Feature selection method, device, network device and computer-readable storage medium

Abstract

The present application proposes a feature selection method, apparatus, network device and computer-readable storage medium, the method comprising: selecting a specified number of features that are most relevant to the classification label from the performance features according to the correlation between the performance features of the communication network and the preset classification label, to obtain a selected feature set; judging whether the selected features in the selected feature set meet feature selection conditions, the feature selection conditions comprising: a feature selection parameter when the maximum likelihood estimation for the classification of the performance features reaches a maximum value is the same as the feature selection parameter when the correlation reaches a minimum value; if the selected features meet the feature selection conditions, selecting the selected features as the features that are most relevant to the classification label.

Description

Technical Field

[0001] The present application relates to the field of information and communication technology, and in particular to a feature selection method, device, network equipment and computer-readable storage medium. Background Art

[0002] With the development of wireless communication technology, a large amount of high-dimensional data information appears in wireless communication networks. The feature data extracted from them by supervised anomaly detection contains many features related to classification, as well as a lot of irrelevant feature information, redundant feature information, feature information of multi-feature collinearity and noise data information. Useless feature information will cause the data model to overfit or weaken the representation ability, and therefore cannot accurately represent the feature information of the data set. Summary of the invention

[0003] The present application provides a method, apparatus, network device and computer-readable storage medium for feature selection.

[0004] An embodiment of the present application provides a feature selection method, including: selecting a specified number of features that are most relevant to the classification label from the performance features according to the correlation between the performance features of the communication network and the preset classification label to obtain a selected feature set; judging whether the selected features in the selected feature set meet feature selection conditions, the feature selection conditions including: a feature selection parameter when the maximum likelihood estimation of the classification of the performance features reaches a maximum value, which is the same as the feature selection parameter when the correlation reaches a minimum value; if the selected features meet the feature selection conditions, selecting the selected features as the features that are most relevant to the classification label.

[0005] An embodiment of the present application provides a feature selection device, including: a feature selection module, used to select a specified number of features that are most relevant to the classification label from the performance features according to the correlation between the performance features of the communication network and the preset classification label, to obtain a selected feature set; a feature judgment module, used to judge whether the selected features in the selected feature set meet the feature selection conditions, the feature selection conditions including: the feature selection parameter when the maximum likelihood estimation of the classification of the performance features reaches the maximum value is the same as the feature selection parameter when the correlation reaches the minimum value; the feature selection module, used to select the selected features as the features most relevant to the classification label if the selected features meet the feature selection conditions.

[0006] An embodiment of the present application provides a network device, including: one or more processors; a memory, on which one or more programs are stored. When the one or more programs are executed by one or more processors, the one or more processors implement any one of the feature selection methods in the embodiment of the present application.

[0007] An embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, any one of the feature selection methods in the embodiments of the present application is implemented.

[0008] According to the feature selection method, apparatus, network device and computer-readable storage medium of the embodiments of the present application, in the process of performing feature selection on the performance characteristics of a communication network, the maximum likelihood estimation rule is applied to the field of feature selection based on information theory, thereby effectively establishing a connection between the maximum likelihood estimation and the information theory feature selection algorithm based on the maximum likelihood estimation rule, relying on the above-mentioned mathematical theory, and optimizing the feature selection parameters based on the performance characteristics of the communication network without human intervention, thereby solving the problem of computational cost of directly selecting the performance characteristics of the communication network through information theory related methods.

[0009] With regard to the above embodiments and other aspects of the present application and their implementation, further description is provided in the accompanying drawings, detailed description and claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A schematic flow chart showing a feature selection method according to an embodiment of the present application is shown.

[0011] Figure 2 A schematic flow chart showing a feature selection method according to another embodiment of the present application.

[0012] Figure 3 A schematic flow chart showing a method for calculating feature correlation in an embodiment of the present application is shown.

[0013] Figure 4 A schematic flow chart showing a method for calculating feature redundancy in an embodiment of the present application is shown.

[0014] Figure 5 A schematic flow chart showing the calculation process of the iterative histogram in an embodiment of the present application is shown.

[0015] Figure 6 A schematic flow chart showing the process of calculating the feature condition mutual information in an embodiment of the present application.

[0016] Figure 7 A schematic diagram of the structure of a feature selection device provided by an embodiment of the present invention is shown.

[0017] Figure 8 is a structural diagram showing an exemplary hardware architecture of a computing device capable of implementing the method and apparatus according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the present application more clear, the embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other arbitrarily without conflict.

[0019] In the embodiments of the present application, the high-dimensional data information in the wireless communication network brings a lot of useful knowledge to supervised anomaly detection, but also brings a series of unprecedented difficulties and challenges. The feature data extracted by supervised anomaly detection at this stage contains many features related to classification, and there are also a lot of irrelevant feature information, redundant feature information, feature information of multi-feature collinearity, and noise data information. These useless feature information will cause the data model to overfit or weaken the representation ability, so it cannot accurately represent the feature information of the data set. Therefore, the task of feature selection is to find "essential features" in high-dimensional data sets that can have strong classification representation power and will not weaken the data information of the data sample.

