40results about How to "Guaranteed interpretability" patented technology

The invention discloses a traffic zone dividing method based on sparse vehicle license identification data. The traffic zone dividing method comprises steps of: analyzing vehicle travel characteristics at various intersections from the vehicle license identification data, determining travel vehicle types, and constructing tensors indicative of intersection traffic conditions; analyzing a correlation among the travel vehicle types, introducing map POI data and social media user signature data, and complementing the sparse intersection traffic tensors by using cooperative tensor decomposition; and performing spatial clustering on a map on the basis of the intersection traffic tensors in order to divide the map into different traffic zones. By dividing the traffic zones, the method may reflect urban traffic conditions in a more visualized way and provides help for urban planning.

The invention, which relates to the cross technology application field of big data mining and traditional Chinese medicine digitalization, provides a traditional Chinese medicine clinical big data storage method based on multi-mark feature selection and classification. Information acquisition is carried out on a patient, quantified grading is carried out based on symptoms of different patients, and corresponding multi-type labels are given; the collected patient data are normalized, vectorization processing is carried out on a marking data set to obtain a standard multi-mark data set, and themulti-mark data set is divided into a training set and a testing set to verify the effectiveness of algorithm; a needed related matrix for feature selection of the training set is calculated and global optimal feature weight distribution is calculated by using a penalty function method; and first K feature sub sets with the largest weights are selected, a testing set prediction result is obtainedbased on an MLkNN method, and an optimal feature sub set is selected to carry out new patient disease prediction.

The embodiment of the invention discloses a knowledge reasoning method for a knowledge graph, which comprises the following steps: acquiring an initial knowledge graph, and generating an alternative rule according to the initial knowledge graph; judging the alternative rules, and determining qualified rules of which the confidence coefficients are greater than a set threshold value; and reasoning and complementing the initial knowledge graph according to a qualification rule, obtaining a new node and a corresponding relation, and adding the new node and the corresponding relation into a graph knowledge set. According to the knowledge reasoning method of the knowledge graph provided by the embodiment of the invention, a rule reasoning method and a graph neural network reasoning method are combined, so that a hybrid reasoning framework based on a generative adversarial model is formed, and the hybrid reasoning framework is designed to realize hybrid reasoning; the nodes and the relations are predicted by utilizing a link prediction method based on a hierarchical structure, and the advantages of different reasoning methods are combined, so that the generalization ability and the calculation efficiency of knowledge reasoning are improved, and the accuracy and the interpretability of a reasoning result are also ensured.

The invention relates to the white spirit making field and the image processing field, discloses an automatic detection method for fermented grain quality in the white spirit making process, and solves the problems that in the white spirit production process, the traditional technology depends on artificial experience in detection and evaluation of fermented grains, has large error and poor reproducibility. The method comprises the following steps of: A, establishing a fermented grain quality evaluation model based on an adaptive fuzzy inference algorithm; B, training a fermented grain qualityevaluation model based on fermented grain samples with known quality grading and corresponding grading results; and C, detecting to-be-detected fermented grains by adopting the trained fermented grain quality evaluation model. The method is suitable for automatic detection and evaluation of the fermented grain quality in a grain preparation link in a full-automatic white spirit making process.

The invention relates to a singer classification method based on a Labeled LDA model. The singer classification method comprises the steps of S1 collecting and preprocessing artificial tags of singers; S2 establishing a singer classification model based on user behaviors and collecting user behavior data; S3 cleaning the user behavior data, and filtering data unfavorable for model training in theuser behavior data; S4 distributing the weight of each singer corresponding to each user in the user behavior data; S5 combining the user behavior data with the artificial label data to generate training data; and S6 based on the training data, referring to the label combination relationship, and carrying out Labeled LDA model training based on optimized Gibbs sampling. According to the invention,the song playing behavior of the user is used as training data, the coverage of the user is high, the preference characteristics of each user group are considered, the change of the user behavior reflects the change of social hotspots and public cognition, the model can be periodically trained to change along with the change, the adaptability is strong, the precision degree is high, the label coverage rate is improved, and the classification is fine enough.

