Classification method and device

[0071] In order to enable those skilled in the art to better understand the solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0072] Some processes described in the specification, claims, and the above-mentioned drawings of the present application include multiple operations appearing in a specific order, and these operations may be performed out of the order in which they appear in this document or performed in parallel. The sequence numbers of operations, such as 101, 102, etc., are only used to distinguish different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions of "first" and "second" in this article are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, nor do they limit the "first" and "second" Are different types.

[0073] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative work shall fall within the protection scope of this application.

[0074] figure 1 It shows a schematic flowchart of a classification method provided by an embodiment of the present application. The method includes:

[0075] 10. Obtain the characteristic value corresponding to the specified data item of the user.

[0076] 11. Input the characteristic value corresponding to the specified data item as input into the target classifier to obtain the classification result of the target classifier.

[0077] 12. Output the classification result.

[0078] In the above 10, the feature value corresponding to the specified data item corresponding to the user may be obtained from the user database; or, the feature value corresponding to the specified data item input by the user on the input interface may be obtained; or, by the specified image input by the user Perform image recognition to get the feature value corresponding to the specified data item. For example: the e-commerce platform will analyze the registration information and historical behavior information of all users on the platform to obtain the characteristic values ​​corresponding to the specified data items corresponding to all users on the platform, and store the characteristic values ​​corresponding to the specified data items corresponding to all users In the user database, in this way, the characteristic value corresponding to the specified data item corresponding to the user can be directly obtained from the user database in the subsequent.

[0079] In the above 11, the feature value corresponding to the specified data item is directly input into the target classifier, and the target classifier will directly output the classification result.

[0080] In the above 12, the classification result may be output through a display screen or output through a voice broadcast, which is not specifically limited in the embodiment of the present application.

[0081] The construction process of the aforementioned target classifier may include the following steps:

[0082] 101. Acquire the first feature item traversed in the first feature set.

[0083] 102. Combine the cross-validation method and the grid search method to test the feature item combination composed of the first feature item and the second feature set to obtain the highest classification accuracy rate of the feature item combination, and the feature item combination The parameter structure corresponding to the highest classification accuracy rate.

[0084] 103. If the highest classification accuracy rate of the feature item combination is higher than the highest classification accuracy rate of the second feature set, add the first feature item to the second feature set.

[0085] 104. Determine whether there is a feature item that has not been traversed in the first feature set, and if there is, continue to traverse the next feature item in the first feature set until the traversal ends.

[0086] 105. Construct a target classifier based on the second feature set at the end of the traversal and the parameter structure corresponding to the highest classification accuracy.

[0087] Wherein, the designated data item includes all data items in the second data set at the end of the traversal.

[0088] In 101 above, in different application scenarios, the content referred to by the feature items in the first feature set may also be different. For example: in the application scenario of e-commerce, the feature items in the first feature set refer to user gender, user age group, user consumption level, user education level, user size, user brand preference, user relationship network, and interest Hobbies, psychological characteristics, etc. For another example: in an application scenario of user occupation classification or tinnitus classification, the feature items in the first feature set refer to different brain regions.

[0089] What needs to be added is that research has found that due to different working environments and different ways of thinking, different professional users have different brain regions (such as brain volume); tinnitus is usually accompanied by Changes in the morphology of the brain area (eg brain area volume)

[0090] In the above 102, if the second feature set is an empty set, only the first feature item is included in the feature item combination; if the second feature set is not an empty set, then the feature item combination includes not only the first feature item, but also the first feature item. 2. All feature items in the feature set.

[0091] Among them, the basic idea of ​​the cross-validation method is to group the original data set, one part is used as a training set (train set), and the other part is used as a validation set or test set. First, use the training set to perform the construction of the classifier Training, and then use the validation set to test the trained classifier, which is used as the performance index for evaluating the classifier. The grid search method divides the kernel function parameters to be searched into grids in a certain space range, and finds the optimal kernel function parameters by traversing all points in the grid.