[0020] Feature selection, as a method that can reduce the size of data sets and improve data quality, is widely used in fields such as machine learning and data mining. Feature selection algorithms can generally include the following three categories, such as: filtering, wrapping, and embedded. Depending on the feature selection process, the evaluation criteria are also different, such as: consistency criteria, distance criteria, dependency criteria, classification error rate criteria, and information quantity criteria. As an example, feature selection algorithms may include: the Relief algorithm based on feature weights, the Max-Relevance and Min-Redundancy (MRMR) algorithm based on maximum relevance and minimum redundancy, etc.

[0021] In some embodiments, the information method is widely used due to its stability and accuracy. However, there are some defects in the algorithms related to the information method. For example, the best individual feature BIF (Best Individual Feature) algorithm is simple and fast, especially suitable for high-dimensional data. However, it does not take into account the correlation between the selected features, which will bring greater redundancy; the MRMR (Minimal Redundancy and Maximal Relevance) algorithm only considers the correlation between a single feature attribute and the category attribute and the redundancy between a single feature attribute and the selected feature in the process of feature selection, ignoring the interaction between multiple feature sets and multiple feature sets; the conditional mutual information maximization (CMIM) algorithm uses conditional mutual information to evaluate the importance of features. The larger the conditional mutual information value, the more new information the feature can provide, but the conditional mutual information calculation cost is high, and the multidimensionality of the sample leads to inaccurate valuation.

[0022] The development of wireless communication networks has led to an explosive growth in the amount of information, and data scaling is a very thorny research issue. For machine learning algorithms, data preprocessing, feature selection, and vectorization require a lot of calculations, which takes a long time to calculate and consumes a lot of memory, which reduces the classification performance of the algorithm.

[0023] In some feature selection methods of parallel binary moth-to-fire algorithms based on the distributed computing framework Spark, the moth-to-fire optimization algorithm can be used to reduce feature redundancy by taking advantage of its good classification ability, and introduce the moth-to-fire optimization algorithm for feature selection. And with the characteristics of Spark distributed platform based on memory and parallel computing, the moth-to-fire optimization algorithm is combined with distributed parallel computing to avoid the algorithm from falling into local optimality and improve the classification performance of the algorithm, while maximizing the classification performance and minimizing the number of features.

[0024] However, the moth-to-fire optimization algorithm has good classification capabilities in reducing feature redundancy, but does not consider the correlation between features; in addition, the moth-to-fire algorithm is prone to falling into local optimality and has low search capabilities, which seriously limit the classification performance and dimensionality reduction capabilities of the algorithm; although the use of the Spark distributed computing framework can improve local optimality and search to a certain extent, it does not fundamentally solve the above-mentioned cost and performance problems.

[0025] In some feature selection and classification methods based on maximum information coefficient, feature selection can be performed by using the idea of ​​conditional mutual information. However, the computational cost of conditional mutual information is high, and in the process of feature selection, when there are a large number of features and the feature dimension is high, the consumption of computing resources is huge.

[0026] Some improved feature filtering methods based on correlation feature selection can solve the problems of poor data quality and low efficiency in preprocessing in software defect prediction, and can improve the performance of software defect prediction models while removing redundant and irrelevant features. However, this feature filtering method also faces the problem of high computational cost in the feature selection process. When there are a large number of features and high feature dimensions, the consumption of computing resources is huge.

[0027] The embodiment of the present application provides a feature selection method, which can reduce the cost of the feature selection process and improve the performance of the machine learning algorithm and the feature selection algorithm.

[0028] Figure 1 The following is a flow chart showing the feature selection method of the embodiment of the present application. Figure 1 As shown, the feature selection method in the embodiment of the present application may include the following steps.

[0029] S110, according to the correlation between the performance characteristics of the communication network and the preset classification labels, a specified number of characteristics that are most relevant to the classification labels are selected from the performance characteristics to obtain a selected characteristic set.

[0030] S120, determining whether the selected features in the selected feature set satisfy feature selection conditions, the feature selection conditions comprising: a feature selection parameter when the maximum likelihood estimation for the classification of the performance feature reaches a maximum value is the same as the feature selection parameter when the correlation reaches a minimum value.

[0031] S130: If the selected feature satisfies the feature selection condition, the selected feature is selected as the feature most relevant to the classification label.

[0032] According to the feature selection method of the embodiment of the present application, in the process of performing feature selection on the performance characteristics of a communication network, the maximum likelihood estimation rule is applied to the field of feature selection based on information theory. Thus, the information theory feature selection method based on the maximum likelihood estimation rule effectively establishes a connection between the maximum likelihood estimation and the information theory feature selection algorithm, thereby reducing the computational cost of directly selecting the performance characteristics of the communication network through information theory related methods.

[0033] The following describes the process of acquiring feature selection conditions through a specific embodiment. The method for acquiring feature selection conditions in the embodiment of the present application may include the following steps.