The invention discloses a data processing method and device, and relates to the technical field of computers. One specific embodiment of the method comprises the following steps: determining a training data set for training a data processing model, wherein the data processing model is used for carrying out data processing according to a plurality of strategies with priority sequences; and according to the sequence of the priorities of the strategies from high to low and a preset strategy combination mode, sequentially generating a decision tree corresponding to each strategy, and taking the decision tree corresponding to the strategy with the lowest priority as a data processing model. According to the embodiment, data processing can be carried out according to a plurality of strategies with priority sequences, so that the problem of single processing value of a traditional decision tree is solved by an analysis method based on target value disassembly and integration and a loss function corresponding to the analysis method, and data processing is more reasonable and scientific.

The invention relates to the field of artificial intelligence, and provides a course recommendation method and device based on root cause analysis, equipment and a medium, which can identify a label of each course in course data, construct a training sample training prediction model according to the label of each course and the course data, and according to the prediction model, an improved ID3 algorithm is adopted to calculate the information entropy of each label, the information gain of each label is calculated according to the information entropy of each label, a course recommendation list is generated according to the information gain of each label, and the ID3 algorithm is combined to perform root cause analysis on the influence of training courses on performance, so that interpretability and accuracy of an analysis result are ensured, and the training efficiency is improved. Training courses having great influence on performance are analyzed in an auxiliary manner, and then automatic recommendation of the courses is realized in combination with an artificial intelligence means, so that continuous tracking, performance improvement and retention are carried out, and training really assists team improvement and individual development. In addition, the invention also relates to a block chain technology, and the prediction model can be stored in a block chain node.

The invention provides a target object label processing method and device, equipment and a storage medium. The method comprises the steps of acquiring multiple word segmentation fragments corresponding to the target object; clustering the plurality of word segmentation fragments to obtain different word classes; determining labels corresponding to different word classes according to the preprocessed label information, wherein the preprocessed label information comprises a preset label input by a user, a seed keyword corresponding to the preset label input by the user and an extended keyword obtained according to the seed keyword, and the preset label represents the attribute characteristics of the seed keyword and the extended keyword; and determining the current label of the target objectaccording to the labels corresponding to the different word classes. According to the method, all the tags possibly involved in the segmented fragments can be fully found out, and explanatory performance of tag processing results is guaranteed; and besides, the workload of inputting keywords by a user can be reduced, and sufficient keywords are ensured, so that the accuracy of a label processingresult is ensured.

The invention discloses a continuous reference line decision-making method and device, and the method comprises the steps: determining the semantic information of a reference line, and enabling the determined semantic information to comprise a first semantic information group for intention decision-making and a second semantic information group for direction decision-making; determining a reference line switching intention according to the first semantic information group; selecting an optimal reference line according to the reference line switching intention and the second semantic information group of each reference line; and determining a target reference line according to the selected optimal reference line. According to the scheme, the reference line decision is performed by adopting the semantic information determined for the reference line, so that the reasonability and reliability of the decision result can be ensured; and part of semantic information is used for decision analysis in each stage of decision making, so that the decision making process is clear, visual and interpretable, evaluation parameters, namely the semantic information, of decision analysis can be flexibly adjusted according to requirements, and the debuggeability and expandability of reference line cost evaluation are improved.

The invention belongs to the related technical field of cutter state monitoring, and discloses a milling cutter wear prediction method based on an improved PCANet model, and the method comprises the following steps: (1) inputting a training set into the improved PCANet model for training; the improved PCANet model is divided into an APCANet-MP model and an SVR model, the APCANet-MP model is further divided into three stages, the data processing modes of the first stage and the second stage are completely consistent, and each stage comprises a preprocessing layer, a PCA convolution layer and an activation layer; the post-processing stage comprises a maximum pooling layer and an output layer; the SVR model adopts a linear kernel function; (2) evaluating the trained improved PCANet model by using an evaluation index, and optimizing structural parameters of the improved PCANet model according to an evaluation result; and (3) inputting signal data of a to-be-predicted milling cutter into the optimized improved PCANet model so as to predict the wear of the milling cutter. According to the invention, the parameter scale is reduced, and the analysis efficiency and prediction precision are improved.

