Derogatory user identification method and device, computer equipment and readable storage medium

Abstract

The invention relates to a derogatory user identification method and device, computer equipment and a computer readable storage medium. Comprising the following steps: determining feedback category characteristics of a user according to a plurality of session feedback data; obtaining a plurality of feature fields for satisfaction scoring; screening the plurality of feature fields based on a reinforcement learning model, and determining input fields corresponding to the plurality of prediction dimensions; according to an input field corresponding to the target prediction dimension, obtaining in-mold data corresponding to the target user, and obtaining a target feedback category feature of the target user; and inputting the obtained data into a preset satisfaction scoring model to obtain a satisfaction scoring result of the target user in the target prediction dimension, and further identifying whether the target user is a derogatory user. According to the method, the session feedback data is turned into wealth, the limitation of missing of key fields of the data in a traditional prediction analysis mode is made up, a plurality of feature fields are selected and purified for the second time, and the quality of the in-mold data and the accuracy of derogatory user identification are improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method, apparatus, computer device and computer-readable storage medium for identifying derogatory users. Background Art

[0002] Customer perception directly affects customers' evaluation of the operator's overall service quality and corporate image, and is prone to public opinion risks. It is one of the important concerns of operators providing network services. Customer perception evaluation applications need to output a list of users with reduced satisfaction to assist front-line personnel in proactively carrying out service maintenance work at customer service hotlines, business halls and other touchpoints, and help troubleshoot various network problems, such as discovering communities with poor mobile network quality, thereby improving customer service perception and customer satisfaction. However, the prediction and analysis of satisfaction and reduced satisfaction users for operators has become increasingly complex, and the prediction effect of traditional algorithm models is still unsatisfactory. Summary of the invention

[0003] Based on this, it is necessary to provide a method, apparatus, computer device and computer-readable storage medium for identifying derogatory users in response to the above-mentioned technical problems, which can improve the accuracy of satisfaction ratings and the accuracy of identifying derogatory users.

[0004] In a first aspect, the present application provides a method for identifying a derogatory user, comprising:

[0005] Get several pieces of conversation feedback data;

[0006] Determine the feedback category feature corresponding to each user according to a number of conversation feedback data;

[0007] Get several characteristic fields for satisfaction rating;

[0008] Based on the reinforcement learning model, several feature fields are screened to determine the input fields corresponding to multiple prediction dimensions;

[0009] According to the input field corresponding to the target prediction dimension, the input data corresponding to the target user is obtained, and the target feedback category features corresponding to the target user are obtained; the target prediction dimension is any one of the multiple prediction dimensions;

[0010] Input the input data and target feedback category features into the preset satisfaction rating model to obtain the satisfaction rating results of the target users in the target prediction dimension;

[0011] According to the satisfaction rating results of the target users in multiple prediction dimensions, it is possible to identify whether the target users are derogatory users.

[0012] In one embodiment, the feedback category feature corresponding to each user is determined based on a plurality of pieces of conversation feedback data, including:

[0013] Identify the topic category corresponding to each conversation feedback data;

[0014] Convert topic categories into vector representations;

[0015] Perform unsupervised clustering based on vector representation to identify multiple topic clusters;

[0016] Hot coding was performed on multiple topic clusters to determine multiple feedback category characteristics;

[0017] According to the corresponding relationship between the user and the conversation feedback data, the feedback category feature corresponding to each user is determined.

[0018] In one embodiment, identifying the topic category corresponding to each piece of conversation feedback data includes:

[0019] Obtaining a target prompt word; the target prompt word is used to prompt the preset large model to perform a subject category recognition task;

[0020] Constructing several input texts according to the target prompt words and several conversation feedback data;

[0021] Use the preset large model to recognize each input text separately and output the topic category corresponding to each conversation feedback data.

[0022] In one embodiment, the conversation feedback data is voice data; and a plurality of input texts are constructed according to the target prompt word and the plurality of conversation feedback data, including:

[0023] Performing text conversion processing on a plurality of pieces of conversation feedback data to obtain a plurality of pieces of text data;

[0024] Construct several input texts according to the target prompt word and several text data.

[0025] In one embodiment, a plurality of feature fields are screened based on a reinforcement learning model to determine input fields corresponding to a plurality of prediction dimensions, including:

[0026] According to a plurality of characteristic fields, a plurality of different subsets to be selected are constructed; the subsets to be selected include at least one characteristic field;

[0027] For each prediction dimension, calculate the reward value of transferring from the current state to the next state; the current state corresponds to the first subset, and the next state corresponds to the second subset, and the first subset and the second subset are any two different subsets to be selected;

[0028] According to the reward values ​​of multiple state transitions, determine the target subset corresponding to the prediction dimension;

[0029] The feature field corresponding to the target subset is determined as the input field corresponding to the predicted dimension.

[0030] In one embodiment, calculating the reward value for transitioning from a current state to a next state includes:

[0031] Obtain training samples corresponding to the predicted dimension;

[0032] Extracting first data from the training sample according to the feature field corresponding to the first subset, and extracting second data from the training sample according to the feature field corresponding to the second subset;

[0033] Using the first data to train a preset model to determine a first accuracy rate, and using the second data to train a preset model to determine a second accuracy rate;

[0034] A reward value for transferring from a current state to a next state is determined according to the first accuracy rate and the second accuracy rate.