[0034] S01, data preprocessing.

[0035] In this step, the relevant features with missing data can be deleted and the feature data can be normalized.

[0036] S02, introduce information theory feature selection rules.

[0037] In the N feature sample values ​​of the embodiment of the present application, the relationship between each sample value and the category can be expressed as (x, y). Wherein, the x feature vector is x=[x1, x2, ..., x d ]; y is the category. For N sample observations, it can be expressed as a feature vector X = {X1, X2, ..., X d} and Y.

[0038] In one embodiment, a binary feature vector θ is used to indicate whether a feature is selected or not. For example, 1 represents selected and 0 represents unselected. For example, x θ Indicates the selected feature; represents the unselected features. Therefore, the data features can be expressed as

[0039] In the feature selection method of the embodiment of the present application, for the performance characteristics of the communication network, there is a subset θ of the feature set * , so that For the feature prediction model q, it can be understood as an approximation of the true mapping model p. The model q has two layers of parameters: one layer of parameters is the selected feature θ, and the other layer is the parameter τ used to predict the feature type.

[0040] Therefore, we need to find a minimum feature set to make our prediction model q infinitely close to the true model. In one embodiment, the maximum likelihood estimation of the parameters in the prediction model q can be expressed by the following expression (1).

[0041]

[0042] In the above expression (1), the maximum likelihood estimation of the parameters in the prediction model q is expressed as: according to the feature parameter θ and the category prediction parameter τ, solve a minimum feature set D so that the selected features in D make the sample category y in the selected features The probability density function under the condition of and parameter τ can approximate the probability density of sample category y under the true value condition as much as possible. and parameter τ are the parameters that need to be estimated.

[0043] In one embodiment, normalizing expression (1) and taking the logarithm can obtain the following expression (2):

[0044]

[0045] In this embodiment, taking the logarithm can convert the multiplication calculation into addition calculation, reduce the absolute value of the data, and facilitate calculation. Taking the logarithm does not change the nature and correlation of the data, but compresses the scale of the variable, making the data more stable.

[0046] For expression (2), by adding the term p(y|x θ ), the above expression (2) can be transformed into the following expression (3):

[0047]

[0048] In the above expression (3), add the term p(y|x θ ) indicates that in the selected feature x θ The true distribution of the post-label y is obtained by adding the term p(y|x) to expression (3) to obtain the following expression (4):

[0049]

[0050] In the above expression (4), p(y|x) is the true distribution of all features. The mathematical expectation is denoted by E xy {·}, then take the logarithm of both sides of expression (4), and then take the negative numbers at both ends of the expression to convert the maximum value into the minimum value, and obtain the following expression (5):

[0051]

[0052] In the above expression (5), after introducing mutual information and information entropy, expression (5) can be transformed into the following expression (6):

[0053]

[0054] In expression (6), the first term is the likelihood ratio of the true distribution to the essential feature distribution, which is independent of the specific feature parameter θ and the prediction parameter τ; the third term H(Y|X) is the conditional information entropy, which is related to all features. Therefore, the maximum likelihood estimation information theory feature selection algorithm is similar to the second term Conditional mutual information is closely related.

[0055] From expression 6, we know that when the second term of expression 6 is the smallest, the maximum likelihood estimate is the largest, that is:

[0056]

[0057] From the above description, we can see that when the conditional mutual information in expression (7) When the minimum value is reached, the optimal feature set can be obtained according to the value of θ.

[0058] In the above embodiment, it is determined whether the selected feature in the selected feature set satisfies the feature selection condition, that is, the above expression (7). If the feature selection condition is satisfied, the selected feature is selected as the feature most relevant to the classification label.

[0059] In one embodiment, the value of the maximum likelihood estimate for the classification of the performance feature is the sum of the likelihood ratio of the real feature distribution of the classification label in the performance feature to the essential feature distribution, the correlation between the performance feature and the classification label, and the information entropy of the performance feature and the classification label. The likelihood ratio represents the ratio of the probability density of the classification label under the selected feature condition to the probability density of the classification label under the selected feature and feature prediction parameter condition.

[0060] In one embodiment, before step S110, the feature selection method further includes: S10, storing the performance characteristics as a feature matrix based on the performance characteristics of the performance data of the communication network; S11, performing a transposition operation on the feature matrix to obtain a transposed feature matrix; S12, calculating the correlation between each performance characteristic in the communication network and a preset classification label based on the performance characteristics in the transposed feature matrix.

[0061] In this embodiment, the performance characteristics of the performance data of the communication network are represented as a characteristic matrix, and a transposition operation is performed on the characteristic data. After the characteristic matrix is ​​transposed, the computing performance is greatly improved.

[0062] In one embodiment, each matrix element in a feature matrix of a performance feature represents feature data corresponding to a specified feature index value and a partition index value, wherein the feature index value is used to index a specified feature of a communication network, and the partition index value is used to index a data storage partition corresponding to a classification label, and wherein feature data of the same specified feature is stored in the same partition in the feature vector matrix.

