User intention grading method, device and electronic equipment
By training a user intention classification model using machine learning methods such as one-to-many decomposition and Fisher discriminant analysis, the problems of low efficiency and difficulty in multi-level intention classification in existing technologies are solved, and fast and accurate user intention classification and precision marketing are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA MOBILE GROUP JIANGSU
- Filing Date
- 2022-04-14
- Publication Date
- 2026-06-12
AI Technical Summary
Existing methods for classifying user intention levels are labor-intensive and time-consuming, inefficient, and traditional binary classification models cannot achieve multi-level intention classification.
A machine learning approach using one-to-many decomposition and Fisher discriminant analysis is employed to train a user intention classification model. The model is trained using multidimensional sample data from multiple operators, transforming it into a binary classification problem to achieve multi-classification and obtain at least three classification types.
It improves the efficiency of user intent grading and the targeting of precision marketing, and enables multi-level grading of user intent.
Smart Images
Figure CN116975751B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence technology, specifically to a user intention classification method, apparatus, and electronic device. Background Technology
[0002] Currently, user intent level segmentation is widely used in the field of precision marketing. Existing user intent level segmentation often relies on manual survey data, with experts stratifying potential users into several levels. Alternatively, machine learning algorithms such as KNN (K-Nearest Neighbor) and random forests are used for binary classification to categorize users into different intent levels.
[0003] However, expert-based stratification methods are costly in terms of manpower and time, have long evaluation cycles, and are inefficient. Binary classification models, which divide users into two categories—those with and without intention—cannot provide further levels of user intention grading. Summary of the Invention
[0004] This application provides a user intention classification method, apparatus, and electronic device to improve the efficiency of user intention classification and to achieve more levels of user intention classification, thereby improving the targeting of precision marketing to potential users.
[0005] In a first aspect, embodiments of this application provide a method for classifying user intent, including:
[0006] Obtain multi-dimensional data about the user's carrier;
[0007] Obtaining user intention classification results: Input the operator's multidimensional data into the user intention classification model to obtain the user intention classification results output by the user intention classification model;
[0008] The user intention classification result includes at least three classification types; and the user intention classification model outputs the user intention classification result of a single classification type at a time.
[0009] The user intention classification model is trained using multidimensional sample data from multiple operators based on a one-to-many decomposition method and Fisher's discriminant method.
[0010] In one embodiment, before the step of acquiring operator multidimensional data, a step of training the user intention classification model is also included;
[0011] Training the user intention classification model includes:
[0012] Obtain multi-dimensional sample data of multiple users from various carriers;
[0013] Based on the one-to-many decomposition method, the multidimensional sample data of multiple users' operators are decomposed to obtain multiple datasets; each dataset includes one multidimensional sample data of operators as the positive class and multiple multidimensional sample data of operators as the negative class.
[0014] Based on the Fisher discriminant method, binary classification is performed on each of the datasets to obtain the confusion matrix of the classifier corresponding to each dataset.
[0015] Multiple binary classification discriminant models are determined based on the confusion matrices of each of the classifiers;
[0016] Multiple binary classification discriminant models are used as the user intention classification model;
[0017] The binary classification model includes at least three models, and the number of binary classification models is the same as the number of classification types in the user intention classification result.
[0018] In one embodiment, determining multiple binary classification discriminant models based on the confusion matrices of each of the classifiers includes:
[0019] The precision and recall of each classifier are determined based on the confusion matrix of each classifier.
[0020] Calculate the balanced F-score of each classifier based on its precision and recall.
[0021] The classifiers with the highest balanced F-scores are determined as the plurality of binary classification discriminant models.
[0022] In one embodiment, the user intent classification method further includes:
[0023] Calculate the probability that each negative class is classified as a positive class under each of the aforementioned binary classification discriminant models;
[0024] If the binary classification model determines that the negative class is classified as positive class when it exceeds the set probability threshold, the number of positive and negative classes in the dataset of the one-to-many decomposition method is readjusted to obtain multiple updated datasets.
[0025] Based on the aforementioned multiple updated datasets, multiple updated binary classification discriminant models are obtained by retraining using the Fisher discriminant method;
[0026] Multiple updated binary classification discriminant models are used as the user intention classification model;
[0027] The updated binary classification model includes at least three models, and the number of updated binary classification models is the same as the number of classification types in the user intention classification result.
[0028] In one embodiment, the step of inputting the operator's multidimensional data into the user intention classification model to obtain the user intention classification result output by the user intention classification model includes:
[0029] If only one of the binary classification models predicts a positive result, the classification result output by the binary classification model shall be used as the user intention classification result.
[0030] If multiple binary classification models predict positive results, the classification result output by the binary classification model with the highest confidence among the multiple binary classification models is taken as the user intention classification result.