The invention discloses a concept drift-oriented interpretable Android malicious software detection method, and belongs to the technical field of information security. The method comprises the following steps: introducing detection features through an artificial Android malicious software analysis report, improving a traditional feature package based on an automatic machine learning algorithm and an interpretable algorithm, and fusing an identically distributed inspection and transfer learning algorithm. According to the method, the interpretability of the Android malicious software detection model is improved, manual verification of the detection model by reverse analysts is facilitated, the influence of the concept drift problem on the accuracy of the detection model is reduced, low-cost long-time maintenance of high accuracy of the detection model is facilitated, and the method is used for detection and analysis of Android malicious application software.

The embodiment of the invention provides a decision tree model construction method and device, electronic equipment and a medium, and the method comprises the steps: constructing a word bag model by using a training text; establishing a first decision tree model according to the first characteristic value of each answer text included in the word bag model and an answer score label set for each answer text, and obtaining an importance degree value of the word characteristic of each answer text output by the first decision tree model; according to the importance degree value of the word featureof each answer text, obtaining an answer text, and screening out keyword features meeting preset conditions from the word features of the answer texts, and establishing a second decision tree model for answer score prediction according to the second feature values of the answer texts obtained by the keyword features and the answer score tags set for the answer texts. With the adoption of the application, the scoring prediction precision can be improved, and meanwhile, the interpretability of the model is ensured.

The invention provides a personal credit evaluation and explanation method and device based on time sequence attribution analysis, equipment and a storage medium. The method comprises the steps of creating a credit scoring model; respectively training the credit scoring model by utilizing a plurality of groups of historical credit investigation data sets with time labels to obtain a plurality of historical credit scoring models; predicting a plurality of future credit scoring models based on a plurality of historical credit scoring models with time labels or a plurality of groups of historicalcredit investigation data sets with time labels according to the types of the credit scoring models; inputting to-be-assessed credit investigation data into the selected historical credit scoring model or future credit scoring model to obtain a credit investigation assessment result of the credit investigation subject corresponding to the to-be-assessed credit investigation data; and explaining the credit investigation evaluation result. According to the invention, a series of credit scoring models are constructed for a plurality of historical time points and a plurality of future time points, credit assessment of a credit investigation subject at a specific time point can be realized by selecting a proper credit scoring model, and explanation with reference value is made for an assessment result, so that personal credit scores are guided to be improved.

The invention provides a user decision behavior quantitative analysis method and device. The invention relates to the field of big data calculation and artificial intelligence in the computer technology. At least one quantized decision factor related to a target decision made by a user is input into a machine learning model, the decision factors are further analyzed by the machine learning model,and finally, a prediction result of the target decision made by the user is output according to the output of the machine learning model. Therefore, the decision factors of the target decision made bythe user can be analyzed to obtain the prediction result of the making of the target decision, and therefore, analysis requirements on user decision behaviors are enriched.

The invention discloses an electricity consumption transaction analysis method based on financial technical indexes, and relates to the field of electricity consumption transaction analysis. At present, the abnormal analysis of the power consumption data is mostly limited to mining the abnormal condition of the power consumption data, and the abnormal condition caused by enterprise production transformation, economic operation and the like is difficult to capture. Through the characteristics of an electrical load sequence are described through indexes of a cross filtering line, a zero-lag similarities and differences moving average line and a Brinell line, the index parameters are adaptively set to meet specific scenes of an abnormal change analysis task, a reasonable threshold value is set according to an index probability distribution diagram in the same industry, and the indexes exceeding the index threshold value are qualitatively determined as transaction; therefore, the load curve characteristics are effectively described, the abnormal phenomenon is effectively grabbed, and it is guaranteed that an abnormal analysis result has certain interpretability. Significant abnormal changes of the power utilization data can be accurately and effectively screened out; the detection of the power utilization mutation is captured through an auxiliary Brinell line index, thereby achieving the capturing of the abnormal movement, and further guaranteeing the effectiveness of the abnormal movement analysis.

