Subject state recognition method and device based on multi-dimensional model fusion framework

Abstract

The application discloses a subject state recognition method and device based on a multi-dimensional model fusion framework and electronic equipment, wherein the recognition method divides the data of a historical subject into multiple dimensional data sets; the data subsets in the data sets are combined in multiple dimensions to obtain multiple combination sets; the recognition models are trained according to the combination sets to obtain a model set containing multiple recognition models; the recognition models in the model set are fused to obtain a fusion recognition model; and the to-be-recognized subject is recognized through the fusion recognition model to determine the state of the to-be-recognized subject. According to the scheme, the recognition models of multiple dimensions are fused to obtain a fusion recognition model, the state of the to-be-recognized subject is determined through the fusion recognition model, the state of the to-be-recognized subject can be quickly and accurately supervised and evaluated, early warning information can be generated in time, and the information safety, data transmission safety and system safety of the subject and the service platform are ensured.

Description

Technical Field

[0001] This invention relates to the field of computer information processing, and more specifically, to a method, apparatus, and electronic device for subject state recognition based on a multi-dimensional model fusion framework. Background Technology

[0002] With the rapid development of the internet, more and more devices are using data on data platforms via the internet. These data platforms also need to manage massive amounts of device data. Providing matching data services to these devices and addressing the data security risks of both devices and internet platforms are pressing issues that need to be resolved. To ensure the security of information on devices and data platforms, various service providers are striving to analyze the current status of devices in order to provide early warnings when user and terminal device information is at risk.

[0003] Current scoring methods primarily employ two approaches: manual judgment and computer model judgment. Manual judgment relies heavily on human experience, is highly subjective, and prone to significant fluctuations. Computer model judgment, on the other hand, tends to be relatively one-dimensional and often lacks accuracy. Summary of the Invention

[0004] This invention aims to solve the problem that existing scoring methods have only one dimension and the results are not accurate enough.

[0005] To address the aforementioned technical problems, the first aspect of this invention proposes a subject state recognition method based on a multi-dimensional model fusion framework, the method comprising:

[0006] The obtained historical subject data is classified and divided into multi-dimensional datasets, which include multiple data subsets;

[0007] The data subsets in the dataset are combined in multiple dimensions to obtain multiple combined sets; each combined set includes at least two data subsets from datasets of different dimensions;

[0008] Train recognition models separately according to the combination set to obtain a model set containing multiple recognition models; fuse the recognition models in the model set to obtain a fused recognition model;

[0009] The fusion recognition model is used to identify the subject to be identified in order to determine the state of the subject.

[0010] According to a preferred embodiment of the present invention, a subset of data in the dataset is combined in multiple dimensions to obtain multiple combined sets, including:

[0011] Obtain at least one data subset from each dataset of different dimensions and use it as a combined subset;

[0012] The combined subsets are merged to form the combined set;

[0013] The combined sets are formed in a loop until a subset of data in each of the datasets is obtained at least once, resulting in multiple combined sets.

[0014] Specifically, the dimensions include: data source dimension, sample dimension, and state dimension.

[0015] According to a preferred embodiment of the present invention, the recognition models in the model set are fused to obtain a fused recognition model, specifically including:

[0016] The model recognition scores of each recognition model are binned at equal frequencies to obtain multiple bin features corresponding to the recognition model;

[0017] The various recognition models are combined to obtain multiple recognition model groups; each recognition model group includes at least two recognition models.

[0018] For each group of recognition models, the bin features of each bin in the recognition model of the model recognition group are cross-combined to generate multiple combined features and form a combined feature sequence.

[0019] The combined features in the combined feature sequence are encoded to obtain a combined feature vector corresponding to the combined feature sequence.

[0020] Based on the binning features and the combined feature vector, the recognition model is fused using a breadth-depth joint model to obtain a fused recognition model.

[0021] According to a preferred embodiment of the present invention, the binning features of each bin in the model recognition group are cross-combined to generate multiple combined features, forming a combined feature sequence, including:

[0022] The binning features of the two recognition models are combined in pairs to obtain multiple combined features;

[0023] The obtained combined features are arranged in order to obtain a combined feature sequence.

[0024] According to a preferred embodiment of the present invention, based on the binning features and the combined feature vector, a fused recognition model is obtained by fusing the recognition model using a breadth-and-depth joint model, including:

[0025] Input the historical data of each combined set into the corresponding recognition model to obtain the recognition results;

[0026] The identification result, the corresponding model identification score, and the corresponding combined feature vector are used as inputs, and the real result corresponding to the data is used as output to train the breadth model in the breadth and depth joint model.

[0027] Using the recognition result and the corresponding model recognition score as input, and the real result corresponding to the data as output, the deep model in the breadth and depth joint model is trained.

[0028] The trained breadth and depth joint model is obtained as the fusion recognition model.