[0031] In one embodiment, after obtaining multi-dimensional sample data of multiple users' operators, the method further includes:
[0032] The data type of the sample is determined based on the number of non-repeating values in each sample data in the operator's multidimensional sample data;
[0033] When the sample data type is continuous sample data, perform variance analysis on the continuous sample data and delete sample data with F value less than a first set threshold;
[0034] When the sample data type is discrete sample data, a chi-square test is performed on the discrete sample data, and sample data with a P-value greater than a second set threshold are deleted.
[0035] Secondly, embodiments of this application provide a user intention classification device, comprising:
[0036] The first acquisition module is used to acquire the user's multi-dimensional operator data;
[0037] The second acquisition module is used to acquire user intention classification results: inputting the operator's multidimensional data into the user intention classification model to obtain the user intention classification results output by the user intention classification model;
[0038] The user intention classification result includes at least three classification types; and the user intention classification model outputs the user intention classification result of a single classification type at a time.
[0039] The user intention classification model is trained using multidimensional sample data from multiple operators based on a one-to-many decomposition method and Fisher's discriminant method.
[0040] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the program to implement the steps of the user intention classification method described in the first aspect.
[0041] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the user intention classification method described in the first aspect.
[0042] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the user intention classification method described in the first aspect.
[0043] The user intention classification method, apparatus, and electronic device provided in this application acquire multi-dimensional operator data of users; input the multi-dimensional operator data into a user intention classification model to obtain the user intention classification result output by the user intention classification model. Compared with the expert stratification method, this application uses a machine learning method of one-to-many decomposition and Fisher discriminant analysis to train the user intention classification model on multi-dimensional operator sample data, enabling the user intention classification model to classify users' intentions more quickly based on their multi-dimensional operator data, thus improving the efficiency of user intention classification. Compared with traditional machine learning binary classification models, this application uses a machine learning method of one-to-many decomposition and Fisher discriminant analysis to train on multi-dimensional operator sample data, transforming the multi-classification problem into a binary classification problem using Fisher discriminant analysis. Since the trained user intention classification model includes at least three classification types, it achieves more levels of user intention classification, improving the targeting of precise marketing to potential users. Attached Figure Description
[0044] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0045] Figure 1 This is one of the flowcharts illustrating the user intent classification method provided in the embodiments of this application;
[0046] Figure 2 This is a second flowchart illustrating the user intention classification method provided in the embodiments of this application;
[0047] Figure 3 This is the third flowchart illustrating the user intention classification method provided in the embodiments of this application;
[0048] Figure 4 This is the fourth flowchart illustrating the user intention classification method provided in the embodiments of this application;
[0049] Figure 5 This is the fifth flowchart illustrating the user intention classification method provided in the embodiments of this application;
[0050] Figure 6 This is a schematic diagram of the user intention grading device provided in the embodiments of this application;
[0051] Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation
[0052] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0053] Please refer to Figure 1 This application provides a user intention classification method, which may include:
[0054] Step 200: Obtain the user's multi-dimensional operator data.
[0055] Electronic devices acquire multi-dimensional data from users' mobile carriers. This multi-dimensional data can include data from various carriers, such as call history, data usage, SMS messages, and app usage.
[0056] Specifically, a user's call details could include the duration or number of calls made within a predetermined timeframe. For example, a user might have made 300 minutes of calls and 35 calls within a month.
[0057] Data usage refers to the amount of data a user consumes within a specified time period. For example, a user might use 10GB of data in a month.
[0058] SMS activity can be measured by the number of SMS messages a user sent and received within a specified timeframe. For example, a user might send 10 SMS messages and receive 70 SMS messages within a month.
[0059] App usage data can be the time a user spends on various types of apps within a predetermined timeframe. For example, the time a user spends on instant messaging apps, shopping apps, and news apps within a month.
[0060] This application embodiment uses four dimensions of data—call activity, data usage, SMS activity, and app usage—as the user's multi-dimensional operator data.
[0061] It should be noted that in other embodiments, operator multidimensional data may also include operator data with three or more dimensions.
[0062] Step 300: Obtain user intention classification results. Input the operator's multidimensional data into the user intention classification model to obtain the user intention classification results output by the user intention classification model. The user intention classification results include at least three classification types; and the user intention classification model outputs a single classification type of user intention classification result at a time; the user intention classification model is trained using multidimensional sample data from multiple operators based on a one-to-many decomposition method and Fisher's discriminant method.
[0063] The electronic device inputs the operator's multidimensional data into the user intention classification model to obtain the user intention classification result output by the user intention classification model. The user intention classification model is trained using multidimensional sample data from multiple operators based on a one-to-many decomposition method and Fisher's discriminant method.