The invention relates to a deep integrated learning model construction method for malicious WebShell detection, and the method comprises the following steps: acquiring normal samples and malicious WebShell samples, randomly disorganizing the normal samples and the malicious WebShell samples, and dividing the normal samples and the malicious WebShell samples into a training set and a test set according to a proportion of 4: 1; extracting dynamic features and static features of each sample in the training set and the test set, and combining the dynamic features and the static features to obtain a training set feature set and a test set feature set; selecting m base classifiers, and training a deep integrated learning model by using the training set feature set and a K-fold cross validation method to obtain a weight value of each base classifier; and performing model testing by using the test set feature set as the input of the deep integrated learning model to obtain a comprehensive average prediction probability value of a test set sample, and evaluating the deep integrated learning model. The method is based on deep ensemble learning, the detection rate can be improved, and the advantages between machine learning and deep learning can be perfectly absorbed.

The invention relates to a method and device for identifying an IP gang, a medium and equipment, and the method for identifying the IP gang comprises the steps of obtaining a URL accessed by each IP and N behavior characteristics based on log data in a preset time period; aggregating all the IPs into different clusters based on the URL accessed by each IP and the N behavior characteristics; and when the cluster satisfies a preset condition, determining that the user corresponding to the IP in the cluster is an IP gang. According to the invention, more accurate gang clustering is realized for user behaviors of web logs, users with similar accessed URLs and similar access behaviors within a period of time are aggregated together, the accuracy of a clustering result is ensured, the method is also effective for low-frequency gangs, and the interpretability and flexibility of an identification result are ensured through specific rule parameters.

The embodiment of the application discloses a method and device for processing goods and steel scrap. According to the embodiment of the present application, the cargo image of the single-layer cargo in the overall cargo is obtained, and the cargo image of the single-layer cargo is identified by cargo recognition, and the single-layer quality data of the single-layer cargo obtained thereby is used as the basis for determining the quality data of the overall cargo. According to the basis, not only the automatic identification of the quality of the whole vehicle is realized, but compared with only obtaining the cargo image of the whole cargo or the image of a partial angle, the solution of this application can fully obtain the quality data corresponding to more sufficient and comprehensive details, which is obviously possible Greatly improve the accuracy of overall cargo quality judgment. Furthermore, this application also determines the weight of the goods to be deducted that do not meet the quality conditions in the overall goods according to the overall quality data of the overall goods, and provides a theoretical basis for deducting miscellaneous items and weight of goods, ensuring that the entire estimation plan plausibility and explainability.

The invention discloses a score card model construction method, device and equipment and a computer readable storage medium, and the method comprises: obtaining the credit behavior data of a customer, taking the credit behavior data as training data, and carrying out the training of a GBDT model based on the training data, wherein the GBDT model comprises a plurality of decision trees; when each decision-making tree in the GBDT model is trained, determining undetermined abnormal nodes in non-leaf nodes of the decision-making trees, verifying the undetermined abnormal nodes, and determining abnormal nodes in the undetermined abnormal nodes; and based on the abnormal node, retraining the GBDT model, and obtaining a corresponding score card model after the GBDT model is trained. The GBDT model is optimized by controlling the internal structure of the GBDT model, so that the score card model not only keeps the excellent effect of the GBDT model, but also ensures the interpretability of the model.

