181results about How to "Good explainability" patented technology

The invention provides a fine-grained word representation model-based sequence labeling model, which is used for performing a sequence labeling task, and belongs to the field of computer application and natural language processing. The structure of the model is mainly composed of three parts including a feature representation layer, a BiLSTM layer and a CRF layer. When the sequence labeling task is performed by utilizing the model, firstly an attention mechanism-based character level word representation model Finger is proposed for fusing morphological information and character information ofwords; secondly the Finger and a BiLSTM-CRF model finish the sequence labeling task jointly; and finally a result with F1 of 91.09% is obtained in a CoNLL 2003 data set in end-to-end and no any feature engineering forms by a method. An experiment shows that the designed Finger model remarkably improves the recall rate of a sequence labeling system, so that the model identification capability is remarkably improved.

The invention discloses a multi-strategy fusion knowledge question and answer method and system, and the method comprises an offline part and an online part, the offline part is mainly used for data preparation and model training, and the online part is mainly used for system service, and the method comprises the steps: receiving a statement input by a user, and correcting a spelling error; Performing word segmentation and part-of-speech tagging on the statement input by the user; Extracting entity information in the user statement, and linking the entity to the knowledge graph node; Obtainingan executable query statement through a multi-strategy fusion semantic analysis step according to the result of the entity recognition and connection process; Executing a query on a knowledge graph by the executable query statement to obtain an answer, and then generating a corresponding natural language reply user according to the answer through a reply generation mode combined by multiple methods, so that the question and answer system is suitable for question query of general and domain knowledge graphs, the system robustness is improved, and meanwhile, the good interpretability and controllability are achieved.

The invention discloses a heterogeneous information network based content providing method. When users subscribe to a recommendation service, content is recommended to the users through an optimal prediction matrix, a similarity matrix of the users and a similarity matrix of projects are obtained by an element path based similarity calculation method according to a heterogeneous information network through the optimal prediction matrix, the fusion is performed on two or three of the user similarity matrix, a user and project evaluation matrix and the project similarity matrix to obtain the internal relation between the users and the projects, the prediction is performed by a collaborative filtering based matrix decomposition prediction method, and results are combined to obtain the optimal prediction matrix. According to the heterogeneous information network based content providing method, the recommendation accuracy is effectively improved, the results which are accord with the will of the users can be recommended to the users through limited time calculation on the basis of the existing data, the cold start problem is partially solved, and the interpretability of the recommended results is improved.

The invention discloses a collaborative filtering recommendation algorithm based on a graph convolution attention mechanism. The method comprises the steps of firstly collecting and processing data and dividing a data set, secondly constructing a GACF model, and finally training the model and recommending by predicting an association score between a user and an item. According to the graph convolution attention mechanism collaborative filtering model provided by the invention, firstly, user-project interaction information is mapped to a vector space by using a graph embedding technology, then, the embedding expression of a user-project interaction graph is learned through a graph convolution network, and then, different weights are allocated to neighbor nodes by using an attention mechanism. By aggregating the feature information of the neighbor nodes, the weight between the neighbor nodes can only depend on the feature expression between the nodes, so that the generalization ability of the model is improved, and finally, a plurality of embedded vectors learned by a graph convolution layer are weighted and aggregated to obtain the association score between the user and the project.

The embodiment of the invention provides a user account abnormity detection method based on time sequence characteristics, and the method comprises the steps: building a time sequence abnormity detection algorithm model with a plurality of characteristic dimensions according to the historical record data of a user account; Obtaining current data of a user account, and according to the plurality ofanomaly detection algorithm models, respectively carrying out anomaly detection on the current data from the plurality of feature dimensions to obtain scores of the current data in the plurality of feature dimensions; Taking the average value of the scores of the plurality of feature dimensions as a comprehensive abnormal score of the current data; And when the comprehensive abnormal score is greater than a preset score threshold, judging that the user account is an abnormal user account. According to the technical scheme, dynamic behavior changes of a user are captured through user account evaluation based on a time sequence model and abnormal feature set matching based on association rule analysis, and the accuracy and interpretability of abnormal account detection in practical application are improved.

The invention discloses a fine-grained action detection method of a convolutional neural network based on multistage condition influence. The method comprises the steps of: establishing a convolutional neural network influenced by multistage conditions; fusing the explicit knowledge added in the visual scene with the multi-level visual features; enabling the multi-level conditional influence convolutional neural network MLCNet to take a conditional influence multi-branch convolutional neural network structure as a main trunk, generating multi-level visual features, encoding additional spatialsemantic information of human body structure and object context information as a condition, dynamically influencing feature extraction of a CNN through affine transformation and an attention mechanism, and finally fusing and modulating multi-mode features to distinguish various interactive actions; and carrying out model training on the convolutional neural network influenced by the multi-level condition, and outputting a fine-grained action detection result by the obtained model. According to the method, the proposed method is evaluated on the basis of two most common references, namely HICO-DET and V-COCO, and experimental results show that the method is superior to the existing method.