[0092] Combine the cross-validation method and the grid search method to test the feature item combination, that is, combine the cross-validation method and the grid search method to test the classifier based on the feature item combination at the same time. The test process includes the training and testing process. In the training process, combine the training set and test set obtained by grouping the original data set with the cross-validation method, and use the training set to train the constructed classifier to obtain the trained classifier; in the test process, use grid search Method of testing the trained classifier on the test set to obtain multiple classification accuracy rates and the kernel function parameter values ​​corresponding to the multiple classification accuracy rates, and the maximum value of the multiple classification accuracy rates and the kernel function parameter values ​​corresponding to the maximum value The highest classification accuracy rate as a feature combination and its corresponding kernel function parameter value.

[0093] For example: a parameter structure determined by the cross-validation method and the grid search method is the parameter structure A. If the classifier constructed based on the parameter structure A has the highest classification accuracy, then the highest classification accuracy of the feature combination The corresponding parameter structure is the parameter structure A.

[0094] In the above 103, the highest classification accuracy rate of the second feature set recorded is obtained. If the highest classification accuracy rate of the feature item combination is higher than the highest classification accuracy rate of the second feature set, it indicates that the addition of the first feature item is beneficial to improve the classification Accuracy, therefore, the first feature item needs to be added to the second feature set.

[0095] It should be noted that when the second feature set is an empty set, the highest classification accuracy rate corresponding to the second feature set may be 0 by default. In an achievable solution, before traversing, the second feature set can be initialized to an empty set.

[0096] In 104 above, if there are still feature items that have not been traversed in the first feature set, it is necessary to continue to traverse the next feature item in the first feature set until the traversal ends; if there is no feature item that has not been traversed in the first feature set The feature items, that is, the feature items in the first feature are all traversed, the traversal ends.

[0097] In 105 above, at the end of the traversal, all the feature items that are helpful to improve the classification accuracy are retained in the second feature set, and the parameter structure corresponding to the highest classification accuracy of the second feature set is also the optimal parameter structure.

[0098] In the technical solutions provided by the embodiments of the present application, the most representative and limited number of feature items are extracted from a large number of feature items through traversal and inspection; moreover, in the process of feature item screening, it is also achieved Optimization of the kernel function parameters of the classifier. The target classifier constructed based on the most representative and limited number of feature items and the optimized kernel function parameters is more scientific and reasonable, and its prediction results have higher reference value.

[0099] In an achievable solution, the target classifier may be a Support Vector Machine (SVM) classifier. The kernel function in the support vector machine can be selected according to specific data. The applicant found in research that the radial basis kernel (RBF) kernel function is the most suitable kernel function in the application scenario of tinnitus. Specifically, the kernel function in the support vector machine is an RBF kernel function; the RBF kernel function includes a parameter C and a parameter γ; the search range of the grid search method for the values ​​of the parameter C and the parameter γ can be defined as: log 2 C={-5,-4,…,4,5} and log 2 γ={-5,-4,...,4,5}.

[0100] The above cross-validation method is also called K-fold cross-validation method. For example: the above cross-validation method is a 5-fold cross-validation method.

[0101] Generally, at least two types of objects are involved in a classification scene. For the convenience of introduction, the following will introduce two types of objects as examples:

[0102] The above method may also include:

[0103] 106. Obtain feature values ​​corresponding to the Q feature items of each of the N first-type objects and M second-type objects.

[0104] Wherein, N, M, and Q are positive integers; the first feature set includes the Q feature items.

[0105] In the above 106, (N+M)*Q data can be obtained. According to the (N+M)*Q data, a matrix of N+M rows and Q columns is generated, and each row corresponds to Q feature items of an object The characteristic value.

[0106] In some application scenarios involving morphological changes of brain regions, such as user occupation classification and tinnitus classification scenarios, the above step 106 can be specifically implemented by the following steps:

[0107] S1. Acquire brain magnetic resonance images of the N objects of the first type and M objects of the second type respectively.