The invention discloses an electrocardio vector reconstruction method based on unsupervised learning. The method carries out the reconstruction of an electrocardio vector of an inputted standard 12-lead electrocardiogram through employing a neural network. In the training process, a method of firstly mapping the standard 12 lead to the electrocardiogram vector and then restoring the 12 lead electrocardiogram by using a projection method is used, so that the problem that the traditional method depends on the corresponding data of the 12 lead and the electrocardiogram vector is solved, the utilization efficiency of the data is obviously improved, and the data cost is reduced. During reconstruction, a neural network is used for recalculating a projection vector to perform reconstruction froman electrocardiogram vector to a 12-lead electrocardiogram, and a regularization term is used for constraining the projection vector in a final loss calculation module, so that the interpretability and the accuracy of a reconstruction process are ensured while the individual difference of the electrocardiogram is solved. The multi-order differential loss is used in a final loss calculation module,so that the problems of low frequency, such as baseline interference and the like, are avoided on the basis of ensuring morphological characteristics.

The embodiment of the invention discloses an interference source positioning processing method and device, and the method comprises the steps: carrying out the conversion of the first position information of each sample point according to a preset coordinate system, and obtaining the corresponding second position information under the preset coordinate system; performing clustering analysis on thesecond position information to obtain an interference center point of each interference source; dividing the preset search range with each interference center point as the center of a circle into regions, calculating the value of each region, and selecting the region with the maximum value as the value region of the current interference center point; and calculating position information of each interference source according to the sample points and the propagation model in the value area of each interference center point. The time cost and the labor cost are saved; the method and device can well work in a place with directional interference or dispersed base station distribution, and has high robustness; the height of an interference signal and the real position of an interference sourcecan be calculated; and meanwhile, a propagation model of the signal in a three-dimensional space is used, so that rigorous and interpretable results are ensured.

The invention provides a data enhancement method based on a data dimension reduction process. The method is mainly applied to the technical field of information, and comprises: firstly, constructing a data set into a matrix for subsequent data processing operation; when a CUR matrix decomposition method is used for carrying out row selection and column selection on a matrix, using a PCA dimension reduction method for extracting eigenvectors of rows and columns of the matrix, determining correlation according to the eigenvectors and correlation coefficients of the row vectors and the column vectors so as to select the rows and the columns which are high in representativeness to construct the matrix. Meanwhile, compared with a recovery matrix obtained by a conventional CUR, the recovery matrix obtained by the method provided by the invention has the advantages that the error between the recovery matrix obtained by the method and an actual original matrix is smaller, and the dimension reduction result is more accurate.

The embodiment of the invention relates to a pain volatility feature selection method and device, a storage medium and equipment, and the method obtains pain volatility data through adding a time interval variable in the processing process of set data, enables the improved pain volatility data to be affected by a time interval, and improves the accuracy of the pain volatility feature selection, the change condition of the pain severity along with the time can be reflected more accurately; feature selection is performed on a feature result output by the LASSO regression prediction model based on an LASSO regression mode of logistic regression to obtain first feature data, and the first feature data is integrated with second feature data, third feature data and fourth feature data which are subjected to feature selection with a feature result output by the random forest prediction model; according to the method, the selected result features have representativeness and universality, the interpretability of the patient pain prediction model is guaranteed, high accuracy of the prediction result is still kept, and the problem that feature selection in an existing patient pain prediction model lacks representativeness and universality is solved.

The invention discloses a partial discharge mode identification method of a deep super learning machine (DSL) in combination with evidence discount. The method mainly comprises the following steps: (S1) signal acquisition: collecting a partial discharge signal through a sensor; (S2) data preprocessing: carrying out denoising and other processing on the original signal; (S3) carrying out primary diagnosis on the partial discharge signal by the deep super learning machine, and constructing basic probability assignment (BPA) of a plurality of evidence bodies; (S4) evidence discounting: performingevidence discounting on the BPA of each evidence body constructed in the step (S3); (S5) constructing a new feature: adding the evidence discounted in the step (S4) to original feature data, and sending the original feature data as a new feature vector to a deep super learning machine for next iteration; and (S6) repeating the step (S3) and the step (S5) until the loss function is not reduced anymore, and obtaining a final pattern recognition result output by the deep super learning machine. Compared with the prior art, the method has the advantages that the identification accuracy is higher; and the method is high in result interpretability, and has a wide market prospect and application value.

