Model training method and device, target detection method and device, equipment and medium

Abstract

The invention discloses a model training method and device, a target detection method and device, equipment and a medium, which are used for improving the performance of a detection model. According to the method, an attribute enhancement graph and a structure enhancement graph are constructed based on multiple views; based on the graph features with the label nodes in the attribute enhancement graph and the graph features with the label nodes in the structure enhancement graph, determining the graph features after weighted fusion of the different view features; inputting the graph features into a classifier, and determining classification loss based on a classification result; based on the image features after weighted fusion, attribute feature reconstruction of label nodes of the attribute enhancement image and structural feature reconstruction of label nodes of the structure enhancement image are carried out, and feature reconstruction loss is determined; regularizing the model based on the graph features of the label-free nodes of the attribute enhancement graph and the graph features of the label-free nodes of the structure enhancement graph, and determining semi-supervised learning loss of the label-free nodes based on a regularization result; and based on each loss, determining the joint training target loss of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of view, and particularly relates to a model training based on multiple views, a target detection method and device, equipment and medium. BACKGROUND

[0002] Graph algorithm is widely used in transaction anti-fraud field due to its efficient representation of complex relationships of transaction network. There are two problems in anomaly detection based on transaction network: one is sparse risk features, and the other is lack of high-quality labels. The two factors jointly limit the performance of detection model. SUMMARY

[0003] The embodiments of the present application provide a model training based on multiple views, a target detection method and device, equipment and medium, to realize sparse feature enhancement based on multiple view reconstruction, effectively alleviate the problem of limited model performance caused by lack of discriminative features, and regularize the model by using unlabeled samples, so that the model can be effectively trained with a small amount of labels.

[0004] The embodiments of the present application provide a model training method based on multiple views, comprising:

[0005] Based on the multiple views of the input model to be trained, an attribute enhancement graph and a structure enhancement graph are constructed;

[0006] Based on the graph features of the labeled nodes in the attribute enhancement graph and the graph features of the labeled nodes in the structure enhancement graph, a first graph feature after weighted fusion of different view features is determined;

[0007] The first graph feature after weighted fusion is input into a classifier to obtain a predicted first classification result, and a classification loss is determined based on the first classification result; and based on the first graph feature after weighted fusion, attribute feature reconstruction of the labeled nodes of the attribute enhancement graph and structure feature reconstruction of the labeled nodes of the structure enhancement graph are performed, and a feature reconstruction loss is determined based on the reconstruction result;

[0008] Based on the graph features of the unlabeled nodes of the attribute enhancement graph and the graph features of the unlabeled nodes of the structure enhancement graph, the model is regularized to obtain a regularization result;

[0009] Based on the regularization result, a semi-supervised learning loss of the unlabeled nodes is determined;

[0010] Based on the classification loss, the reconstruction loss and the semi-supervised learning loss of the unlabeled nodes, a joint training target loss of the model is determined.

[0011] The embodiment of the application constructs an attribute enhanced graph and a structure enhanced graph based on multiple views of an input to-be-trained model, so that multi-view representation learning can be performed based thereon, that is, a first graph feature after weighted fusion of different view features is determined based on the graph features of the labeled nodes in the attribute enhanced graph and the graph features of the labeled nodes in the structure enhanced graph; since the importance of features from different views to the final target detection task is different, the embodiment of the application realizes the fusion of different view features. Further, the first classification result of prediction can be obtained by inputting the first graph feature after the weighted fusion into a classifier, and a classification loss is determined based on the first classification result; further, the attribute feature reconstruction of the labeled nodes in the attribute enhanced graph and the structure feature reconstruction of the labeled nodes in the structure enhanced graph are performed based on the first graph feature after the weighted fusion, and a feature reconstruction loss is determined based on the reconstruction result. Therefore, after the above multi-view representation learning, it can be seen that the embodiment of the application effectively realizes the supplement of information by using the multi-view enhancement strategy and the feature fusion based on reconstruction, that is, the sparse feature enhancement based on multi-view reconstruction is realized, and the problem of limited detection model performance caused by the lack of discriminative features is effectively alleviated. On this basis, in order to further solve the problem of limited model performance caused by few labels, the unlabeled nodes in the attribute enhanced graph and the structure enhanced graph constructed above are processed as follows: the model is regularized based on the graph features of the unlabeled nodes in the attribute enhanced graph and the graph features of the unlabeled nodes in the structure enhanced graph, to obtain a regularization result; and a semi-supervised learning loss of the unlabeled nodes is determined based on the regularization result. It can be seen that the embodiment of the application considers that the feature distribution of different views has diversity, while the label has consistency, so the generalization ability of the model is improved by the feature distribution and the label distribution of the unlabeled samples in different views, so that the model can be effectively trained under a small amount of labels. Finally, the joint training target loss of the model can be determined based on the classification loss, the reconstruction loss, and the semi-supervised learning loss of the unlabeled nodes, so the performance of the entire model is improved.

[0012] In some embodiments, the model is regularized based on the graph features of the unlabeled nodes in the attribute enhanced graph and the graph features of the unlabeled nodes in the structure enhanced graph to obtain a regularization result, including:

[0013] The graph features of the unlabeled nodes in the attribute enhanced graph are obtained by a first encoder for extracting features, and the graph features of the unlabeled nodes in the structure enhanced graph are obtained by a second encoder for extracting features;

[0014] determine a different view feature diversity regularization result based on the graph feature of the unlabeled node of the attribute enhanced graph and the graph feature of the unlabeled node of the structure enhanced graph.

[0015] In some embodiments, based on the graph feature of the unlabeled node of the attribute enhanced graph and the graph feature of the unlabeled node of the structure enhanced graph, the model is regularized to obtain a regularization result, and the method further comprises:

[0016] determine a second graph feature after weighted fusion of different view features based on the graph feature of the unlabeled node of the attribute enhanced graph and the graph feature of the unlabeled node of the structure enhanced graph;

[0017] input the second graph feature after weighted fusion into a classifier to obtain a predicted second classification result;

[0018] input the graph feature of the unlabeled node of the attribute enhanced graph into a classifier to obtain a predicted third classification result, and input the graph feature of the unlabeled node of the structure enhanced graph into a classifier to obtain a predicted fourth classification result;

[0019] determine a label distribution result of the enhanced unlabeled node based on the third classification result and the fourth classification result;

[0020] determine a label consistency regularization result between the second classification result and the label distribution result of the enhanced unlabeled node.

[0021] In some embodiments, based on the regularization result, a semi-supervised learning loss of the unlabeled node is determined, comprising:

[0022] determine a semi-supervised learning loss of the unlabeled node based on the different view feature diversity regularization result and the label consistency regularization result.

[0023] In some embodiments, based on the third classification result and the fourth classification result, the label distribution result of the enhanced unlabeled node is determined using the following formula:

[0024]

[0025] wherein, the label distribution result of the enhanced unlabeled node is represented by y;

[0026] the third classification result is represented by y;

[0027] the fourth classification result is represented by y;

[0028] α μ[:,0] represents the weight of the third classification result;

[0029] α μ [:,1] represents the weight of the fourth classification result.

[0030] In some implementations, the label consistency regularization result between the second classification result and the enhanced label distribution result of the unlabeled nodes is determined using the following formula:

[0031]

[0032] in, This indicates the result of the label consistency regularization;

[0033] This represents the second classification result for the unlabeled node i;

[0034] This represents the label distribution result of the enhanced unlabeled node i.

[0035] In some implementations, based on the graph features of labeled nodes in the attribute enhancement graph and the graph features of labeled nodes in the structure enhancement graph, a first graph feature after weighted fusion of different view features is determined, including:

[0036] For the labeled nodes in the attribute augmentation graph, a first encoder for feature extraction is used to obtain the graph features of the labeled nodes in the attribute augmentation graph; and for the labeled nodes in the structure augmentation graph, a second encoder for feature extraction is used to obtain the graph features of the labeled nodes in the structure augmentation graph.

