User donation intention processing method and device, service equipment and storage medium

Abstract

The application provides a user's donation intention processing method and device, a service device and a storage medium. The method comprises the following steps: obtaining portrait data, signaling data and network data of a target user to obtain first structured data; obtaining voice data and text data of the target user to obtain unstructured data; converting the unstructured data into second structured data, and performing data fusion on the second structured data and the first structured data to obtain user feature data of the target user; inputting the user feature data of the target user into a trained donation intention recognition model to obtain a donation intention score of the target user; determining the donation intention of the target user according to the donation intention score; and outputting the donation intention of the target user. By combining structured data and unstructured data as data sources, multi-dimensional user feature data is formed, and the problem of low accuracy caused by single data source is avoided.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of big data processing, and specifically provides a user donation intention processing method and device, a service equipment and a storage medium. BACKGROUND

[0002] In the digital information era, data tagging has become a development trend. Among them, user tags reflect the personal characteristics of users on network platforms by describing the characteristics and behaviors of users. In particular, the user data generated by various social network platforms has also increased, making it particularly important to analyze the donation intention of users based on user data generated by social network platforms to build user tags.

[0003] Currently, in the prior art, structured user data is mainly obtained, and a built classification model is used for prediction to directly obtain user intention.

[0004] However, in the prior art, only structured user data still has the problem of low accuracy. SUMMARY

[0005] The application provides a user donation intention processing method and device, a service equipment and a storage medium, which solves the problem of low accuracy caused by obtaining structured user data and directly obtaining user intention by using a built classification model for prediction in the prior art.

[0006] In a first aspect, the application provides a user donation intention processing method applied to a service equipment, comprising:

[0007] Obtaining portrait data, signaling data and network data of a target user to obtain first structured data;

[0008] Obtaining voice data and text data of the target user to obtain unstructured data;

[0009] Converting the unstructured data into second structured data, and performing data fusion on the second structured data and the first structured data to obtain user feature data of the target user;

[0010] Inputting the user feature data of the target user into a trained donation intention recognition model to obtain a donation intention score of the target user;

[0011] Determining the donation intention of the target user according to the donation intention score;

[0012] Outputting the donation intention of the target user.

[0013] In a possible design, the converting the unstructured data into second structured data and the data fusion of the second structured data and the first structured data to obtain the user feature data of the target user comprises: converting the voice data in the unstructured data into structured industry classification confidence data and stay-in intention confidence data; converting the text data in the unstructured data into third structured data; performing data fusion on the industry classification confidence data, the stay-in intention confidence data, the third structured data and the first structured data by using a data fusion method; and determining the fused data as the user feature data of the target user.

[0014] In a possible design, the trained stay-in intention recognition model is a DeepFM model, and the DeepFM model comprises an FM module and a DNN module. Correspondingly, the inputting the user feature data of the target user into the trained stay-in intention recognition model to obtain a stay-in intention score of the target user comprises: performing feature conversion on the user feature data to obtain a converted feature set; inputting the converted feature set into the FM module in the DeepFM model, so that the FM module extracts low-order features in the feature set; inputting the converted feature set into the DNN module in the DeepFM model, so that the DNN module extracts high-order features in the feature set; fusing the low-order features and the high-order features; and performing nonlinear conversion on the fused features by using a preset activation function to obtain the stay-in intention score.

[0015] In a possible design, before the obtaining of the portrait data, the signaling data and the network data of the target user, the method further comprises: obtaining the user feature data of a plurality of users of a historical project as sample data to form a sample data set; screening positive samples and negative samples from the sample data set; and performing iterative training on the initial stay-in intention recognition model according to the positive samples and the negative samples to obtain the trained stay-in intention recognition model.

[0016] In a possible design, the screening of the positive samples from the sample data set comprises: obtaining the industry classification confidence data and the stay-in intention confidence data from the sample data set; screening, from the industry classification confidence data, the industry classification confidence data whose industry classification confidence exceeds a first preset limit value as a first positive sample set; screening, from the stay-in intention confidence data, the stay-in intention confidence data whose stay-in intention confidence exceeds a second preset limit value as a second positive sample set; and taking an intersection of the first positive sample set and the second positive sample set to obtain a final positive sample set.

[0017] In a possible design, the negative samples are screened out from the sample data set, including: marking the screened positive samples to obtain a set of marked data; obtaining unmarked data from the sample data set to obtain a set of unmarked data; obtaining reliable negative samples from the marked data and the unmarked data by using a machine learning method; and obtaining the negative samples from the unmarked data by using the machine learning method according to the positive samples and the reliable negative samples.

[0018] In a possible design, the obtaining of the reliable negative samples from the marked data and the unmarked data by using the machine learning method includes: obtaining a set of spy samples from the marked data according to a preset probability; obtaining a set of negative sample samples from the unmarked data according to a preset sampling rate; determining initial marked data according to the marked data and the set of spy samples; determining initial unmarked data according to the set of spy samples and the set of negative sample samples; constructing a first binary classifier according to the initial marked data and the initial unmarked data; and performing prediction on the initial unmarked data by using the first binary classifier, and determining the reliable negative samples if a prediction probability of the target user in the initial unmarked data is less than a preset threshold.

