A method for all-scene service intelligent sensing and accurate classification

Abstract

The application discloses a panoramic service intelligent sensing and accurate classification method, comprising the following steps: a sensing probe collects characteristic data of different service types of service end information, and divides the characteristic data into labeled data and unlabeled data; a feature subset of the labeled data is constructed based on a mixed feature selection method, and a global optimal feature subset is obtained; based on a semi-supervised learning model of an improved K-means clustering algorithm, service end infection data is output; the sensing probe collects, stores and transmits service request information of a user end, and the service request information is used as prior information; a trained convolutional neural network model is used to output high-level features of the prior information and the service end infection data; and a trained random forest service sensing and identification model is used to classify the high-level features. The panoramic service intelligent sensing and accurate classification method can improve the accuracy of mixed service identification.

Description

Technical Field

[0001] This invention relates to the field of information network technology, specifically to a method for intelligent perception and accurate classification of services across all scenarios. Background Technology

[0002] The integrated space-ground information network is formed by the interconnection and integration of various heterogeneous networks, including space-based networks (such as space-based backbone networks, space-based access networks, and ground-based node networks), terrestrial internet, and terrestrial mobile communication networks, and adopts a unified technical system and standards. Among these, terrestrial networks such as the terrestrial internet and mobile communication networks primarily provide internet users with access to satellite networks.

[0003] The topology is dynamic and the network elements are heterogeneous, resulting in a wide variety of services. Different services have different requirements for Quality of Service (QoS). How to accurately and efficiently distinguish between different types of service flows involves the field of service awareness.

[0004] Networks based on IP (Internet Protocol) packet technology offer a rich variety of services. Service awareness technology is used to distinguish between different service flows within the network. Service awareness encompasses three concepts: the object of perception, the method of perception, and the result of perception. The object of perception can be data packets or service flows. The method of perception can be through protocol parsing, traffic feature analysis, or a combination of both. The result of perception is the outcome required by the perception action, i.e., the corresponding classification requirement. Service awareness technology includes service classification techniques and service feature extraction techniques, corresponding to the determination of the perception subject and the specific means of perception, respectively.

[0005] Integrated air-space-ground information networks require timely perception of service information within the current network environment. This information will be used in subsequent cognitive processing processes such as planning and decision-making. Currently, most service flow classification algorithms primarily focus on service perception and recognition in single scenarios. Furthermore, services in integrated air-space-ground networks are often dynamically changing and require rapid response; existing service perception and recognition algorithms struggle to meet the demands for accurate service identification. A suitable intelligent service recognition model can enhance the adaptability of network service dynamic management and configuration to changes in various user services, improving the accuracy of mixed service perception and recognition. Therefore, current service perception and recognition algorithms require further optimization.

[0006] In existing technologies, Wang ZJ, Dong YN, Zhang H, et al. A multimedia traffic classification method based on improved Hidden Markov Model[J]. Journal of Electronics & Information Technology, 2015, 37(2): 499-503, uses packet size and packet arrival time interval as features to establish a Markov Model (HMM) to achieve coarse-grained classification of traffic flows, which can optimize the classification accuracy of traffic flows. However, there are some technical bottlenecks that have not been effectively resolved, including:

[0007] (1) The problem of imbalanced samples was not considered: The number and proportion of categories in the training samples have a great influence on the classification results, so the requirements for sample collection are very high.

[0008] (2) Failure to consider the use of unlabeled data: In many real-world problems, on the one hand, the cost of manually labeling samples is very high, resulting in very few labeled samples. On the other hand, unlabeled samples are easy to collect, and their number is often hundreds of times that of labeled samples. Ignoring unlabeled data means missing a large number of potential training samples.

[0009] (3) The impact of dynamic changes in business on the accuracy of classification is not considered. Summary of the Invention

[0010] The technical problem to be solved by this invention is to address the impact of failing to consider sample imbalance, neglecting the use of unlabeled data, and the dynamic changes in business operations on the prediction accuracy of classification.

[0011] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0012] A method for intelligent perception and accurate classification of panoramic business operations includes the following steps:

[0013] S100, the perception probe collects feature data of different business types from the business terminal information, divides the feature data into labeled data and unlabeled data, and constructs a feature subset of the labeled data based on a hybrid feature selection method, and obtains the globally optimal feature subset;

[0014] S200, Simultaneously, the labeled data is divided into labeled training data and labeled test data, and a semi-supervised learning model based on the improved K-means clustering algorithm is trained using the labeled training data and the labeled test data. Based on the trained semi-supervised learning model based on the improved K-means clustering algorithm, the labeled training data and the unlabeled data are clustered, the unlabeled data is infected, and the business-side infection data is output.