[0035] In one embodiment, identifying whether a target user is a derogatory user based on satisfaction rating results of the target user in multiple prediction dimensions includes:

[0036] Determine the comprehensive satisfaction score of the target user based on the satisfaction rating results of the target user in multiple prediction dimensions;

[0037] Based on the comprehensive satisfaction score, identify whether the target user is a derogatory user.

[0038] In a second aspect, the present application further provides a derogatory user identification device, comprising:

[0039] An acquisition module is used to acquire several pieces of conversation feedback data;

[0040] A determination module, used to determine the feedback category feature corresponding to each user according to a number of conversation feedback data;

[0041] The acquisition module is also used to obtain several characteristic fields used for satisfaction rating;

[0042] The determination module is also used to screen several feature fields based on the reinforcement learning model to determine the input fields corresponding to multiple prediction dimensions;

[0043] The acquisition module is further used to acquire the input data corresponding to the target user according to the input field corresponding to the target prediction dimension, and acquire the target feedback category features corresponding to the target user; the target prediction dimension is any one of the multiple prediction dimensions;

[0044] The scoring module is used to input the input data and target feedback category features into the preset satisfaction scoring model to obtain the satisfaction scoring results of the target users in the target prediction dimension;

[0045] The identification module is used to identify whether the target user is a derogatory user based on the satisfaction score results of the target user in multiple prediction dimensions.

[0046] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method described in the first aspect above are implemented.

[0047] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described in the first aspect above are implemented.

[0048] The above-mentioned derogatory user identification method, device, computer equipment and computer-readable storage medium obtain several conversation feedback data; determine the feedback category features corresponding to each user according to the several conversation feedback data; obtain several feature fields for satisfaction scoring; screen several feature fields based on the reinforcement learning model to determine the input fields corresponding to multiple prediction dimensions; obtain the input data corresponding to the target user according to the input field corresponding to the target prediction dimension, and obtain the target feedback category features corresponding to the target user; the target prediction dimension is any one of the multiple prediction dimensions; input the input data and the target feedback category features into the preset satisfaction scoring model to obtain the satisfaction scoring result of the target user in the target prediction dimension; identify whether the target user is a derogatory user according to the satisfaction scoring results of the target user in multiple prediction dimensions. Through the above-mentioned method, the conversation feedback data is "turned into treasure" and important features related to satisfaction are extracted from it, which makes up for the limitation of missing key fields of data in the traditional satisfaction and derogatory user prediction and analysis methods, thereby improving the accuracy of the satisfaction scoring model. Secondary selection and purification of several feature fields selected by experts and elimination of fields irrelevant to satisfaction can improve the quality of input data, thereby improving the accuracy of satisfaction scores and the accuracy of identifying derogatory users. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0050] FIG1 is a schematic diagram of a flow chart of a method for identifying derogatory users in one embodiment;

[0051] FIG2 is a schematic flow chart of a step of generating feedback category features in one embodiment;

[0052] FIG3 is a schematic diagram of a flow chart of a feature field screening step in one embodiment;

[0053] FIG4 is a schematic diagram of a data processing flow in one embodiment;

[0054] FIG5 is a block diagram of a derogatory user identification device in one embodiment;

[0055] FIG. 6 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0057] It is understandable that the current prediction analysis of satisfaction and derogatory users for operators still has the following pain points:

[0058] 1. The traditional M / B / O domain numerical data comes from many sources and has complex logic. It is unreasonable to rely on expert experience to manually select the input fields from massive fields. It is easy to have redundant field features, and some of the selected fields will not have any effect on model prediction. In fact, some of the selected fields will even interfere with the prediction performance of the model.

[0059] 2. The limitations of traditional M / B / O domain numerical data, such as missing key fields, deviations in data quality, and inability to directly represent customer perception, also severely restrict the accuracy of the prediction model.

[0060] 3. The massive amount of voice conversation data received by the customer service hotline has not been effectively utilized. For example, voice conversation data such as customer complaints and fault reports have an important impact on the prediction analysis of satisfaction deterioration users.

[0061] Among them, the M domain (Management Support System) has location information, such as crowd flow trajectory, map information, etc. The Management Support System (MSS) includes all non-core business processes required to support the enterprise, covering the formulation of corporate strategy and development direction, enterprise risk management, audit management, public publicity and image management, financial and asset management, human resource management, knowledge and R&D management, shareholder and external relationship management, procurement management, enterprise performance evaluation, government policies and laws, etc.

[0062] Domain B (Business Support System): contains user data and business data, such as user consumption habits, terminal information, ARPU grouping, business content, business target groups, etc. The Business Support System (BSS) mainly implements the management of telecommunications services, telecommunications rates, telecommunications marketing, as well as the management and service processes for customers.

[0063] O domain (Operation Support System): contains network data, such as signaling, alarms, faults, network resources, etc. The Operation Support System (OSS) is mainly a backend support system for resources (networks, equipment, computing systems), including professional network management systems, integrated network management systems, resource management systems, business activation systems, service assurance systems, etc., providing support for reliable, secure and stable network operation.

[0064] Based on this, the present application provides a method, apparatus, computer device and computer-readable storage medium for identifying derogatory users, which can solve the above-mentioned technical problems to a certain extent and achieve the effect of improving the accuracy of satisfaction ratings and the accuracy of derogatory user identification.