[0019] In a possible design, the obtaining of the negative samples from the unmarked data by using the machine learning method according to the positive samples and the reliable negative samples includes: constructing a second binary classifier according to the positive samples and the reliable negative samples; performing cyclic prediction on the unmarked data by using the second binary classifier according to a preset cyclic stop condition, and screening out the negative samples.

[0020] In a second aspect, the present application provides a user's donation intention processing apparatus, applied to a service device, including:

[0021] a data acquisition and processing unit, configured to obtain portrait data, signaling data and network data of a target user to obtain first structured data, and obtain voice data and text data of the target user to obtain unstructured data;

[0022] a multi-modal data fusion unit, configured to convert the unstructured data into second structured data, and perform data fusion on the second structured data and the first structured data to obtain user feature data of the target user;

[0023] a donation intention recognition unit, configured to input the user feature data of the target user into a trained donation intention recognition model to obtain a donation intention score of the target user;

[0024] The fund retention intention output unit is configured to determine the fund retention intention of the target user according to the fund retention intention score, and output the fund retention intention of the target user.

[0025] In a third aspect, the present application provides a service device, comprising at least one processor and a memory;

[0026] The memory stores computer-executable instructions.

[0027] The at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the fund retention intention processing method of the user as described in the first aspect and various possible designs of the first aspect.

[0028] In a fourth aspect, the present application provides a computer storage medium, wherein the computer storage medium stores computer-executable instructions, and when the processor executes the computer-executable instructions, the fund retention intention processing method of the user as described in the first aspect and various possible designs of the first aspect is implemented.

[0029] The fund retention intention processing method, device, service device and storage medium provided by the present application analyze structured data (including portrait data, signaling data and network data) and unstructured data (including voice data and text data) through the service device, and convert the unstructured data into structured data, fuse the data to obtain user feature data. The user feature data is input into the trained fund retention intention recognition model through feature conversion to obtain a fund retention intention score of a target user, determine the fund retention intention of the target user according to the fund retention intention score, and output the fund retention intention of the target user. By combining structured data and unstructured data as data sources, multi-dimensional user feature data is formed, and the problem of low accuracy caused by single data source is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0031] Figure 1 The scene schematic diagram of the fund retention intention processing system of the user provided by the embodiments of the present application;

[0032] Figure 2 The flowchart of the fund retention intention processing method of the user provided by an embodiment of the present application;

[0033] Figure 3A flowchart of a user's donation intention processing method according to another embodiment of the present application is provided;

[0034] Figure 4 A structural diagram of a user's donation intention processing device according to an embodiment of the present application is provided;

[0035] Figure 5 A hardware structural diagram of a service device according to an embodiment of the present application is provided. DETAILED DESCRIPTION

[0036] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0037] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards, and provide corresponding operation portal for user to choose authorization or refusal.

[0038] The following are explanations of terms:

[0039] Donation intention: the intention of a user to leave personal information.

[0040] In the digital information era, data tagging has become a development trend. Among them, user tags reflect the personal characteristics of users on network platforms through the marks describing the characteristics and behaviors of users. In particular, the user data generated by various social network platforms also increases accordingly, so it is particularly important to analyze the donation intention of users according to the user data generated by social network platforms to build user tags. In the prior art, structured user data is used to predict directly by using a well-constructed classification model, and the user intention is obtained directly. Due to the single data source, direct application to the classification model, the sample selection may be biased, the prediction result may be inclined or single, and thus the problem of low accuracy may be caused.

[0041] In order to solve the above technical problems, the technical idea of the present application is as follows: considering that the data source is single in the prior art, structured data (including portrait data, signaling data and network data) is obtained, and then unstructured data (including voice data and text data) is obtained. By combining structured data and unstructured data as a data source, multi-dimensional user feature data is formed, avoiding the problem of low accuracy caused by single data source. The following detailed embodiments are described in detail.

[0042] Figure 1 The scene schematic diagram of the user's donation intention processing system provided by the embodiment of the present application is shown in FIG. 1. Figure 1 As shown in FIG. 1, the system provided by the embodiment of the present application includes a network operator device 101, a service device 102 and a platform device 103.

[0043] The network operator device 101 can be one or more of a base station, a routing device and a switching device. The network operator device 101 includes but is not limited to a macro base station, a micro base station, a pico base station, a flying base station, a wired router, a wireless router, a telephone switch and a transmission gateway, etc.

[0044] The service device 102 can be a server or a cluster composed of multiple servers.

[0045] The platform device 103 can be used for storage, display or marketing. The platform device 103 includes but is not limited to a hard disk, a display and a large screen, etc.

[0046] Referring to FIG. 1, Figure 1 The network operator device 101 is used to collect terminals of different users to obtain various user-related data, and send the user-related data to the service device 102. The service device 102 is used to process a series of user-related data to obtain the user's donation intention, and output the donation intention of each user to the platform device 103. The platform device 103 is used to store, display or market the donation intention of each user.