[0064] The multi-dimensional sample data from multiple operators can be multi-dimensional sample data from multiple users, used as the training data for the user intention classification model. The multi-dimensional sample data includes sample data from four dimensions: user call activity, data usage, SMS activity, and app usage. Table 1 represents the multi-dimensional sample data. Each row in the feature layer of Table 1 represents a single multi-dimensional sample data point from each operator, and the label layer represents the user intention classification corresponding to each multi-dimensional sample data point. It should be noted that the user intention classification for each multi-dimensional sample data point from each operator is unknown before being input into the user intention classification model. In this embodiment, the user intention classification is set to four categories: no intention, low intention, medium intention, and high intention, i.e., y... i ∈{0, 1, 2, 3}. In this application embodiment, 0 represents no intention; 1 represents low intention; 2 represents medium intention; and 3 represents high intention.
[0065]
[0066] Table 1
[0067] This application embodiment uses a one-to-many decomposition method and Fisher's discriminant method to train a user intention classification model on multi-dimensional sample data from multiple operators. Training with these methods transforms a multi-classification problem into a binary classification problem, enabling the user intention classification model to classify user intentions at multiple levels. Compared to the highly subjective expert-based classification method, the user intention classification model constructed using machine learning methods such as one-to-many decomposition and Fisher's discriminant method in this application embodiment can more quickly and accurately classify customer intention levels, achieving speed and accuracy while avoiding the subjectivity inherent in expert-based methods.
[0068] This application embodiment acquires multi-dimensional operator data of users; inputs the multi-dimensional operator data into a user intention classification model, and obtains the user intention classification result output by the user intention classification model. Compared with the expert stratification method, this application embodiment uses a machine learning method of one-to-many decomposition and Fisher discriminant analysis to train the user intention classification model on multi-dimensional operator sample data, enabling the user intention classification model to classify users' intentions more quickly based on the multi-dimensional operator data, thus improving the efficiency of user intention classification. Compared with traditional machine learning binary classification models, this application embodiment uses a machine learning method of one-to-many decomposition and Fisher discriminant analysis to train on multi-dimensional operator sample data, transforming the multi-classification problem into a binary classification problem using Fisher discriminant analysis. Since the trained user intention classification model includes at least three classification types, it achieves more levels of user intention classification, improving the targeting of precise marketing to potential users.
[0069] For other aspects of the embodiments of this application, please refer to Figure 2 Before step 200, which involves acquiring multidimensional data from the operator, the process also includes step 100, which involves training a user intention grading model.
[0070] Please refer to Figure 3 Step 100, training the user intention classification model, includes:
[0071] Step 110: Obtain multi-dimensional sample data of multiple users from the operator.
[0072] Electronic devices acquire multi-dimensional sample data from multiple users' carriers. This multi-dimensional sample data includes sample data from four dimensions: user call activity, data usage, SMS activity, and app usage.
[0073] Step 130: Based on the one-to-many decomposition method, decompose the multidimensional sample data of multiple users' operators to obtain multiple datasets; each dataset includes one multidimensional sample data of operators as the positive class and multiple multidimensional sample data of operators as the negative class.
[0074] The electronic device decomposes the multidimensional sample data of multiple users' operators based on the one-to-many decomposition method, resulting in multiple datasets. Each dataset includes one multidimensional sample data of operators as the positive class and multiple multidimensional sample data of operators as the negative class.
[0075] Specifically, this embodiment of the application, based on the idea of multi-class OvR (one-to-many decomposition method), transforms multi-class into a binary classification problem. The idea of the one-to-many decomposition method is to train four classifiers by taking one class's samples as the positive class and all other classes' samples as the negative classes each time. That is, the transformation is performed according to the method in Table 2 below, which represents multiple datasets in one embodiment. Each column in Table 2 represents a dataset. +1 represents the positive class, -1 represents the negative class, and f1, f2, f3, and f4 represent four classifiers. As shown in Table 2, assume that there are four multi-dimensional sample data of operators, represented by labels 0, 1, 2, and 3 respectively. Taking the column containing f1 as an example, taking the classifier f1 corresponding to the first dataset as an example, the four multi-dimensional sample data of operators are decomposed into four datasets with label 0 as the positive class and labels 1, 2, and 3 as the negative classes. The datasets in the columns containing f2, f3, and f4 can be decomposed in the same way.
[0076] <![CDATA[y i ]]> f1 f2 f3 f4 0 +1 -1 -1 -1 1 -1 +1 -1 -1 2 -1 -1 +1 -1 3 -1 -1 -1 +1
[0077] Table 2
[0078] Step 140: Perform binary classification on each dataset based on the Fisher discriminant method; obtain the confusion matrix of the classifier corresponding to each dataset.
[0079] The electronic device performs binary classification on each of the datasets based on the Fisher discriminant method, thereby obtaining the confusion matrix of the classifier corresponding to each dataset.
[0080] The electronic device performs binary classification on each dataset based on the Fisher discriminant method, which can be done as follows:
[0081] Step 1. Calculate the mean vector of each class.
[0082] Step 2. Calculate the within-class discrete matrix Swi for each class;
[0083] Step 3. Calculate the total discrete matrix Sw within the class;
[0084] Step 4. Calculate the inverse of matrix Sw.