[0037] Based on the graph features of the labeled nodes of the attribute-enhanced graph and the graph features of the labeled nodes of the structure-enhanced graph, a first graph feature is determined after weighted fusion of different view features.

[0038] In some implementations, based on the features of the first graph after weighted fusion, attribute feature reconstruction of the labeled nodes of the attribute-enhanced graph and structural feature reconstruction of the labeled nodes of the structure-enhanced graph are performed, and feature reconstruction loss is determined based on the reconstruction results, including:

[0039] Based on the first graph features after weighted fusion, a reconstructed set of attribute features of labeled nodes of the attribute-enhanced graph is obtained through a first decoder for feature recovery; and a reconstructed set of structural features of labeled nodes of the structure-enhanced graph is obtained through a second decoder for feature recovery.

[0040] determine a feature reconstruction loss based on the original attribute feature set of the labeled node of the attribute enhanced graph, the reconstructed attribute feature set, the original structure feature set of the labeled node of the structure enhanced graph, and the reconstructed structure feature set.

[0041] In some embodiments, a joint training target loss of the model is determined based on the classification loss, the reconstruction loss, and the semi-supervised learning loss of the unlabeled node, according to the following formula:

[0042] L = L CE + βL Rec + λL Semi

[0043] wherein L represents the joint training target loss of the model.

[0044] L CE represents the classification loss.

[0045] L Rec represents the reconstruction loss.

[0046] L Semi represents the semi-supervised learning loss of the unlabeled node.

[0047] β and λ are preset coefficients.

[0048] In some embodiments, the attribute enhanced graph is determined by using a principal component analysis (PCA) algorithm.

[0049] The structure enhanced graph is determined by using a K-nearest neighbor algorithm.

[0050] Accordingly, a target detection method provided by an embodiment of the present application comprises:

[0051] inputting a target detection model into a target detection model trained in advance by any of the model training methods;

[0052] performing target detection on the target detection model, and outputting a detection result.

[0053] Accordingly, a model training device based on multiple views provided by an embodiment of the present application comprises:

[0054] a multiple view construction unit configured to construct an attribute enhanced graph and a structure enhanced graph based on multiple views of an input model to be trained;

[0055] a fusion unit configured to determine a first graph feature after weighted fusion of different view features based on graph features of labeled nodes in the attribute enhanced graph and graph features of labeled nodes in the structure enhanced graph;

[0056] a classification loss and a reconstruction loss determination unit configured to input the weighted fused first graph feature into a classifier to obtain a predicted first classification result, and determine a classification loss based on the first classification result; and perform attribute feature reconstruction of the labeled nodes of the attribute enhanced graph and structure feature reconstruction of the labeled nodes of the structure enhanced graph based on the weighted fused first graph feature, and determine a feature reconstruction loss based on the reconstruction results;

[0057] a regularization unit configured to regularize the model based on the graph features of the unlabeled nodes of the attribute enhanced graph and the graph features of the unlabeled nodes of the structure enhanced graph to obtain a regularization result;

[0058] a semi-supervised learning loss determination unit configured to determine a semi-supervised learning loss of the unlabeled nodes based on the regularization result;

[0059] a target loss determination unit configured to determine a joint training target loss of the model based on the classification loss, the reconstruction loss, and the semi-supervised learning loss of the unlabeled nodes.

[0060] Correspondingly, the embodiment of the present application provides a target detection device, which comprises:

[0061] an input target object unit configured to input a target object into a target detection model trained in advance by any one of the model training methods;

[0062] a detection output unit configured to perform target detection on the target object by the target detection model and output a detection result.

[0063] Another embodiment of the present application provides an electronic device comprising a memory and a processor, wherein the memory is configured to store program instructions, and the processor is configured to call the program instructions stored in the memory to execute any one of the above methods according to the obtained program.

[0064] Another embodiment of the present application provides a computer readable storage medium storing computer executable instructions, and the computer executable instructions are configured to make the computer execute any one of the above methods. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0066] Figure 1A flowchart of a model training method based on multiple views provided by an embodiment of the present application is shown in FIG. 1.

[0067] Figure 2 A schematic diagram of a model training principle based on multiple views provided by an embodiment of the present application is shown in FIG. 2.

[0068] Figure 3 A schematic diagram of the overall flow of a model training method based on multiple views provided by an embodiment of the present application is shown in FIG. 3.

[0069] Figure 4 A data enhancement flowchart of a labeled node provided by an embodiment of the present application is shown in FIG. 4.

[0070] Figure 5 A data enhancement flowchart of an unlabeled node provided by an embodiment of the present application is shown in FIG. 5.

[0071] Figure 6 A flowchart of a target detection method provided by an embodiment of the present application is shown in FIG. 6.

[0072] Figure 7 A structural diagram of an electronic device provided by an embodiment of the present application is shown in FIG. 7.

[0073] Figure 8 A structural diagram of a model training device based on multiple views provided by an embodiment of the present application is shown in FIG. 8.

[0074] Figure 9 A structural diagram of a target detection device provided by an embodiment of the present application is shown in FIG. 9. DETAILED DESCRIPTION

[0075] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0076] The embodiments of the present application provide a model training method, a target detection method, a device, and a medium based on multiple views, to realize sparse feature enhancement based on multiple view reconstruction, effectively alleviate the problem of limited model performance caused by the lack of discriminative features, and regularize the model by using unlabeled samples, so that the model can be effectively trained with a small amount of labels.

[0077] The method and the device, the equipment, and the medium are based on the same application concept. Since the principles of the method and the device, the equipment, and the medium for solving problems are similar, the implementation of the device, the equipment, and the medium can be mutually referred to, and the repeated parts will not be described again.

[0078] The terms "first", "second", etc. (if any) in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the described embodiments can be implemented in an order other than that illustrated or described. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0079] The following examples and embodiments will only be understood as illustrative examples. Although the present specification may refer to "a", "one", or "some" example or embodiment, this does not mean that each such reference relates to the same example or embodiment, nor does it mean that the feature only applies to a single example or embodiment. Individual features of different embodiments can also be combined to provide other embodiments. In addition, terms such as "include" and "contain" should be understood as not limiting the described embodiments to only those features mentioned; such examples and embodiments can also include features, structures, units, modules, etc. that are not specifically mentioned.

[0080] The various embodiments of the present application will be described in detail below with reference to the accompanying drawings of the specification. It should be noted that the order of presentation of the embodiments of the present application only represents the order of the embodiments and does not represent the superiority or inferiority of the technical solutions provided by the embodiments.

[0081] Referring to Figure 1 The model training method based on multiple views provided by the embodiments of the present application comprises:

[0082] S101, based on the multiple views of the input to-be-trained model, constructing an attribute enhancement graph and a structure enhancement graph;

[0083] The to-be-trained model, for example, an abnormal transfer bank card identification model applied to an abnormal transfer scenario, is a target detection model, and the target to be detected is an abnormal transfer bank card. Thus, the problem of few labels and sparse features caused by high difficulty in label acquisition and sparse risk behavior in the abnormal transfer scenario can be effectively alleviated, and the target detection accuracy of the model is improved.

[0084] The multiple views of the input to-be-trained model, for example, a weak information graph Using it directly as input to the model is difficult to obtain optimal results due to the lack of information. Therefore, this application's embodiments first analyze the weak information graph from the perspectives of node features and topological structure. Enhancement is performed, that is, the formation of an attribute enhancement graph. and structural enhancement diagram

[0085] In some implementations, the Principal Components Analysis (PCA) algorithm is used to determine the attribute enhancement graph;

[0086] The K-Nearest Neighbors (KNN) algorithm is used to determine the structure enhancement graph.

[0087] In some implementations, when enhancing each view, enhancement is performed from only one perspective, while the other remains unchanged. For example, when enhancing the node features of a graph for the same view, the graph's topology remains unchanged.