[0047] Embodiment one

[0048] Figure 2 The flowchart of the user's donation intention processing method provided by an embodiment of the present application is shown in FIG. 2. The execution subject of the embodiment can be Figure 1 The service device 102 in the embodiment can also be other service devices with similar functions, which are not particularly limited in the embodiment. As shown in FIG. 2, Figure 2 The method includes the following steps:

[0049] S201: Obtain the portrait data, signaling data and network data of the target user to obtain the first structured data.

[0050] The image data can be a labeled user model data abstracted according to user information, including but not limited to gender, age, occupation, preference, behavior, and preference.

[0051] The signaling data is communication data between the mobile phone user and the base station. The spatial position of the user can be determined through the communication data, and a very resident place set of the user is recorded. The very resident place set is a set of positions other than the resident place of the user.

[0052] Specifically, for the signaling data in the first structured data, a density-based clustering algorithm is used to generate a user daily trajectory and a resident point. The work address and the residence address of the user are obtained according to a monthly time window. The work address and the residence address are excluded from the trajectory and the resident point of the user to obtain the very resident place set of the user. Further, according to the area of interest (AOI) data, a subset of interest points through which the user trajectory passes is obtained. Optionally, the density-based clustering algorithm is a DBSCAN algorithm. The time window can be a rolling time window or a sliding time window.

[0053] The network data can be a data set of the use of the App by the user in different time dimensions.

[0054] Specifically, the use of the App by the user in different time dimensions is counted to generate a data set. The data set includes but is not limited to the frequency of accessing the App, the duration of accessing the App, the number of active days of the App, the traffic of accessing the App, and the number of using the same type of App.

[0055] The structured data refers to data managed in the form of a relational database table.

[0056] Specifically, the first structured data of the target user sent by the network operator device can be acquired through a wireless communication mode, or the first structured data of the target user sent by the network operator device can be acquired through a wired communication mode.

[0057] S202: Acquire voice data and text data of the target user to obtain unstructured data.

[0058] The voice data refers to the call record of the user, and the text data refers to the data generated by the user reading.

[0059] In this embodiment, the unstructured data refers to data without a fixed mode.

[0060] Specifically, the unstructured data of the target user sent by the network operator device can be acquired through a wireless communication mode, or the unstructured data of the target user sent by the network operator device can be acquired through a wired communication mode.

[0061] S203: converting the unstructured data into second structured data, and data fusing the second structured data with the first structured data to obtain user feature data of the target user.

[0062] Specifically, S203 specifically includes Sa-Sd:

[0063] Sa: converting speech data in the unstructured data into structured industry classification confidence data and stay intention confidence data.

[0064] The industry classification confidence data refers to a data set of probabilities of being classified into a specific industry, and the stay intention confidence refers to a data set of probabilities of leaving personal information.

[0065] Specifically, the speech data is converted into outbound text data by using a speech recognition technology, the outbound text data is analyzed by using a classification model, a business scenario is recognized to obtain the industry classification confidence data, and the stay intention of the user is analyzed by using an initial stay intention recognition model to obtain the stay intention confidence data. Optionally, the classification model is a Textcnn text classification model.

[0066] Sb: converting text data in the unstructured data into third structured data.

[0067] Specifically, the text data in the unstructured data can be reading data generated by the user reading an article. First, an encoder constructed by using a language model is used to disassemble the reading data into a plurality of feature labels, and a word vector of the reading data is obtained through the feature labels; second, the reading data is classified similarly to form a corpus; then, the corpus and the corresponding word vector of the corpus are selected according to the requirement, and the word vector is mapped to a high-dimensional semantic space to obtain the third structured data. Optionally, the language model is a BERT model.

[0068] Specifically, the method for converting the text data in the unstructured data into structured data includes but is not limited to: segmentation and splitting, text classification and clustering, regular expressions, natural language processing, and information extraction.

[0069] Sc: data fusing the industry classification confidence data, the stay intention confidence data, the third structured data, and the first structured data by using a data fusion method.

[0070] Specifically, features of the industry classification confidence data, the stay intention confidence data, and the third structured data are extracted respectively, and different features are combined into new features by using a data fusion method, such as a feature-based fusion method, to obtain fused data.

[0071] Sd: determining the fused data as the user feature data of the target user.

[0072] The user feature data refers to a set of data representing the user features of a user. The user features refer to variables used for model training, which need to be converted before machine learning.

[0073] S204: input the user feature data of the target user into the trained fund retention intention recognition model to obtain the fund retention intention score of the target user.

[0074] The fund retention intention recognition model is a DeepFM model. DeepFM is an improved version of Wide&Deep, which replaces the LR in the Wide part of the Wide&Deep model with FM. The DeepFM model includes an FM module and a DNN module.

[0075] Specifically, S204 specifically includes Se-Si:

[0076] Se: perform feature conversion on the user feature data to obtain a converted feature set.

[0077] The feature conversion is a conversion operation on the data in the original data to obtain user feature data suitable for algorithm model construction.