[0108] S2. Perform data preprocessing on the brain magnetic resonance image to obtain a processed brain magnetic resonance image.

[0109] S3. Extract volume data corresponding to Q brain regions of each subject from the processed brain magnetic resonance image.

[0110] Among them, Q brain regions are Q feature items, and volume data corresponding to Q brain regions are feature values ​​corresponding to Q feature items.

[0111] The Q brain regions and the corresponding brain positions of each brain region can be obtained through literature study, summarizing the previously published results. The applicant summarized 61 brain regions for idiopathic tinnitus. Table 1 shows the names of some brain regions in these 61 brain regions and their corresponding peak coordinates. Note: Some brain areas in Table 1 correspond to multiple different peak coordinates. This is because some brain areas have a larger range, so there are multiple peak coordinates.

[0112] Table 1:

[0113]

[0114]

[0115]

[0116] In order to achieve the evaluation of the target classifier, in the specific implementation, the feature values ​​corresponding to the Q feature items of each object in the N+n first type objects and M+m ​​second type objects can be obtained, and the n Two objects of the first type and m objects of the second type are used as the final test set to evaluate the effectiveness of the target classifier; alternatively, ROC (receiver operating characteristic curve) is used to evaluate the effectiveness of the target classifier. If ROC The larger the area under the curve (AUC value), the greater the classification accuracy of the target classifier. Usually, if 0.5 <1, it means that the prediction result obtained by applying this classifier is better than the subjective guess of tinnitus state.

[0117] Take the tinnitus classification scene as an example: N+n is 46, M+m is 56, which means that it includes 46 first-class objects (the first-class objects can be patients with idiopathic tinnitus) and 56 age- and gender-matched The second type of object (the second type of object can be a normal person), a total of 102 objects.

[0118] Image acquisition: The magnetic resonance image was acquired by a GE magnetic resonance scanner with a magnetic field strength of 3.0T. The high-resolution 3D whole brain T1-weighted structure image is obtained by 3D-BRAVO sequence scanning. After scanning, 196 scanned images were obtained for each object.

[0119] Data preprocessing: completed using the VBM8 toolbox, including spatial standardization, image segmentation, smoothing and other steps. The subsequent analysis is based on the preprocessed gray matter image.

[0120] Select the target brain area: A summary of the published tinnitus morphology research literature, a total of 61 target brain areas were extracted for further feature extraction and morphological change analysis.

[0121] Making the target brain area: The results of the morphological changes of the same brain area in different studies are different. In order to facilitate the analysis, it is necessary to make the target brain area. Use MarsBar to make circular target brain regions, the center of the circle is the MNI peak coordinates of each target brain region, the radius is 5mm, and a total of 61 target brain regions are obtained.

[0122] Extract brain area volume: Use MATLAB self-made program to extract the target brain area volume. For each object, calculate and record the volume data of 61 target brain regions. Each target brain area is a feature item, that is, the feature value corresponding to 61 feature items is extracted for each object, and finally a matrix of size 102*61 is obtained.

[0123] In the subsequent grouping, the 102 subjects (including 46 idiopathic tinnitus patients and 56 normal persons) can be divided into five groups, and the 46 idiopathic tinnitus patients can be divided into 10, 9, 9, 9, 9 , Divide 56 normal people into 12, 11, 11, 11, 11; fuse tinnitus patients and normal people one-to-one, and finally get sample sizes (that is, the number of objects) of 22, 20, 20, 20, and 20, respectively Five groups. One of these five groups can be used as the final test set.

[0124] When the cross-validation method is a 5-fold cross-validation method, the combination of the cross-validation method and the grid search method in 102 above is used to test the combination of the feature items composed of the first feature item and the second feature set to obtain The highest classification accuracy rate of the feature item combination and the parameter structure corresponding to the highest classification accuracy rate of the feature item combination" can be implemented in the following steps:

[0125] 1021. The N objects of the first type and the M objects of the second type are randomly divided into five groups of objects.