[0065] In an exemplary embodiment, as shown in FIG1 , a method for identifying derogatory users is provided. This embodiment uses the method applied to a server as an example for illustration. It is understandable that the method can also be applied to a computer device, and can also be applied to a system including a computer device and a server, and is implemented through the interaction between the computer device and the server. In this embodiment, the method includes the following steps:

[0066] Step 102: Obtain several pieces of conversation feedback data.

[0067] The conversation feedback data refers to conversation data used for complaints or fault reporting, and may be data received in advance through a customer service hotline (e.g., service hotline 10000) or online customer service. The conversation feedback data may be voice data (e.g., recorded data of a customer service hotline) or conversation text data (e.g., recorded data of online customer service), which is not limited in this embodiment.

[0068] Step 104: Determine the feedback category feature corresponding to each user according to the plurality of conversation feedback data.

[0069] Among them, by analyzing each conversation feedback data, important features related to satisfaction are extracted from it, and the feedback category features corresponding to each user are determined. It is understandable that the current massive complaint and fault reporting conversation data has high redundancy and low information density. Useful information is covered by a large amount of irrelevant data, and its utilization rate is low. This embodiment uses these conversation data to enrich the input features of the algorithm model.

[0070] Step 106, obtaining several characteristic fields for satisfaction rating.

[0071] Among them, the characteristic field can be a numerical field for satisfaction scoring artificially selected from the M / B / O domain based on expert experience. Optionally, feature engineering processing is performed on the original data corresponding to several characteristic fields, including outlier processing, missing value processing, standard normalization and other operations, to obtain user data for satisfaction scoring.

[0072] Step 108, screening a number of feature fields based on the reinforcement learning model to determine input fields corresponding to a plurality of prediction dimensions.

[0073] Among them, in order to solve the problem of unreasonable method of manually selecting input fields from massive fields based on expert experience, this embodiment designs a reinforcement learning model to perform secondary purification and optimization on the M / B / O domain fields selected based on expert experience, and screen out input fields that are beneficial to the output of the algorithm model in terms of quantity and quality. Optionally, the satisfaction rating task can be understood as predicting multiple tasks at the same time, and the server performs satisfaction rating from multiple preset prediction dimensions, which together constitute a comprehensive evaluation of user satisfaction.

[0074] Step 110, according to the input field corresponding to the target prediction dimension, obtain the input data corresponding to the target user, and obtain the target feedback category features corresponding to the target user; the target prediction dimension is any one of the multiple prediction dimensions.

[0075] Among them, the user data corresponding to the target user is queried, including the field values ​​under multiple fields; according to the input field corresponding to the target prediction dimension, the field values ​​under each input field are filtered out from the user data to obtain the input data corresponding to the target user. The target user is any user to be evaluated, and the user data refers to the basic information about the user recorded in the server and the relevant information generated by the user during the use of the product or service. According to the feedback category features corresponding to each user determined in advance, the target feedback category features corresponding to the target user are queried.

[0076] Step 112, input the input data and target feedback category features into a preset satisfaction rating model to obtain the satisfaction rating result of the target user in the target prediction dimension.

[0077] The preset satisfaction rating model is a pre-trained algorithm model, such as a machine learning model or a deep learning model. This embodiment does not limit the model architecture of the preset satisfaction rating model. The preset satisfaction rating model is used to identify the integrated input data and target feedback category features to determine the satisfaction rating result of the target user in the target prediction dimension.

[0078] In the specific implementation, by inputting the input data and target feedback category features corresponding to different prediction dimensions into the preset satisfaction scoring model, the preset satisfaction scoring model is enabled to predict multiple tasks simultaneously or multiple times, and perform satisfaction scoring from the preset multiple prediction dimensions, thereby obtaining the satisfaction scoring results of the target users in multiple prediction dimensions.

[0079] Step 114 , identifying whether the target user is a derogatory user based on the satisfaction rating results of the target user in multiple prediction dimensions.

[0080] Among them, based on the preset evaluation system, the satisfaction score results of the target user in multiple prediction dimensions are analyzed to identify whether the target user is a derogatory user. Optionally, if the satisfaction score of the target user in any prediction dimension is lower than a certain value (a set parameter, which can be adjusted according to actual conditions), the target user is identified as a derogatory user. Optionally, a comprehensive satisfaction score is determined based on the satisfaction score results of the target user in multiple prediction dimensions. If the comprehensive satisfaction score is lower than a certain value (a set parameter, which can be adjusted according to actual conditions), the target user is identified as a derogatory user. Derogatory users refer to people who are dissatisfied with the products or services provided by the company and may complain or complain frequently.

[0081] In an optional implementation, the method provided in this embodiment can be used to score the satisfaction of each user on the user list, thereby identifying whether each user is a derogatory user. Optionally, a derogatory user list is generated based on the identified derogatory users, and the server outputs the derogatory user list.

[0082] In the above method for identifying derogatory users, several pieces of conversation feedback data are obtained; according to the several pieces of conversation feedback data, the feedback category features corresponding to each user are determined; several feature fields for satisfaction scoring are obtained; based on the reinforcement learning model, several feature fields are screened to determine the input fields corresponding to multiple prediction dimensions; according to the input fields corresponding to the target prediction dimension, the input data corresponding to the target user is obtained, and the target feedback category features corresponding to the target user are obtained; the target prediction dimension is any one of the multiple prediction dimensions; the input data and the target feedback category features are input into the preset satisfaction scoring model to obtain the satisfaction scoring result of the target user in the target prediction dimension; according to the satisfaction scoring results of the target user in multiple prediction dimensions, whether the target user is a derogatory user is identified. Through the above method, the conversation feedback data is "turned into treasure" and important features related to satisfaction are extracted from it, which makes up for the limitation of missing key fields of data in the traditional satisfaction and derogatory user prediction and analysis methods, thereby improving the accuracy of the satisfaction scoring model. The secondary selection and purification of several feature fields selected by experts and the elimination of fields unrelated to satisfaction can improve the quality of the input data, thereby improving the accuracy of satisfaction scoring and the accuracy of derogatory user identification.