[0078] Specifically, for the category type features in the user feature data: first, use character encoding for encoding, then use an encryption algorithm for encryption, and finally determine the final value according to the pre-set hash size. Optionally, the character encoding is UTF-8, and the encryption algorithm is MD5.

[0079] For example, the feature value is "high", which is converted to "395f4ac03c4455a23fed274f2e8ea07b" by MD5, and when the hash_size is set to 16, the final value is 6.

[0080] Specifically, for the numerical type features in the user feature data: when there are more than three different values, perform binning processing. Binning processing is a process of converting continuous values into discrete values.

[0081] For example, the user's monthly exercise frequency is binned, 0-5 times are classified as very poor, 6-10 times are classified as poor, 11-15 times are classified as general, 16-20 times are classified as good, and 21-25 times are classified as good.

[0082] Specifically, for the Boolean type features in the user feature data: when the value is not more than three, perform one-hot encoding processing. One-hot encoding is to use an N-bit status register to encode N states, each state has its own independent register bit, and only one bit is valid at any time.

[0083] Exemplarily, the gender features are ["male", "female"], N is 2, "male" is represented as 10, and "female" is represented as 01; the region features are ["A city", "B city", "C city"], N is 3, "A city" is represented as 100, "B city" is represented as 010, and "C city" is represented as 001.

[0084] Sf: input the converted feature set into the FM module in the DeepFM model, so that the FM module extracts low-order features in the feature set.

[0085] In the DeepFM model, the FM module is responsible for extracting low-order features.

[0086] In this embodiment, the low-order features output by the FM module are the accumulation of the first-order features of the Addition unit and the second-order features of the InnerProduct unit. The Addition unit is used to represent the influence of each first-order feature, and the InnerProduct is used to represent the influence of the interaction of the second-order features.

[0087] Sg: input the converted feature set into the DNN module in the DeepFM model, so that the DNN module extracts high-order features in the feature set.

[0088] In the DeepFM model, the DNN module is responsible for extracting high-order features.

[0089] Specifically, the feature set is converted into dense features, and the dense features are combined through two fully connected layers to output high-order features.

[0090] Sh: fuse low-order features and high-order features.

[0091] Specifically, a feature fusion method is used to fuse low-order features and high-order features. Optionally, the feature fusion method is feature splicing.

[0092] Si: use a preset activation function to perform nonlinear transformation on the fused features to obtain a stay intention score.

[0093] The activation function mainly provides the neural network with nonlinear modeling capability.

[0094] Optionally, the preset activation function is Sigmoid. Specifically, the Sigmoid activation function is used to perform nonlinear transformation on the fused features, and the output result is mapped to [0, 1]; the output result includes the probability of predicting the target user as 1 and the probability of predicting the target user as 0, and the sum of the two parts is 1; the probability of predicting the target user as 1 is taken as the stay intention score of the target user.

[0095] S205: determine the stay intention of the target user according to the stay intention score.

[0096] In the embodiment, the higher the fund retention intention score is, the greater the fund retention intention of the target user is.

[0097] S206: output the fund retention intention of the target user.

[0098] Specifically, the fund retention intention of the target user can be output to a platform device for display.

[0099] From the above embodiment description, first, the structured data (including portrait data, signaling data and network data) and unstructured data (including voice data and text data) of the target user are obtained, and the unstructured data is converted into structured data, the data is fused to obtain user feature data; then, the user feature data is input into the trained fund retention intention recognition model for feature conversion to obtain the fund retention intention score of the target user; finally, the fund retention intention of the target user is determined according to the fund retention intention score, and the fund retention intention of the target user is output. By combining structured data and unstructured data as data sources, multi-dimensional user feature data is formed, avoiding the problem of low accuracy caused by single data source.

[0100] Embodiment Two

[0101] Figure 3 The flowchart of the user's fund retention intention processing method provided by another embodiment of the application is shown in Figure 2 Based on the embodiment, the embodiment focuses on describing the specific process of how to establish a fund retention intention recognition model. As Figure 3 shown, the details are as follows:

[0102] S301: obtain user feature data of a plurality of users of a historical project as sample data to form a sample data set.

[0103] The historical project includes historical voice data.

[0104] Specifically, the portrait data, signaling data and network data of the plurality of users are obtained to obtain first structured data; the voice data and text data of the plurality of users are obtained to obtain unstructured data; the unstructured data is converted into second structured data, and the second structured data is fused with the first structured data to obtain the user feature data of the plurality of users. The user feature data of the plurality of users is taken as sample data to form a sample data set.

[0105] S302: filter out positive samples in the sample data set.

[0106] Specifically, S302 includes Sj~Sm:

[0107] Sj: obtaining industry classification confidence data and stay intention confidence data in the sample data set.

[0108] Specifically, unstructured data in the sample data set is obtained, including voice data and text data, and the voice data in the unstructured data is converted into structured industry classification confidence data and stay intention confidence data.