[0088] See Figure 2 Among them, regarding attribute enhancement graphs

[0089] To retain useful information to the maximum extent and eliminate noise, this application employs Principal Component Analysis (PCA) algorithm to extract the original node features. Mapped to a low-dimensional embedding space. Finally, the attribute enhancement graph can be represented as shown in Equation 1 below:

[0090]

[0091] in, Represents an attribute-enhanced graph;

[0092] Represents a set of nodes;

[0093] Represents the adjacency matrix;

[0094] Represents the feature set, It is the representation of a node in a low-dimensional embedding space, where N represents the number of nodes, and the features of node i in the node set V are... After PCA, we obtained...

[0095] Indicates the label of the corresponding node;

[0096] fPCA (*) is a principal component analysis function;

[0097] is the compression ratio of the node feature dimension, the larger the value of δ is, the more node feature components are retained.

[0098] Regarding the structure-enhanced graph

[0099] Since the K-neighbor graph captures the local similarity and the clustering structure is widely used in Hypergraph Neural Network (HGNN), the original weak information graph is converted into a K-neighbor graph, i.e., a structure-enhanced graph, by using a K-Nearest Neighbors (KNN) algorithm, which can be represented as shown in the following formula two:

[0100]

[0101] wherein, is the structure-enhanced graph.

[0102] is the adjacency matrix of the K-neighbor graph obtained by the K-Nearest Neighbors algorithm.

[0103] f KNN (*) is a K-Nearest Neighbors (KNN) algorithm.

[0104] It can be seen that by the above step S101, the attribute-enhanced graph and the structure-enhanced graph are constructed from two aspects of the node attributes (i.e., node features) and the topological structure of the graph, respectively, to lay a foundation for further mining of diversified information. In some embodiments, whether it is PCA or the construction of a K-neighbor graph, only one calculation is required, i.e., for the same view, either the attribute-enhanced graph is calculated by the above formula one or the structure-enhanced graph

[0105] On the basis of step S101, the following steps S102 and S103 of the multi-view representation learning process are performed for the labeled nodes in the graph:

[0106] S102, determining a first graph feature after weighted fusion of different view features based on the graph features of the labeled nodes in the attribute-enhanced graph and the graph features of the labeled nodes in the structure-enhanced graph;

[0107] In some embodiments, step S102 includes:

[0108] ​For the labeled nodes in the attribute-enhanced graph, a first encoder for extracting features is used to obtain the graph features of the labeled nodes in the attribute-enhanced graph; and for the labeled nodes in the structure-enhanced graph, a second encoder for extracting features is used to obtain the graph features of the labeled nodes in the structure-enhanced graph.

[0109] Based on the graph features of the labeled nodes in the attribute-enhanced graph and the graph features of the labeled nodes in the structure-enhanced graph, a first graph feature after weighted fusion of different view features is determined.

[0110] Referring to Figure 2 , for example:

[0111] Regarding the feature encoder:

[0112] Since the graph neural network has achieved remarkable success in many graph data mining tasks (such as node classification and recommendation system). Therefore, for the given attribute-enhanced graph and structure-enhanced graph, the embodiment of the present application adopts an encoder with GNN as the backbone to realize graph feature extraction, as shown in the following formula three:

[0113]

[0114] Among them, and respectively represent the encoder for feature extraction, and its backbone can be a basic model such as Graph Convolution Networks (GCN), Graph Attention Networks (GAT) and Graph Sample and Aggregate (GraphSAGE).

[0115] H PCA and H KNN are respectively the graph features extracted by the encoder and , that is, the graph features of the labeled nodes in the attribute-enhanced graph and the graph features of the labeled nodes in the structure-enhanced graph.

[0116] It should be noted that, for ease of description, the embodiment of the present application uses capital letters to represent the set of features, for example, represents the graph feature representation of all nodes in the attribute-enhanced graph.

[0117] Regarding the feature fusion based on reconstruction:

[0118] Through the above operation, the embodiment of the present application obtains the exclusive features of different views.

[0119] Since the importance of features from different views to the final task is different, the embodiment of the present application realizes the fusion of different view features through an attention mechanism.

[0120] For example, the embodiment of the present application calculates the corresponding weight by splicing H PCA and H KNN together, and realizes the weighted fusion of different view features, and the specific implementation is shown in the following formula four:

[0121]

[0122] wherein, weight is a matrix with a dimension of (N, 2), and weight.sum(1) is the sum by row.

[0123] H is the node feature representation after fusing different view features, i.e., the first graph feature after the above weighted fusion (i.e., the node feature H after the weighted fusion of different view features).

[0124] is the importance weight of different views.

[0125] W and b are model learnable parameters.

[0126] S103, input the weighted fused first graph feature into a classifier to obtain a predicted first classification result, and determine a classification loss based on the first classification result; and based on the weighted fused first graph feature, perform attribute feature reconstruction of the labeled nodes of the attribute enhanced graph and structure feature reconstruction of the labeled nodes of the structure enhanced graph, and determine a feature reconstruction loss based on the reconstruction result;

[0127] In some embodiments, based on the weighted fused first graph feature, performing attribute feature reconstruction of the labeled nodes of the attribute enhanced graph and structure feature reconstruction of the labeled nodes of the structure enhanced graph, and determining a feature reconstruction loss based on the reconstruction result, comprises:

[0128] Based on the weighted fused first graph feature, obtaining a reconstructed attribute feature set of the labeled nodes of the attribute enhanced graph through a first decoder for restoring features, and obtaining a reconstructed structure feature set of the labeled nodes of the structure enhanced graph through a second decoder for restoring features;

[0129] Based on the original attribute feature set of the labeled nodes of the attribute enhanced graph, the reconstructed attribute feature set, the original structure feature set of the labeled nodes of the structure enhanced graph, and the reconstructed structure feature set, determining a feature reconstruction loss.

[0130] To mitigate the impact of missing information on model performance, this embodiment aims to ensure that the first graph feature H, obtained from the weighted fusion of features from different views, simultaneously contains information from both the attribute-enhanced graph and the structure-enhanced graph. To explicitly achieve this goal, this embodiment decodes each of the weighted fusion features H to recover their original features. This is because if H can effectively recover its original features, it indicates that H simultaneously possesses information from different views.

[0131] See Figure 2 For example, H is input to the first decoder for feature recovery and the second decoder for feature recovery, respectively, so as to decode and recover the features of each node, as shown in Formula 5 below:

[0132]

[0133] in, and These are the first decoder corresponding to the attribute-enhanced graph and the second decoder corresponding to the structure-enhanced graph, respectively.

[0134] and These are respectively after being decoded and The reconstructed node feature set is the reconstructed attribute feature set of the labeled nodes in the attribute-enhanced graph and the reconstructed structural feature set of the labeled nodes in the structure-enhanced graph.

[0135] The objective of this application is to allow the fused graph features H to simultaneously possess valid information from different views, rather than requiring completely identical reconstruction. Therefore, mutual information between the input node features and the decoder output features is constrained to allow H to simultaneously possess multi-view information.

[0136] Specifically, taking attribute-enhanced graphs as an example, the embodiments of this application will use node v i The original attribute features and the reconstructed attribute features are considered to be directly opposite, that is... Furthermore, the original attribute features and the reconstructed attribute features of different nodes should be different, therefore... Treat this as a negative pair. The same applies to the structurally enhanced view, so I won't elaborate further.

[0137] Therefore, the constraints of the entire feature reconstruction can be expressed as shown in Formula Six below:

[0138]

[0139] Formula 6 is a contrast-based reconstruction loss function used to constrain the reconstruction of different view features.

[0140] where B = {PCA, KNN} is a set of different view categories. b belongs to B, which means that the summation process has two parts because B = {PCA, KNN}.

[0141] S<.,.> and τ are cosine similarity and temperature coefficient respectively.

[0142] L Rec is the mutual information reconstruction loss, i.e., the feature reconstruction loss, minimizing which is equivalent to maximizing the lower bound of mutual information features.