[0029] A second aspect of this invention proposes a subject state recognition device based on a multi-dimensional model fusion framework, the device comprising:

[0030] The dimension partitioning module is used to classify the acquired historical subject data into datasets with multiple dimensions, and the datasets include multiple data subsets;

[0031] The recognition model training module is used to combine data subsets in the dataset in multiple dimensions to obtain multiple combination sets; each combination set includes at least two data subsets from datasets of different dimensions; and to train recognition models according to the combination sets to obtain a model set containing multiple recognition models.

[0032] The model fusion module is used to fuse the recognition models in the model set to obtain a fused recognition model;

[0033] The subject status recognition module is used to identify the subject to be identified through the fusion recognition model in order to determine the status of the subject to be identified.

[0034] According to a preferred embodiment of the present invention, a subset of data in the dataset is combined in multiple dimensions to obtain multiple combined sets, including:

[0035] Obtain at least one data subset from each dataset of different dimensions and use it as a combined subset;

[0036] The combined subsets are merged to form the combined set;

[0037] The combined sets are formed in a loop until a subset of data in each of the datasets is obtained at least once, resulting in multiple combined sets.

[0038] Specifically, the dimensions include: data source dimension, sample dimension, and state dimension.

[0039] According to a preferred embodiment of the present invention, the recognition models in the model set are fused to obtain a fused recognition model, specifically including:

[0040] The model recognition scores of each recognition model are binned at equal frequencies to obtain multiple bin features corresponding to the recognition model;

[0041] The various recognition models are combined to obtain multiple recognition model groups; each recognition model group includes at least two recognition models.

[0042] For each group of recognition models, the bin features of each bin in the recognition model of the model recognition group are cross-combined to generate multiple combined features and form a combined feature sequence.

[0043] The combined features in the combined feature sequence are encoded to obtain a combined feature vector corresponding to the combined feature sequence.

[0044] Based on the binning features and the combined feature vector, the recognition model is fused using a breadth-depth joint model to obtain a fused recognition model.

[0045] According to a preferred embodiment of the present invention, the binning features of each bin in the model recognition group are cross-combined to generate multiple combined features, forming a combined feature sequence, including:

[0046] The binning features of the two recognition models are combined in pairs to obtain multiple combined features;

[0047] The obtained combined features are arranged in order to obtain a combined feature sequence.

[0048] According to a preferred embodiment of the present invention, based on the binning features and the combined feature vector, a fused recognition model is obtained by fusing the recognition model using a breadth-and-depth joint model, including:

[0049] Input the historical data of each combined set into the corresponding recognition model to obtain the recognition results;

[0050] The identification result, the corresponding model identification score, and the corresponding combined feature vector are used as inputs, and the real result corresponding to the data is used as output to train the breadth model in the breadth and depth joint model.

[0051] Using the recognition result and the corresponding model recognition score as input, and the real result corresponding to the data as output, the deep model in the breadth and depth joint model is trained.

[0052] The trained breadth and depth joint model is obtained as the fusion recognition model.

[0053] A third aspect of the present invention provides an electronic device, wherein the electronic device comprises:

[0054] Processor; and,

[0055] A memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the method.

[0056] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores one or more programs that, when executed by a processor, implement the method.

[0057] This technical solution involves dividing data into multiple dimensions, building a recognition model based on these dimensions, and then fusing these models to obtain a fusion model. This significantly improves the accuracy of score calculations and enhances user satisfaction. Attached Figure Description

[0058] To make the technical problems solved by this invention, the technical means adopted, and the technical effects achieved clearer, specific embodiments of this invention will be described in detail below with reference to the accompanying drawings. However, it should be noted that the drawings described below are merely drawings of exemplary embodiments of this invention. Those skilled in the art can obtain drawings of other embodiments based on these drawings without any creative effort.

[0059] Figure 1 This is a flowchart illustrating the subject state recognition method based on a multi-dimensional model fusion framework in an embodiment of the present invention.

[0060] Figure 2 This is a schematic diagram of the main state recognition device based on a multi-dimensional model fusion framework in an embodiment of the present invention. Detailed Implementation

[0061] Exemplary embodiments of the invention will now be described more fully with reference to the accompanying drawings. While these exemplary embodiments can be implemented in various specific ways, they should not be construed as limiting the invention to the embodiments set forth herein. Rather, these exemplary embodiments are provided to make the content of the invention more complete and to facilitate a full communication of the inventive concept to those skilled in the art.

[0062] The same reference numerals in the accompanying drawings denote the same or similar elements, components, or parts, and therefore, repeated descriptions of the same or similar elements, components, or parts may be omitted below. It should also be understood that although terms such as first, second, third, etc., indicating numbers may be used herein to describe various devices, elements, components, or parts, these devices, elements, components, or parts should not be limited by these terms. That is, these terms are only used to distinguish one from another. For example, a first device may also be referred to as a second device, without departing from the essential technical solution of the invention. Furthermore, the terms "and / or" and "and / or" refer to all combinations including any one or more of the listed items.