[0015] S300, the sensing probe collects, stores and transmits service request information from the user terminal, and uses the service request information as prior information. The prior information and the service terminal infection data are used to train the convolutional neural network model. The trained convolutional neural network model extracts and mines service features from the prior information and the service terminal infection data, and outputs high-level features.

[0016] S400, the prior information and the business terminal infection data are then used to train the random forest business perception and recognition model, and the trained random forest business perception and recognition model is used to classify the high-level features to achieve the recognition of multiple types of client services.

[0017] Advantages: By processing a large number of unlabeled data samples through a semi-supervised learning model based on an improved K-means clustering algorithm and including them in the same cluster with a small number of labeled data samples, the utilization of unlabeled data is achieved, and a balanced distribution of samples within the same category is maintained. By adding prior information input and periodically training the convolutional neural network model, the high-level features output can improve recognition accuracy in subsequent random forest business perception and recognition models, avoiding the impact of dynamic business changes on classification.

[0018] In one embodiment of the present invention, the construction of a feature subset of the labeled data based on a hybrid feature selection method and the acquisition of the globally optimal feature subset include the following steps:

[0019] S110, using five feature selection algorithms—FCBF, InfoGain, GainRatio, Chi.square, and Consistency—to output feature subsets of the labeled data respectively;

[0020] S120: The KNN algorithm is used to evaluate the effectiveness of the five feature subsets output in step S110, and the top three feature subsets with the highest accuracy are obtained.

[0021] S130, the union of the top three feature subsets in terms of accuracy is obtained to get the frequency of each feature, and features with too low frequency are removed to obtain the remaining feature subsets;

[0022] S140, the remaining feature subset is used as the output result of selectively integrating feature subsets;

[0023] S150, starting from an empty set, select one feature from the unselected features each time, and combine this feature with the selectively integrated feature subset to obtain a feature combination subset. Then, use the KNN algorithm again to evaluate the accuracy of the feature combination subset. When the accuracy no longer increases after adding features, the search process ends.

[0024] S160 selects the feature combination subset with the highest accuracy as the globally optimal feature subset.

[0025] In one embodiment of the present invention, the semi-supervised learning model based on the improved K-means clustering algorithm, which has been trained, clusters the labeled training data and the unlabeled data, infects the unlabeled data, and outputs business-side infected data, including the following steps:

[0026] S210, remove noisy samples;

[0027] S220, perform clustering of minority sample classes into similar clusters, and obtain the sampling weight of each minority sample cluster based on the number of samples contained in each cluster in each minority sample class; and obtain the number of interpolations required for each minority sample cluster based on the sampling weight and the difference between the number of samples in the most sample class remaining after removing the noise samples and the number of samples in the minority sample class, so as to satisfy the balanced distribution among similar samples.

[0028] S230, based on the sample weights of each sample in the minority sample cluster, and based on the sample weights, select a minority sample class sample that is difficult to learn but contains important information as a seed sample, set the weight of the selected sample to a fixed value, and obtain the weight gain; and based on the weight gain and the weight of the selected sample, normalize it to obtain the probability of each sample in the cluster being sampled.

[0029] S240, each time a portion of reference samples are selected from the sample set of the minority sample class, and the interpolation position for interpolating the minority sample cluster is obtained. The reference samples are then interpolated at the interpolation position in combination with the interpolation number.

[0030] S250: Determine the number of clusters based on the maximum value of the transformed cluster profile coefficients, and perform K-means clustering.

[0031] S260: Determine the class without label data based on the class pairs with labeled data in the cluster, and complete the infection of the unlabeled data.

[0032] S270, re-evaluate the number of samples in each category. If the ratio of the number of samples in the most frequent category to the number of samples in the least frequent category is greater than the threshold, then execute steps S210-S240 again; otherwise, execute step S300.