[0143] In some embodiments, referring to Figure 2 the weighted fused first graph features are input into a classifier to obtain a predicted first classification result, and a classification loss is determined based on the first classification result, including:

[0144] Based on the reconstructed feature fusion, H contains rich information in the attribute-enhanced graph and the structure-enhanced graph, and therefore, the embodiments of the present application take it as the final node feature representation. Finally, a final predicted classification result is obtained through a classification head (i.e., a classifier c) composed of a fully connected layer. Figure 2 as shown in the following formula seven:

[0145]

[0146] wherein, represents the final prediction result, i.e., the first classification result. Formula seven represents a classifier as a whole, and W and b are parameters of the classifier.

[0147] The fraud detection task based on a graph is an imbalanced binary classification task, and therefore, the embodiments of the present application use weighted cross-entropy to optimize the model, as shown in the following formula eight:

[0148]

[0149] wherein, L CE represents a classification loss calculated using cross-entropy;

[0150] η c is an imbalance weight corresponding to the category c;

[0151] represents the true label of node i in the category c;

[0152] is the predicted probability of node i in the category c.

[0153] ​It can be seen that, by the above steps, the embodiment of the application realizes sparse feature enhancement based on multi-view reconstruction, effectively enhances the sparse features by combining different features in multiple different views, and effectively alleviates the problem of limited model performance caused by the lack of discriminative features.

[0154] After the above learning, the embodiment of the application effectively realizes information supplement by using the multi-view enhancement strategy and the reconstruction-based feature fusion.

[0155] To further solve the problem of limited model performance caused by few labels, the embodiment of the application considers that the feature distribution of different views is diverse, while the label is consistent. Therefore, the embodiment of the application improves the generalization ability of the model by the feature distribution and the label distribution of the unlabeled samples in different views.

[0156] Therefore, further based on step S101, the following steps S104 and S105 of the regularization process of the diversity of the feature distribution and the consistency of the label are performed for the unlabeled nodes in the graph:

[0157] S104, based on the graph features of the unlabeled nodes of the attribute enhanced graph and the graph features of the unlabeled nodes of the structure enhanced graph, regularizing the model to obtain a regularization result;

[0158] In some embodiments, step S104 comprises:

[0159] For the unlabeled nodes in the attribute enhanced graph, the graph features of the unlabeled nodes of the attribute enhanced graph are obtained by the first encoder for extracting features; and for the unlabeled nodes in the structure enhanced graph, the graph features of the unlabeled nodes of the structure enhanced graph are obtained by the second encoder for extracting features;

[0160] Based on the graph features of the unlabeled nodes of the attribute enhanced graph and the graph features of the unlabeled nodes of the structure enhanced graph, a different view feature diversity regularization result is determined.

[0161] In some embodiments, step S104 further comprises:

[0162] Based on the graph features of the unlabeled nodes of the attribute enhanced graph and the graph features of the unlabeled nodes of the structure enhanced graph, a second graph feature after weighted fusion of different view features is determined.

[0163] The second graph feature after weighted fusion is input into a classifier to obtain a predicted second classification result.

[0164] inputting the graph features of the unlabeled nodes of the attribute enhanced graph into the classifier to obtain a predicted third classification result, and inputting the graph features of the unlabeled nodes of the structure enhanced graph into the classifier to obtain a predicted fourth classification result;

[0165] determining a label distribution result of the enhanced unlabeled nodes based on the third classification result and the fourth classification result;

[0166] determining a label consistency regularization result between the second classification result and the label distribution result of the enhanced unlabeled nodes.

[0167] Referring to Figure 2 , for example:

[0168] Regarding feature diversity regularization:

[0169] For the unlabeled nodes in the attribute enhanced graph and the structure enhanced graph constructed in the above steps, the embodiments of the present application obtain their exclusive features in different views through the above formula three (that is, the graph features of the unlabeled nodes of the attribute enhanced graph) and (that is, the graph features of the unlabeled nodes of the structure enhanced graph).

[0170] Only the exclusive features of different views have diversity, can better play the advantages of multi-view feature fusion. In other words, the exclusive features of different views should be low redundancy. Among them, the diversity of feature distribution is used to measure the redundancy degree of different views, that is, the different view feature diversity regularization result. Specifically, the redundancy degree between different view feature distributions (that is, the different view feature diversity regularization result) can be represented as shown in the following formula nine:

[0171]

[0172] Among them, D Diversity is a distance function defined in a vector space, such as Euclidean distance and cosine distance, which can represent the redundancy between features specific to a view. The smaller it is, the lower the redundancy is, and the higher the diversity is.

[0173] Regarding label consistency regularization:

[0174] For the feature H μ (that is, the second graph feature after weighted fusion of different view features) fused through formula four, the predicted second classification result

[0175] The label distribution of the nodes in different enhanced views should be consistent. Therefore, for the node embedding set of the unlabeled sample in different views and The prediction set corresponding to the view is obtained through the above Formula Seven And That is, the graph feature of the unlabeled node of the attribute enhanced graph Input the classifier to obtain a predicted third classification result And the graph feature of the unlabeled node of the structure enhanced graph Input the classifier to obtain a predicted fourth classification result

[0176] Unlike the common direct constraint on the consistency of pairwise label distribution through the mean square error loss, the embodiment of the application helps And And the importance score α of the unlabeled sample in the above Formula Four μ Form a weakly enhanced label distribution, that is, the label distribution result of the enhanced unlabeled node, as shown in the following Formula Ten:

[0177]

[0178] Among them, Indicates the label distribution result of the weakly enhanced unlabeled node;

[0179] Indicates the third classification result;

[0180] Indicates the fourth classification result;

[0181] α μ [:,0] indicates the weight of the third classification result;

[0182] α μ [:,1] indicates the weight of the fourth classification result.

[0183] Weakly enhanced label distribution Should be consistent with Therefore, for example, the embodiment of the application can use the KL divergence (Kullback-Leibler Divergence, KLD, also known as relative entropy) to constrain their consistency, thereby determining the label consistency regularization result between the second classification result and the label distribution result of the enhanced unlabeled node, as shown in the following Formula Eleven:

[0184]

[0185] Among them, D Consis Is a prediction consistency distance defined in the label space. For example, the embodiment of the application uses the KL divergence to measure it.

[0186] the second classification result of the unlabeled node i;

[0187] the label distribution result of the enhanced unlabeled node i.

[0188] S105, based on the regularization result, determine the semi-supervised learning loss of the unlabeled node;

[0189] In some embodiments, based on the regularization result, determining the semi-supervised learning loss of the unlabeled node comprises:

[0190] Based on the different view feature diversity regularization result, the label consistency regularization result, determine the semi-supervised learning loss of the unlabeled node.

[0191] For example, in the different view embedding, the feature diversity space of the unlabeled center node and the consistency of the prediction result constitute the entire semi-supervised learning loss, as shown in the following formula twelve:

[0192] L Semi = D Consis + D Diversity Formula twelve

[0193] Wherein, L Semi is the semi-supervised learning loss function of the unlabeled node. That is, the calculation result of the above formula nine and the calculation result of the above formula eleven are added to obtain the semi-supervised learning loss function of the unlabeled node.

[0194] As can be seen, through the above steps, the present application embodiment proposes regularization based on multi-view feature diversity and label consistency, which realizes the regularization of the model by using a large number of unlabeled samples in the feature space diversity and the label consistency in the label space, so that the model can be effectively trained under a small number of labels.

[0195] S106, based on the classification loss, the reconstruction loss, and the semi-supervised learning loss of the unlabeled node, determine the joint training target loss of the model.

[0196] As can be seen, the model optimization target proposed by the present application embodiment has three parts, namely the feature reconstruction loss L Rec , the node classification loss L CE and the semi-supervised learning loss L Semi of the unlabeled node (unlabeled data). Therefore, the overall optimization target can be represented as shown in the following formula thirteen:

[0197] L = L CE + βL Rec + λLSemi Equation Thirteen

[0198] i.e. the calculation result L of Equation Six above Rec , the calculation result L of Equation Eight CE , the calculation result L of Equation Twelve Semi , and finally obtain the overall optimization target loss value L of the model.