[0063] In this embodiment, in the feature matrix of the performance feature, the column row represents the feature, the column col represents the object, and the matrix is ​​represented by a triple (k, part.index, matrix(k)), where k represents the feature index, part.index represents the partition index, and matrix(k) represents the feature data, thereby all data of the same feature of all objects are allocated to a specific partition, for example, in one partition, thereby reducing computational costs.

[0064] In one embodiment, if the selected features do not meet the feature selection conditions, the feature selection method also includes: S140, calculating the feature redundancy between the selected features, performance features and classification labels in the selected feature set; S141, using the calculated feature redundancy information to update the selected feature set; S142, using the updated selected feature set as the most relevant feature data selected from the performance features.

[0065] In this embodiment, feature redundancy among selected features, performance features, and classification labels in the selected feature set is calculated until a sufficient set of selected features is obtained.

[0066] In order to better understand this application, Figure 2 A feature selection method according to another embodiment of the present application is described. Figure 2 FIG. 2 is a flow chart showing a feature selection method according to another embodiment of the present application. Figure 2 As shown, the feature selection method may include the following steps.

[0067] S201, performing data preprocessing on input performance characteristic data of the communication network.

[0068] S202, performing a transposition operation on the feature matrix of the preprocessed performance feature data to obtain a transposed feature matrix.

[0069] In this step, the feature matrix can maintain the original partitioning of the data, thereby suppressing data dispersion. After the feature matrix is ​​transposed, the resulting data will be pooled and the data can be reused to reduce computational costs.

[0070] S203, calculating the correlation between the performance feature and the label.

[0071] S204: Initialize the selected feature set according to the feature correlation result.

[0072] In this step, the feature correlation results include: the most relevant feature P of the booking quantity best , according to the feature correlation results, it is added to the selected feature set, an initial level of features is created, and the iterative process of S205-S207 described below is performed.

[0073] S205, determining the most relevant feature P selected best Whether the preset feature selection conditions are met, if yes, execute step S208, if not, execute step S206.

[0074] In this step, it is determined whether the selected feature in the selected feature set meets the feature selection condition, that is, the above expression (7). If the feature selection condition is met, the selected feature is selected as the feature most relevant to the classification label. In other words, the most relevant feature P can be selected based on the selected feature. best Whether the preset feature selection conditions are met, update the intermediate result pool of the selected feature set.

[0075] In each iteration, the termination condition of the iteration is that the set of elements in the selected feature set and the set of unselected elements satisfy the above expression (7) or there are no features remaining. In other words, it is determined whether the number of features in the selected feature set satisfies the above expression (7). If yes, it jumps to S208; if not, it jumps to S206.

[0076] S206, calculating the redundancy between features.

[0077] S207, updating the selected feature set according to the calculation result of the redundancy between the features.

[0078] S208, obtaining a selected feature set.

[0079] Through the above steps S201-S208, the maximum likelihood estimation rule can be introduced and applied to the field of feature selection based on information theory, providing a selection rule with a theoretical basis for the information theory feature selection algorithm.

[0080] In one embodiment, the above step S12 may specifically include the following steps.

[0081] S21, broadcasting classification labels to multiple preset distributed computing nodes; S22, instructing each distributed computing node to calculate the mutual information between the alternative features and the classification labels to obtain the feature correlation between the alternative features and the classification labels, where the alternative features are performance features pre-assigned to each distributed computing node; S23, determining the correlation between each performance feature and the classification label based on the feature correlation calculated by each distributed computing node.

[0082] In this embodiment, features with high correlation with labels are selected from large-scale feature data in a distributed manner, thereby effectively solving the problem of high computational complexity of algorithms related to information theory.

[0083] Figure 3 A schematic diagram showing a flow chart of a method for calculating feature correlation in an embodiment of the present application is shown. Figure 3 As shown, in one embodiment, a method for calculating feature correlation may include the following steps.

[0084] S301, input characteristic data of the communication network.

[0085] In this step, a matrix of transposed feature data is obtained and input. In the matrix of feature data, each matrix element represents feature data corresponding to a specified feature index value and a partition index value. The feature index value is used to index a specified feature of the communication network, and the partition index value is used to index a data storage partition corresponding to a classification label. Feature data of the same specified feature is stored in the same partition in the feature vector matrix.

[0086] S302, broadcast the label vector.

[0087] In this step, the label vector is broadcasted so that all nodes in the Spark cluster can obtain the label vector. For example, in the matrix of the transposed feature data, the classification labels of all data storage partitions are obtained to form a vector and broadcasted.

[0088] S303, calculating the histogram of all candidate features and classification labels.

[0089] In this step, the candidate characteristics may be at least a portion of performance characteristics of the communication network pre-assigned to each distributed computing node.

[0090] S304, calculating the joint probability distribution of the classification label and the candidate features by means of a histogram.

[0091] S305, calculating the marginal probability distribution of the classification labels and the candidate features by means of a histogram.