[0033] In one embodiment of the present invention, the sampling weight of each minority sample cluster is obtained by the following formula:

[0034]

[0035] In the formula, W i Let N be the sampling weight of the i-th minority sample cluster, and N be the number of clusters in each minority sample class. i n is the number of samples contained in the i-th cluster of the minority sample class. j This represents the number of samples contained in the j-th cluster of the minority sample class;

[0036] The number of interpolations required for each minority sample cluster is obtained using the following formula:

[0037] num i =(N max -N min W i ;

[0038] In the formula, num i N represents the number of interpolations required for the i-th minority sample cluster; Nmax represents the maximum number of sample classes; Nmix represents the minority sample classes.

[0039] The weight gain is obtained using the following formula:

[0040]

[0041] In the formula, p represents the weight gain, n represents the number of non-similar samples among the Q nearest neighbors of the sample point, and d i This is the Euclidean distance between the sample point and each of its neighboring non-class samples;

[0042] The interpolation position for interpolating the minority sample cluster is obtained using the following formula:

[0043]

[0044] In the formula, x i The value represents the interpolation position of the i-th interpolation point; m represents the number of reference samples, and d represents the interpolation position of the i-th interpolation point. c Let d represent the Euclidean distance from the seed sample point to the c-th reference sample. j Let P be the Euclidean distance of the j-th reference sample. ic It is represented as the feature value of the c-th reference sample corresponding to the i-th interpolation point.

[0045] In one embodiment of the present invention, in step S250, the distance between samples in K-means clustering is obtained by the following formula:

[0046]

[0047] In the formula, X and Y represent any two sample points, and X = (x1, x2, ..., xy). n ); Y = (y1, y2, ..., y n ), dist ed (X,Y) represents the distance between two sample points;

[0048] The sum of squared errors is used as the objective function for K-means clustering, which is obtained by the following formula:

[0049]

[0050] In the formula, SSE represents the objective function, K represents the number of clusters selected for clustering, and C represents the number of clusters selected for clustering. i Let x represent the cluster center and x represent the cluster it belongs to.

[0051] In one embodiment of the present invention, in step S260, the cluster category mapping of infected unlabeled data is divided into the following three cases based on the number of categories of samples with labeled data in the same cluster:

[0052] The first type: The same cluster contains only one type of labeled sample, and the categories of all other unlabeled samples are mapped to that category;

[0053] The second method is as follows: If the same cluster contains labeled samples of different categories, count the number of each category in the labeled samples in the cluster. The number of the most labeled category and the second most n2 are counted. Only when n1 ≥ 2n2 are the remaining unlabeled samples in the cluster mapped to the primary category. Otherwise, the n1-category cluster is discarded and awaits manual verification.

[0054] The third type: The cluster does not contain any labeled samples and requires subsequent manual verification.

[0055] In one embodiment of the present invention, the high-level feature extraction includes the following steps:

[0056] S310 creates a one-hot code for the service;

[0057] S320 sums the one-hot encoded data of the business requests made by the user end to create a priori information feature vector.

[0058] S330, the prior information feature vector is expanded to have the same spatial dimension as the business-end infection data, and then concatenated with the business-end infection data along the channel dimension to obtain a new input tensor containing the business-end infection data and prior information;

[0059] S340, The new input tensor is input into the convolutional neural network model for training, and high-level features are extracted.

[0060] In one embodiment of the present invention, in step S400, during the training phase of the decision tree of the random forest business perception and recognition model, the prediction class performance of each decision tree is evaluated, and higher weights are assigned to decision trees that can accurately predict a small number of samples, and the final prediction result is obtained through weighted voting.

[0061] In one embodiment of the present invention, the prediction result of the weighted random forest business perception and identification model is defined as:

[0062]

[0063] In the formula, H(T) represents the prediction result, T is the number of decision trees; I(·) is the indicator function; f t (T) represents the prediction result of the t-th decision tree; y represents the category; ω t Let t be the voting weight of the t-th decision tree; the indicator function is 1 when the prediction result of the decision tree is true, and 0 otherwise.

[0064] In one embodiment of the present invention, assuming TP represents classifying a stable sample as a stable sample, FN represents classifying a stable sample as an unstable sample, FP represents classifying an unstable sample as a stable sample, and TN represents classifying an unstable sample as an unstable sample, the voting weight of each decision tree is defined as follows: (The text then goes on to describe the voting weight of each decision tree, which is not directly related to the previous sentence.)

[0065]

[0066]

[0067]

[0068] In the formula, F1 represents the harmonic mean of recall, and P re Represented as accuracy, R ec This is expressed as recall rate.