[0199] Wherein, β and λ are balance coefficients of different loss functions, and are preset values. Embodiments of the present application can adjust them so that different loss function values L CE , L Rec , L Semi are located on the same numerical scale.

[0200] It can be seen that the embodiments of the present application not only realize sparse feature enhancement based on multi-view reconstruction, but also realize data enhancement based on multi-view regularization, so that the model trained based thereon can achieve good detection effect in the case of few labels and sparse features.

[0201] In summary, referring to Figure 3 , the model training method provided by the embodiments of the present application as a whole comprises:

[0202] S301, constructing an attribute enhancement graph and a structure enhancement graph based on the multi-view of the input to-be-trained model and the structure enhancement graph The attribute enhancement graph includes labeled nodes and / or unlabeled nodes, and the structure enhancement graph includes labeled nodes and / or unlabeled nodes;

[0203] S302, data enhancement of the labeled nodes;

[0204] S303, data enhancement of the unlabeled nodes;

[0205] S304, determining the overall optimization target of the model based on the data enhancement result of the labeled nodes and the data enhancement result of the unlabeled nodes.

[0206] In some embodiments, referring to Figure 4 , the step S302 comprises:

[0207] S401, for the labeled nodes in the attribute enhancement graph, obtaining the graph features (H PCA ) of the labeled nodes of the attribute enhancement graph through a first encoder for extracting features; and for the labeled nodes in the structure enhancement graph, obtaining the graph features (H KNN ) of the labeled nodes of the structure enhancement graph through a second encoder for extracting features;

[0208] S402, determine a first graph feature (H) of a weighted fusion of different view features based on the graph feature of the labeled nodes of the attribute enhanced graph and the graph feature of the labeled nodes of the structure enhanced graph;

[0209] S403, obtain a reconstructed attribute feature set of the labeled nodes of the attribute enhanced graph through a first decoder for recovering features based on the first graph feature of the weighted fusion obtain a reconstructed structure feature set of the labeled nodes of the structure enhanced graph through a second decoder for recovering features

[0210] S404, determine a feature reconstruction loss (L Rec ) based on the original attribute feature set, the reconstructed attribute feature set of the labeled nodes of the attribute enhanced graph, and the original structure feature set, the reconstructed structure feature set of the labeled nodes of the structure enhanced graph

[0211] S405, input the first graph feature of the weighted fusion into a classifier to obtain a predicted first classification result and determine a classification loss (L CE ) based on the first classification result.

[0212] In some embodiments, referring to Figure 5 , the step S303 comprises:

[0213] S501, obtain a graph feature of unlabeled nodes of the attribute enhanced graph through a first encoder for extracting features and obtain a graph feature of unlabeled nodes of the structure enhanced graph through a second encoder for extracting features

[0214] S502, determine a different view feature diversity regularization result based on the graph feature of the unlabeled nodes of the attribute enhanced graph and the graph feature of the unlabeled nodes of the structure enhanced graph

[0215] S503, determine a second graph feature (H μ ) of a weighted fusion of different view features based on the graph feature of the unlabeled nodes of the attribute enhanced graph and the graph feature of the unlabeled nodes of the structure enhanced graph

[0216] S504, input the second graph feature of the weighted fusion into a classifier to obtain a predicted second classification result

[0217] S505, inputting the graph feature of the unlabeled node of the attribute enhanced graph inputting the classifier to obtain a predicted third classification result and inputting the graph feature of the unlabeled node of the structure enhanced graph inputting the classifier to obtain a predicted fourth classification result

[0218] S506, determining a label distribution result of the enhanced unlabeled node based on the third classification result and the fourth classification result

[0219] S507, determining a label consistency regularization result between the second classification result and the label distribution result of the enhanced unlabeled node

[0220] S508, determining the different view feature diversity regularization result the label consistency regularization result determining a semi-supervised learning loss (L Semi ) of the unlabeled node;

[0221] In some embodiments, the step S304 comprises:

[0222] determining a joint training target loss (L) of the model based on the classification loss (L CE ), the reconstruction loss (L Rec ), and the semi-supervised learning loss (L Semi ) of the unlabeled node.

[0223] Correspondingly, referring to Figure 6 , the target detection method provided by the embodiments of the present application comprises:

[0224] S601, inputting a to-be-detected object into a target detection model pre-trained by the model training method provided by the embodiments of the present application;

[0225] The model trained by the model training method based on multiple views provided by the embodiments of the present application can be applied to the process of identifying abnormal transfer bank cards, and can effectively alleviate the problem of few labels and sparse features caused by high difficulty in label acquisition and sparse risk behaviors in the abnormal transfer scenario. Therefore, the accuracy of identifying abnormal transfer bank cards can be improved.

[0226] The to-be-detected object, for example, includes account information of a bank card involved in the transfer process.

[0227] S602, performing target detection on the to-be-detected object by the target detection model, and outputting a detection result.

[0228] The target detection, such as the identification of abnormal bank cards for transfers, can output the relevant information of the abnormal bank cards as detection results, such as outputting it to an alarm platform.

[0229] The following describes the device or apparatus provided in the embodiments of this application, and the explanations or examples of the same or corresponding technical features as those described in the above methods will not be repeated hereafter.

[0230] An electronic device is provided in an embodiment of this application, see [link to example]. Figure 7 For example, including:

[0231] Processor 600 is used to read the program from memory 620 and execute the following procedures:

[0232] Based on multiple views of the input model to be trained, construct attribute augmentation graphs and structure augmentation graphs;

[0233] Based on the graph features of labeled nodes in the attribute enhancement graph and the graph features of labeled nodes in the structure enhancement graph, the first graph feature after weighted fusion of different view features is determined.

[0234] The weighted and fused first graph features are input into a classifier to obtain a predicted first classification result, and a classification loss is determined based on the first classification result; furthermore, based on the weighted and fused first graph features, attribute feature reconstruction of the labeled nodes of the attribute enhancement graph and structural feature reconstruction of the labeled nodes of the structure enhancement graph are performed, and feature reconstruction loss is determined based on the reconstruction results.

[0235] Based on the graph features of the unlabeled nodes of the attribute-enhanced graph and the graph features of the unlabeled nodes of the structure-enhanced graph, the model is regularized to obtain the regularization result.

[0236] Based on the regularization results, determine the semi-supervised learning loss for unlabeled nodes;

[0237] Based on the classification loss, the reconstruction loss, and the semi-supervised learning loss of the unlabeled nodes, the joint training objective loss of the model is determined.

[0238] In some implementations, the model is regularized based on the graph features of the unlabeled nodes of the attribute-enhanced graph and the graph features of the unlabeled nodes of the structure-enhanced graph to obtain a regularization result, including:

[0239] For unlabeled nodes in the attribute augmentation graph, graph features of the unlabeled nodes in the attribute augmentation graph are obtained by a first encoder for feature extraction; and for unlabeled nodes in the structure augmentation graph, graph features of the unlabeled nodes in the structure augmentation graph are obtained by a second encoder for feature extraction.

[0240] determine different view feature diversity regularization results based on the graph features of the unlabeled nodes of the attribute enhanced graph and the graph features of the unlabeled nodes of the structure enhanced graph.

[0241] In some embodiments, based on the graph features of the unlabeled nodes of the attribute enhanced graph and the graph features of the unlabeled nodes of the structure enhanced graph, the model is regularized to obtain a regularization result, and the method further comprises:

[0242] determine a second graph feature after weighted fusion of different view features based on the graph features of the unlabeled nodes of the attribute enhanced graph and the graph features of the unlabeled nodes of the structure enhanced graph;

[0243] input the second graph feature after weighted fusion into a classifier to obtain a predicted second classification result;

[0244] input the graph features of the unlabeled nodes of the attribute enhanced graph into a classifier to obtain a predicted third classification result, and input the graph features of the unlabeled nodes of the structure enhanced graph into a classifier to obtain a predicted fourth classification result;

[0245] determine a label distribution result of the enhanced unlabeled nodes based on the third classification result and the fourth classification result;

[0246] determine a label consistency regularization result between the second classification result and the label distribution result of the enhanced unlabeled nodes.