[0085] Step 5. Find the vector (Here, the optimal projection direction for Fisher's criterion is determined using the Lagrange multiplier method.)
[0086] Step 6. Discriminant function
[0087] Step 7. Determine the probability of belonging to each category based on y.
[0088] After the electronic device performs binary classification on each of the datasets based on the Fisher discriminant method, the confusion matrix of the classifier corresponding to each dataset is obtained.
[0089] The confusion matrix of a single classifier is shown in Table 3:
[0090]
[0091] Table 3
[0092] In Table 3, TP represents the number of times the classifier identifies the true value as +1 when it is +1; FP represents the number of times the classifier identifies the true value as -1 when it is -1; FN represents the number of times the classifier identifies the true value as +1 when it is +1; and TN represents the number of times the classifier identifies the true value as -1 when it is -1.
[0093] Step 150: Determine multiple binary classification discriminant models based on the confusion matrix of each classifier.
[0094] The electronic device determines multiple binary classification models based on the confusion matrices of each classifier. Further, the electronic device determines the precision and recall of each classifier based on the confusion matrices of each classifier. Then, multiple binary classification models are derived based on the precision and recall of each classifier.
[0095] Specifically, in one embodiment, step 150, determining multiple binary classification discriminant models based on the confusion matrices of each classifier, includes:
[0096] Step 151: Determine the precision and recall of each classifier based on the confusion matrix of each classifier;
[0097] The electronic device determines the precision and recall of each classifier based on the confusion matrix of each classifier. Specifically, the precision and recall of each classifier are calculated using the following formulas:
[0098]
[0099]
[0100] TP represents the number of times the classifier identifies the true value as +1 when it is +1; FP represents the number of times the classifier identifies the true value as -1 when it is -1; FN represents the number of times the classifier identifies the true value as +1 when it is +1; TN represents the number of times the classifier identifies the true value as -1 when it is -1.
[0101] Step 152: Calculate the balance F-score of each classifier based on its precision and recall.
[0102] The electronic device calculates the balance F-score for each classifier based on their precision and recall. The formula for calculating the balance F-score is as follows:
[0103]
[0104] Step 153: Determine the multiple classifiers with the highest balanced F-scores as the multiple binary classification discriminant models.
[0105] The electronic device determines the multiple classifiers with the highest balanced F-scores as the multiple binary classification discriminant models. The binary classification discriminant models include at least three, and the number of binary classification discriminant models is the same as the number of classification types in the user intention classification result.
[0106] Specifically, this application embodiment implements a four-level user intention classification. Therefore, the electronic device determines the four classifiers with the highest balanced F-scores as the four binary classification discriminant models.
[0107] It should be noted that the binary classification model can also include other numbers, such as 3, 5, 6, etc. The number of binary classification models is the same as the number of classification types in the user intention classification result.
[0108] Step 160: Use multiple binary classification discriminant models as the user intention classification model.
[0109] The electronic device uses the four binary classification discriminant models obtained in step 150 as the user intention classification model. The four binary classification discriminant models are used to classify the user's multi-dimensional operator data into four categories: no intention, low intention, medium intention, and high intention.
[0110] After classifying users into four categories based on their intentions, different marketing methods (such as marketing data and call plans via SMS or mobile pop-ups) are used for users with different intention levels to improve the targeting of precise marketing to potential users. This application embodiment integrates multi-dimensional data from operators to classify user intention levels, establishes a user information database, and performs multi-classification based on a one-to-many decomposition method and Fisher's discriminant method, dividing users into: no intention, low intention, medium intention, and high intention. Different marketing methods are used for users with different intention levels. This approach not only allows for classification at a lower cost but also provides a more accurate and objective classification of user intentions.
[0111] By acquiring multi-dimensional operator data of users, and inputting this multi-dimensional operator data into a user intention classification model, the user intention classification results output by the user intention classification model are obtained. Compared with the expert stratification method, this application embodiment uses a machine learning method of one-to-many decomposition and Fisher discriminant analysis to train a user intention classification model on multi-dimensional operator sample data. This allows the user intention classification model of this application to classify users' intentions based on multi-dimensional operator data more quickly, improving the efficiency of user intention classification. Compared with traditional machine learning binary classification models, this application embodiment uses a machine learning method of one-to-many decomposition and Fisher discriminant analysis to train on multi-dimensional operator sample data, transforming the multi-classification problem into a binary classification problem using Fisher discriminant analysis. Since the trained user intention classification model includes at least three classification types, it can achieve more levels of user intention classification, improving the targeting of precise marketing to potential users.
[0112] For other aspects of the embodiments of this application, please refer to Figure 4 The user intention classification method further includes:
[0113] Step 400: Calculate the probability that each negative class is classified as a positive class under each of the binary classification discriminant models;
[0114] Step 500: If the binary classification discrimination model determines the positive class among the negative classes that exceed the set probability threshold, the number of positive and negative classes in the dataset of the one-to-many splitting method is readjusted to obtain multiple updated datasets.