[0069] Compared with the prior art, the beneficial effects of the present invention are:

[0070] This invention takes into account the large differences in sample size among different business categories, which can lead to significant biases in the analysis results of classification-related modeling problems. Therefore, it adopts a semi-supervised learning model based on an improved K-means clustering algorithm to solve the sample imbalance problem through oversampling.

[0071] This invention takes into account the training value of a large number of unlabeled data samples. Therefore, it adopts a semi-supervised learning model based on an improved K-means clustering algorithm. By using a large number of unlabeled data samples and a small number of labeled data samples for training the classifier through semi-supervised learning, the problem of insufficient labeled samples is solved.

[0072] This invention takes into account the rapid response requirements of dynamic services in space-ground integrated networks. Therefore, it adopts a convolutional neural network method to combine service feature data with prior information from the user end to extract and mine service features, enhance the adaptability of network service dynamic management and configuration to changes in multiple types of user services, and improve the accuracy of mixed service perception and identification.

[0073] To improve the prediction ability of random forests for minority class samples, the decision trees of the random forest business perception and recognition model evaluate the prediction class performance of each decision tree during the training phase, and assign higher weights to decision trees that can accurately predict minority samples, and obtain the final prediction results through weighted voting. Attached Figure Description

[0074] Figure 1 This is a flowchart illustrating a method for intelligent perception and precise classification of panoramic services according to an embodiment of the present invention.

[0075] Figure 2 This is a flowchart illustrating the process of obtaining the globally optimal feature subset according to an embodiment of the present invention.

[0076] Figure 3 This is a flowchart illustrating the output of infected data from the service terminal in an embodiment of the present invention.

[0077] Figure 4 This is a flowchart of high-level feature extraction according to an embodiment of the present invention. Detailed Implementation

[0078] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0079] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0080] Please see Figure 1 As shown, this invention provides a method for intelligent perception and accurate classification of panoramic services, comprising the following steps:

[0081] S100, the perception probe collects feature data of different business types from the business end information, divides the feature data into labeled data and unlabeled data, constructs feature subsets of the labeled data based on a hybrid feature selection method, and obtains the globally optimal feature subset.

[0082] S200, simultaneously, the labeled data is divided into labeled training data and labeled test data, and a semi-supervised learning model based on the improved K-means clustering algorithm is trained using the labeled training data and the labeled test data. Based on the trained semi-supervised learning model based on the improved K-means clustering algorithm, the labeled training data and the unlabeled data are clustered, the unlabeled data is infected, and the business-side infection data is output.

[0083] S300, the sensing probe collects, stores, and transmits service request information from the user terminal, and uses the service request information as prior information. The prior information and the service terminal infection data are used to train a convolutional neural network model. The trained convolutional neural network model extracts and mines service features from the prior information and the service terminal infection data, and outputs high-level features.

[0084] S400, the prior information and the business terminal infection data are then used to train the random forest business perception and recognition model, and the trained random forest business perception and recognition model is used to classify the high-level features to achieve the recognition of multiple types of client services.

[0085] Please see Figure 1 and Figure 2As shown, in one embodiment of the present invention, in a space-ground integrated network scenario, the system uses perception probes to collect, store, transmit, and process feature data of different service types. The collected data is called a "five-tuple," which consists of an IP address, source port number, destination IP address, destination port number, and protocol type. Features are extracted from the collected "five-tuples." To improve the accuracy of feature extraction, a hybrid feature selection algorithm is used in this step. The hybrid feature extraction method mainly consists of two parts: the first part uses a selective ensemble method to integrate multiple feature selection algorithms to obtain a more stable feature subset than a single classifier. The second part uses a heuristic search to remove redundant and irrelevant features from the features obtained by the selective ensemble in the first part. Specifically, the construction of a feature subset based on the hybrid feature selection method for the labeled data and the acquisition of the globally optimal feature subset includes the following steps:

[0086] S110 uses five feature selection algorithms—FCBF, InfoGain, GainRatio, Chi.square, and Consistency—to output feature subsets of the labeled data.

[0087] S120: The KNN algorithm is used to evaluate the effectiveness of the five feature subsets output in step S110, and the top three feature subsets with the highest accuracy are obtained.