[0092] S306, obtaining a feature mutual information value of each candidate feature.

[0093] Through the above steps S301 to S306, the calculation process of feature correlation in large-scale data is optimized based on the distributed computing framework, so that the feature selection algorithm based on information theory can be used in the current scenario, improving the performance of the machine learning algorithm and the feature selection algorithm.

[0094] In one embodiment, S22 may specifically include the following steps.

[0095] S221, calculating the histogram information of the classification labels and the histogram information of the candidate features.

[0096] S222, calculating the marginal probability distribution of the classification labels, the marginal probability distribution of the candidate features, and the joint probability distribution of the classification labels and the candidate features according to the histogram information of the classification labels and the histogram information of the candidate features.

[0097] S223, calculating the probability distribution product of the marginal probability distribution of the classification label and the marginal probability distribution of the candidate feature, and calculating the similarity between the joint probability distribution and the probability distribution product, to obtain the mutual information between the candidate feature and the classification label.

[0098] S224, calculating the mutual information between the candidate features and the classification labels.

[0099] In this embodiment, the mutual information between candidate features and classification labels is calculated based on a distributed computing framework, which can greatly reduce the computational cost of information theory related algorithms.

[0100] In one embodiment, step S140 may specifically include: S141, broadcasting the classification label and the selected feature set to a preset plurality of distributed computing nodes; S142, instructing each distributed computing node to calculate the feature redundancy of the alternative features, the selected features in the selected feature set, and the classification label, where the alternative features are performance features pre-assigned to each distributed computing node; S143, determining the feature redundancy between each performance feature and the classification label based on the feature redundancy calculated by each distributed computing node.

[0101] In this embodiment, redundant information between features is calculated in large-scale feature data in a distributed manner, thereby effectively solving the problem of high computational complexity of algorithms related to information theory.

[0102] Figure 4 A schematic diagram showing a flow chart of a method for calculating feature redundancy in an embodiment of the present application is shown. Figure 4 As shown, in one embodiment, a method for calculating feature redundancy may include the following steps.

[0103] S401, input characteristic data of the communication network.

[0104] In this step, the transposed feature data is obtained and input.

[0105] S402: Broadcast the label vector Y to the Spark cluster.

[0106] S403: broadcast the selected feature vector to the Spark cluster.

[0107] In this step, for example, the selected optimal feature P is broadcast to the Spark cluster. best .

[0108] S404, calculating the candidate features and the selected optimal feature P best , feature iteration histogram between classification labels Y.

[0109] In this step, the candidate characteristics may be at least a portion of performance characteristics of the communication network pre-assigned to each distributed computing node.

[0110] S405, calculating feature redundancy information of all candidate features.

[0111] In this step, feature redundancy among features in the selected feature set, candidate features, and classification labels is calculated.

[0112] Through the above steps S401-S405, the calculation process of feature redundancy in large-scale data is optimized based on the distributed computing framework, so that the feature selection algorithm based on information theory can be used in the current scenario, improving the performance of the machine learning algorithm and the feature selection algorithm.

[0113] In one embodiment, S142 may specifically include the following steps.

[0114] S142-01, calculating histogram information of candidate features, histogram information of selected features in a selected feature set, and histogram information of classification labels.

[0115] S142-02, based on the histogram information of the candidate features, the histogram information of the selected features, and the histogram information of the classification labels, the conditional mutual information between the candidate features, the selected features, and the classification labels is calculated to obtain the feature redundancy of the candidate features, the selected features, and the classification labels.

[0116] In this embodiment, conditional mutual information, conditional mutual information and feature redundancy among candidate features, selected features and classification labels are calculated through histogram information, thereby improving calculation efficiency.

[0117] Figure 5 FIG. 1 is a flow chart showing the calculation process of the iterative histogram in the embodiment of the present application. Figure 5 As shown, in this embodiment, the calculation process of the iterative histogram may include the following steps.

[0118] S501, input characteristic data of the communication network.

[0119] In this step, the transposed feature data may be input.

[0120] S502: Broadcast label Y to the Spark cluster.

[0121] S503: broadcast the selected feature vector to the Spark cluster.

[0122] For example, broadcast the selected optimal feature P to the Spark cluster best .

[0123] S504: Generate a corresponding feature iteration histogram on each computing node in the Spark cluster.

[0124] In this step, at each computing node, the candidate features and the selected optimal features P are calculated. best , feature iteration histogram between classification labels Y.

[0125] S505 , according to the acquired feature iteration histogram generated at each computing node, merge and sort the acquired feature iteration histograms.

[0126] Through the above steps S501-S505, according to the candidate features, the selected optimal feature P best and label Y, calculate the three-dimensional histogram of the features.

[0127] Figure 6 The flowchart of the calculation process of the feature condition mutual information in the embodiment of the present application is shown. Figure 5 As shown, in this embodiment, the calculation process of the feature condition mutual information may include the following steps.

[0128] S601, input histogram information.