[0247] In some embodiments, based on the regularization result, a semi-supervised learning loss of the unlabeled nodes is determined, comprising:

[0248] determine a semi-supervised learning loss of the unlabeled nodes based on the different view feature diversity regularization results and the label consistency regularization result.

[0249] In some embodiments, based on the third classification result and the fourth classification result, the label distribution result of the enhanced unlabeled nodes is determined using the following formula:

[0250]

[0251] wherein, represents the label distribution result of the enhanced unlabeled nodes;

[0252] represents the third classification result;

[0253] represents the fourth classification result;

[0254] alpha μ [:, 0] represents the weight of the third classification result;

[0255] alpha μ [:, 1] represents the weight of the fourth classification result.

[0256] In some embodiments, the label consistency regularization result between the second classification result and the enhanced label distribution result of the unlabeled node is determined by using the following formula:

[0257]

[0258] wherein, represents the label consistency regularization result;

[0259] represents the second classification result of the unlabeled node i;

[0260] represents the enhanced label distribution result of the unlabeled node i.

[0261] In some embodiments, the first graph feature after weighted fusion of different view features is determined based on the graph feature of the labeled node in the attribute enhanced graph and the graph feature of the labeled node in the structure enhanced graph, comprising:

[0262] obtaining the graph feature of the labeled node in the attribute enhanced graph through a first encoder for extracting features, and obtaining the graph feature of the labeled node in the structure enhanced graph through a second encoder for extracting features;

[0263] determining the first graph feature after weighted fusion of different view features based on the graph feature of the labeled node in the attribute enhanced graph and the graph feature of the labeled node in the structure enhanced graph.

[0264] In some embodiments, based on the first graph feature after weighted fusion, attribute feature reconstruction of the labeled node in the attribute enhanced graph and structure feature reconstruction of the labeled node in the structure enhanced graph are performed, and a feature reconstruction loss is determined based on the reconstruction result, comprising:

[0265] obtaining a reconstructed attribute feature set of the labeled node in the attribute enhanced graph through a first decoder for restoring features based on the first graph feature after weighted fusion, and obtaining a reconstructed structure feature set of the labeled node in the structure enhanced graph through a second decoder for restoring features;

[0266] determine a feature reconstruction loss based on the original attribute feature set of the labeled node of the attribute enhanced graph, the reconstructed attribute feature set, the original structure feature set of the labeled node of the structure enhanced graph, and the reconstructed structure feature set.

[0267] In some embodiments, a joint training target loss of the model is determined based on the classification loss, the reconstruction loss, and the semi-supervised learning loss of the unlabeled node, according to the following formula:

[0268] L = L CE + βL Rec + λL Semi

[0269] wherein L represents the joint training target loss of the model.

[0270] L CE represents the classification loss.

[0271] L Rec represents the reconstruction loss.

[0272] L Semi represents the semi-supervised learning loss of the unlabeled node.

[0273] β and λ are preset coefficients.

[0274] In some embodiments, the attribute enhanced graph is determined by using a principal component analysis (PCA) algorithm.

[0275] The structure enhanced graph is determined by using a K-nearest neighbor algorithm.

[0276] The transceiver 610 is configured to receive and send data under the control of the processor 600.

[0277] The processor 600 can further read programs in the memory 620 and perform the following processes.

[0278] The target detection model is pre-trained by any of the model training methods.

[0279] The target detection model is used to detect the target of the to-be-tested object, and output a detection result.

[0280] wherein, in Figure 7In particular embodiments, bus architecture can include any number of interconnecting buses and bridges, depending on the specific application of processor 600 and the overall design constraints. Bus architecture can link together various circuits such as one or more processors represented by processor 600, the various circuits that represent the memory, 620, and can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, all of which are well known in the art, and therefore, will not be described further. Bus interface provides an interface to the bus architecture. Transceiver 610 can be a plurality of elements including a transmitter and a receiver, providing a means for communicating with various other apparatus over a transmission medium, including a wireless channel, a wired channel, optical cable, and the like. User interface 630 can also be a means for interfacing with a user of the various devices, including but not limited to a keypad, a display, a speaker, a microphone, a joystick, and the like.

[0281] Processor 600 is responsible for managing the bus architecture and general processing, and memory 620 can store data used by processor 600 during execution of operations.

[0282] In some embodiments, processor 600 can be a CPU (Central Processor Unit), an ASIC (Application Specific Integrated Circuit), a FPGA (Field-Programmable Gate Array), or a CPLD (Complex Programmable Logic Device), and the processor can also be a multi-core architecture.

[0283] The processor can call a computer program stored in the memory to execute any of the methods provided by the embodiments of the present application according to the executable instructions obtained. The processor and the memory can also be arranged physically separately.

[0284] It should be noted that the above-mentioned device provided by the embodiments of the present application can realize all the method steps realized by the above-mentioned method embodiments, and can achieve the same technical effects. Therefore, the same parts and beneficial effects of the method embodiments will not be described in detail.

[0285] Referring to Figure 8 The model training device provided by the embodiments of the present application comprises:

[0286] The multi-view construction unit 11 is configured to construct an attribute enhancement graph and a structure enhancement graph based on the multi-view of the input to-be-trained model.

[0287] the fusion unit 12 is configured to determine a first graph feature of different view features after weighted fusion based on the graph feature of the labeled nodes in the attribute enhanced graph and the graph feature of the labeled nodes in the structure enhanced graph;

[0288] the classification loss and reconstruction loss determination unit 13 is configured to input the first graph feature after weighted fusion into a classifier to obtain a predicted first classification result, and determine a classification loss based on the first classification result; and perform attribute feature reconstruction of the labeled nodes in the attribute enhanced graph and structure feature reconstruction of the labeled nodes in the structure enhanced graph based on the first graph feature after weighted fusion, and determine a feature reconstruction loss based on the reconstruction result;

[0289] the regularization unit 14 is configured to perform regularization on the model based on the graph feature of the unlabeled nodes in the attribute enhanced graph and the graph feature of the unlabeled nodes in the structure enhanced graph to obtain a regularization result;

[0290] the semi-supervised learning loss determination unit 15 is configured to determine a semi-supervised learning loss of the unlabeled nodes based on the regularization result;

[0291] the target loss determination unit 16 is configured to determine a joint training target loss of the model based on the classification loss, the reconstruction loss, and the semi-supervised learning loss of the unlabeled nodes.

[0292] In some embodiments, performing regularization on the model based on the graph feature of the unlabeled nodes in the attribute enhanced graph and the graph feature of the unlabeled nodes in the structure enhanced graph to obtain a regularization result comprises:

[0293] obtaining the graph feature of the unlabeled nodes in the attribute enhanced graph through a first encoder for feature extraction, and obtaining the graph feature of the unlabeled nodes in the structure enhanced graph through a second encoder for feature extraction;

[0294] determining a diversity regularization result of different view features based on the graph feature of the unlabeled nodes in the attribute enhanced graph and the graph feature of the unlabeled nodes in the structure enhanced graph.

[0295] In some embodiments, performing regularization on the model based on the graph feature of the unlabeled nodes in the attribute enhanced graph and the graph feature of the unlabeled nodes in the structure enhanced graph to obtain a regularization result further comprises:

[0296] determining a second graph feature after weighted fusion of different view features based on the graph feature of the unlabeled nodes in the attribute enhanced graph and the graph feature of the unlabeled nodes in the structure enhanced graph.

[0297] inputting the weighted fused second graph feature into a classifier to obtain a predicted second classification result;

[0298] inputting the graph feature of the unlabeled node of the attribute enhanced graph into a classifier to obtain a predicted third classification result, and inputting the graph feature of the unlabeled node of the structure enhanced graph into a classifier to obtain a predicted fourth classification result;

[0299] determining a label distribution result of the enhanced unlabeled node based on the third classification result and the fourth classification result;

[0300] determining a label consistency regularization result between the second classification result and the label distribution result of the enhanced unlabeled node.