[0063] Figure 1 This is a flowchart illustrating a subject state recognition method based on a multi-dimensional model fusion framework according to the present invention, as shown below. Figure 1 As shown, the method of the present invention comprises the following steps:

[0064] S101. Classify the obtained historical subject data into multi-dimensional datasets, wherein the datasets include multiple data subsets.

[0065] Model training is based on historical subject data that has already been obtained. However, since there are many types of data, it is necessary to classify the data in order to better train the model.

[0066] Based on the above technical solution, further, the data subsets in the dataset are combined in multiple dimensions to obtain multiple combined sets, including:

[0067] Obtain at least one data subset from each dataset of different dimensions and use it as a combined subset;

[0068] The combined subsets are merged to form the combined set;

[0069] The combined sets are formed in a loop until a subset of data in each of the datasets is obtained at least once, resulting in multiple combined sets.

[0070] Specifically, these multiple dimensions include data source dimensions, sample dimensions, and state dimensions. In this embodiment, the data is divided into three dimensions: the data source dimension, which is mainly distinguished by different data sources, such as video websites, news websites, financial websites, etc.; the sample dimension, which is mainly distinguished by different user identity information, such as age, gender, education, etc., filled in when registering on a website; and the state dimension, which is the user's status after applying for the service, such as not applying, application approved and used, application approved but not used, etc. In other embodiments, it can also be divided into four, five, or more dimensions. Relatively speaking, the more dimensions there are, the more accurate the judgment results of the trained model.

[0071] In this implementation, dimensionality partitioning is performed using a dimensionality partitioning model, specifically the TextCNN deep learning-based network model. This model is trained using historical main datasets. A portion of the data is manually indexed, with corresponding dimensional labels assigned to the indexed data. The indexed data is then divided into training and validation data.

[0072] Train the dimension partitioning model using the training data, adjusting the model parameters until the session analysis model converges. Then, validate the dimension partitioning model using validation samples. If the validation passes, the dimension partitioning model training is complete; otherwise, retrain the dimension partitioning model.

[0073] S102. Combine the data subsets in the dataset in multiple dimensions to obtain multiple combination sets; each combination set includes at least two data subsets from datasets of different dimensions; train the recognition model according to the combination set to obtain a model set containing multiple recognition models.

[0074] In this implementation, since there are multiple data subsets for each dimension, one subset is selected from each dimension when training the recognition model. For example, the data source dimension selects a subset of video website data, the sample dimension selects a subset of educational background data, and the status dimension selects a subset of application approval and usage data. The selection method is arbitrary, and the combination method is also arbitrary.

[0075] Based on the above technical solution, the training recognition model is further defined as follows:

[0076]

[0077] Among them, X i For any subset of data in the data source dimension, S j For any subset of data in the sample dimension, Y k For any subset of data in the state dimension, The mathematical notation for tensor product is . In mathematics, tensor product can be applied to different contexts such as vectors, matrices, tensors, vector spaces, algebra, topological vector spaces, and modulo.

[0078] In this embodiment, the Cartesian product is applied to the data subsets of the three dimensions to obtain the recognition model set {m1, m2, m3…m n}

[0079] S103. The recognition models in the model set are fused to obtain a fused recognition model.

[0080] In this implementation, a subset of data from three dimensions is selected based on the business requirements scenario, and a fusion model is trained based on the obtained recognition model set.

[0081] Based on the above technical solution, further, the fusion of the recognition models in the model set to obtain the fused recognition model specifically includes:

[0082] S1031. Perform equal-frequency binning on the model recognition scores of each recognition model to obtain multiple binning features corresponding to the recognition model.

[0083] In this embodiment, the model recognition score of the recognition model can be the accuracy of the recognition model in recognizing the data, or the AUC value of the recognition model; this solution does not impose any particular limitation on this. The model recognition score of the recognition model is binned using equal frequency binning. In other embodiments, equal-interval binning can also be used. Relatively speaking, models trained using equal frequency binning produce more accurate judgment results.

[0084] In this embodiment, for each recognition model m i The model identification is divided into 10 equal-frequency bins. For example, the binned values ​​of model m1 are: (-inf, 0.988], (0.996, 0.997], (0.997, 0.998], (0.988, 0.991], (0.995, 0.996], (0.993, 0.995], (0.991, 0.993], (0.998, in f). The values ​​of the m2 model after binning are: (-inf,0.959],(0.982,0.985],(0.985,0.988],(0.988,0.991],(0.959,0.969],(0.969,0.975],(0.975,0.979],(0.979,0.982],(0.991,inf).

[0085] S1032. Combine the various recognition models to obtain multiple recognition model groups; each recognition model group includes at least two recognition models.