[0126] 1022. The five groups of objects are arranged and combined to obtain five types of object combinations, and each of the five types of object combinations includes a group of objects as a test set and four groups of objects as a training set.

[0127] 1023. Training according to the data related to the feature item combination in the training set in each object combination, the number of objects in the training set in each object combination, and the type label training corresponding to all objects in the training set in each object combination Obtain the parameter structure to be optimized corresponding to each object combination.

[0128] 1024. Use the grid search method to test the classifier to be optimized constructed based on the parameter structure to be optimized corresponding to each object combination, and obtain multiple classification accuracy rates corresponding to each object combination.

[0129] 1025. Compare multiple classification accuracy rates corresponding to each object combination in the five object combinations, and determine the highest classification accuracy rate corresponding to the feature item combination.

[0130] 1026. Optimize the parameter structure to be optimized corresponding to the highest classification accuracy of the feature item combination to obtain the highest classification accuracy of the feature item combination based on the parameter C and the parameter γ value corresponding to the highest classification accuracy rate of the feature item combination The parameter structure corresponding to the rate.

[0131] In the above 1021 and 1022, the number of objects in each group may be the same or different. Choose one of these five groups as the training set (the remaining four groups as the test set), and there are five selection methods. Therefore, five object combinations are available. The set of subjects to be tested in each combination of subjects is different.

[0132] In the above 1023, the data related to the feature item combination in the training set in each object combination includes: feature values ​​corresponding to each feature item in the feature item combination of all objects in the corresponding training set. The parameter structure to be optimized corresponding to each object combination includes: the number of objects corresponding to the training set, all object type labels corresponding to the training set, the feature values ​​corresponding to each feature item in the feature item combination of all objects in the corresponding training set, and the The assigned parameter C and parameter γ.

[0133] It should be noted that in the above 1023, the parameter C and the parameter γ in the structure are not assigned values, but each has a numerical search range. For example, the search ranges of the parameter C and the parameter γ value are respectively: log 2 C={-5,-4,…,4,5} and log 2 γ={-5,-4,...,4,5}.

[0134] For example, the multiple object combinations include a first object combination, and the first object combination includes a first training set and a first test set. The data related to the feature item combination in the first training set in the first object combination includes: feature values ​​corresponding to each feature item in the feature item combination of all objects in the first training set. The parameter structure to be optimized corresponding to the first object combination includes: the number of objects in the first training set, all object type labels in the first training set, and feature values ​​corresponding to each feature item in the feature item combination of all objects in the first training set And the parameter C and parameter γ to be assigned.

[0135] In order to facilitate understanding, the following will take tinnitus as an example to introduce: the combination of feature items includes the left hypothalamus and the left superior frontal gyrus. The five types of object combinations include object combination Z, which includes: training set (including object A and object B) and test set (including object C and object D), among which, object A and object C are objects of the first type. Object B and Object D are objects of the second type. Then, the parameter structure to be optimized corresponding to the first object combination includes: the number of objects in the training set is 2, the type label of object A is the first type, the type label of object B is the second type, and the left hypothalamus of object A The characteristic values ​​of and the characteristic values ​​of the left superior frontal gyrus, the characteristic values ​​of the left hypothalamus and the left superior frontal gyrus of subject B, and the parameter C and parameter γ to be assigned.

[0136] In the above 1024, the multiple object combinations include a first object combination, and the first object combination includes a first test set and a first training set. The grid search method is used to test the to-be-optimized classifier constructed based on the parameter-to-be-optimized structure corresponding to the first object combination to obtain multiple classification accuracy rates corresponding to the first object combination. Specifically, the following steps can be adopted:

[0137] S1. Determine multiple numerical combinations of the parameter C and the parameter γ based on the search range of the parameter C and the parameter γ value.

[0138] S2. Determine the parameter structure corresponding to each numerical combination based on each numerical combination of the multiple numerical combinations and the parameter structure to be optimized corresponding to the first object combination.