[0115] Step 600: Based on the multiple updated datasets, retrain multiple updated binary classification models using the Fisher discriminant method to obtain multiple updated binary classification models;
[0116] Step 700: Use multiple updated binary classification discriminant models as the user intention classification model; wherein, the updated binary classification discriminant model includes at least 3 models, and the number of updated binary classification discriminant models is the same as the number of classification types in the user intention classification result.
[0117] Please refer to Table 4 for the probability P that each negative class is classified as a positive class under each of the aforementioned binary classification discriminant models. ij =M ij / N ij i∈{0,1,2,3},j∈{1,2,3,4}, where i represents the i-th class and j represents the j-th classifier.
[0118] <![CDATA[y i ]]> f1 f2 f3 f4 0 - <![CDATA[P 02 ]]> <![CDATA[P 03 ]]> <![CDATA[P 04 ]]> 1 <![CDATA[P 11 ]]> - <![CDATA[P 13 ]]> <![CDATA[P 14 ]]> 2 <![CDATA[P 21 ]]> <![CDATA[P 22 ]]> - <![CDATA[P 24 ]]> 3 <![CDATA[P 31 ]]> <![CDATA[P 32 ]]> <![CDATA[P 33 ]]> -
[0119] Table 4
[0120] Where M ij N represents the number of times the negative class i is classified as a positive class under the j-th classifier. ij This represents the number of samples in class i that are the anticlass of class j under the j-th classifier.
[0121] During the user intent level sampling process, an imbalance in the samples across categories can occur, leading to poor classification performance for some categories in the classifier. Therefore, the following method is used to reclassify the positive and negative samples in the one-to-many splitting method for binary classification.
[0122] P under the first classifier f1 i1 >50% i∈{1,2,3}
[0123] P under the second classifier f2 i2 >50% i∈{0,2,3}
[0124] P under the third classifier f3 i3 >50% i∈{0,1,3}
[0125] P under the fourth classifier f4 i4 >50% i∈{0,1,2}
[0126] That is, when more than half of the data in the i-th class of the negative class are classified as positive under the j-th classifier, the i-th class is converted to the positive class, and a binary classification model is performed. That is, when P... 21 >50% involves the following transformation: converting Table 2 into Table 5. Table 5 is shown below:
[0127] <![CDATA[y i ]]> f1 f2 f3 f4 0 +1 -1 -1 -1 1 -1 +1 -1 -1 2 +1 -1 +1 -1 3 -1 -1 -1 +1
[0128] Table 5
[0129] At this time the matrix
[0130] By readjusting the number of positive and negative classes in the dataset using the improved one-to-many decomposition method of matrix A, multiple updated datasets are obtained.
[0131] Based on the improved one-to-many decomposition method, assume there are four multidimensional sample data points from four operators, represented by labels 0, 1, 2, and 3. Taking the column containing f1 as an example, using the classifier f1 corresponding to the first dataset as an example, the four multidimensional sample data points from the four operators are decomposed into four datasets with labels 0 and 2 as positive classes and labels 1 and 3 as negative classes. The decomposition method for the datasets in the columns containing f2, f3, and f4 remains unchanged, resulting in multiple updated datasets. Each column in Table 5 represents one updated dataset.
[0132] Multiple updated binary classification models are obtained by retraining on multiple updated datasets using the Fisher discriminant method. Specific steps are detailed in steps 130 and 140, and will not be repeated here.
[0133] By readjusting the number of positive and negative classes in the dataset obtained from the one-to-many decomposition method, multiple updated datasets are obtained. Then, based on these updated datasets, multiple updated binary classification models are retrained using the Fisher discriminant method. This avoids the problem of imbalanced samples occurring during the original user data collection process, which leads to poor classification performance for some categories in the classifier, thus improving the classification accuracy of the trained user intention grading model.
[0134] In other aspects of the embodiments of this application, step 300, the step of inputting the operator's multidimensional data into the user intention classification model to obtain the user intention classification result output by the user intention classification model, includes:
[0135] Step 310: If only one of the binary classification discriminant models predicts a positive result, the classification result output by the binary classification discriminant model shall be used as the user intention classification result.
[0136] Step 320: When multiple binary classification models predict positive results, the classification result output by the binary classification model with the highest confidence among the multiple binary classification models is taken as the user intention classification result.
[0137] Based on step 100, four binary classification models were trained and used to classify users' multidimensional operator data. During testing, if only one binary classification model predicts a positive class, the corresponding classification result is taken as the user intention classification result. If multiple binary classification models predict a positive class, the prediction confidence of each model is considered, and the classification result with the highest confidence is selected as the user intention classification result. Finally, the user intention classification model identifies customer intention levels as 0, 1, 2, and 3, which represent one of the following: no intention, low intention, medium intention, and high intention.