[0083] In an exemplary embodiment, step 104 includes: identifying the topic category corresponding to each conversation feedback data; converting the topic category into a vector representation; performing unsupervised clustering based on the vector representation to determine multiple topic clusters; performing hot encoding on the multiple topic clusters to determine multiple feedback category features; and determining the feedback category feature corresponding to each user based on the correspondence between the user and the conversation feedback data.

[0084] The topic category refers to one or more core topics or fields determined based on the conversation content, such as the business category of complaints and the business category of fault reports.

[0085] In an optional implementation, a pre-trained semantic recognition model is used to identify the topic category corresponding to each conversation feedback data. In an optional implementation, a large model is used to identify the topic category corresponding to each conversation feedback data.

[0086] Unsupervised clustering is performed based on the vector representation corresponding to each topic category, and various topic clusters are automatically mined. Hot encoding is performed on the various topic clusters formed by mining to form important category features that can be used for modeling, namely feedback category features. The conversation feedback data includes user identification (ID), which can indicate the corresponding relationship between the user and the conversation feedback data. The conversation feedback data has a corresponding relationship with the topic category, vector representation, and topic cluster. Based on this, the corresponding relationship between the user and the feedback category features can be established, and one or more feedback category features corresponding to each user can be determined.

[0087] In an exemplary embodiment, the topic category corresponding to each conversation feedback data is identified respectively, including: obtaining a target prompt word; the target prompt word is used to prompt a preset large model to perform a topic category identification task; according to the target prompt word and a number of conversation feedback data, a number of input texts are constructed; each input text is identified respectively using the preset large model, and the topic category corresponding to each conversation feedback data is output.

[0088] The preset large model (also called the foundation model) refers to a machine learning model with a large number of parameters and complex structures, which can process massive data and complete various complex tasks, such as natural language processing, computer vision, speech recognition, etc. The preset large model can be any large language model. In this embodiment, the semantic understanding and element extraction of key features are realized by combining the large language model with the prompt project, and the topic category corresponding to each conversation feedback data is determined.

[0089] In an exemplary embodiment, the conversation feedback data is voice data; based on the target prompt word and the plurality of conversation feedback data, a plurality of input texts are constructed, including: performing text conversion processing on the plurality of conversation feedback data to obtain a plurality of text data; based on the target prompt word and the plurality of text data, a plurality of input texts are constructed.

[0090] Among them, the massive complaint and fault reporting voice conversation data is converted into text data, and the semantics of the text data is understood using a preset large model. It is understandable that in response to the challenges of missing some key fields of M / B / O domain numerical data, deviations in data quality, and inability to directly represent customer perception, the embodiment of this application uses speech recognition, large models, vector representation generation, unsupervised learning and other technologies to capture important features that are directly related to customer satisfaction from the massive voice conversation data of the customer service hotline, turning the massive voice conversation data of the customer service hotline into "treasure", greatly improving the performance of the algorithm model.

[0091] In an optional implementation, referring to FIG. 2 , the feedback category feature generating step includes:

[0092] 1. Take the massive voice conversation data of business complaints, fault repairs, etc. from the customer service hotline as input, and use the new modality voice data to enhance the satisfaction score and the accuracy of derogatory user identification.

[0093] 2. The Fast Whisper model converts massive voice conversations into text data.

[0094] 3. ChatGLM3 semantically understands each text according to the prompt and determines which category of complaint or which category of business assurance the text topic belongs to.

[0095] 4. Use the Sentence-Transformer model to generate vectors corresponding to each topic category for representation.

[0096] 5. Use the DBSCAN model to automatically cluster all the vectors in step 4. There is no need to pre-specify the number of clusters. The topic clusters of any shape can be mined according to the actual situation.

[0097] 6. In order to facilitate modeling, each topic cluster formed by clustering needs to be represented by hot coding.

[0098] Among them, voice conversation data such as complaints and guarantees from the customer service hotline express the customers' true feelings, and often more directly reflect the customers' perception of the product or service. They are key modeling indicators for predicting customer satisfaction or derogatory users, and combined with algorithm models can greatly improve the prediction performance.

[0099] In an exemplary embodiment, as shown in FIG3 , step 108 includes:

[0100] Step 302: construct a plurality of different subsets to be selected based on a plurality of characteristic fields; each subset to be selected includes at least one characteristic field.

[0101] Among them, the reinforcement learning model is mainly composed of an agent, an environment, a state, an action, and a reward. After the agent performs an action, the environment will switch to a new state, and the environment will give a reward signal (positive reward or negative reward) for the new state. The agent performs new actions according to a certain strategy based on the new state and the reward fed back by the environment. The goal of reinforcement learning is to automatically find the optimal strategy in a continuous time series. The multiple different subsets to be selected constructed in this embodiment constitute a state set: S={s 1 ,s 2 ,s 3 ,…,s n},s i is the i-th subset to be selected, where i is an integer greater than or equal to 1 and less than or equal to n.