The invention realizes a set of differential privacy recommendation method based on heterogeneous information network embedding. The differential privacy recommendation method comprises the followingfour steps of: performing network representation learning by using HAN, and calculating heterogeneous attention sensitivity by using characterizations of HAN and an attention weight result; based on adifferential privacy definition, using the heterogeneous attention sensitivity to generate corresponding random noise, and generating a random noise matrix through using a heterogeneous attention random disturbance mechanism; constructing an objective function of differential privacy recommendation embedded with heterogeneous information for learning to obtain a prediction score matrix; and outputting the score matrix as a prediction score capable of keeping privacy. Therefore, the original scoring data is protected for the recommendation system scene under the heterogeneous information network, an attacker is prevented from improving the reasoning attack capability by utilizing the heterogeneous information network data acquired by other channels, and the original scoring data can be guessed or learned again with high probability by observing the recommendation result change of the score.

The invention relates to an immunohistochemical PD-L1 membrane staining pathological section image processing method, device and equipment. The image processing method comprises the following steps: acquiring a digital section full-field image of a to-be-diagnosed immunohistochemical PD-L1 (SP263) membrane staining pathological section; adopting a region segmentation network to identify and segment a tumor cell region in the digital slice full-field image under the first visual field multiplying power to obtain a tumor cell region probability graph of the whole digital slice full-field image;identifying and segmenting cells in each digital slice full-field graph, taking the tumor cell region probability graph as a weight matrix to carry out region constraint on the cell positioning network, identifying cell characteristics on the digital slice full-field graph, and positioning and classifying various cells on the digital slice full-field graph; and marking the cell position, the celltype and the immunohistochemical PD-L1 (SP263) index on the full-field diagram of the digital slice. By designing a multi-level feature collaborative diagnosis strategy, the tumor proportion score isaccurately evaluated in a mode of constraining cell features by using regional features.

The invention belongs to the field of data mining, and relates to an information interactive network-based criminal individual recognition method. The method comprises the following steps of: (1) obtaining a data set which comprises criminal activity contents, and pre-processing the data set; (2) extracting keyword descriptions of criminal topics; (3) determining the number of subjects of a subject model LDA on the basis of perplexity; (4) extracting interactive subjects between individuals in the pre-processed data set on the basis of LDA, wherein the interactive subjects are as follows: an association probability matrix of the interactive subjects and keywords, and an association probability matrix of interactive edges of the interactive subjects; (5) calculating weights of the interactive edges; (6) calculating local criminal suspects of the individuals on the basis of a structure of a weighted information interactive network; and (7) calculating global criminal suspects of the individuals on the basis of a fuzzy K-means cluster and distance density cluster combined method, and recognizing the criminal individuals. The method is independent of prior information, and can be used for analyzing the most possible suspected person according to communication contents so that the case handling efficiency is improved.

The invention discloses an industrial process soft measurement method based on an xgboost model. The method firstly performs independent re-sampling on historical data, obtains a training sample set and a verification data set respectively after preprocessing, establishes an xgboost model by using training samples, and then selects the best model parameters by cross validation to determine the soft-measurement model for the target variable. Compared with other nonlinear models, the xgboost model may greatly improve the accuracy and speed of soft measurement, and better fit the variable relationship in complex production processes.

The invention discloses a robust tensor maintenance based child community-acquired pneumonia (CAP) data processing system and method. The system comprises a CAP electronic medical record system, a data preprocessing module and an etiologic analysis module, wherein the CAP electronic medical record system extracts original data of a CAP child patient from an electronic health record system by using an SQL and outputs the original data to the data preprocessing module; the data preprocessing module performs calculations such as data cleaning, format conversion, rule check and the like, and outputs standardized child CAP data to the etiologic analysis module; and the etiologic analysis module performs robust tensor analysis to obtain a child CAP pathologic data model. The system and method place emphasis on data cleaning and mining in the existing electronic health record system, realize data standardization and etiologic analysis based on a robust tensor maintenance algorithm, and can provide decision support for child respiratory physicians to perform diagnosis and drug use scheme selection for patients.

The invention discloses a multi-relationship collaborative filtering recommendation method based on a dynamic graph attention network. The method comprises the following steps: S1, performing data acquisition and processing; S2, dividing a data set; S3, constructing a fusion model; and S4, model training and project recommendation. According to the invention, a recurrent neural network (RNN) is used for modeling behaviors of a user in a session, the current interest of the user is captured through RNN potential representation, the influence of friends related to the user is captured through agraph attention network, the influence of each friend is weighed by measuring the characteristics of movement along each side according to an attention mechanism, and the current user representation and the social friend representation are combined; the project relationship is obtained from the interaction data of the user and the project, the project relationship and the user dynamic social relationship are fused into the learning process of the user and the project interaction, the influence of multiple relationships on the user and the project interaction is learned, and the recommendationaccuracy is improved, so that the model can better model the user preference.