[0109] The description of converting the voice data in the unstructured data into structured industry classification confidence data and stay intention confidence data in Sj in step S302 is consistent with the description of Sa in step S201, and specific description is referred to related content, which will not be repeated here.

[0110] Sk: filtering out industry classification confidence data with industry classification confidence exceeding a first preset limit value as a first positive sample set in the industry classification confidence data.

[0111] Wherein, the first preset limit is a preset limit of the industry classification confidence, denoted as A1.

[0112] Specifically, if the industry classification confidence A in the industry classification confidence data is greater than or equal to A1, it is determined as the first positive sample set.

[0113] Sl: filtering out stay intention confidence data with stay intention confidence exceeding a second preset limit value as a second positive sample set in the stay intention confidence data.

[0114] Wherein, the second preset limit is a preset limit of the stay intention confidence, denoted as B1.

[0115] Specifically, if the stay intention confidence B in the stay intention confidence data is greater than or equal to B1, it is determined as the second positive sample set.

[0116] Sm: taking the intersection of the first positive sample set and the second positive sample set to obtain the final positive sample set.

[0117] Specifically, the same users in the first positive sample set and the second positive sample set are obtained, the same users are the intersection part, and are determined as the final positive sample set.

[0118] Exemplarily, users with industry theme of "home decoration" and industry classification confidence higher than 0.9 are filtered out as set D1, and users with stay intention confidence higher than 0.7 are filtered out as set D2; finally, the intersection of set D1 and set D2 is taken to obtain the final positive sample set.

[0119] S303: filtering out negative samples in the sample data set.

[0120] Specifically, S303 includes Sn~Sq:

[0121] Sn: Mark the positive samples that have been screened to obtain a set of marked data.

[0122] Specifically, the screened positive samples are marked to obtain a set of marked data.

[0123] So: Obtain unmarked data in the sample data set to obtain a set of unmarked data.

[0124] Specifically, other data in the sample set except the marked data is obtained to obtain a set of unmarked data.

[0125] Sp: Obtain reliable negative samples from the marked data and the unmarked data by a machine learning method.

[0126] Specifically, according to a preset probability, a set of spy samples is obtained from the marked data, according to a preset sampling rate, a set of negative sample samples is obtained from the unmarked data, and initial marked data is determined according to the marked data and the set of spy samples; the set of spy samples and the set of negative sample samples are used to determine initial unmarked data. A first binary classifier is constructed according to the initial marked data and the initial unmarked data; the initial unmarked data is predicted using the first binary classifier, and if the prediction probability of the target user in the initial unmarked data is less than a preset threshold, it is determined as a reliable negative sample.

[0127] Among them, the set of spy samples is a part of data randomly selected from the marked data, and the set of negative sample samples is a part of data randomly selected from the unmarked data. The binary classifier is a machine learning model for classifying input samples into two categories; the probability of predicting the data in the set of spy samples as positive samples is used as the preset threshold, denoted as C1.

[0128] Specifically, the marked data is D + , the unmarked data is D U , the preset probability is α, and the set of spy samples is:

[0129] D SP = α * D +

[0130] The preset sampling rate is β, and the set of negative sample samples is:

[0131] D US = β * D U

[0132] At this time, the initial marked data is represented as (1-α)*D + , that is, D + -D SP , and the initial unmarked data is represented as D US ∪DSP .

[0133] Specifically, a binary classifier using the Expectation Maximization (EM) algorithm is constructed based on the initial labeled data and the initial unlabeled data. This classifier predicts the value of the initial unlabeled data. If the probability C ≤ C1 of predicting a data point as a positive sample in the initial unlabeled data is considered a reliable negative sample, then it is determined to be a reliable negative sample. Optionally, the EM algorithm can be a Naive Bayes algorithm.

[0134] Sq: Obtain negative samples from unlabeled data using machine learning methods based on positive samples and reliable negative samples.

[0135] Specifically, a second binary classifier is constructed based on positive samples and reliable negative samples; according to a preset loop stopping condition, the second binary classifier is used to perform cyclic prediction on unlabeled data to filter out negative samples.

[0136] The preset loop stopping conditions can be selected in several ways, including but not limited to: reaching the maximum number of loops and the positive sample prediction probability.

[0137] Specifically, an EM algorithm binary classifier is constructed based on positive samples and reliable negative samples. Iterative prediction is performed on unlabeled data, with only negative samples added in each iteration while positive samples remain unchanged, until the iteration stops, thus obtaining negative samples. Optionally, the EM algorithm is a Naive Bayes algorithm.

[0138] S304: Iteratively train the initial investment intention recognition model based on positive and negative samples to obtain a trained investment intention recognition model.

[0139] Specifically, positive and negative samples are used as datasets. Ray parameter tuning is used, distributed computing is employed to search the hyperparameter space, find the optimal combination of hyperparameters, and iteratively train the model to obtain a well-trained lead retention intention recognition model.

[0140] In this embodiment, the description of steps S305-S310 is the same as that of steps S201-S206 above. For a detailed description, please refer to the relevant content, which will not be repeated here.