[0086] In this implementation, the model set {m1,m2,m3…m} is identified. n There are n recognition models. Any two models can be randomly selected and combined, for example, m1 & m2, m1 & m3, ..., m n-1 &m n .

[0087] In other implementations, three or more recognition models can be combined.

[0088] S1033. For each group of recognition models, the bin features of each bin in the recognition model of the model recognition group are cross-combined to generate multiple combined features and form a combined feature sequence.

[0089] In this embodiment, the recognition model set {m1,m2,m3…m n The binned values ​​of these n recognition models are combined in pairs. Taking m1 and m2 as an example, the feature values ​​of m1 & m2 are combined, such as: (-inf,0.988] & (-inf,0.959], (-inf,0.988] & (0.982,0.985], etc. Since different orders are considered as the same combination, a total of 100 values ​​are obtained by 10*10.

[0090] Therefore, the model set {m1,m2,m3…m} is identified. n There are a total of n 2 The possible values.

[0091] S1034. Encode the combined features in the combined feature sequence to obtain the combined feature vector corresponding to the combined feature sequence.

[0092] In this embodiment, the combined features are one-hot encoded. Taking the m1&m2 features as an example, since there are 100 possible values, they can be converted into 100-dimensional features with values ​​of 0 and 1. The other m... i &m j They have similar characteristics.

[0093] S1035. Based on the features of the recognition model and / or the combined features after encoding processing, the recognition model is fused using a breadth-depth joint model to obtain a fused model.

[0094] This implementation uses a joint breadth-and-depth model to fuse the recognition model. Other implementations may employ different fusion methods, such as using n recognition models as features to obtain a column vector V. m [m1,m2,m3…m n Train the fusion model M, M = XGboost(V i S j Y k ).

[0095] Based on the above technical solution, further, based on the binning features and the combined feature vector, a fused recognition model is obtained by fusing the recognition model using a breadth-and-depth joint model, including:

[0096] Input the historical data of each combined set into the corresponding recognition model to obtain the recognition results;

[0097] The identification result, the corresponding model identification score, and the corresponding combined feature vector are used as inputs, and the real result corresponding to the data is used as output to train the breadth model in the breadth and depth joint model.

[0098] Using the recognition result and the corresponding model recognition score as input, and the real result corresponding to the data as output, the deep model in the breadth and depth joint model is trained.

[0099] The trained breadth-and-depth joint model is obtained and used as the fusion recognition model. In this embodiment, the breadth model is a linear regression model.

[0100] In this implementation, the breadth model possesses memory capabilities, which can be interpreted as learning frequently co-occurring features and uncovering co-occurrences in historical data. Therefore, the breadth model remembers common, high-frequency models from the historical data center; specifically, it remembers the relationship between the binning combination of the two recognition models in the training data and credit security.

[0101] In this embodiment, the breadth model is:

[0102]

[0103] Among them, W wide b is the weight vector of the breadth-first model. wide This represents the offset of the breadth model.

[0104] X wide The input to the breadth model is the recognition result of the recognition model for the historical subject data in the above real-time example, the recognition model score of the recognition model, and the combined feature vector corresponding to the recognition model. y1 is the label corresponding to the real result of the historical subject data.

[0105] In this embodiment, the input features are the recognition results of the recognition model for historical subject data, the recognition model's segmentation and combination features, i.e., [m1, m2, m3, ... m n ,m1&m2,m1&m3,…m n-1 &m n ], with dimensions n+n 2 W wide The dimension is the same as the input feature. wide The initial value of the offset is a preset constant, which is iterated to a new value each time during the training process.

[0106] In this embodiment, the deep model is a neural network model.

[0107] In this embodiment, the deep model has the ability to expand, which is based on transfer relevance and explores new feature combinations that have hardly appeared before.

[0108] In this embodiment, the depth model is:

[0109] y2=ReLU(W deep X deep +b deep ),

[0110] Among them, W deep b is the weight vector of the deep model. deep This represents the offset of the depth model.

[0111] X deep y1 is the input of the deep model, that is, the recognition result of the recognition model for the historical subject data in the above real-time example, the recognition model score of the recognition model, y2 is the label corresponding to the real result of the historical subject data, and ReLU() is the linear rectification function, also known as the corrected linear unit, which is a commonly used activation function in artificial neural networks, usually referring to nonlinear functions represented by ramp functions and their variants.

[0112] In this embodiment, the input features of the deep model are the recognition results of the recognition model for historical subject data and the model recognition score of the recognition model, i.e., [m1,m2,m3,…,m n ], with a dimension of n. W deep The dimension is the same as the input feature.

[0113] In this embodiment, the fusion model is:

[0114] P = sigmoid(y1) + y2.

[0115] Where P is the recognition result obtained by fusing the recognition results of the breadth model and the depth model; sigmoid(y1) is the Sigmoid function calculation performed on y1.