[0129] In this step, the input histogram information includes: histogram information of candidate feature vectors, histogram information of selected feature vectors, and histogram information of classification label vectors.

[0130] S602, broadcast the edge distribution of the selected feature vector to all computing nodes.

[0131] S603: Broadcast the edge distribution of the label vector to all computing nodes.

[0132] S604: broadcast the joint distribution of feature vectors to all computing nodes.

[0133] In this step, the joint probability distribution of the selected feature vectors is broadcasted, for example, to all computing nodes.

[0134] The first element in each histogram tuple can be a feature index, and the second element is a three-dimensional matrix.

[0135] S605, calculating feature mutual information values ​​and conditional mutual information values.

[0136] In this step, according to the histogram information of the candidate feature vectors, the histogram information of the selected feature vectors, and the histogram information of the classification labels.

[0137] Through the above steps S601-S605, the histogram is an accurate graphical representation of the distribution of numerical data. Calculating the mutual information of the candidate features through the histogram can more clearly describe the distribution state of the performance characteristics, which is beneficial to the calculation process of the feature mutual information in the embodiment of the present application.

[0138] Through the feature selection method of the embodiment of the present application, the feature selection process in large-scale data can be optimized based on a distributed computing framework, so that the feature selection algorithm based on information theory can be used in the current scenario, thereby improving the performance of the machine learning algorithm and the feature selection algorithm.

[0139] The device according to the embodiment of the present invention is described in detail below with reference to the accompanying drawings. Figure 7 FIG. 2 is a schematic diagram showing the structure of a feature selection device provided by an embodiment of the present invention. Figure 7 As shown, the feature selection device may include the following modules.

[0140] The feature selection module 710 is used to select a specified number of features that are most relevant to the classification label from the performance features according to the correlation between the performance features of the communication network and the preset classification label, to obtain a selected feature set.

[0141] The feature judgment module 720 is used to judge whether the selected features in the selected feature set meet the feature selection conditions, and the feature selection conditions include: the feature selection parameter when the maximum likelihood estimation of the classification of the performance feature reaches the maximum value is the same as the feature selection parameter when the correlation reaches the minimum value.

[0142] The feature selection module 730 is configured to select the selected feature as the feature most relevant to the classification label if the selected feature satisfies the feature selection condition.

[0143] In one embodiment, the value of the maximum likelihood estimate for the classification of the performance feature is the sum of the likelihood ratio of the true feature distribution of the classification label in the performance feature to the essential feature distribution, the correlation between the performance feature and the classification label, and the information entropy of the performance feature and the classification label; wherein the likelihood ratio represents: the ratio of the probability density of the classification label under the selected feature conditions to the probability density of the classification label under the selected feature and feature prediction parameter conditions.

[0144] In one embodiment, the feature selection device is also used to store the performance characteristics as a feature matrix based on the performance characteristics of the performance data of the communication network before calculating the correlation between the performance characteristics of the communication network and the preset classification label; transpose the feature matrix to obtain a transposed feature matrix; and calculate the correlation between each performance characteristic in the communication network and the preset classification label based on the performance characteristics in the transposed feature matrix.

[0145] In one embodiment, when the feature selection device is used to calculate the correlation between each performance feature in a communication network and a preset classification label, it is specifically used to: broadcast the classification label to a preset plurality of distributed computing nodes; instruct each distributed computing node to calculate the mutual information between the alternative feature and the classification label to obtain the feature correlation between the alternative feature and the classification label, where the alternative feature is a performance feature pre-assigned to each distributed computing node; and determine the correlation between each performance feature and the classification label based on the feature correlation calculated by each distributed computing node.

[0146] In one embodiment, when the feature selection device is used to calculate the mutual information between the candidate features and the classification label, it is specifically used to include: calculating the histogram information of the classification label and the histogram information of the candidate features; calculating the marginal probability distribution of the classification label, the marginal probability distribution of the candidate features, and the joint probability distribution of the classification label and the alternative features based on the histogram information of the classification label and the histogram information of the alternative features; calculating the probability distribution product of the marginal probability distribution of the classification label and the marginal probability distribution of the alternative features, and calculating the similarity between the joint probability distribution and the probability distribution product to obtain the mutual information between the alternative features and the classification label.

[0147] In one embodiment, if the selected features do not meet the feature selection conditions, the feature selection device also includes: a feature redundancy calculation module, which is used to calculate the feature redundancy between the selected features, performance features and classification labels in the selected feature set; using the calculated feature redundancy information to update the selected feature set; and using the updated selected feature set as the most relevant feature data selected from the performance features.

[0148] In one embodiment, the feature redundancy calculation module, when used to calculate the feature redundancy between selected features in a selected feature set, performance features, and classification labels, is specifically used to: broadcast the classification labels and the selected feature set to a preset plurality of distributed computing nodes; instruct each distributed computing node to calculate the feature redundancy of alternative features, selected features in the selected feature set, and classification labels, where the alternative features are performance features pre-assigned to each distributed computing node; and determine the feature redundancy between each performance feature and the classification label based on the feature redundancy calculated by each distributed computing node.