[0102] Step 304, for each prediction dimension, calculate the reward value for transferring from the current state to the next state; the current state corresponds to the first subset, the next state corresponds to the second subset, and the first subset and the second subset are any two different subsets to be selected.

[0103] Step 306: Determine the target subset corresponding to the prediction dimension according to the reward values ​​of multiple state transitions.

[0104] Among them, the stopping conditions are set for the reinforcement learning model, including reaching a certain number of training rounds, the agent has no obvious improvement within a period of time, or the agent's performance reaches a preset threshold, etc. According to the reward values ​​of multiple state transitions, the optimal state is selected, and the candidate subset corresponding to the optimal state is determined as the target subset corresponding to the prediction dimension.

[0105] Step 308: determine the feature field corresponding to the target subset as the input field corresponding to the predicted dimension.

[0106] In this embodiment, high-quality M / B / O domain numerical data features are secondary selected based on the reinforcement learning model, and fields that are not related to satisfaction are eliminated, which can improve the quality of input data.

[0107] In an exemplary embodiment, calculating the reward value for transferring from a current state to a next state includes: obtaining training samples corresponding to the prediction dimension; extracting first data from the training samples according to the feature field corresponding to the first subset, and extracting second data from the training samples according to the feature field corresponding to the second subset; training a preset model using the first data to determine a first accuracy rate, and training the preset model using the second data to determine a second accuracy rate; determining the reward value for transferring from the current state to the next state based on the first accuracy rate and the second accuracy rate.

[0108] The preset model is an algorithm model set in advance, and this embodiment does not limit its composition. According to the feature fields before and after the transfer, the same model is trained and the accuracy index is evaluated to determine the reward value for transferring from the current state to the next state.

[0109] Exemplarily, the implementation method of secondary selection of high-quality M / B / O domain value fields based on the reinforcement learning model is as follows:

[0110] ① State: The state is a subset of the features present in the M / B / O dataset. For example, the current dataset contains three features: dw_pdcp_packet_loss_rate_t2 (PDCP layer downlink user plane packet loss rate in T-2 months), erab_qos_flow_drop_rate_t1 (E-RAB|QoS Flow drop rate in T-1 months), and bill_query_num_percetion_num (actual value of bill query times). The possible states are as follows:

[0111] []-->Emptyset;

[0112] [dw_pdcp_packet_loss_rate_t2],[erab_qos_flow_drop_rate_t1],[bill_query_num_percetion_num]-->1-featureset;

[0113] [dw_pdcp_packet_loss_rate_t2,erab_qos_flow_drop_rate_t1],…,-->2-featureset;

[0114] [dw_pdcp_packet_loss_rate_t,erab_qos_flow_drop_rate_t1,bill_query_num_percetion_num]-->AII-featureset;

[0115] The order of fields in a state is not important; treat the whole state as a set.

[0116] ② Action: Go from one subset to any subset of features that have not been explored. In the feature selection problem, the action is to select features that have not been explored in the current state and add them to the next state. Here are some possible actions:

[0117] [dw_pdcp_packet_loss_rate_t2]-->[dw_pdcp_packet_loss_rate_t2,erab_qos_flow_drop_rate_t1];

[0118] [dw_pdcp_packet_loss_rate_t2,erab_qos_flow_drop_rate_t1]-->[dw_pdcp_packet_loss_rate_t2,erab_qos_flow_drop_rate_t1,bill_query_num_percetion_num].

[0119] For example, the following are impossible actions:

[0120] [dw_pdcp_packet_loss_rate_t2,erab_qos_flow_drop_rate_t1]-->[dw_pdcp_packet_loss_rate_t2];

[0121] [dw_pdcp_packet_loss_rate_t2]-->[dw_pdcp_packet_loss_rate_t2,erab_qos_flow_drop_rate_t1,bill_query_num_percetion_num].

[0122] ③ Reward: Reward is a real number used to evaluate the quality of the state. In the feature selection problem, the reward can be an indicator of the accuracy of the same model improved by adding new features. For example, the reward is calculated as follows:

[0123] [dw_pdcp_packet_loss_rate_t2]-->Accuracy=0.65;

[0124] [dw_pdcp_packet_loss_rate_t2,erab_qos_flow_drop_rate_t1]-->Accuracy=0.77;

[0125] Reward(erab_qos_flow_drop_rate_t1)=0.77-0.65=0.12.

[0126] ④ Optionally, there is no need to train a model in all states. You can set a stopping condition for the feature selection problem. For example, randomly select the next action from the current state with a probability of epsilon (between 0 and 1, usually around 0.2). Otherwise, select the action that maximizes the function, that is, select the action with the highest expected reward in the current state with a probability of 1-epsilon. In multiple experiments, the average reward for each feature on the model accuracy is used to evaluate the importance of each feature, thereby helping feature selection.

[0127] In an optional implementation, referring to FIG. 4 , for numerical input, feature engineering processing is performed using a traditional feature selection set, and then feature selection is performed based on a reinforcement learning model; for voice input, feedback category features are obtained using a speech-to-text algorithm, a large language model and prompt processing, a semantic vector generation algorithm, an unsupervised clustering algorithm, and hot encoding processing; the data determined by the feature selection and the feedback category features are integrated, and the Xgboost algorithm is used to predict the user's satisfaction scores in each dimension and the comprehensive satisfaction score, to identify dissatisfied and derogatory users, and output a list of satisfied and derogatory users.