[0116] Breadth-based models have a memory capacity; for example, they remember that eagles and sparrows have wings and can fly. However, breadth-based models only remember features from historical data; they won't make judgments about data or features not present in historical data. Deep models, on the other hand, have an expansion capacity. For example, based on the fact that eagles and sparrows have wings and can fly, they can expand to include all winged animals that can fly. Deep models can extend the scope of judgment from historical data to new data. However, often the expansion is incorrect; for example, ostriches have wings but cannot fly. Therefore, breadth-based models are needed to limit and correct the expansion of deep models. Combining breadth-based and deep models effectively addresses the shortcomings of both. The resulting fusion model can expand based on historical data without leading to erroneous conclusions.

[0117] S104. The subject to be identified is identified using the fusion recognition model to determine the state of the subject to be identified.

[0118] In this embodiment, when a new user requests a service, the new user's data is obtained and input into the recognition model set {m1,m2,m3…m}. n Each recognition model m in} i Obtain the set of recognition model scores [m1, m2, m3…m n Then, based on the recognition model, the breadth model y1 and the depth model y2 are calculated respectively. The fusion model is used to obtain P, which is the user's credit security score. Services are then allocated to the user based on the score.

[0119] Figure 2 This is a schematic diagram of the main state recognition device based on a multi-dimensional model fusion framework in an embodiment of the present invention, as shown below. Figure 2 As shown, the present invention provides a subject state recognition device 200 based on a multi-dimensional model fusion framework, comprising:

[0120] The dimension partitioning module 201 is used to classify the obtained historical subject data into a multi-dimensional dataset, which includes multiple data subsets.

[0121] Model training is based on historical subject data that has already been obtained. However, since there are many types of data, it is necessary to classify the data in order to better train the model.

[0122] Based on the above technical solution, further, the data subsets in the dataset are combined in multiple dimensions to obtain multiple combined sets, including:

[0123] Obtain at least one data subset from each dataset of different dimensions and use it as a combined subset;

[0124] The combined subsets are merged to form the combined set;

[0125] The combined sets are formed in a loop until a subset of data in each of the datasets is obtained at least once, resulting in multiple combined sets.

[0126] Specifically, these multiple dimensions include data source dimensions, sample dimensions, and state dimensions. In this embodiment, the data is divided into three dimensions: the data source dimension, which is mainly distinguished by different data sources, such as video websites, news websites, financial websites, etc.; the sample dimension, which is mainly distinguished by different user identity information, such as age, gender, education, etc., filled in when registering on a website; and the state dimension, which is the user's status after applying for the service, such as not applying, application approved and used, application approved but not used, etc. In other embodiments, it can also be divided into four, five, or more dimensions. Relatively speaking, the more dimensions there are, the more accurate the judgment results of the trained model.

[0127] In this implementation, dimensionality partitioning is performed using a dimensionality partitioning model, specifically the TextCNN deep learning-based network model. This model is trained using historical main datasets. A portion of the data is manually indexed, with corresponding dimensional labels assigned to the indexed data. The indexed data is then divided into training and validation data.

[0128] Train the dimension partitioning model using the training data, adjusting the model parameters until the session analysis model converges. Then, validate the dimension partitioning model using validation samples. If the validation passes, the dimension partitioning model training is complete; otherwise, retrain the dimension partitioning model.

[0129] The recognition model training module 202 is used to combine data subsets in the dataset in multiple dimensions to obtain multiple combination sets; each combination set includes at least two data subsets from datasets of different dimensions; and to train recognition models according to the combination sets to obtain a model set containing multiple recognition models.

[0130] In this implementation, since there are multiple data subsets for each dimension, one subset is selected from each dimension when training the recognition model. For example, the data source dimension selects a subset of video website data, the sample dimension selects a subset of educational background data, and the status dimension selects a subset of application approval and usage data. The selection method is arbitrary, and the combination method is also arbitrary.

[0131] Based on the above technical solution, the training recognition model is further defined as follows:

[0132]

[0133] Among them, X iFor any subset of data in the data source dimension, S j For any subset of data in the sample dimension, Y k For any subset of data in the state dimension, The mathematical notation for tensor product is . In mathematics, tensor product can be applied to different contexts such as vectors, matrices, tensors, vector spaces, algebra, topological vector spaces, and modulo.

[0134] In this embodiment, the Cartesian product is applied to the data subsets of the three dimensions to obtain the recognition model set {m1, m2, m3…m}. n}

[0135] The model fusion module 203 is used to fuse the recognition models in the model set to obtain a fused recognition model.

[0136] In this implementation, a subset of data from three dimensions is selected based on the business requirements scenario, and a fusion model is trained based on the obtained recognition model set.

[0137] Based on the above technical solution, further, the fusion of the recognition models in the model set to obtain the fused recognition model specifically includes:

[0138] The binning module 2031 performs equal-frequency binning on the model recognition scores of each recognition model to obtain multiple binning features corresponding to the recognition model.