[0141] S305: Obtain the target user's profile data, signaling data, and network data to obtain the first structured data.

[0142] S306: Obtain the target user's voice and text data to obtain unstructured data.

[0143] S307: convert the unstructured data into second structured data, and perform data fusion between the second structured data and the first structured data to obtain user feature data of the target user.

[0144] S308: input the user feature data of the target user into the trained stay intention recognition model to obtain a stay intention score of the target user.

[0145] S309: determine the stay intention of the target user according to the stay intention score.

[0146] S310: output the stay intention of the target user.

[0147] As can be known from the above embodiment description, by using a semi-supervised learning method to select samples from labeled data and unlabeled data, representative samples can be selected, the generalization ability and accuracy of the model can be improved, and the stay intention recognition model is pre-trained to assist iterative optimization of the stay intention of the user. By improving the generalization ability and accuracy of the model and iteratively training the stay intention recognition model, the accuracy can be improved.

[0148] Figure 4 A structural schematic diagram of a user's stay intention processing apparatus provided by the embodiment is shown in FIG. 4. Figure 4 As shown in the figure, the user's stay intention processing apparatus 40 includes a data acquisition and processing unit 401, a multi-modal data fusion unit 402, a stay intention recognition unit 403, and a stay intention output unit 404.

[0149] The data acquisition and processing unit 401 is configured to obtain portrait data, signaling data, and network data of a target user to obtain first structured data, and obtain voice data and text data of the target user to obtain unstructured data.

[0150] The multi-modal data fusion unit 402 is configured to convert the unstructured data into second structured data, and perform data fusion between the second structured data and the first structured data to obtain user feature data of the target user.

[0151] The stay intention recognition unit 403 is configured to input the user feature data of the target user into a trained stay intention recognition model to obtain a stay intention score of the target user.

[0152] The stay intention output unit 404 is configured to determine the stay intention of the target user according to the stay intention score, and output the stay intention of the target user.

[0153] The apparatus provided by the embodiment can be used to execute the technical solutions of the above method embodiments, and has similar implementation principles and technical effects, which will not be described here again.

[0154] In a possible design, the unstructured data is converted into second structured data, and the second structured data is data fused with the first structured data to obtain the user feature data of the target user. Accordingly, the multi-modal data fusion unit 402 is specifically configured to: convert the voice data in the unstructured data into structured industry classification confidence data and stay-in intention confidence data; convert the text data in the unstructured data into third structured data; perform data fusion on the industry classification confidence data, the stay-in intention confidence data, the third structured data, and the first structured data by using a data fusion method; and determine the fused data as the user feature data of the target user.

[0155] In a possible design, the trained stay-in intention recognition model is a DeepFM model, and the DeepFM model includes an FM module and a DNN module. Accordingly, the user feature data of the target user is input into the trained stay-in intention recognition model to obtain a stay-in intention score of the target user. Accordingly, the stay-in intention recognition unit 403 is specifically configured to: perform feature conversion on the user feature data to obtain a converted feature set; input the converted feature set into the FM module in the DeepFM model, so that the FM module extracts low-order features in the feature set; input the converted feature set into the DNN module in the DeepFM model, so that the DNN module extracts high-order features in the feature set; fuse the low-order features and the high-order features; and perform nonlinear conversion on the fused features by using a preset activation function to obtain the stay-in intention score.

[0156] In a possible design, the module further includes a model training unit 405 configured to: obtain the user feature data of a plurality of users of historical projects as sample data to form a sample data set; filter out positive samples and negative samples from the sample data set; and perform iterative training on an initial stay-in intention recognition model according to the positive samples and the negative samples to obtain the trained stay-in intention recognition model.

[0157] In a possible design, the module further includes a sample selection unit 406, configured to: obtain the industry classification confidence data and the stay intention confidence data in the sample data set; filter, in the industry classification confidence data, the industry classification confidence data with an industry classification confidence exceeding a first preset limit value as a first positive sample set; filter, in the stay intention confidence data, the stay intention confidence data with a stay intention confidence exceeding a second preset limit value as a second positive sample set; and obtain an intersection of the first positive sample set and the second positive sample set as a final positive sample set.

[0158] In a possible design, the sample selection unit 406 is further configured to: mark the filtered positive samples to obtain a set of marked data; obtain, in the sample data set, unmarked data to obtain a set of unmarked data; obtain reliable negative samples by using a machine learning method according to the marked data and the unmarked data; and obtain the negative samples from the unmarked data by using the machine learning method according to the positive samples and the reliable negative samples.

[0159] In a possible design, the sample selection unit 406 is further configured to: obtain a spy sample set from the marked data according to a preset probability; obtain a negative sample sampling set from the unmarked data according to a preset sampling rate; determine initial marked data according to the marked data and the spy sample set; determine initial unmarked data according to the spy sample set and the negative sample sampling set; construct a first binary classifier according to the initial marked data and the initial unmarked data; and use the first binary classifier to predict the initial unmarked data, and determine the reliable negative samples if a prediction probability of the target user in the initial unmarked data is less than a preset threshold.