[0128] In an exemplary embodiment, step 114 includes: determining a comprehensive satisfaction score of the target user based on satisfaction rating results of the target user in multiple prediction dimensions; and identifying whether the target user is a derogatory user based on the comprehensive satisfaction score.

[0129] The target user's comprehensive satisfaction score is determined by integrating the target user's satisfaction rating results in multiple prediction dimensions. The integration method can be to find the average of the satisfaction ratings in multiple prediction dimensions, or to perform weighted summation according to the weights of multiple prediction dimensions, which is not limited in this embodiment.

[0130] Determine whether the comprehensive satisfaction score of the target user is lower than a certain value (a set parameter that can be adjusted according to actual conditions). If so, identify the target user as a derogatory user.

[0131] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0132] Based on the same inventive concept, the embodiment of the present application also provides a derogatory user identification device for implementing the derogatory user identification method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more derogatory user identification device embodiments provided below can refer to the limitations of the derogatory user identification method above, and will not be repeated here.

[0133] In an exemplary embodiment, as shown in FIG5 , a derogatory user identification device is provided, comprising:

[0134] The acquisition module 502 is used to acquire a number of conversation feedback data.

[0135] The determination module 504 is used to determine the feedback category feature corresponding to each user according to a plurality of conversation feedback data.

[0136] The acquisition module 502 is also used to acquire several characteristic fields for satisfaction rating.

[0137] The determination module 504 is further used to screen a number of feature fields based on the reinforcement learning model to determine the input fields corresponding to the multiple prediction dimensions.

[0138] The acquisition module 502 is also used to acquire the input data corresponding to the target user and the target feedback category features corresponding to the target user according to the input field corresponding to the target prediction dimension; the target prediction dimension is any one of the multiple prediction dimensions.

[0139] The scoring module 506 is used to input the input data and target feedback category features into a preset satisfaction scoring model to obtain the satisfaction scoring result of the target user in the target prediction dimension.

[0140] The identification module 508 is used to identify whether the target user is a derogatory user according to the satisfaction rating results of the target user in multiple prediction dimensions.

[0141] In the above-mentioned derogatory user identification device, several conversation feedback data are obtained; feedback category features corresponding to each user are determined according to the several conversation feedback data; several feature fields for satisfaction scoring are obtained; several feature fields are screened based on the reinforcement learning model to determine the input fields corresponding to multiple prediction dimensions; according to the input fields corresponding to the target prediction dimension, the input data corresponding to the target user is obtained, and the target feedback category features corresponding to the target user are obtained; the target prediction dimension is any one of the multiple prediction dimensions; the input data and the target feedback category features are input into the preset satisfaction scoring model to obtain the satisfaction scoring result of the target user in the target prediction dimension; according to the satisfaction scoring results of the target user in multiple prediction dimensions, whether the target user is a derogatory user is identified. Through the above-mentioned method, the conversation feedback data is "turned into treasure", and important features related to satisfaction are extracted from it, which makes up for the limitation of missing key fields of data in the traditional satisfaction and derogatory user prediction and analysis methods, thereby improving the accuracy of the satisfaction scoring model. The secondary selection and purification of several feature fields selected by experts and the elimination of fields unrelated to satisfaction can improve the quality of the input data, thereby improving the accuracy of satisfaction scoring and the accuracy of derogatory user identification.

[0142] In an exemplary embodiment, the determination module 504 is further used to respectively identify the topic category corresponding to each conversation feedback data; convert the topic category into a vector representation; perform unsupervised clustering based on the vector representation to determine multiple topic clusters; perform hot encoding on the multiple topic clusters to determine multiple feedback category features; and determine the feedback category feature corresponding to each user based on the correspondence between the user and the conversation feedback data.

[0143] In an exemplary embodiment, the determination module 504 is also used to obtain a target prompt word; the target prompt word is used to prompt the preset large model to perform a topic category recognition task; according to the target prompt word and a number of conversation feedback data, a number of input texts are constructed; each input text is recognized separately using the preset large model, and the topic category corresponding to each conversation feedback data is output.

[0144] In an exemplary embodiment, the conversation feedback data is voice data; the determination module 504 is further used to perform text conversion processing on the plurality of conversation feedback data to obtain a plurality of text data; and construct a plurality of input texts according to the target prompt words and the plurality of text data.

[0145] In an exemplary embodiment, the determination module 504 is also used to construct multiple different candidate subsets based on several feature fields; the candidate subset includes at least one feature field; for each prediction dimension, the reward value for transferring from the current state to the next state is calculated; the current state corresponds to the first subset, the next state corresponds to the second subset, and the first subset and the second subset are any two different candidate subsets; according to the reward values ​​of multiple state transitions, the target subset corresponding to the prediction dimension is determined; the feature field corresponding to the target subset is determined as the input field corresponding to the prediction dimension.

[0146] In an exemplary embodiment, the determination module 504 is also used to obtain training samples corresponding to the prediction dimension; extract first data from the training samples according to the feature field corresponding to the first subset, and extract second data from the training samples according to the feature field corresponding to the second subset; use the first data to train the preset model to determine the first accuracy rate, and use the second data to train the preset model to determine the second accuracy rate; determine the reward value for transferring from the current state to the next state based on the first accuracy rate and the second accuracy rate.