[0139] In this embodiment, the model recognition score of the recognition model can be the accuracy of the recognition model in recognizing the data, or the AUC value of the recognition model; this solution does not impose any particular limitation on this. The model recognition score of the recognition model is binned using equal frequency binning. In other embodiments, equal-interval binning can also be used. Relatively speaking, models trained using equal frequency binning produce more accurate judgment results.

[0140] In this embodiment, for each recognition model m iThe model identification is divided into 10 equal-frequency bins. For example, the binned values ​​of model m1 are: (-inf, 0.988], (0.996, 0.997], (0.997, 0.998], (0.988, 0.991], (0.995, 0.996], (0.993, 0.995], (0.991, 0.993], (0.998, in f). The values ​​of the m2 model after binning are: (-inf,0.959],(0.982,0.985],(0.985,0.988],(0.988,0.991],(0.959,0.969],(0.969,0.975],(0.975,0.979],(0.979,0.982],(0.991,inf).

[0141] The model combination submodule 2032 combines the various recognition models to obtain multiple recognition model groups; each recognition model group includes at least two recognition models.

[0142] In this implementation, the model set {m1,m2,m3…m} is identified. n There are n recognition models. Any two models can be randomly selected and combined, for example, m1 & m2, m1 & m3, ..., m n-1 &m n .

[0143] In other implementations, three or more recognition models can be combined.

[0144] Feature combination submodule 2033: For each group of recognition models, cross-combine the bin features of each recognition model in the model recognition group to generate multiple combined features and form a combined feature sequence.

[0145] In this embodiment, the recognition model set {m1,m2,m3…m n The binned values ​​of these n recognition models are combined in pairs. Taking m1 and m2 as an example, the feature values ​​of m1 & m2 are combined, such as: (-inf,0.988] & (-inf,0.959], (-inf,0.988] & (0.982,0.985], etc. Since different orders are considered as the same combination, a total of 100 values ​​are obtained by 10*10.

[0146] Therefore, the model set {m1,m2,m3…m} is identified. n There are a total of n 2 The possible values.

[0147] The encoding processing submodule 2034 encodes the combined features in the combined feature sequence to obtain the combined feature vector corresponding to the combined feature sequence.

[0148] In this embodiment, the combined features are one-hot encoded. Taking the m1&m2 features as an example, since there are 100 possible values, they can be converted into 100-dimensional features with values ​​of 0 and 1. The other m... i &m j They have similar characteristics.

[0149] The fusion submodule 2035, based on the features of the recognition model and / or the combined features after encoding processing, uses a breadth-depth joint model to fuse the recognition model to obtain a fusion model.

[0150] This implementation uses a joint breadth-and-depth model to fuse the recognition model. Other implementations may employ different fusion methods, such as using n recognition models as features to obtain a column vector V. m [m1,m2,m3…m n Train the fusion model M, M = XGboost(V i S j Y k ).

[0151] Based on the above technical solution, further, based on the binning features and the combined feature vector, a fused recognition model is obtained by fusing the recognition model using a breadth-and-depth joint model, including:

[0152] Input the historical data of each combined set into the corresponding recognition model to obtain the recognition results;

[0153] The identification result, the corresponding model identification score, and the corresponding combined feature vector are used as inputs, and the real result corresponding to the data is used as output to train the breadth model in the breadth and depth joint model.

[0154] Using the recognition result and the corresponding model recognition score as input, and the real result corresponding to the data as output, the deep model in the breadth and depth joint model is trained.

[0155] The trained breadth-and-depth joint model is obtained and used as the fusion recognition model. In this embodiment, the breadth model is a linear regression model.

[0156] In this implementation, the breadth model possesses memory capabilities, which can be interpreted as learning frequently co-occurring features and uncovering co-occurrences in historical data. Therefore, the breadth model remembers common, high-frequency models from the historical data center; specifically, it remembers the relationship between the binning combination of the two recognition models in the training data and credit security.

[0157] In this embodiment, the breadth model is:

[0158]

[0159] Among them, W wide b is the weight vector of the breadth-first model. wide This represents the offset of the breadth model.

[0160] X wide The input to the breadth model is the recognition result of the recognition model for the historical subject data in the above real-time example, the recognition model score of the recognition model, and the combined feature vector corresponding to the recognition model. y1 is the label corresponding to the real result of the historical subject data.

[0161] In this embodiment, the input features are the recognition results of the recognition model for historical subject data, the recognition model's segmentation and combination features, i.e., [m1, m2, m3, ... m n ,m1&m2,m1&m3,…m n-1 &m n ], with dimensions n+n 2 W wide The dimension is the same as the input feature. wide The initial value of the offset is a preset constant, which is iterated to a new value each time during the training process.

[0162] In this embodiment, the deep model is a neural network model.