[0160] In a possible design, the sample selection unit 406 is further configured to: construct a second binary classifier according to the positive samples and the reliable negative samples; and use the second binary classifier to perform cyclic prediction on the unmarked data according to a preset cyclic stop condition, and filter the negative samples.

[0161] The apparatus provided in this embodiment can be used to execute the technical solutions of the method embodiments, and has similar implementation principles and technical effects, which will not be repeated here.

[0162] In the embodiment, the data acquisition and processing unit, the multi-modal data fusion unit, the sample selection unit, the model training unit and the stay intention recognition unit are processed by PipeLine to form modular units; according to the pre-set configuration, the user's stay intention can be automatically output without manual intervention; the user's stay intention is used as the basis for reaching the user, and the voice data generated by reaching the user is used as historical voice data, which is again used as the data source of the stay intention recognition model, the stay intention recognition model is iteratively trained to form a data closed loop.

[0163] Figure 5 The hardware structure schematic diagram of the service equipment provided in the embodiment of the present application is shown in the figure. Figure 5 As shown in the figure, the service equipment 50 of the embodiment includes a processor 501 and a memory 502; wherein

[0164] The memory 502 is used to store computer execution instructions.

[0165] The processor 501 is used to execute the computer execution instructions stored in the memory to realize each step performed in the above-mentioned method embodiment. For details, please refer to the related description in the foregoing method embodiment.

[0166] Optionally, the memory 502 can be independent or integrated with the processor 501.

[0167] When the memory 502 is independently set, the service equipment further includes a bus 503 for connecting the memory 502 and the processor 501.

[0168] The embodiment of the present application also provides a computer storage medium, which stores computer execution instructions, and when the processor executes the computer execution instructions, the user's stay intention processing method as described above is realized.

[0169] The embodiment of the present application also provides a computer program product, which includes a computer program stored in a computer storage medium, and at least one processor can read the computer program from the computer storage medium, and when the at least one processor executes the computer program, the user's stay intention processing method as above can be realized.

[0170] In several embodiments of the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative, and the division of the modules is merely a logical function division. In actual implementation, another division manner can be used, for example, a plurality of modules can be combined or integrated into another system, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the modules shown or discussed can be indirect coupling or communication connection through some interfaces, devices or modules, and can be electrical, mechanical or other forms.

[0171] The modules described as separate components can or can not be physically separated, and the components shown as modules can or can not be physical units, that is, can be located in one place or distributed on a plurality of network units. Part or all of the modules can be selected to implement the embodiment scheme according to actual needs.

[0172] In addition, the functional modules in each embodiment of the present application can be integrated in one processing unit, or each module can be physically present separately, or two or more modules can be integrated in one unit. The unit formed by the above modules can be realized in the form of hardware or in the form of hardware plus software functional unit.

[0173] The integrated modules realized in the form of software functional modules can be stored in a computer readable storage medium. The software functional modules stored in a storage medium include a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute part of the steps of the method described in each embodiment of the present application.

[0174] It should be understood that the processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), etc. The general-purpose processor can be a microprocessor or the processor can be any conventional processor. The steps of the method disclosed in the application can be directly embodied as hardware processor execution or combined execution by hardware and software modules in the processor.

[0175] The memory can include a high-speed RAM memory, and can also include a non-volatile storage NVM, for example, at least one disk memory, and can also be a U disk, a mobile hard disk, a read-only memory, a magnetic disk or an optical disk, etc.

[0176] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of the present application does not limit to only one bus or one type of bus.

[0177] The storage medium described above can be realized by any type of volatile or nonvolatile storage devices or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic storage, a flash memory, a magnetic disk or an optical disk. The storage medium can be any available medium that can be accessed by a general or special purpose computer.

[0178] An exemplary storage medium is coupled to the processor so that the processor can read information from, and write information to, the storage medium. Of course, the storage medium can be part of the processor. The processor and the storage medium can be located in an application specific integrated circuits (ASIC). Of course, the processor and the storage medium can exist as discrete components in the electronic device or host device.

[0179] Those of ordinary skill in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by relevant hardware instructed by programs. The foregoing programs can be stored in a computer readable storage medium. When the programs are executed, the steps of the above-mentioned method embodiments are executed; and the foregoing storage medium includes various storage media that can store program codes, such as ROM, RAM, magnetic disks or optical disks.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some or all of the technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for processing user lead generation intentions, characterized in that, Applied to service equipment, including: Acquire target user profile data, signaling data, and network data to obtain the first structured data; Acquire the target user's voice and text data to obtain unstructured data; The voice data in the unstructured data is converted into structured industry classification confidence data and lead retention intention confidence data; the text data in the unstructured data is converted into third-party structured data. The industry classification confidence data, the lead retention intention confidence data, the third structured data, and the first structured data are fused using a data fusion method to obtain the user characteristic data of the target user. The user feature data of the target user is input into the trained lead retention intention recognition model to obtain the lead retention intention score of the target user; Based on the lead retention intention score, the lead retention intention of the target user is determined; Output the lead retention intention of the target user; The lead retention intention recognition model is trained based on positive and negative samples. The positive samples are those whose industry classification confidence exceeds a first preset limit and whose lead retention intention confidence exceeds a second preset limit. The negative samples are selected by using a second binary classifier built based on the positive samples and reliable negative samples to perform cyclic prediction on unlabeled data according to a preset loop stopping condition. The positive samples are labeled data. The reliable negative samples were obtained based on the following method: Based on a preset probability, a set of spy samples is obtained from the labeled data; Based on a preset sampling rate, a negative sample set is obtained from the unlabeled data; Use the data in the labeled data excluding the spy sample set as the initial labeled data; Use the union of the spy sample set and the negative sample sample set as the initial unlabeled data; Construct a first binary classifier based on the initial labeled data and the initial unlabeled data; If the predicted probability of the target user in the initial unlabeled data is less than a preset threshold when the first binary classifier is used to predict the initial unlabeled data, then it is determined to be a reliable negative sample.