[0147] In an exemplary embodiment, the identification module 508 is further used to determine the target user's comprehensive satisfaction score based on the target user's satisfaction rating results in multiple prediction dimensions; and to identify whether the target user is a derogatory user based on the comprehensive satisfaction score.

[0148] Each module in the above-mentioned derogatory user identification device can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each module.

[0149] In an exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be shown in FIG6. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store conversation feedback data, feature fields, intermediate processing data of a preset satisfaction rating model, etc. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for identifying derogatory users is implemented.

[0150] Those skilled in the art will understand that the structure shown in FIG. 6 is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0151] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the following steps when executing the computer program: obtaining a number of conversation feedback data; determining feedback category features corresponding to each user based on the number of conversation feedback data; obtaining a number of feature fields for satisfaction scoring; screening the number of feature fields based on a reinforcement learning model to determine input fields corresponding to a plurality of prediction dimensions; obtaining input data corresponding to a target user and obtaining target feedback category features corresponding to the target user based on the input fields corresponding to a target prediction dimension; the target prediction dimension is any one of a plurality of prediction dimensions; inputting the input data and the target feedback category features into a preset satisfaction scoring model to obtain a satisfaction scoring result of the target user in the target prediction dimension; and identifying whether the target user is a derogatory user based on the satisfaction scoring results of the target user in the plurality of prediction dimensions.

[0152] In one embodiment, when the processor executes the computer program, the following steps are also implemented: respectively identifying the topic category corresponding to each piece of conversation feedback data; converting the topic category into a vector representation; performing unsupervised clustering based on the vector representation to determine multiple topic clusters; performing hot encoding on the multiple topic clusters to determine multiple feedback category features; and determining the feedback category feature corresponding to each user based on the correspondence between the user and the conversation feedback data.

[0153] In one embodiment, when the processor executes the computer program, the following steps are also implemented: obtaining a target prompt word; the target prompt word is used to prompt a preset large model to perform a topic category recognition task; constructing a plurality of input texts according to the target prompt word and a plurality of conversation feedback data; using the preset large model to recognize each input text respectively, and outputting the topic category corresponding to each conversation feedback data.

[0154] In one embodiment, when the processor executes the computer program, the following steps are further implemented: performing text conversion processing on a plurality of conversation feedback data to obtain a plurality of text data; and constructing a plurality of input texts according to the target prompt words and the plurality of text data.

[0155] In one embodiment, when the processor executes the computer program, the following steps are also implemented: constructing multiple different subsets to be selected based on several feature fields; the subset to be selected includes at least one feature field; for each prediction dimension, calculating the reward value for transferring from the current state to the next state; the current state corresponds to the first subset, the next state corresponds to the second subset, and the first subset and the second subset are any two different subsets to be selected; determining the target subset corresponding to the prediction dimension based on the reward values ​​of multiple state transitions; determining the feature field corresponding to the target subset as the input field corresponding to the prediction dimension.

[0156] In one embodiment, when the processor executes the computer program, the following steps are also implemented: obtaining training samples corresponding to the predicted dimension; extracting first data from the training samples according to the feature field corresponding to the first subset, and extracting second data from the training samples according to the feature field corresponding to the second subset; using the first data to train the preset model to determine the first accuracy rate, and using the second data to train the preset model to determine the second accuracy rate; determining the reward value for transferring from the current state to the next state based on the first accuracy rate and the second accuracy rate.

[0157] In one embodiment, when the processor executes the computer program, the following steps are also implemented: determining a comprehensive satisfaction score of the target user based on the satisfaction rating results of the target user in multiple prediction dimensions; and identifying whether the target user is a derogatory user based on the comprehensive satisfaction score.

[0158] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented: obtaining a number of conversation feedback data; determining feedback category features corresponding to each user based on the number of conversation feedback data; obtaining a number of feature fields for satisfaction scoring; screening the number of feature fields based on a reinforcement learning model to determine input fields corresponding to a plurality of prediction dimensions; obtaining input data corresponding to a target user based on the input field corresponding to a target prediction dimension, and obtaining target feedback category features corresponding to the target user; the target prediction dimension is any one of a plurality of prediction dimensions; inputting the input data and the target feedback category features into a preset satisfaction scoring model to obtain a satisfaction scoring result of the target user in the target prediction dimension; and identifying whether the target user is a derogatory user based on the satisfaction scoring results of the target user in the plurality of prediction dimensions.

[0159] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: respectively identifying the topic category corresponding to each piece of conversation feedback data; converting the topic category into a vector representation; performing unsupervised clustering based on the vector representation to determine multiple topic clusters; performing hot encoding on the multiple topic clusters to determine multiple feedback category features; and determining the feedback category feature corresponding to each user based on the correspondence between the user and the conversation feedback data.

[0160] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: obtaining a target prompt word; the target prompt word is used to prompt a preset large model to perform a topic category recognition task; constructing a plurality of input texts based on the target prompt word and a plurality of conversation feedback data; using the preset large model to recognize each input text separately, and outputting the topic category corresponding to each conversation feedback data.

[0161] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: performing text conversion processing on a plurality of conversation feedback data to obtain a plurality of text data; and constructing a plurality of input texts according to the target prompt words and the plurality of text data.