[0163] In this embodiment, the deep model has the ability to expand, which is based on transfer relevance and explores new feature combinations that have hardly appeared before.

[0164] In this embodiment, the depth model is:

[0165] y2=ReLU(W deep X deep +b deep ),

[0166] Among them, W deep b is the weight vector of the deep model. deep This represents the offset of the depth model.

[0167] X deepy1 is the input of the deep model, that is, the recognition result of the recognition model for the historical subject data in the above real-time example, the recognition model score of the recognition model, y2 is the label corresponding to the real result of the historical subject data, and ReLU() is the linear rectification function, also known as the corrected linear unit, which is a commonly used activation function in artificial neural networks, usually referring to nonlinear functions represented by ramp functions and their variants.

[0168] In this embodiment, the input features of the deep model are the recognition results of the recognition model for historical subject data and the model recognition score of the recognition model, i.e., [m1,m2,m3,…,m n ], with a dimension of n. W deep The dimension is the same as the input feature.

[0169] In this embodiment, the fusion model is:

[0170] P = sigmoid(y1) + y2.

[0171] Where P is the recognition result obtained by fusing the recognition results of the breadth model and the depth model; sigmoid(y1) is the Sigmoid function calculation performed on y1.

[0172] Breadth-based models have a memory capacity; for example, they remember that eagles and sparrows have wings and can fly. However, breadth-based models only remember features from historical data; they won't make judgments about data or features not present in historical data. Deep models, on the other hand, have an expansion capacity. For example, based on the fact that eagles and sparrows have wings and can fly, they can expand to include all winged animals that can fly. Deep models can extend the scope of judgment from historical data to new data. However, often the expansion is incorrect; for example, ostriches have wings but cannot fly. Therefore, breadth-based models are needed to limit and correct the expansion of deep models. Combining breadth-based and deep models effectively addresses the shortcomings of both. The resulting fusion model can expand based on historical data without leading to erroneous conclusions.

[0173] The subject state recognition module 204 is used to recognize the subject to be identified through the fusion recognition model in order to determine the state of the subject to be identified.

[0174] In this embodiment, when a new user requests a service, the new user's data is obtained and input into the recognition model set {m1,m2,m3…m}. n Each recognition model m in} i Obtain the set of recognition model scores [m1, m2, m3…m nThen, based on the recognition model, the breadth model y1 and the depth model y2 are calculated respectively. The fusion model is used to obtain P, which is the user's credit security score. Services are then allocated to the user based on the score.

[0175] An embodiment of the present invention provides an electronic device comprising a processor and a memory, the memory being used to store a computer-executable program, wherein when the computer program is executed by the processor, the processor executes a configurable data query platform.

[0176] The electronic device is manifested in the form of a general-purpose computing device. It may contain one or more processors that work collaboratively. This invention also does not preclude distributed processing, meaning that processors may be distributed across different physical devices. The electronic device of this invention is not limited to a single entity, but may also be the sum of multiple physical devices.

[0177] The memory stores a computer-executable program, typically machine-readable code. The computer-readable program can be executed by the processor to enable the electronic device to perform the method of the present invention, or at least some steps of the method.

[0178] The memory includes volatile memory, such as random access memory (RAM) and / or cache memory, and may also be non-volatile memory, such as read-only memory (ROM).

[0179] Optionally, in this embodiment, the electronic device further includes an I / O interface for exchanging data with external devices. The I / O interface can represent one or more of several bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0180] It should be understood that the electronic device is merely one example of the present invention, and the electronic device of the present invention may also include elements or components not shown in the above examples. For example, some electronic devices also include display units such as displays, and some electronic devices also include human-computer interaction elements such as buttons and keyboards. As long as the electronic device can execute a computer-readable program in its memory to implement the method of the present invention or at least some steps of the method, it can be considered as an electronic device covered by the present invention.

[0181] An embodiment of the present invention provides a computer-readable recording medium. The computer-readable recording medium stores a computer-executable program, which, when executed, implements the configurable data query platform described above. The computer-readable storage medium may include a data signal propagated in baseband or as part of a carrier wave, carrying readable program code. This propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0182] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0183] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the present invention is not inherently related to any specific computer, virtual device, or electronic device, and various general-purpose devices can also implement the present invention. The above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A subject state recognition method based on a multi-dimensional model fusion framework, characterized in that, The methods include: The obtained historical subject data is classified and divided into multi-dimensional datasets, which include multiple data subsets; The data subsets in the dataset are combined in multiple dimensions to obtain multiple combined sets; each combined set includes at least two data subsets from datasets of different dimensions; Train recognition models according to the combination set to obtain a model set containing multiple recognition models; The recognition models in the model set are fused to obtain a fused recognition model, specifically including: The model recognition scores of each recognition model are binned at equal frequencies to obtain multiple bin features corresponding to the recognition model; The various recognition models are combined to obtain multiple sets of recognition model groups; The identification model group includes at least two identification models; For each group of recognition models, the bin features of each bin in the recognition model of the model recognition group are cross-combined to generate multiple combined features and form a combined feature sequence. The combined features in the combined feature sequence are encoded to obtain a combined feature vector corresponding to the combined feature sequence. Based on the binning features and the combined feature vector, the recognition model is fused using a breadth-and-depth joint model to obtain a fused recognition model; The fusion recognition model is used to identify the subject to be identified in order to determine the state of the subject.