[0301] In some embodiments, based on the regularization result, a semi-supervised learning loss of the unlabeled node is determined, including:

[0302] based on the different view feature diversity regularization result and the label consistency regularization result, determining a semi-supervised learning loss of the unlabeled node.

[0303] In some embodiments, based on the third classification result and the fourth classification result, the label distribution result of the enhanced unlabeled node is determined using the following formula:

[0304]

[0305] wherein, the label distribution result of the enhanced unlabeled node is represented by y;

[0306] the third classification result is represented by y;

[0307] the fourth classification result is represented by y;

[0308] α μ [:, 0] represents the weight of the third classification result;

[0309] α μ [:, 1] represents the weight of the fourth classification result.

[0310] In some embodiments, the label consistency regularization result between the second classification result and the label distribution result of the enhanced unlabeled node is determined using the following formula:

[0311]

[0312] wherein, representing the label consistency regularization result of the label;

[0313] representing the second classification result of the unlabeled node i;

[0314] representing the label distribution result of the enhanced unlabeled node i.

[0315] In some embodiments, based on the graph features of the labeled nodes in the attribute enhanced graph and the graph features of the labeled nodes in the structure enhanced graph, a first graph feature after weighted fusion of different view features is determined, including:

[0316] For the labeled nodes in the attribute enhanced graph, graph features of the labeled nodes in the attribute enhanced graph are obtained through a first encoder for extracting features; and for the labeled nodes in the structure enhanced graph, graph features of the labeled nodes in the structure enhanced graph are obtained through a second encoder for extracting features;

[0317] Based on the graph features of the labeled nodes in the attribute enhanced graph and the graph features of the labeled nodes in the structure enhanced graph, a first graph feature after weighted fusion of different view features is determined.

[0318] In some embodiments, based on the first graph feature after weighted fusion, attribute feature reconstruction of the labeled nodes in the attribute enhanced graph and structure feature reconstruction of the labeled nodes in the structure enhanced graph are performed, and a feature reconstruction loss is determined based on the reconstruction results, including:

[0319] Based on the first graph feature after weighted fusion, a reconstructed attribute feature set of the labeled nodes in the attribute enhanced graph is obtained through a first decoder for restoring features; and a reconstructed structure feature set of the labeled nodes in the structure enhanced graph is obtained through a second decoder for restoring features;

[0320] Based on the original attribute feature set of the labeled nodes in the attribute enhanced graph, the reconstructed attribute feature set, the original structure feature set of the labeled nodes in the structure enhanced graph, and the reconstructed structure feature set, a feature reconstruction loss is determined.

[0321] In some embodiments, based on the classification loss, the reconstruction loss, and the semi-supervised learning loss of the unlabeled nodes, a joint training target loss of the model is determined using the following formula:

[0322] L = L CE + βL Rec + λL Semi

[0323] L represents a joint training target loss of the model;

[0324] L CE represents the classification loss;

[0325] L Rec represents the reconstruction loss;

[0326] L Semi represents the semi-supervised learning loss of the unlabeled node;

[0327] β and λ are preset coefficients.

[0328] In some embodiments, a principal component analysis (PCA) algorithm is used to determine the attribute enhancement graph;

[0329] A K-nearest neighbor algorithm is used to determine the structure enhancement graph.

[0330] Referring to Figure 9 , the target detection device provided by the embodiments of the present application comprises:

[0331] The input object unit 21 is configured to input a to-be-detected object into a target detection model pre-trained by the model training method provided by the embodiments of the present application.

[0332] The detection output unit 22 is configured to perform target detection on the to-be-detected object by using the target detection model, and output a detection result.

[0333] It should be noted that the division of the units in the embodiments of the present application is illustrative, and is only a logical function division. In actual implementation, another division mode can be used. In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present alone, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0334] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor execute all or part of the steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0335] Any of the devices or apparatuses provided in the embodiments of the present application can be specifically a desktop computer, a portable computer, a smart phone, a tablet computer, a personal digital assistant (Personal Digital Assistant, PDA), etc. It can include a central processing unit (Center Processing Unit, CPU), a memory, an input / output device, etc. The input device can include a keyboard, a mouse, a touch screen, etc. The output device can include a display device, such as a liquid crystal display (Liquid Crystal Display, LCD), a cathode ray tube (Cathode Ray Tube, CRT), etc.

[0336] The memory can include a read-only memory (ROM) and a random access memory (RAM), and provide the processor with program instructions and data stored in the memory. In the embodiments of the present application, the memory can be used to store the programs of any of the methods provided in the embodiments of the present application.

[0337] The processor calls the program instructions stored in the memory, and the processor is used to execute any of the methods provided in the embodiments of the present application according to the obtained program instructions.

[0338] The embodiments of the present application further provide a computer program product or computer program, which comprises computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to enable the computer device to perform any of the methods described in the above embodiments. The program product can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium may, for example, be but is not limited to an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination of the above. More specific examples (non-exhaustive list) of the readable storage medium include an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0339] The embodiments of the present application provide a computer readable storage medium for storing computer program instructions for the apparatus provided in the above embodiments of the present application, which contains programs for executing any of the methods provided in the above embodiments of the present application. The computer readable storage medium can be a non-transitory computer readable medium.

[0340] The computer readable storage medium can be any available medium or data storage device that can be accessed by a computer, including but not limited to a magnetic storage (such as a floppy disk, a hard disk, a magnetic tape, a magneto-optical disk (MO), etc.), an optical storage (such as a CD, a DVD, a BD, a HVD, etc.), and a semiconductor storage (such as a ROM, an EPROM, an EEPROM, a non-volatile memory (NAND FLASH), a solid state disk (SSD)), etc.

[0341] It should be understood that:

[0342] The access technology via which entities in a communication network communicate traffic to and from each other can be any suitable current or future technology, such as WLAN (Wireless Local Access Network), WiMAX (Worldwide Interoperability for Microwave Access), LTE, LTE-A, 5G, Bluetooth, infrared, etc. can be used; in addition, embodiments can also apply wired technologies, for example, IP-based access technologies, such as wired networks or fixed lines.

[0343] Embodiments suitable for being implemented as software code or parts thereof and being run using a processor or processing functionality are independent of the software code and can be specified using any known or future developed programming language, such as a high-level programming language, such as objective-C, C, C++, C#, Java, Python, Javascript, other scripting languages, etc., or a low-level programming language, such as a machine language or assembler.

[0344] Embodiments are independent of hardware and can be implemented using any known or future developed hardware technology or any hybrid of these, such as microprocessors or CPUs (Central Processing Units), MOS (Metal Oxide Semiconductor), CMOS (Complementary MOS), BiMOS (Bipolar MOS), BiCMOS (Bipolar CMOS), ECL (Emitter Coupled Logic), and / or TTL (Transistor-Transistor Logic).

[0345] Embodiments can be implemented as individual devices, units, components, or functions, or in a distributed manner, e.g., using or sharing one or more processors or processing functionality in processing, or using and sharing one or more processing segments or processing portions, wherein one physical processor or more than one physical processor can be used for implementing one or more processing portions dedicated to specific processing as described.

[0346] A device can be implemented by a semiconductor chip, a chipset, or a (hardware) module including such chip or chipset.

[0347] Embodiments can also be implemented as any combination of hardware and software, such as ASIC (Application Specific IC (Integrated Circuit)) components, FPGA (Field-programmable Gate Array) or CPLD (Complex Programmable Logic Device) components, or DSP (Digital Signal Processor) components.

[0348] Embodiments can also be implemented as a computer program product comprising a computer usable medium having a computer readable program code embodied therein, the computer readable program code adapted to be executed by a computer to perform processes as described in embodiments, wherein the computer usable medium can be a non-transitory medium.

[0349] Those skilled in the art will appreciate that embodiments of the present application can be supplied as a method, a system, or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer readable storage media (including, but not limited to, magnetic disks or optical storage) embodying computer readable program code, for example.

