learn the patterns of minority class nodes. Then, we design the Attribute-Structure Dual-Channel Hybrid Module so that the model adaptively combines the representations of two channels for detection. A graph disparity convolutional network is also introduced to model the dissimilarity between nodes to solve the inconsistency problem. Experiment results on the benchmark datasets demonstrate the effectiveness of our model.

ABSTRACT. In recent years, linked data has been introduced into the field of geographic information science, and has become an important way to publish and share geographic information on the network. In order to unify the expression form of spatial information in linked data, the Open Geospatial Alliance proposed GeoSPARQL standard to describe geospatial linked data. However, not all geospatial linked data published for spatial objects and spatial relationships are complete. The incompleteness of data is not conducive to maintenance and management, as well as the interaction between various datasets. At the same time, the published linked data may also have spatial semantic consistency problems (such as inconsistent topological relationships). In order to improve the quality of geospatial linked data, it is necessary to establish a spatial integrity constraint mechanism and an integrity verification method for linked data under this constraint. We first extended the definition of integrity constraints for spatial data type attributes and spatial object attributes, and proposed GeoSHACL+; Furthermore, the GeoSHACL+language oriented spatial constraint satisfiability determination method and the integrity verification method of geographical linked data under spatial constraints are given. Finally, we use SLIPO open source data validation to verify the effectiveness of our proposed method.

ABSTRACT. Non-negative Matrix Factorization (NMF) is a data dimensionality reduction method, which can process large scale and high dimensional data more efficiently. Among them, kernel-based non-negative matrix factorization is essentially NMF after mapping data to high-dimensional space, which is a nonlinear NMF method and is mainly used to extract high-order features of image data. The feature information obtained by different kernel functions has different characteristics and makes different contributions to image classification. However, most of the researches only consider single-kernel NMF, and some multi-kernel NMF methods also have some problems such as uneven weight distribution, which loses the feature information of some kernel functions. Therefore, it is necessary to study the NMF method of multi-kernel fusion with adaptive weights. In this paper, a non-negative matrix factorization method based on mixed Gaussian kernels is proposed. By fusing multiple Gaussian kernels with different weights and updating weights adaptively, the feature information of different Gaussian kernels can complement each other, which has good performance on public datasets FERET, ORL and Yale Face, indicating that the algorithm can effectively improve the classification performance.

ABSTRACT. In the healthcare field, predicting future diagnoses of patients based on their medical history is a critical task. Electronic health records (EHR) have facilitated the development of many deep-learning models for prediction in this field. However, two issues have not been addressed simultaneously, which can impact the accuracy of the prediction model. The first issue is that rare diseases have little opportunity to be learned during the training process, while the second issue is that short medical visit record sequences pose a challenge for sequential models. To address these challenges, we propose using a graph neural network to encode medical ontology and co-occurrence information into diagnosis representation. We also use a pre-trained vanilla prediction model to obtain similarities between visit contexts and extend visit records by referencing similar visit contexts in the training dataset. Our experiments show that the proposed model performs better than the current state-of-the-art in diagnostic prediction.

ABSTRACT. The migrativity is a very important property of uninorms. This paper continues to study migrativity for uninorms. In previous studies, this kind of migrativity for the most usual classes of uninorms over uninorms with continuous underlying operators is missing. In this paper, we get α-migrative property under some conditions for two uninorms, where a uninorm is one of the most usual classes of uninorms and the other is a uninorm with continuous underlying operators, they have different neutral elements. The solutions improve migrativity for uninorms.

ABSTRACT. The weighted quasi-arithmetic mean associated with a continuous triangular norm is proposed. It generalizes the quasi-arithmetic mean. By the means of numerical example, a new multi-attribute decision model is presented to explain the application of this weighted quasi-arithmetic mean.

ABSTRACT. The analysis of different algebraic properties is an important issue for studying fuzzy implications, as it contributes to locate appropriate implication functions more easily in applications. In approximate reasoning, it is of great significance to combine the invariance of fuzzy implications regarding the powers of t-norm T (T-power invariance for short) with other algebraic properties. Certain kinds of fuzzy implications satisfying T-power invariance with respect to continuous t-norms T have been proposed in recent years, with the relevant algebraic properties having been drilled. This paper, based upon the previous studies, aims to propose a new kind of fuzzy implications: nilpotent T-power invariant implications, through studying the structure of fuzzy implications which meet the T-power invariance with respect to the nilpotent t-norms T. As a supplement, the continuity and some other algebraic properties of such fuzzy implications will also be discussed.

ABSTRACT. In the fuzzy field, Zadeh proposed the concept of fuzzy sets and the CRI algorithm to deal with fuzzy reasoning. Later, Wang discovered the limitations of the CRI algorithm and proposed the triple I algorithm. After studying the triple I algorithm, Tang found that the harsh conditions of the triple identical implication operators lead to low system performance, so the α-universal triple I algorithm is proposed, and on this basis, the universal triple I chain algorithm is developed. In practical application, the quality of fuzzy inference algorithm directly depends on the robustness of inference algorithm. This study studies the interval robustness of the α-universal triple I algorithm under the interval-valued environment. The interval perturbation performance of the α-universal triple I algorithm is respectively studied for the R-interval implication, the triangular modulus interval implication, the S-interval implication and the QL-interval implication. It is found that the α-universal triple I algorithm has good interval perturbation performance for single rule or multi-rules, in which unified interval implications are employed.