[0149] In one embodiment, the feature redundancy calculation module, when used to calculate the feature redundancy of alternative features, selected features in a selected feature set, and classification labels, is specifically used to: calculate histogram information of the alternative features, histogram information of the selected features in a selected feature set, and histogram information of the classification labels; calculate the conditional mutual information between the alternative features, the selected features, and the classification labels based on the histogram information of the alternative features, the histogram information of the selected features, and the histogram information of the classification labels, and obtain the feature redundancy of the alternative features, the selected features, and the classification labels.

[0150] In one embodiment, each matrix element in a feature matrix of a performance feature represents feature data corresponding to a specified feature index value and a partition index value, wherein the feature index value is used to index a specified feature of a communication network, and the partition index value is used to index a data storage partition corresponding to a classification label, and wherein feature data of the same specified feature is stored in the same partition in the feature vector matrix.

[0151] Through the feature selection device of the embodiment of the present application, the feature selection process in large-scale data can be optimized based on a distributed computing framework, so that the feature selection algorithm based on information theory can be used in the current scenario, thereby improving the performance of the machine learning algorithm and the feature selection algorithm.

[0152] It should be clear that the present invention is not limited to the specific configurations and processes described in the above embodiments and shown in the figures. For the convenience and brevity of description, a detailed description of the known methods is omitted here, and the specific working processes of the systems, modules and units described above can refer to the corresponding processes in the aforementioned method embodiments, which will not be repeated here.

[0153] Figure 8 is a structural diagram showing an exemplary hardware architecture of a computing device capable of implementing the method and apparatus according to an embodiment of the present invention.

[0154] like Figure 8 As shown, the computing device 800 includes an input device 801, an input interface 802, a central processing unit 803, a memory 804, an output interface 805, and an output device 806. The input interface 802, the central processing unit 803, the memory 804, and the output interface 805 are connected to each other via a bus 810, and the input device 801 and the output device 806 are connected to the bus 810 via the input interface 802 and the output interface 805, respectively, and then connected to other components of the computing device 800.

[0155] Specifically, the input device 801 receives input information from the outside and transmits the input information to the central processing unit 803 through the input interface 802; the central processing unit 803 processes the input information based on the computer executable instructions stored in the memory 804 to generate output information, temporarily or permanently stores the output information in the memory 804, and then transmits the output information to the output device 806 through the output interface 805; the output device 806 outputs the output information to the outside of the computing device 800 for user use.

[0156] In one embodiment, Figure 8 The computing device shown may be implemented as a network device, which may include: a memory configured to store a program; and a processor configured to run the program stored in the memory to execute the feature selection method described in the above embodiment.

[0157] The above is only an exemplary embodiment of the present application and is not intended to limit the scope of protection of the present application. In general, various embodiments of the present application can be implemented in hardware or special circuits, software, logic or any combination thereof. For example, some aspects can be implemented in hardware, while other aspects can be implemented in firmware or software that can be executed by a controller, a microprocessor or other computing device, although the present application is not limited thereto.

[0158] Embodiments of the present application may be implemented by executing computer program instructions by a data processor of a mobile device, for example in a processor entity, or by hardware, or by a combination of software and hardware. The computer program instructions may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages.

[0159] The block diagram of any logic flow in the accompanying drawings of the present application can represent program steps, or can represent interconnected logic circuits, modules and functions, or can represent a combination of program steps and logic circuits, modules and functions. The computer program can be stored on a memory. The memory can have any type suitable for the local technical environment and can be implemented using any suitable data storage technology, such as but not limited to read-only memory (ROM), random access memory (RAM), optical memory device and system (digital versatile disc DVD or CD disc), etc. Computer-readable media may include non-transient storage media. The data processor can be any type suitable for the local technical environment, such as but not limited to a general-purpose computer, a special-purpose computer, a microprocessor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a programmable logic device (FGPA) and a processor based on a multi-core processor architecture.

[0160] By way of exemplary and non-limiting examples, a detailed description of exemplary embodiments of the present application has been provided above. However, various modifications and adjustments to the above embodiments will be apparent to those skilled in the art when considered in conjunction with the accompanying drawings and claims, but will not depart from the scope of the present invention. Therefore, the proper scope of the present invention will be determined according to the claims.