2. The method according to claim 1, characterized in that, The trained lead retention intention recognition model is a DeepFM model, which includes an FM module and a DNN module. Accordingly, the step of inputting the user feature data of the target user into the trained lead retention intention recognition model to obtain the lead retention intention score of the target user includes: The user feature data is transformed to obtain a transformed feature set; The transformed feature set is input into the FM module of the DeepFM model, so that the FM module extracts low-order features from the feature set; The transformed feature set is input into the DNN module of the DeepFM model, so that the DNN module extracts high-order features from the feature set. The low-order features and the high-order features are fused together; The fused features are nonlinearly transformed using a preset activation function to obtain the investment intention score.

3. The method according to claim 1 or 2, characterized in that, Before acquiring the target user's profile data, signaling data, and network data, the process also includes: The user characteristic data of multiple users from historical projects are obtained as sample data to form a sample data set. Positive and negative samples are selected from the sample dataset; The initial investment intention recognition model is iteratively trained based on the positive and negative samples to obtain the trained investment intention recognition model.

4. The method according to claim 3, characterized in that, Selecting positive samples from the sample dataset includes: Obtain the industry classification confidence data and the lead retention intention confidence data from the sample dataset; The industry classification confidence data that exceeds a first preset limit is selected from the industry classification confidence data and used as the first positive sample set. From the confidence data of the intention to retain funds, those data whose confidence level of the intention to retain funds exceeds a second preset limit are selected as the second positive sample set; The intersection of the first positive sample set and the second positive sample set is used to obtain the final positive sample set.

5. The method according to claim 3, characterized in that, Negative samples are selected from the sample dataset, including: Label the filtered positive samples to obtain a set of labeled data; Unlabeled data is obtained from the sample dataset to obtain a set of unlabeled data; Based on the labeled data and the unlabeled data, reliable negative samples are obtained through machine learning methods; Based on the positive samples and the reliable negative samples, the negative samples are obtained from the unlabeled data using the machine learning method.

6. A device for processing user lead retention intentions, characterized in that, Applied to service equipment, including: The data acquisition and processing unit is used to acquire the target user's profile data, signaling data, and network data to obtain the first structured data; and to acquire the target user's voice data and text data to obtain unstructured data. A multimodal data fusion unit is used to convert the voice data in the unstructured data into structured industry classification confidence data and lead retention intention confidence data; convert the text data in the unstructured data into third structured data; and fuse the industry classification confidence data, lead retention intention confidence data, third structured data, and first structured data through a data fusion method to obtain the user characteristic data of the target user. The lead retention intention recognition unit is used to input the user feature data of the target user into the trained lead retention intention recognition model to obtain the lead retention intention score of the target user; The lead retention intention output unit is used to determine the lead retention intention of the target user based on the lead retention intention score; and output the lead retention intention of the target user. The lead retention intention recognition model is trained based on positive and negative samples. The positive samples are those whose industry classification confidence exceeds a first preset limit and whose lead retention intention confidence exceeds a second preset limit. The negative samples are selected by using a second binary classifier built based on the positive samples and reliable negative samples to perform cyclic prediction on unlabeled data according to a preset loop stopping condition. The positive samples are labeled data. A sample selection unit is configured to: obtain a spy sample set from the labeled data according to a preset probability; obtain a negative sample set from the unlabeled data according to a preset sampling rate; use the data in the labeled data excluding the spy sample set as initial labeled data; use the union of the spy sample set and the negative sample set as initial unlabeled data; construct a first binary classifier based on the initial labeled data and the initial unlabeled data; use the first binary classifier to predict the initial unlabeled data; and if the predicted probability of the target user in the initial unlabeled data is less than a preset threshold, then determine it as a reliable negative sample.

7. A service device, characterized in that, It includes at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, causing the at least one processor to perform the user's lead retention intention processing method as described in any one of claims 1-5.

8. A computer storage medium, characterized in that, The computer storage medium stores computer execution instructions. When the processor executes the computer execution instructions, it implements the user's information retention intention processing method as described in any one of claims 1-5.

Images