[0088] Among them, the five feature selection algorithms, FCBF, InfoGain, GainRatio, Chi.square, and Consistency, and the KNN algorithm are all existing technologies.

[0089] S130, the union of the top three feature subsets in terms of accuracy is obtained to get the frequency of each feature, and features with too low frequency are removed to obtain the remaining feature subset.

[0090] S140, the remaining feature subset is used as the output result of selectively integrating feature subsets.

[0091] S150, starting from the empty set, select one feature from the unselected features each time, and combine this feature with the selectively integrated feature subset to obtain a feature combination subset. Then, use the KNN algorithm again to evaluate the accuracy of the feature combination subset. When the accuracy no longer increases after adding features, the search process ends.

[0092] S160 selects the feature combination subset with the highest accuracy as the globally optimal feature subset.

[0093] Please see Figure 1 and Figure 3As shown, in one embodiment of the present invention, the K-means clustering algorithm is an iterative clustering analysis algorithm that classifies and organizes data members in a dataset that are similar in some aspects. Traditional K-Means clustering algorithms may suffer from imbalanced samples, resulting in minority class samples not being accurately clustered. Therefore, this step improves the existing K-Means clustering algorithm. The semi-supervised learning model based on the improved K-means clustering algorithm is as follows:

[0094] S210, remove noisy samples.

[0095] In the original dataset, if the proportion of similar samples among the Q nearest neighbors of a sample is less than 20%, the sample can be considered a noise sample point.

[0096] S220, perform clustering of minority sample classes, and obtain the sampling weight of each minority sample cluster based on the number of samples contained in each cluster in each minority sample class. Then, obtain the number of interpolations required for each minority sample cluster based on the sampling weight and the difference between the number of samples in the largest sample class remaining after removing the noise samples and the number of samples in the minority sample class, so as to satisfy the balanced distribution among samples of the same class.

[0097] The sampling weights of each minority sample cluster are obtained using the following formula:

[0098]

[0099] In the formula, W i Let N be the sampling weight of the i-th minority sample cluster, and N be the number of clusters in each minority sample class. i n is the number of samples contained in the i-th cluster of the minority sample class. j This represents the number of samples contained in the j-th cluster of the minority sample class.

[0100] The more samples a cluster of a minority sample class has, the greater its proportion of the total number of samples in that minority sample class. This results in a smaller sampling weight, meaning a smaller assigned weight, and thus fewer synthesized samples in the later stages, ultimately achieving a balanced distribution among samples of the same class.

[0101] The number of interpolations required for each minority sample cluster is obtained using the following formula:

[0102] num i =(N max -N min W i ;

[0103] In the formula, num iNmax represents the number of interpolations required for the i-th minority sample cluster; Nmax represents the maximum number of sample classes; and Nmix represents the minority sample classes.

[0104] S230, based on the sample weights of each sample in the minority sample cluster, and based on the sample weights, select a minority sample class sample that is difficult to learn but contains important information as a seed sample, set the weight of the selected sample to a fixed value, obtain the weight gain, and normalize it based on the weight gain and the weight of the selected sample to obtain the probability of each sample in the cluster being sampled.

[0105] In classification tasks, minority class samples, which are closer to the decision boundary, are more prone to misclassification, thus increasing the learning difficulty for these samples. Therefore, it's necessary to screen oversampled samples. This is because the sample weights of each sample within the minority cluster must also be considered. Based on these weights, difficult-to-learn but information-rich minority class samples are selected as seed samples to ensure the quality of the synthesized samples. These seed samples are reference samples from the minority class set, used as the basis for interpolation.

[0106] The weight gain is obtained using the following formula:

[0107]

[0108] In the formula, p represents the weight gain, n represents the number of non-similar samples among the Q nearest neighbors of the sample point, and d i This represents the Euclidean distance between the sample point and each of its neighboring non-class samples.

[0109] The probability of each sample being selected is determined by its distance from the boundaries of other classes. The more neighboring samples from other classes a sample is, the higher its probability of being selected. This takes into account the distribution characteristics of the samples and effectively expands the minority class decision boundary. The initial sample weight of the selected sample is set to a fixed value of 1.

[0110] S240, each time a portion of reference samples are selected from the sample set of the minority sample class, and the interpolation position for interpolating the minority sample cluster is obtained. The reference samples are then interpolated at the interpolation position in combination with the interpolation number.