Claims

1. A feature selection method, characterized in that: The method comprises: According to the correlation between the performance characteristics of the communication network and the preset classification labels, a specified number of characteristics that are most relevant to the classification labels are selected from the performance characteristics to obtain a selected characteristic set; Determining whether a selected feature in the selected feature set satisfies a feature selection condition, wherein the feature selection condition comprises: a feature when a maximum likelihood estimate for classification of the performance feature reaches a maximum value is the same as a feature when the correlation reaches a minimum value; If the selected feature satisfies the feature selection condition, selecting the selected feature as the feature most relevant to the classification label; The value of the maximum likelihood estimate for the classification of the performance feature is the sum of the likelihood ratio of the true feature distribution and the essential feature distribution of the classification label in the performance feature, the correlation between the performance feature and the classification label, and the information entropy of the performance feature and the classification label; Wherein, the likelihood ratio represents the ratio of the probability density of the classification label under the selected feature condition to the probability density of the classification label under the selected feature and feature prediction parameter condition; The step of calculating the correlation between the performance characteristics of the communication network and the preset classification labels includes: According to the performance characteristics of the performance data of the communication network, the performance characteristics are stored as a characteristic matrix; Performing a transposition operation on the characteristic matrix to obtain a transposed characteristic matrix; The correlation between the performance features in the communication network and the preset classification labels is calculated based on the performance features in the transposed feature matrix.

2. The method according to claim 1, characterized in that The calculation of the correlation between each performance feature in the communication network and a preset classification label includes: Broadcasting the classification label to a plurality of preset distributed computing nodes; Instructing each distributed computing node to calculate mutual information between a candidate feature and the classification label to obtain a feature correlation between the candidate feature and the classification label, wherein the candidate feature is a performance feature pre-assigned to each distributed computing node; The correlation between each performance feature and the classification label is determined according to the feature correlation calculated by each distributed computing node.

3. The method according to claim 2, characterized in that The calculating the mutual information between the candidate features and the classification labels includes: Calculating histogram information of the classification label and histogram information of the candidate features; Calculating the marginal probability distribution of the classification label, the marginal probability distribution of the candidate feature, and the joint probability distribution of the classification label and the candidate feature according to the histogram information of the classification label and the histogram information of the candidate feature; The probability distribution product of the marginal probability distribution of the classification label and the marginal probability distribution of the candidate feature is calculated, and the similarity between the joint probability distribution and the probability distribution product is calculated to obtain the mutual information between the candidate feature and the classification label.

4. The method according to claim 1, characterized in that: If the selected feature does not satisfy the feature selection condition, the method further includes: calculating feature redundancy among selected features in the selected feature set, the performance features, and the classification labels; Using the calculated feature redundancy information, updating the selected feature set; The updated selected feature set is used as the most relevant feature data selected from the performance features.

5. The method according to claim 4, characterized in that The calculating feature redundancy among the selected features in the selected feature set, the performance features, and the classification labels comprises: Broadcasting the classification label and the selected feature set to a plurality of preset distributed computing nodes; Instructing each distributed computing node to calculate feature redundancy of candidate features, selected features in the selected feature set, and the classification label, wherein the candidate features are performance features pre-assigned to each distributed computing node; The feature redundancy between each performance feature and the classification label is determined according to the feature redundancy calculated by each distributed computing node.

6. The method according to claim 5, characterized in that The calculating the feature redundancy of the candidate features, the selected features in the selected feature set, and the classification label includes: Calculating histogram information of the candidate features, histogram information of the selected features in the selected feature set, and histogram information of the classification labels; According to the histogram information of the candidate features, the histogram information of the selected features, and the histogram information of the classification labels, the conditional mutual information among the candidate features, the selected features, and the classification labels is calculated to obtain feature redundancy of the candidate features, the selected features, and the classification labels.

7. The method according to claim 1, characterized in that Each matrix element in the feature matrix of the performance feature represents feature data corresponding to a specified feature index value and a partition index value, wherein the feature index value is used to index a specified feature of the communication network, and the partition index value is used to index a data storage partition corresponding to the classification label, and wherein feature data of the same specified feature is stored in the same partition in the feature matrix.

8. A feature selection device, characterized in that: The device comprises: A feature selection module, configured to select a specified number of features that are most relevant to the classification label from the performance features according to the correlation between the performance features of the communication network and the preset classification label, to obtain a selected feature set; A feature judgment module is used to judge whether the selected features in the selected feature set meet the feature selection conditions, and the feature selection conditions include: the feature when the maximum likelihood estimate of the classification of the performance feature reaches the maximum value is the same as the feature when the correlation reaches the minimum value; wherein the value of the maximum likelihood estimate of the classification of the performance feature is the sum of the likelihood ratio of the real feature distribution and the essential feature distribution of the classification label in the performance feature, the correlation between the performance feature and the classification label, and the information entropy of the performance feature and the classification label; the likelihood ratio represents the ratio of the probability density of the classification label under the selected feature condition to the probability density of the classification label under the selected feature and the feature prediction parameter condition; and, according to the performance characteristics of the performance data of the communication network, the performance characteristics are stored as a feature matrix; the feature matrix is ​​transposed to obtain a transposed feature matrix; according to the performance characteristics in the transposed feature matrix, the correlation between the performance characteristics in the communication network and the preset classification label is calculated; A feature selection module is used to select the selected feature as the feature most relevant to the classification label if the selected feature meets the feature selection condition.

9. A network device comprising: one or more processors; A memory having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the feature selection method described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the feature selection method according to any one of claims 1 to 7 is implemented.

Images