[0139] S3. Perform a test on the first test set on a classifier constructed based on the parameter structure corresponding to each numerical value combination to obtain a classification accuracy rate corresponding to each numerical value combination.

[0140] In the above S1, the search ranges of the grid search method for the parameter C and the parameter γ value are respectively: log 2 C={-5,-4,…,4,5} and log 2 γ={-5,-4,...,4,5}, that is, both parameter C and parameter γ have 11 values. Therefore, there are 121 numerical combinations of parameter C and parameter γ.

[0141] In the above S2, the plurality of numerical combinations include the first numerical combination. The parameter structure corresponding to the first numerical combination includes: the number of objects in the first training set, the type labels of all objects in the first training set, and the features corresponding to each feature item in the feature item combination of all objects in the first training set Value and the parameter C and parameter γ corresponding to the combination of the first value. That is, the parameter C and the parameter γ in the parameter structure corresponding to each numerical combination are assigned.

[0142] In the above S3, the first test set is tested on the classifier constructed based on the parameter structure corresponding to the first numerical combination to obtain the classification accuracy rate corresponding to the first numerical combination.

[0143] For example, the first test set includes: Object A: Object of the first type; Object B: Object of the second type; Object C: Object of the first type. The first test set is input into a classifier constructed based on the structure corresponding to the first numerical combination. The output result of the classifier is: Object A is the second type of object; Object B is the second type of object; Object C is the first type A class of objects. It can be seen that the classification accuracy rate corresponding to the first value combination is 2/3.

[0144] Since there are 121 numerical combinations of parameter C and parameter γ, 121 classification accuracy rates will be obtained in the above S3, that is, the first object combination corresponds to 121 classification accuracy rates.

[0145] In the above 1025, since there are five types of object combinations, 5*121=605 classification accuracy rates are finally obtained, that is, the above combination of feature items corresponds to 605 classification accuracy rates, and the maximum value of these 605 classification accuracy rates is taken as the above feature The highest classification accuracy rate of the item combination, and the parameter C and parameter γ corresponding to the highest classification accuracy rate of the feature item combination and the parameter structure to be optimized corresponding to the highest classification accuracy rate of the feature item combination are obtained.

[0146] In the above 1026, the values ​​of parameter C and parameter γ to be assigned in the parameter structure to be optimized corresponding to the highest classification accuracy rate of the above feature item combination are respectively assigned to the parameters C and parameter corresponding to the highest classification accuracy rate of the above feature item combination γ value, so as to obtain the parameter structure corresponding to the highest classification accuracy of the above-mentioned feature item combination.

[0147] For example: the feature item combination F corresponds to five object combinations, including the object combination Z1, the parameter structure to be optimized for the object combination Z1 is S0, and a numerical combination D of the parameter C and the parameter γ determined by the network search method is C=1, γ=2, the parameter C and parameter γ to be assigned in the parameter structure S0 to be optimized are respectively assigned to C=1 and γ=2, and the parameter structure S corresponding to the numerical combination D is obtained. The classifier constructed by the parameter structure S corresponding to D is tested to obtain the classification accuracy. If the classification accuracy rate is the highest classification accuracy rate of the feature item combination F, the parameter C and the parameter γ corresponding to the highest classification accuracy rate of the feature item combination F are C=1, γ=2, and the highest classification of the feature item combination F The parameter structure to be optimized corresponding to the accuracy rate is S0.

[0148] In order to avoid certain important or representative feature items from being eliminated during the traversal process (that is, not being screened into the second feature set), the feature items in the first feature set can be identified from high to high before the traversal Low ranking, so that subsequent traversals are traversed according to the order of the feature items in the first feature set. In this way, feature items with higher recognition will be tested earlier, which can avoid some important or representative items to a certain extent The feature item of is eliminated, which can effectively improve the representativeness of the second feature set finally obtained. That is, the above method may also include:

[0149] 107. Use F-score to sort all the feature items in the first feature set from high to low recognition.

[0150] 108. Traverse each feature item in the first data according to the sequence of the feature items in the first feature set.