[0111] The steps are as follows: 1. Select a seed sample point based on probability. 2. Select the m nearest samples to the seed sample point as reference samples. 3. Obtain the interpolation position. 4. Repeat steps 1-3 multiple times for interpolation.

[0112] The interpolation position for interpolating the minority sample cluster is obtained using the following formula:

[0113]

[0114] In the formula, x i The value represents the interpolation position of the i-th interpolation point; m represents the number of reference samples, and d represents the interpolation position of the i-th interpolation point. c Let d represent the Euclidean distance from the seed sample point to the c-th reference sample. j Let P be the Euclidean distance of the j-th reference sample. ic It is represented as the feature value of the c-th reference sample corresponding to the i-th interpolation point.

[0115] S250: Based on the maximum value of the transformed cluster profile coefficients, determine the number of clusters and perform K-means clustering.

[0116] In step S250, the distance between samples in K-means clustering is obtained using the following formula:

[0117]

[0118] In the formula, X and Y represent any two sample points, and X = (x1, x2, ..., xy). n ); Y = (y1, y2, ..., y n ), dist ed (X,Y) represents the distance between two sample points.

[0119] The goal of K-means clustering is to minimize the differences between samples within the same cluster and maximize the differences between samples in different clusters. It uses the sum of squared errors as the objective function, which is obtained using the following formula:

[0120]

[0121] In the formula, SSE represents the objective function, K represents the number of clusters selected for clustering, and C represents the number of clusters selected for clustering. i Let x represent the cluster center and x represent the cluster it belongs to. Calculate the Euclidean distance from the cluster to the cluster center, and sum the distances of all sample points to their respective cluster centers to obtain the objective function. Determine the number of centroids for K-means clustering using the silhouette coefficient transformation formula, and achieve clustering by minimizing the objective function.

[0122] S260: Determine the class of unlabeled data based on the class pairs with labeled data in the cluster, and complete the infection of the unlabeled data.

[0123] In this clustering algorithm, a large number of unlabeled samples and a small number of labeled samples are grouped into the same cluster. The cluster category mapping for unlabeled data is specifically divided into three cases based on the number of categories of labeled samples within the same cluster:

[0124] The first type: The same cluster contains only one type of labeled sample, and the categories of all other unlabeled samples are mapped to that category.

[0125] The second method is as follows: If the same cluster contains labeled samples of different categories, count the number of each category in the labeled samples in the cluster. The number of the most labeled category and the second most n2 are counted. Only when n1 ≥ 2n2 are the remaining unlabeled samples in the cluster mapped to the primary category. Otherwise, the n1-category cluster is discarded and awaits manual verification.

[0126] The third type: The cluster does not contain any labeled samples and requires subsequent manual verification.

[0127] The labeled samples mentioned here refer to samples with labeled data, while the unlabeled samples refer to samples without labeled data.

[0128] S270, re-evaluate the number of samples in each category. If the ratio of the number of samples in the most frequent category to the number of samples in the least frequent category is greater than the threshold, then execute steps S210-S240 again; otherwise, execute step S300.

[0129] In this step, the threshold is 1.1. Thus, the initial small set of network flow samples with class labels and the large set of samples without class labels have been expanded into a large set of network flow samples with accurate class labels, awaiting classification by the subsequent random forest business perception and recognition model. Furthermore, in K-means clustering, the reference sample's sample points are data within the globally optimal feature subset.

[0130] Please see Figure 1 and Figure 4 As shown, in one embodiment of the present invention, during actual user interaction, the patterns of business operations change over time. The proposed intelligent business recognition model needs to consider these changes in the recognition target and make adjustments accordingly. Therefore, prior information from the user end is required to assist in the recognition process. Specifically, the high-level feature extraction includes the following steps:

[0131] S310 creates a one-hot code for the business.

[0132] S320 sums the one-hot encoded data of the services requested by the user and creates a priori information feature vector.

[0133] S330, the prior information feature vector is expanded to have the same spatial dimension as the business-side infection data, and then concatenated with the business-side infection data along the channel dimension to obtain a new input tensor containing the business-side infection data and prior information.

[0134] S340, The new input tensor is input into the convolutional neural network model for training, and high-level features are extracted.

[0135] Among them, the prior information is the historical business request information of the user terminal. By adding prior information input, the convolutional neural network model is trained regularly to avoid the impact of dynamic changes in business on classification. This prevents the high-level features output from reducing the recognition accuracy in the subsequent random forest business perception and recognition model.