The invention discloses a cancer targeted marker mapping method based on a coexpression network. The method comprises the following steps: 1) building a coexpression basic network, according to gene expression data of a feature gene, calculating an adjacent matrix and a topology matrix; 2) extracting features of the coexpression basic network, namely converting each gene node of the topology network to a feature vector which is used as a feature value of the network; 3) training a neural network model, according to a migration sequence, executing the training of neutral network model parameters; and 4) executing the cancer targeted marker mapping, according to a clustering center self-adaptive algorithm based on a density peak, executing the automatic discovery of a targeted gene community. The provided method has the good universality and precision, and is capable of realizing the target gene mapping by using the building of the coexpression basic network, the extraction of the node feature vector and the automatic discovery of the gene community.

The invention provides a video cover image extraction method and device, an electronic device and a storage medium, and relates to the technical field of artificial intelligence. The video cover imageextraction method comprises the following steps: carrying out wonderful degree evaluation on framed images of a target video to obtain a first metric value used for representing the wonderful degreeof each framed image; carrying out wonderful segment identification on the target video to obtain wonderful segments in the target video and second metric values used for representing wonderful degrees of the wonderful segments; for each framed image, calculating a third metric value of the framed image according to the first metric value of the framed image and the second metric value of the wonderful fragment to which the framed image belongs; and extracting a cover image from the video according to the third metric value of each framed image. According to the invention, the accuracy and stability of video cover image extraction can be improved.

The invention provides a multi-class pneumonia screening deep learning system based on CT images. The method comprises the following steps of a CT image data preprocessing part preprocesses a CT imageto obtain a preprocessed CT image; the slice-level pneumonia dichotomy part analyzes the basic depth features to obtain a dichotomy result, the weakly supervised lesion positioning part obtains a lesion position map according to the class labels, and the slice-level pneumonia four-class part detects four classes of pneumonia to obtain a four-class pneumonia classification task result; the case-level pneumonia classification part performs analysis to obtain a multi-class pneumonia classification task result; and the pneumonia diagnosis comprehensive evaluation part outputs a pneumonia diagnosis comprehensive evaluation result and a focus positioning distribution map based on the slice-level four-classification result, the focus position map and the case-level multi-classification pneumoniaclassification task result. The method is advantaged in that the COVID-19 and other pneumonia diseases can be quickly and accurately distinguished on the premise of no medical staff besides obtainingthe common pneumonia diseases, which is helpful for screening the epidemic situation of the COVID-19; the COVID-19 and other pneumonia diseases can be quickly and accurately distinguished.

The invention provides a man-machine conversation method, a man-machine conversation device, a storage medium and a computer program product. The method comprises the steps of determining a current conversation theme and current utterance information of a user; determining a current utterance representation vector of the user according to the current utterance information; in combination with thecurrent utterance information and the current utterance representation vector, performing graph reasoning calculation on the heterogeneous knowledge graph corresponding to the current dialogue theme,and selecting current knowledge corresponding to the current utterance information from the heterogeneous knowledge graph; acquiring current utterance information according to the current utterance information and current knowledge; generating a reply statement corresponding to the current statement, wherein the heterogeneous knowledge graph is created on the basis of structured knowledge and unstructured knowledge and can generate reply statements with rich contents. In addition, the accuracy of knowledge selection can be improved by adopting a graph reasoning algorithm, so that the knowledgeselection process has very good interpretability and generalization ability. Meanwhile, the dependence of the whole scheme on corpora with labels is reduced.

The invention relates to an uncertain data provenance query processing method based on D-S evidence theory. The method comprises the following steps of: taking selection, projection and connection query operation related to an uncertain data table as a representative, acquiring elementary probability assignment of each input data item to a result data item from a provenance expression which describes SPJ query operation; based on an evidence combining rule in the D-S evidence theory, calculating the combined influence of the uncertainty of a plurality of input data items on the uncertainty of each result data item, and acquiring the probability assignment of each result data item; and performing standardization according to the probability assignment of each result data item, and calculating the belief value and the likelihood value of each result data item, so that the uncertainty of the result data item is determined, and if the uncertainty of the result data item accords with the result obtained on the basis of an input uncertain data-based probable world example, demonstration and evaluation can be performed on the basis of the pair of provenance query results.