[0151] F-score is a simple and effective filtering hybrid feature selection algorithm, which can sort the selected feature items according to their recognition ability and importance. For a given sample set, the sample set can include n + Objects of the first type (can be regarded as positive samples) and n - An object of the second type (can be regarded as a negative sample).

[0152] The F-score calculation formula of the i-th feature item is as follows:

[0153]

[0154] among them, Represents the mean value of the feature value of the i-th feature item over the entire sample set, with Respectively represent the mean value of the feature value of the i-th feature item on the positive and negative sample sets; x k,i (+) And x k,i (-) They are the eigenvalues ​​of the i-th feature item of the k-th sample of the positive class and the k-th sample of the negative class. F i The larger the value of, the stronger the classification recognition ability of the corresponding i-th feature item, and the more it may contribute to the classification. According to the calculated F-score value, the features are ranked in descending order of importance (that is, the recognition degree is ranked from high to low). It should be noted that in the above F-score calculation formula, the numerator represents the difference between classes, and the denominator represents the difference within the class. The greater the difference between the classes, the smaller the difference within the class, indicating that the recognition degree of the corresponding feature item is better. Bigger.

[0155] It should be noted that when traversing feature items from the first feature set, a floating order forward selection algorithm can be used to search for feature items (ie traversal).

[0156] In the tinnitus classification application scenario, 8 representative brain regions were selected from the above 61 brain regions using the technical solutions provided in the embodiments of this application: bilateral hypothalamus, left superior frontal gyrus ( leftsuperior frontal gyrus (SFG), right rostal middle frontal gyrus (RMF), right insula (right insula), right inferior parietal lobule (IPL) and left two Two adjacent upper temporal gyrus (left superior temporal gyrus, STG).

[0157] figure 2 Shows the ROC curve diagram of the classifier constructed using the technical solution provided in the embodiments of the present application, and the AUC value is 0.72. It can be seen that the classification accuracy of the classifier constructed by the technical solution provided in the embodiments of the present application is better, and the classification result has a higher reference value.

[0158] image 3 It shows a structural block diagram of a classification device provided by an embodiment of the present application. Such as image 3 As shown, the device includes: an acquisition module 301, an input module 302, an output module 303, and a construction module 304; among them,

[0159] The obtaining module 301 is used to obtain the characteristic value corresponding to the specified data item of the user;

[0160] The input module 302 is configured to input the characteristic value corresponding to the specified data item as input into the target classifier to obtain the classification result of the target classifier;

[0161] The output module 303 is configured to output the classification result;

[0162] The construction module 304 is used to construct the target classifier; the construction module 304 includes: a first acquisition unit, a verification unit, a joining unit, a judgment unit, and a construction unit, wherein:

[0163] The first acquiring unit is configured to acquire the first feature item traversed in the first feature set;

[0164] The inspection unit is configured to inspect the feature item combination composed of the first feature item and the second feature set in combination with a cross-validation method and a grid search method to obtain the highest classification accuracy rate of the feature item combination, and The parameter structure corresponding to the highest classification accuracy of the feature item combination;

[0165] The adding unit is configured to add the first feature item to the second feature set if the highest classification accuracy rate of the feature item combination is higher than the highest classification accuracy rate of the second feature set;

[0166] The judging unit is configured to judge whether there is a feature item that has not been traversed in the first feature set, and if there is, continue to traverse the next feature item in the first feature set until the traversal ends;

[0167] The construction unit is configured to construct the target classifier based on the second feature set at the end of the traversal and the parameter structure corresponding to the highest classification accuracy;

[0168] Wherein, the designated data item includes all data items in the second data set at the end of the traversal.

[0169] In the technical solutions provided by the embodiments of the present application, the most representative and limited number of feature items are extracted from a large number of feature items through traversal and inspection; moreover, in the process of feature item screening, it is also achieved Optimization of the kernel function parameters of the classifier. The target classifier constructed based on the most representative and limited number of feature items and the optimized kernel function parameters is more scientific and reasonable, and its classification results have higher reference value.