[0350] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure One one or more flow or blocks Figure One one or more flow or blocks

[0351] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure One one or more flow or blocks Figure One one or more flow or blocks

[0352] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure One one or more flow or blocks Figure One one or more flow or blocks

[0353] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A method for training a model based on multi-view, characterized in that, The method comprises: based on the input of the multi-view of the model to be trained, the attribute enhanced graph and the structure enhanced graph are constructed; based on the graph features of the labeled nodes in the attribute enhanced graph and the graph features of the labeled nodes in the structure enhanced graph, a first graph feature after weighted fusion of different view features is determined; the first graph feature after weighted fusion is input into a classifier to obtain a predicted first classification result, and a classification loss is determined based on the first classification result; and based on the first graph feature after weighted fusion, the attribute feature reconstruction of the labeled nodes in the attribute enhanced graph and the structure feature reconstruction of the labeled nodes in the structure enhanced graph are performed, and a feature reconstruction loss is determined based on the reconstruction result; based on the graph features of the unlabeled nodes in the attribute enhanced graph and the graph features of the unlabeled nodes in the structure enhanced graph, the model is regularized to obtain a regularization result; based on the regularization result, a semi-supervised learning loss of the unlabeled nodes is determined; based on the classification loss, the reconstruction loss and the semi-supervised learning loss of the unlabeled nodes, a joint training target loss of the model is determined.

2. The method of claim 1, wherein, based on the graph features of the unlabeled nodes in the attribute enhanced graph and the graph features of the unlabeled nodes in the structure enhanced graph, the model is regularized to obtain a regularization result, comprising: for the unlabeled nodes in the attribute enhanced graph, a first encoder for extracting features is used to obtain the graph features of the unlabeled nodes in the attribute enhanced graph; and for the unlabeled nodes in the structure enhanced graph, a second encoder for extracting features is used to obtain the graph features of the unlabeled nodes in the structure enhanced graph; based on the graph features of the unlabeled nodes in the attribute enhanced graph and the graph features of the unlabeled nodes in the structure enhanced graph, a diversity regularization result of different view features is determined.

3. The method of claim 2, wherein, based on the graph features of the unlabeled nodes in the attribute enhanced graph and the graph features of the unlabeled nodes in the structure enhanced graph, the model is regularized to obtain a regularization result, further comprising: based on the graph features of the unlabeled nodes in the attribute enhanced graph and the graph features of the unlabeled nodes in the structure enhanced graph, a second graph feature after weighted fusion of different view features is determined; the second graph feature after weighted fusion is input into a classifier to obtain a predicted second classification result; the graph features of the unlabeled nodes in the attribute enhanced graph are input into a classifier to obtain a predicted third classification result, and the graph features of the unlabeled nodes in the structure enhanced graph are input into a classifier to obtain a predicted fourth classification result; based on the third classification result and the fourth classification result, a label distribution result of the enhanced unlabeled nodes is determined; a label consistency regularization result between the second classification result and the label distribution result of the enhanced unlabeled nodes is determined.

4. The method of claim 3, wherein, based on the regularization result, a semi-supervised learning loss of the unlabeled nodes is determined, comprising: based on the diversity regularization result of different view features and the label consistency regularization result, a semi-supervised learning loss of the unlabeled nodes is determined.

5. The method of claim 3, wherein, Based on the third classification result and the fourth classification result, a label distribution result of an enhanced unlabeled node is determined by using the following formula: wherein, represents the label distribution result of the enhanced unlabeled node; representing the third classification result; representing the fourth classification result; alpha μ [:,0] represents the weight of the third classification result; alpha μ [:,1] represents the weight of the fourth classification result.

6. The method of claim 5, wherein, A label consistency regularization result between the second classification result and the label distribution result of the enhanced unlabeled node is determined by using the following formula: wherein, denotes the label consistency regularization result; denotes the second classification result of the label-free node i; represents the label distribution result of the enhanced unlabelled node i.

7. The method of claim 1, wherein, Based on the graph features of the labeled nodes in the attribute enhanced graph and the graph features of the labeled nodes in the structure enhanced graph, a first graph feature after weighted fusion of different view features is determined, including: For the labeled nodes in the attribute enhanced graph, a graph feature of the labeled nodes in the attribute enhanced graph is obtained through a first encoder for extracting features, and for the labeled nodes in the structure enhanced graph, a graph feature of the labeled nodes in the structure enhanced graph is obtained through a second encoder for extracting features; Based on the graph features of the labeled nodes in the attribute enhanced graph and the graph features of the labeled nodes in the structure enhanced graph, a first graph feature after weighted fusion of different view features is determined.

8. The method of claim 1, wherein, Based on the first graph feature after weighted fusion, attribute feature reconstruction of the labeled nodes in the attribute enhanced graph and structure feature reconstruction of the labeled nodes in the structure enhanced graph are performed, and a feature reconstruction loss is determined based on the reconstruction results, including: Based on the first graph feature after weighted fusion, a reconstructed attribute feature set of the labeled nodes in the attribute enhanced graph is obtained through a first decoder for restoring features, and a reconstructed structure feature set of the labeled nodes in the structure enhanced graph is obtained through a second decoder for restoring features; Based on the original attribute feature set of the labeled nodes in the attribute enhanced graph, the reconstructed attribute feature set, the original structure feature set of the labeled nodes in the structure enhanced graph, and the reconstructed structure feature set, a feature reconstruction loss is determined.

9. The method of claim 1, wherein, Based on the classification loss, the reconstruction loss, and the semi-supervised learning loss of the unlabeled nodes, a joint training target loss of the model is determined by using the following formula: L = L CE + βL Rec + λL Semi Wherein, L represents the joint training target loss of the model; L CE represents the classification loss; L Rec represents the reconstruction loss; L Semi denotes the semi-supervised learning loss of the label-free node; β and λ are preset coefficients.

10. The method of claim 1, wherein, The attribute enhanced graph is determined by using a principal component analysis (PCA) algorithm. The structure enhanced graph is determined by using a K-nearest neighbor algorithm.

11. A target detection method characterized by, The method comprises: Inputting a to-be-tested object into a target detection model trained in advance by the model training method in any one of claims 1 to 10; Performing target detection on the to-be-tested object by the target detection model, and outputting a detection result.

12. A multi-view based model training apparatus, characterized by comprising: The device comprises: A multi-view construction unit configured to construct an attribute enhanced graph and a structure enhanced graph based on a multi-view of an input to-be-trained model; A fusion unit configured to determine a first graph feature after weighted fusion of different view features based on graph features of labeled nodes in the attribute enhanced graph and graph features of labeled nodes in the structure enhanced graph. The classification loss and reconstruction loss determination unit is configured to input the weighted and fused first graph feature into a classifier to obtain a predicted first classification result, and determine a classification loss based on the first classification result; and perform attribute feature reconstruction of the labeled nodes of the attribute enhanced graph and structure feature reconstruction of the labeled nodes of the structure enhanced graph based on the weighted and fused first graph feature, and determine a feature reconstruction loss based on the reconstruction results; The regularization unit is configured to regularize the model based on the graph features of the unlabeled nodes of the attribute enhanced graph and the graph features of the unlabeled nodes of the structure enhanced graph to obtain a regularization result; The semi-supervised learning loss determination unit is configured to determine a semi-supervised learning loss of the unlabeled nodes based on the regularization result; The target loss determination unit is configured to determine a joint training target loss of the model based on the classification loss, the reconstruction loss, and the semi-supervised learning loss of the unlabeled nodes.

13. A target detection apparatus characterized by comprising: The device comprises: The input object determination unit is configured to input a to-be-detected object into a target detection model trained by the model training method of any one of claims 1 to 10; The detection output unit is configured to perform target detection on the to-be-detected object by the target detection model and output a detection result.

14. An electronic device, comprising: The device comprises: A memory configured to store program instructions; A processor configured to invoke the program instructions stored in the memory to execute the method of any one of claims 1 to 11.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing the computer to execute the method of any one of claims 1 to 11.

Images