[0136] Please see Figures 1 to 4 As shown, in one embodiment of the present invention, traditional random forests treat all decision trees indiscriminately, resulting in decision trees with different prediction performances having the same voting weight. To improve the prediction ability of random forests for minority class samples, a weight-based random forest algorithm is proposed. During the training phase of the random forest business perception and recognition model, the prediction performance of each decision tree is evaluated, and decision trees that can accurately predict minority samples are assigned higher weights. The final prediction result is obtained through weighted voting. The prediction result of the weight-based random forest business perception and recognition model is defined as follows:

[0137]

[0138] In the formula, H(T) represents the prediction result, T is the number of decision trees, I(·) is the indicator function, and f t (T) represents the prediction result of the t-th decision tree, y represents the category, and ω t Let represent the voting weight of the t-th decision tree. argmax is a function that indicates the value of the function is 1 when the prediction result of the decision tree is true, and 0 otherwise.

[0139] Assuming TP represents classifying a stable sample as stable, FN represents classifying a stable sample as unstable, FP represents classifying an unstable sample as stable, and TN represents classifying an unstable sample as unstable, the voting weight of each decision tree is defined as follows: (The text then goes on to describe the voting weight of each decision tree, which is not directly related to the previous sentence.)

[0140]

[0141]

[0142]

[0143] In the formula, F1 represents the harmonic mean of recall, and P re Represented as accuracy, R ec This is expressed as recall. The higher the F1 score, the better the prediction performance of the decision tree for a small number of samples.

[0144] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0145] The above embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for panoramic service intelligent sensing and accurate classification, characterized in that, The method comprises the following steps: S100, collecting feature data of different service types of the business end information by the sensing probe, dividing the feature data into labeled data and unlabeled data, constructing a feature subset of the labeled data based on a hybrid feature selection method, and obtaining a globally optimal feature subset; S200, dividing the labeled data into labeled training data and labeled test data, training a semi-supervised learning model based on an improved K-means clustering algorithm by using the labeled training data and the labeled test data, clustering the labeled training data and the unlabeled data based on the trained semi-supervised learning model based on the improved K-means clustering algorithm, infecting the unlabeled data, and outputting business end infection data, comprising: S210, removing noise samples; S220, clustering the same class of a small number of samples, obtaining a sampling weight of each small sample cluster according to the number of samples contained in each cluster of each small sample class, and obtaining the number of interpolation required for each small sample cluster according to the sampling weight and the difference between the number of samples of the most sample class and the number of samples of the small sample class after removing the noise samples, so as to balance the distribution between the same samples; S230, selecting a small number of sample classes containing important information as seed samples according to the sample weight of each sample in the small sample cluster and according to the sample weight, setting the selected sample weight to a fixed value, obtaining a weight gain, and normalizing the selected sample weight according to the weight gain, obtaining the probability of sampling each sample in the cluster; S240, selecting part of the reference samples from the sample set of the small sample class each time, obtaining the interpolation position of the small sample cluster, interpolating the reference samples at the interpolation position, and interpolating the reference samples at the interpolation position; S250, determining the number of clusters of the clustering according to the maximum value point of the converted clustering contour coefficient, and performing K-means clustering; S260, determining the class of the unlabeled data according to the class of the labeled data in the cluster, and completing the infection of the unlabeled data; S270, rejudging the number of samples of each class, if the ratio of the number of samples of the most class to the number of samples of the least class is greater than a threshold value, then steps S210-S240 are executed again; otherwise, step S300 is executed; The sampling weight of each small sample cluster is obtained by the following formula: ； In the formula, W i For the first i The sampling weights of a few sample clusters N The number of clusters in each minority sample class. n i For the minority sample class The number of samples contained in a cluster. n j For the minority sample class j The number of samples contained in each cluster; The number of interpolation required for each small sample cluster is obtained by the following formula: ； where num i is the number of clusters for the minority class; is the number of clusters for the minority class; N is the number of clusters for the minority class; The weight gain is obtained by the following formula: ； wherein p is represented as a weight gain, n is represented as the number of non-similar samples among the Q nearest neighbors of the sample point Q, d i is the Euclidean distance between the sample point and each of the nearest non-similar samples. The interpolation position of the small sample cluster is obtained by the following formula: ； In the formula, x i Represented as the first i The interpolation position of the interpolation point; m This is represented as the reference sample size. d c Represented as seed sample point to the th c Euclidean distance of each reference sample d j Represented as the first j Euclidean distance of each reference sample P ic Represented as the first i The interpolation point corresponding to the first c Feature values ​​of one reference sample; And the sample point of the reference sample is the data in the globally optimal feature subset. S300, the perception probe collects, stores and transmits the service request information of the user end, and the service request information is used as prior information, the prior information and the service end infection data are used for training a convolutional neural network model, the trained convolutional neural network model is used for extracting and mining service characteristics from the prior information and the service end infection data, and high-level features are output; S400, the prior information and the service end infection data are used for training a random forest service perception and identification model, and the trained random forest service perception and identification model is used for classifying the high-level features, so as to realize identification of multiple types of client services.