[0138] Since the confidence level of a binary classification model indicates its reliability, when multiple binary classification models predict positive results, the classification result output by the model with the highest confidence level among them is used as the user intention classification result. This helps improve the accuracy of the user intention classification model in classifying user intentions.
[0139] For other aspects of the embodiments of this application, please refer to Figure 5 After step 110, obtaining multi-dimensional sample data of multiple users from various operators, the following steps are also included:
[0140] Step 121: Determine the data type of the sample based on the number of non-repeating values in each sample data in the operator's multidimensional sample data;
[0141] Specifically, in one embodiment, the electronic device determines the value range of each sample data in the operator's multidimensional sample data. If the value range contains more than 30 non-repeating values, it is judged as a continuous feature; otherwise, it is classified as a discrete feature.
[0142] Taking user call history as an example, if the number of non-repeating values in a user's call time in the operator's multidimensional sample data exceeds 30, then the user's call time is judged as a continuous feature; otherwise, the user's call time is classified as a discrete feature.
[0143] Step 122: When the sample data type is continuous sample data, perform variance analysis on the continuous sample data and delete sample data with F value less than the first set threshold.
[0144] Step 123: When the sample data type is discrete sample data, perform a chi-square test on the discrete sample data and delete sample data with a P value greater than the second set threshold.
[0145] In analysis of variance (ANOVA), the F-value reflects the inter-group variation in continuous sample data. A larger F-value indicates greater inter-group variation, meaning that sample data with F-values less than a first set threshold are noise and should be deleted. Similarly, in a chi-square test, the P-value reflects the deviation between observed and theoretically predicted values in discrete sample data. A larger P-value indicates greater deviation, meaning that sample data with P-values greater than a second set threshold are noise and should be deleted. For example, in a multi-dimensional sample data set of a telecom operator regarding user data usage, if sample data has an F-value less than the first set threshold or a P-value greater than the second set threshold, the corresponding data in the data usage data should be deleted.
[0146] When the sample data type is continuous sample data, the electronic device performs variance analysis on the continuous sample data and deletes sample data with F values less than a first set threshold. When the sample data type is discrete sample data, it deletes sample data with large P values in the discrete sample data, thereby reducing the impact of noise on the user intention grading model and accelerating the fitting speed of the user intention grading model.
[0147] The first and second threshold values can be set according to actual conditions. In one embodiment, the first threshold value can be 2 and the second threshold value can be 0.5.
[0148] The user intention classification device provided in the embodiments of this application is described below. The user intention classification device described below can be referred to in correspondence with the user intention classification method described above.
[0149] Please refer to Figure 6 A user intention grading device, comprising:
[0150] The first acquisition module 201 is used to acquire the user's multi-dimensional operator data;
[0151] The second acquisition module 202 is used to acquire user intention classification results: inputting the operator's multidimensional data into the user intention classification model to obtain the user intention classification results output by the user intention classification model;
[0152] The user intention classification result includes at least three classification types; and the user intention classification model outputs the user intention classification result of a single classification type at a time.
[0153] The user intention classification model is trained using multidimensional sample data from multiple operators based on a one-to-many decomposition method and Fisher's discriminant method.
[0154] The user intention classification device in this application embodiment acquires multi-dimensional operator data of users; inputs the multi-dimensional operator data into a user intention classification model, and obtains the user intention classification result output by the user intention classification model. Compared with the expert stratification method, since this application embodiment uses a machine learning method of one-to-many decomposition and Fisher discriminant analysis to train a user intention classification model on multi-dimensional operator sample data, the user intention classification model of this application can classify users' intentions based on multi-dimensional operator data more quickly, improving the efficiency of user intention classification. Compared with the traditional machine learning binary classification model, this application embodiment uses a machine learning method of one-to-many decomposition and Fisher discriminant analysis to train on multi-dimensional operator sample data, transforming the multi-classification problem into a binary classification problem using Fisher discriminant analysis. Since the trained user intention classification model has at least three classification types, it can achieve more levels of user intention classification, improving the targeting of precise marketing to potential users.
[0155] In one embodiment, the user intention classification device further includes a model training module for training the user intention classification model.
[0156] The model training module includes:
[0157] The third acquisition module is used to acquire multi-dimensional sample data of multiple users from the operator.
[0158] The decomposition module is used to decompose the multidimensional sample data of multiple users based on the one-to-many decomposition method to obtain multiple datasets; each dataset includes one multidimensional sample data of the operator as the positive class and multiple multidimensional sample data of the operator as the negative class.
[0159] The Fisher discriminant module is used to perform binary classification on each of the datasets based on the Fisher discriminant method, and to obtain the confusion matrix of the classifier corresponding to each dataset;
[0160] The binary classification model determination module determines multiple binary classification models based on the confusion matrix of each classifier.