[0170] Further, the target classifier is a support vector machine classifier;

[0171] The kernel function in the support vector machine is an RBF kernel function; the RBF kernel function includes a parameter C and a parameter γ; the search range of the grid search method for the values ​​of the parameter C and the parameter γ is defined as: log 2 C={-5,-4,…,4,5} and log 2 γ={-5,-4,...,4,5}.

[0172] Further, the cross-validation method is a 5-fold cross-validation method;

[0173] The aforementioned building module 304 may also include:

[0174] The second acquiring unit is configured to acquire the feature values ​​corresponding to the Q feature items of each of the N first-type objects and M second-type objects; wherein, N, M, and Q are positive integers; the first feature Including the Q feature items collectively;

[0175] The inspection unit includes:

[0176] A grouping subunit, configured to randomly divide the N objects of the first type and the M objects of the second type into five groups of objects;

[0177] The combination subunit is used to arrange and combine the five groups of objects to obtain five types of object combinations, and each of the five types of object combinations includes a group of objects as a test set and four groups of objects as a training set;

[0178] The training subunit is used to correspond to the data related to the feature item combination in the training set in each object combination, the number of objects in the training set in each object combination, and all objects in the training set in each object combination. Type label training to obtain the parameter structure to be optimized corresponding to each object combination;

[0179] The inspection subunit is used to inspect the to-be-optimized classifier constructed based on the parameter-to-be-optimized structure corresponding to each object combination by using a grid search method to obtain multiple classification accuracy rates corresponding to each object combination;

[0180] The comparison subunit is used to compare multiple classification accuracy rates corresponding to each of the five object combinations, and determine the highest classification accuracy rate of the feature item combination;

[0181] The optimization subunit is used to optimize the parameter structure to be optimized corresponding to the highest classification accuracy rate of the feature item combination based on the parameter C and the parameter γ corresponding to the highest classification accuracy rate of the feature item combination to obtain the feature item combination The parameter structure corresponding to the highest classification accuracy rate.

[0182] Further, the multiple object combinations include a first object combination, and the first object combination includes a first test set and a first training set;

[0183] The test subunit is specifically used for:

[0184] Determine multiple numerical combinations of the parameter C and the parameter γ based on the search range of the parameter C and the parameter γ;

[0185] Determine the parameter structure corresponding to each numerical value combination based on each numerical value combination of the multiple numerical value combinations and the parameter structure to be optimized corresponding to the first object combination;

[0186] The first test set is tested on a classifier constructed based on the parameter structure corresponding to each numerical value combination to obtain the classification accuracy rate corresponding to each numerical value combination.

[0187] Further, the second acquiring unit includes:

[0188] An acquiring subunit for acquiring brain magnetic resonance images of the N first-type objects and M second-type objects respectively;

[0189] The preprocessing subunit is used to perform data preprocessing on the brain magnetic resonance image to obtain a processed brain magnetic resonance image;

[0190] An extracting subunit for extracting volume data corresponding to Q brain regions of each subject from the processed brain magnetic resonance image;

[0191] Wherein, Q brain regions are the Q feature items, and volume data corresponding to the Q brain regions are feature values ​​corresponding to the Q feature items.

[0192] Further, the device may further include:

[0193] A sorting module, configured to use F-score to sort all feature items in the first feature set from high to low recognition;

[0194] The traversal module is configured to traverse the feature items in the first data according to the sequence of the feature items in the first feature set.

[0195] Further, the device may further include:

[0196] The initialization module is used to initialize the second feature set to an empty set before traversing.

[0197] It should be noted that the classification device provided in this embodiment can implement the technical solutions provided in the foregoing classification method embodiments. For specific implementation principles and beneficial effects, refer to the corresponding content in the foregoing embodiments, and details are not repeated here.

[0198] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the application, not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: The technical solutions recorded in the foregoing embodiments are modified, or some of the technical features are equivalently replaced; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.