The invention discloses a group-oriented rental house multi-dimensional identification method based on a knowledge graph and principal component analysis. The method comprises the following steps: 1, constructing a human-house knowledge graph; 2, calculating the weight of each judgment index based on principal component analysis; 3, calculating a group-oriented rental house judgment threshold value; and 4, identifying group-oriented rental houses. A to-be-checked house is judged based on a judgment formula and a judgment threshold, and if the to-be-checked house is judged to be a group-rented house, the human-house knowledge graph is queried to mine multi-dimensional data of the to-be-checked house as a judgment basis, so that the interpretability of a judgment result is ensured.

The invention relates to the technical field of data processing. The embodiment of the invention provides an entity extraction-based job record generation method, device and equipment. The entity extraction-based job record generation method comprises the following steps of: obtaining brief introduction text data; performing data labeling on the brief introduction text data by adopting a named entity recognition model to obtain a job company entity and a company position entity in the brief introduction text data, wherein the named entity recognition model comprises a bidirectional long-short-term memory network and a conditional random field network, and the named entity recognition model is obtained by training a marked training sample; and combining the job company entity and the company position entity to generate a job record in a preset format. According to the embodiment of the invention, the interpretability of the generated job record can be improved.

The invention is suitable for the technical field of image processing, and provides a low-illumination image enhancement model and method, an electronic device and a storage medium, the low-illumination image enhancement model comprises an initialization module, an optimization module, an illumination adjustment module and an image reconstruction module which are connected in sequence, and the initialization module is used for performing initialization decomposition on an input image; the initialization module is used for initializing the illumination layer and the reflection layer to obtain an initialized illumination layer and an initialized reflection layer, the optimization module is used for carrying out alternative iteration optimization on the initialized illumination layer and the initialized reflection layer by adopting a unfolding algorithm to obtain an optimized illumination layer and an optimized reflection layer, and the illumination adjustment module is used for carrying out illumination adjustment on the optimized illumination layer to obtain a target illumination layer. And the image reconstruction module is used for carrying out image reconstruction according to the target illumination layer and the optimized reflection layer to obtain a target illumination image, so that the flexibility and the interpretability of the low-illumination image enhancement model are ensured, and the robustness of the low-illumination image enhancement model is improved.

The invention discloses an automatic feature online processing method for log type data, which comprises the following steps: acquiring original log type data of a target object, and performing feature extraction on the original log type data to obtain one or more original feature columns; performing feature processing on the one or more original feature columns to obtain one or more target feature columns; configuring one or more constraint rules; configuring the one or more constraint rules into one or more composite conditions through a rule tree; and performing feature screening on the oneor more target feature columns by utilizing the one or more composite conditions. The device can replace an automatic feature processing product system of artificial feature engineering. According tothe invention, the efficiency of log type data artificial feature engineering can be improved, it is guaranteed that the processed features have interpretability, and the processed data can be directly used for model training.

The invention discloses a method for stall monitoring of a wind generating set. The method comprises the following steps that the running state data and guarantee power curve data of the wind generating set are acquired, and a wind speed interval is selected according to the actual running condition of the set; a guarantee power curve trend function in the selected wind speed interval is solved; operation state data of the wind generating set in the selected wind speed interval are acquired, and edge data positioning is performed; stall boundary preliminary solving is conducted based on the edge data; fine solution on the stall boundary is conducted aiming at the stall of the wind generating set under different conditions; and stall discrimination is performed according to the stall boundary. According to the technical scheme, in the stall boundary construction process, the output trend of the unit is considered, and the actual operation data of the unit are combined, so that the method has better flexibility and reliability, and compared with a single data monitoring method, the method takes data distribution monitoring as a main body and carries out verification through stall related physical quantity monitoring. And the effectiveness and the interpretability of the stall monitoring results are ensured.