[0161] The first conversion module is used to use multiple binary classification discriminant models as the user intention classification model;
[0162] The binary classification model includes at least three models, and the number of binary classification models is the same as the number of classification types in the user intention classification result.
[0163] In one embodiment, the binary classification discrimination model determination module includes:
[0164] The first calculation module is used to determine the precision and recall of each classifier based on the confusion matrix of each classifier.
[0165] The second calculation module is used to calculate the balance F-score of each classifier based on the precision and recall of each classifier.
[0166] The binary classification discriminant model acquisition module is used to determine the multiple classifiers with the highest balanced F scores as the multiple binary classification discriminant models.
[0167] In one embodiment, the user intention rating device further includes:
[0168] The third calculation module is used to calculate the probability that each negative class is judged as a positive class under each of the binary classification discrimination models;
[0169] The updated dataset acquisition module is used to readjust the number of positive and negative classes in the dataset of the one-to-many decomposition method when the binary classification discrimination model determines that the negative class is determined to be the positive class in the negative class that exceeds the set probability threshold, so as to obtain multiple updated datasets.
[0170] The module for obtaining updated binary classification discriminant models is used to retrain multiple updated binary classification discriminant models based on the multiple updated datasets using the Fisher discriminant method;
[0171] The second conversion module is used to use the multiple updated binary classification discriminant models as the user intention classification model;
[0172] The updated binary classification model includes at least three models, and the number of updated binary classification models is the same as the number of classification types in the user intention classification result.
[0173] In one embodiment, the second acquisition module 202 is used to:
[0174] If only one of the binary classification models predicts a positive result, the classification result output by the binary classification model will be used as the user intention classification result; or
[0175] If multiple binary classification models predict positive results, the classification result output by the binary classification model with the highest confidence among the multiple binary classification models is taken as the user intention classification result.
[0176] In one embodiment, after the user intention rating device, the system further includes:
[0177] The sample data type determination module is used to determine the sample data type based on the number of non-repeating values of each sample data in the operator's multidimensional sample data.
[0178] The first deletion module is used to perform variance analysis on the continuous sample data when the sample data type is continuous sample data, and delete sample data whose F value is less than a first set threshold.
[0179] The second deletion module is used to perform a chi-square test on the discrete sample data when the sample data type is discrete sample data, and delete sample data whose P value is greater than a second set threshold.
[0180] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740. The processor 710 can call a computer program in the memory 730 to execute steps of a user intention classification method, such as: acquiring multi-dimensional operator data of the user; acquiring user intention classification results: inputting the multi-dimensional operator data into a user intention classification model to obtain the user intention classification results output by the user intention classification model; wherein the user intention classification results include at least three classification types; and the user intention classification model outputs a single classification type of user intention classification result at a time; the user intention classification model is trained using multiple multi-dimensional operator sample data based on a one-to-many decomposition method and a Fisher discriminant method.
[0181] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0182] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the steps of the user intention classification method provided in the above embodiments, such as: acquiring multi-dimensional operator data of users; acquiring user intention classification results: inputting the multi-dimensional operator data into a user intention classification model to obtain the user intention classification results output by the user intention classification model; wherein, the user intention classification results include at least three classification types; and the user intention classification model outputs the user intention classification results of a single classification type at a time; the user intention classification model is trained using multiple multi-dimensional operator sample data based on a one-to-many decomposition method and a Fisher discriminant method.
[0183] On the other hand, embodiments of this application also provide a processor-readable storage medium storing a computer program for causing a processor to execute the steps of the user intention classification method provided in the above embodiments, such as including:
[0184] Obtain multi-dimensional operator data of users; obtain user intention classification results: input the multi-dimensional operator data into the user intention classification model to obtain the user intention classification results output by the user intention classification model; wherein, the user intention classification results include at least 3 classification types; and the user intention classification model outputs the user intention classification results of a single classification type at a time; the user intention classification model is trained based on a one-to-many decomposition method and Fisher discriminant method using multi-dimensional sample data from multiple operators.
[0185] The processor-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical memory (e.g., CD, DVD, BD, HVD), and semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).