The invention discloses a fan gearbox performance detection and health evaluation method based on temperature parameters. According to the method, fan state detection data are analyzed, the operationcharacteristics of a wind turbine generator unit are combined, state detection and fault prediction are carried out on the temperature of a gearbox and the temperature of a gearbox bearing by adoptinga model established by combining KECA and GRNN, an evaluation decision is performed on the running state of the fan gearbox through ideas of threshold discrimination and statistical process control according to predicted residual error, and a gearbox health state result is obtained by adopting comprehensive fuzzy discrimination. When the method is applied to an actual wind turbine generator unitcase with small sample data, the health state of the wind turbine generator unit can be accurately predicted and evaluated. The method is applied to an actual wind turbine generator unit with small sample data, and the health state of the wind turbine generator unit can be accurately predicted and evaluated.

The invention discloses a social media friend recommendation method based on mixing of blog articles and user relationships. According to the method, user preferences are dug in users' Weibo text dataaccording to an LDA theme model, similarity of the users' blog articles is calculated, importance of Weibo social relations is considered, similarly of social relations among users is calculated, andthe comprehensive similarity of the users is finally acquired. Most of ordinary Weibo users have few information for digging, and have simple and reliable relations; few users have many blog articleson main pages, sufficient text information for digging and complex social relations, the number of followed is much more than the number of following, and there are many useless noise data in their social relations. The social media friend recommendation method has the advantages that influences of two different attribute messages on recommendation results are measured through linear weighting, user recommendation lists are finally acquired through experimental learning of weight parameters, and quality of recommendation results is improved.

The embodiment of the invention discloses a method and device for predicting stopping accuracy in the precise stopping stage during ATO control. The method comprises the steps that according to recorddata, corresponding to a first running process of a target train, in various sub-systems of a train control system, target record data of the precise stopping stage during ATO control is extracted; characteristic values of all characteristics in a characteristic element set corresponding to a running path are calculated through the target record data, and a first characteristic value combinationis obtained; the first characteristic value combination serves as an input parameter of a pre-trained prediction model, the stopping accuracy of the target train in the first running process is predicted through the prediction model. The prediction model is obtained through machine learning and training, automatic prediction of the stopping accuracy is achieved, and dependence on work experience is got rid of. The prediction model is obtained through training of a lot of training samples, and the prediction model has a reliable data foundation, so that the prediction result is high in interpretability and reliability.

The invention belongs to the field of computational neurosciences and intelligent robots, particularly relates to an intelligent robot control method, system and device based on a biological neural network, and aims to solve the problem that an existing hippocampus heuristic pulse biological neural network model cannot realize complex mode control of an intelligent robot. The method comprises thefollowing steps: constructing a feature extraction and feature association learning neural network based on the structure and function heuristics of a biological hippocampus sub-region, and processingan intelligent robot environment image; obtaining the behavior category of the intelligent robot based on the obtained feature vector by adopting a classification neural network; and obtaining an intelligent robot control command through the intelligent robot behavior category-control command relationship. According to the method, the anti-noise performance of the biological neural network modelis greatly improved, the correct rate of object sensing under a complex background and object recognition under a high-noise background is improved, the robustness of the network model is good, and aneffective decision-making method is provided for NAO intelligent robot control.

The invention discloses a recommendation system and method based on user and project coupling relationship analysis. The method comprises the steps of data acquisition and processing, data set division, coupling model and training model construction and project recommendation. According to the method, the very microscopic coupling relationship between the user characteristics and the project characteristics is considered, and when the scoring information is sparse, the coupling relationship can recommend favorite projects to the user, so that the recommendation quality is improved; and an Attention mechanism is adopted to capture preference degrees of the user for different features of the project, so that the recommendation system has a better recommendation effect and interpretability. Moreover, the explicit features of the user and the project are extracted from the comment text by utilizing Doc2vec; the dimension of the user/project explicit features is reduced, the model operationspeed is increased, the recommendation accuracy is improved, and compared with matrix decomposition, the nonlinearity of the convolutional neural network and the deep neural network adopted by the method is beneficial to interaction between learning features at a deeper level.

The invention discloses a Web service quality prediction method based on random walk. The Web service quality prediction method comprises the following steps: on the basis of a user position, calculating a distance between users; selecting adjacent users for each user to form a neighborhood, and connecting neighborhood users to construct a user relationship network; calculating a similarity between the users and the similarity between services in the user relationship network; calculating the weight of each edge in the user relationship network; aiming at a target service requested by a source user to independently look up the random walk about the target service i for multiple times from the source user, randomly selecting a next-hop node by each hop in each-time random walk according to a certain probability, returning a QoS value after each-time walk is finished; and synthesizing all QoS values to calculate a quality prediction value, which is relative to the target service, of the source user. The Web service quality prediction method can balance coverage rate and the prediction precision of Web service quality prediction and has the advantages of being accurate in prediction, high in success rate, good in universality and favorable in evaluable confidence coefficient and interpretability.