2. The method as described in claim 1, characterized in that, By combining subsets of data from the dataset in multiple dimensions, several combined sets are obtained, including: Obtain at least one data subset from each dataset of different dimensions and use it as a combined subset; The combined subsets are merged to form the combined set; The combined sets are formed in a loop until a subset of data in each of the datasets is obtained at least once, resulting in multiple combined sets. Specifically, the dimensions include: data source dimension, sample dimension, and state dimension.

3. The method as described in claim 1 or 2, characterized in that, The binning features of each bin in the model recognition group are cross-combined to generate multiple combined features, forming a combined feature sequence, including: The binning features of the two recognition models are combined in pairs to obtain multiple combined features; The obtained combined features are arranged in order to obtain a combined feature sequence.

4. The method as described in claim 1 or 2, characterized in that, Based on the binning features and the combined feature vector, a fused recognition model is obtained by fusing the recognition model using a breadth-and-depth joint model, including: Input the historical data of each combined set into the corresponding recognition model to obtain the recognition results; The identification result, the corresponding model identification score, and the corresponding combined feature vector are used as inputs, and the real result corresponding to the data is used as output to train the breadth model in the breadth and depth joint model. Using the recognition result and the corresponding model recognition score as input, and the real result corresponding to the data as output, the deep model in the breadth and depth joint model is trained. The trained breadth and depth joint model is obtained as the fusion recognition model.

5. A subject state recognition device based on a multi-dimensional model fusion framework, characterized in that, The device includes: The dimension partitioning module is used to classify the acquired historical subject data into datasets with multiple dimensions, and the datasets include multiple data subsets; The recognition model training module is used to combine data subsets from the dataset in multiple dimensions to obtain multiple combined sets; each combined set includes at least two data subsets from datasets of different dimensions; Train recognition models according to the combination set to obtain a model set containing multiple recognition models; The model fusion module is used to fuse the recognition models in the model set to obtain a fused recognition model. Specifically, it includes: performing equal-frequency binning on the model recognition scores of each recognition model to obtain multiple binning features corresponding to the recognition model. The various recognition models are combined to obtain multiple recognition model groups; each recognition model group includes at least two recognition models; for each recognition model group, the binning features of each recognition model in the model recognition group are cross-combined to generate multiple combined features, forming a combined feature sequence; the combined features in the combined feature sequence are encoded to obtain a combined feature vector corresponding to the combined feature sequence; based on the binning features and the combined feature vector, the recognition models are fused using a breadth-and-depth joint model to obtain a fused recognition model; The subject status recognition module is used to identify the subject to be identified through the fusion recognition model in order to determine the status of the subject to be identified.

6. The apparatus as claimed in claim 5, characterized in that, By combining subsets of data from the dataset in multiple dimensions, several combined sets are obtained, including: Obtain at least one data subset from each dataset of different dimensions and use it as a combined subset; The combined subsets are merged to form the combined set; The combined sets are formed in a loop until a subset of data in each of the datasets is obtained at least once, resulting in multiple combined sets. Specifically, the dimensions include: data source dimension, sample dimension, and state dimension.

7. The apparatus as described in claim 5 or 6, characterized in that, The binning features of each bin in the model recognition group are cross-combined to generate multiple combined features, forming a combined feature sequence, including: The binning features of the two recognition models are combined in pairs to obtain multiple combined features; The obtained combined features are arranged in order to obtain a combined feature sequence.

8. The apparatus as described in claim 5 or 6, characterized in that, Based on the binning features and the combined feature vector, a fused recognition model is obtained by fusing the recognition model using a breadth-and-depth joint model, including: Input the historical data of each combined set into the corresponding recognition model to obtain the recognition results; The identification result, the corresponding model identification score, and the corresponding combined feature vector are used as inputs, and the real result corresponding to the data is used as output to train the breadth model in the breadth and depth joint model. Using the recognition result and the corresponding model recognition score as input, and the real result corresponding to the data as output, the deep model in the breadth and depth joint model is trained. The trained breadth and depth joint model is obtained as the fusion recognition model.

9. An electronic device, wherein, The electronic device includes: Processor; and, A memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the method according to any one of claims 1 to 4.

10. A computer-readable storage medium, wherein, The computer-readable storage medium stores one or more programs that, when executed by a processor, implement the method of any one of claims 1 to 4.

Images