[0186] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0187] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A user intention ranking method, characterized by, include: Obtain multi-dimensional carrier data from users; The operator's multidimensional data includes the user's call history, data usage, SMS activity, and app usage. The call history is the user's call duration or number of calls within a predetermined time. The data usage is the user's data consumption within a predetermined time. The SMS activity is the number of SMS messages sent and received by the user within a predetermined time. The app usage is the user's usage time for various types of apps within a predetermined time. Obtaining user intention classification results: Input the operator's multidimensional data into the user intention classification model to obtain the user intention classification results output by the user intention classification model; The user intention classification result includes at least three classification types; and the user intention classification model outputs the user intention classification result of a single classification type at a time. The user intention classification model is trained using multi-dimensional sample data from multiple operators based on a one-to-many decomposition method and Fisher's discriminant method. Before the step of acquiring the user's multi-dimensional operator data, the method further includes a step of training the user intention classification model. Training the user intention classification model includes: acquiring multi-dimensional operator sample data from multiple users; decomposing the multi-dimensional operator sample data from multiple users based on the one-to-many decomposition method to obtain multiple datasets; each dataset includes one multi-dimensional operator sample data as the positive class and multiple multi-dimensional operator sample data as the negative class; performing binary classification on each dataset based on the Fisher discriminant method; obtaining the confusion matrix of the classifier corresponding to each dataset; determining multiple binary classification models based on the confusion matrix of each classifier; and using the multiple binary classification models as the user intention classification model; wherein the number of binary classification models is at least three, and the number of binary classification models is the same as the number of classification types in the user intention classification result. The user intention classification method further includes: calculating the probability that each negative class is classified as a positive class under each binary classification model; when a negative class in the binary classification model is classified as a positive class if it exceeds a set probability threshold, readjusting the number of positive and negative classes in the dataset of the one-to-many decomposition method to obtain multiple updated datasets; retraining multiple updated binary classification models based on the multiple updated datasets using the Fisher discriminant method; and using the multiple updated binary classification models as the user intention classification model; wherein, the updated binary classification model includes at least 3 models, and the number of updated binary classification models is the same as the number of classification types in the user intention classification result.
2. The user intention ranking method of claim 1, wherein, The determination of multiple binary classification discriminant models based on the confusion matrices of each of the classifiers includes: The precision and recall of each classifier are determined based on the confusion matrix of each classifier. Calculate the balanced F-score of each classifier based on its precision and recall. The classifiers with the highest balanced F-scores are determined as the plurality of binary classification discriminant models.
3. The user intent classification method of claim 1, wherein, The step of inputting the operator's multidimensional data into the user intention classification model to obtain the user intention classification result output by the user intention classification model includes: If only one of the binary classification models predicts a positive result, the classification result output by the binary classification model shall be used as the user intention classification result. If multiple binary classification models predict positive results, the classification result output by the binary classification model with the highest confidence among the multiple binary classification models is taken as the user intention classification result.
4. The user intent classification method of claim 1, wherein, After obtaining multi-dimensional sample data of multiple users from various operators, the process also includes: The data type of the sample is determined based on the number of non-repeating values in each sample data in the operator's multidimensional sample data; When the sample data type is continuous sample data, perform variance analysis on the continuous sample data and delete sample data with F value less than a first set threshold; When the sample data type is discrete sample data, a chi-square test is performed on the discrete sample data, and sample data with a P-value greater than a second set threshold are deleted.
5. A user intention grading device, characterized in that, include: The first acquisition module is used to acquire the user's multi-dimensional operator data; The operator's multidimensional data includes the user's call history, data usage, SMS activity, and app usage. The call history is the user's call duration or number of calls within a predetermined time. The data usage is the user's data consumption within a predetermined time. The SMS activity is the number of SMS messages sent and received by the user within a predetermined time. The app usage is the user's usage time for various types of apps within a predetermined time. The second acquisition module is used to acquire user intention classification results: inputting the operator's multidimensional data into the user intention classification model to obtain the user intention classification results output by the user intention classification model; The user intention classification result includes at least three classification types; and the user intention classification model outputs the user intention classification result of a single classification type at a time. The user intention classification model is trained using multi-dimensional sample data from multiple operators based on a one-to-many decomposition method and Fisher's discriminant method. Before acquiring the user's multi-dimensional operator data, the method further includes training the user intention classification model. Training the user intention classification model includes: acquiring multi-dimensional operator sample data from multiple users; decomposing the multi-dimensional operator sample data from multiple users using the one-to-many decomposition method to obtain multiple datasets; each dataset includes one multi-dimensional operator sample data as the positive class and multiple multi-dimensional operator sample data as the negative class; performing binary classification on each dataset using the Fisher discriminant method; obtaining the confusion matrix of the classifier corresponding to each dataset; determining multiple binary classification models based on the confusion matrix of each classifier; and using the multiple binary classification models as the user intention classification model; wherein the number of binary classification models is at least three, and the number of binary classification models is the same as the number of classification types in the user intention classification result. The user intention classification device is further configured to: calculate the probability that each negative class is classified as a positive class under each binary classification model; when a negative class in the binary classification model is classified as a positive class if it exceeds a set probability threshold, readjust the number of positive and negative classes in the dataset of the one-to-many splitting method to obtain multiple updated datasets; retrain multiple updated binary classification models based on the multiple updated datasets using the Fisher discriminant method; and use the multiple updated binary classification models as the user intention classification model; wherein the updated binary classification model includes at least 3 models, and the number of updated binary classification models is the same as the number of classification types in the user intention classification result.
6. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the user intention classification method according to any one of claims 1 to 4.
7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the user intention classification method as described in any one of claims 1 to 4.
8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the user intention classification method according to any one of claims 1 to 4.