2. The method for panoramic business intelligence perception and accurate classification according to claim 1, characterized in that, The labeled data is subjected to feature subset construction based on the hybrid feature selection method, and a globally optimal feature subset is obtained, including the following steps: S110, five feature selection algorithms of FCBF, InfoGain, GainRatio, Chi. square and Consistency are used to output feature subsets of the labeled data respectively; S120, the effectiveness of the five feature subsets output in step S110 is evaluated respectively by using the KNN algorithm, and the feature subsets with top three accuracy rates are obtained; S130, the feature subsets with top three accuracy rates are combined to obtain the frequency of each feature, and the features with low frequency are removed to obtain the remaining feature subsets; S140, the remaining feature subsets are used as the output results of the selective integrated feature subsets; S150, starting from an empty set, one feature is selected from the unselected features each time, and the feature is added to the combination of the selective integrated feature subsets to obtain a feature combination subset, and the KNN algorithm is used again to evaluate the accuracy rate of the feature combination subset, and when the accuracy rate does not increase after adding the feature, the search process is ended; S160, the feature combination subset with the highest accuracy rate obtained is used as the globally optimal feature subset.

3. The method for panoramic business intelligence perception and accurate classification according to claim 1, characterized in that, In step S250, the distance between samples of K-means clustering is obtained by the following formula: ； wherein X and Y is expressed as any two sample points, and ; , is expressed as the distance between two sample points; And the sum of squared errors is used as the objective function of K-means clustering, which is obtained by the following formula: ； wherein SSE is expressed as an objective function, K is expressed as a cluster number of cluster selection, is expressed as a cluster center, x is expressed as a cluster in which the data is located.

4. The method for panoramic business intelligence perception and accurate classification according to claim 1, characterized in that, The high-level feature extraction includes the following steps: S310, a one-hot encoding is created for the service; S320, the one-hot encoding of the user end request service is summed to create a prior information feature vector; S330, the prior information feature vector is expanded to have the same spatial dimension as the service end infection data, and is spliced along the dimension of the service end infection data to obtain a new input tensor containing the service end infection data and the prior information; S340, the new input tensor is input into the convolutional neural network model for training to extract high-level features.

5. The method for panoramic business intelligence perception and accurate classification according to claim 1, characterized in that, In step S400, the decision tree of the random forest service perception and identification model evaluates the prediction class performance of each decision tree in the training stage, and gives higher weight to the decision tree which can accurately predict a small number of samples, and obtains the final prediction result by weighted voting.

6. The method for panoramic business intelligence perception and accurate classification according to claim 5, characterized in that, Wherein, The prediction result of the weight-based random forest business perception and recognition model is defined as: ； In the formula, H(T) represents a prediction result, and T is the number of decision trees; is an indicator function; is a prediction result of the tth decision tree; t is a prediction result of the tth decision tree; y represents a category; is a voting weight of the tth decision tree; when the prediction result of the decision tree is true, the value of the indicator function is 1, and otherwise, the value of the indicator function is 0.

7. The method for panoramic business intelligence perception and accurate classification according to claim 6, characterized in that, Assume represents that the stable sample is discriminated as a stable sample, represents that the stable sample is discriminated as an unstable sample, represents that the unstable sample is discriminated as a stable sample, represents that the unstable sample is discriminated as an unstable sample, and the voting weight of each decision tree is defined as: ； ； ； In the formula, The harmonic mean expressed as recall rate, The precision expressed as precision, The recall expressed as recall.

Images