[0162] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: based on a number of feature fields, a plurality of different candidate subsets are constructed; the candidate subsets include at least one feature field; for each prediction dimension, a reward value for transitioning from a current state to a next state is calculated; the current state corresponds to a first subset, and the next state corresponds to a second subset, and the first subset and the second subset are any two different candidate subsets; based on the reward values ​​of multiple state transitions, a target subset corresponding to the prediction dimension is determined; and the feature field corresponding to the target subset is determined as the input field corresponding to the prediction dimension.

[0163] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: obtaining training samples corresponding to the predicted dimension; extracting first data from the training samples according to the feature field corresponding to the first subset, and extracting second data from the training samples according to the feature field corresponding to the second subset; using the first data to train the preset model to determine the first accuracy rate, and using the second data to train the preset model to determine the second accuracy rate; determining the reward value for transferring from the current state to the next state based on the first accuracy rate and the second accuracy rate.

[0164] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: determining a comprehensive satisfaction score of the target user based on the satisfaction rating results of the target user in multiple prediction dimensions; and identifying whether the target user is a derogatory user based on the comprehensive satisfaction score.

[0165] In one embodiment, a computer program product is provided, including a computer program, which implements the following steps when executed by a processor: obtaining a number of conversation feedback data; determining feedback category features corresponding to each user based on the number of conversation feedback data; obtaining a number of feature fields for satisfaction scoring; screening the number of feature fields based on a reinforcement learning model to determine input fields corresponding to a plurality of prediction dimensions; obtaining input data corresponding to a target user and obtaining target feedback category features corresponding to the target user based on the input fields corresponding to a target prediction dimension; the target prediction dimension is any one of a plurality of prediction dimensions; inputting the input data and the target feedback category features into a preset satisfaction scoring model to obtain a satisfaction scoring result of the target user in the target prediction dimension; and identifying whether the target user is a derogatory user based on the satisfaction scoring results of the target user in the plurality of prediction dimensions.

[0166] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: respectively identifying the topic category corresponding to each piece of conversation feedback data; converting the topic category into a vector representation; performing unsupervised clustering based on the vector representation to determine multiple topic clusters; performing hot encoding on the multiple topic clusters to determine multiple feedback category features; and determining the feedback category feature corresponding to each user based on the correspondence between the user and the conversation feedback data.

[0167] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: obtaining a target prompt word; the target prompt word is used to prompt a preset large model to perform a topic category recognition task; constructing a plurality of input texts based on the target prompt word and a plurality of conversation feedback data; using the preset large model to recognize each input text separately, and outputting the topic category corresponding to each conversation feedback data.

[0168] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: performing text conversion processing on a plurality of conversation feedback data to obtain a plurality of text data; and constructing a plurality of input texts according to the target prompt words and the plurality of text data.

[0169] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: based on a number of feature fields, a plurality of different candidate subsets are constructed; the candidate subsets include at least one feature field; for each prediction dimension, a reward value for transitioning from a current state to a next state is calculated; the current state corresponds to a first subset, and the next state corresponds to a second subset, and the first subset and the second subset are any two different candidate subsets; based on the reward values ​​of multiple state transitions, a target subset corresponding to the prediction dimension is determined; and the feature field corresponding to the target subset is determined as the input field corresponding to the prediction dimension.

[0170] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: obtaining training samples corresponding to the predicted dimension; extracting first data from the training samples according to the feature field corresponding to the first subset, and extracting second data from the training samples according to the feature field corresponding to the second subset; using the first data to train the preset model to determine the first accuracy rate, and using the second data to train the preset model to determine the second accuracy rate; determining the reward value for transferring from the current state to the next state based on the first accuracy rate and the second accuracy rate.

[0171] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: determining a comprehensive satisfaction score of the target user based on the satisfaction rating results of the target user in multiple prediction dimensions; and identifying whether the target user is a derogatory user based on the comprehensive satisfaction score.

[0172] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0173] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0174] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0175] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for identifying derogatory users, characterized in that: The method comprises: Get several pieces of conversation feedback data; Determining a feedback category feature corresponding to each user according to the plurality of pieces of conversation feedback data; Get several characteristic fields for satisfaction rating; Screening the plurality of feature fields based on the reinforcement learning model to determine input fields corresponding to the plurality of prediction dimensions; According to the input field corresponding to the target prediction dimension, the input data corresponding to the target user is obtained, and the target feedback category feature corresponding to the target user is obtained; the target prediction dimension is any one of the multiple prediction dimensions; Input the input data and the target feedback category features into a preset satisfaction rating model to obtain the satisfaction rating result of the target user in the target prediction dimension; According to the satisfaction rating results of the target user in the multiple prediction dimensions, it is identified whether the target user is a derogatory user.

2. The method according to claim 1, characterized in that The determining, according to the plurality of pieces of conversation feedback data, a feedback category feature corresponding to each user includes: Identify the topic category corresponding to each conversation feedback data; Converting the subject categories into vector representations; Performing unsupervised clustering based on the vector representation to determine a plurality of topic clusters; Hot encoding the multiple topic clusters to determine multiple feedback category features; According to the corresponding relationship between the user and the conversation feedback data, the feedback category feature corresponding to each user is determined.

3. The method according to claim 2, characterized in that The step of respectively identifying the topic category corresponding to each piece of conversation feedback data includes: Obtaining a target prompt word; the target prompt word is used to prompt a preset large model to perform a subject category recognition task; Constructing a plurality of input texts according to the target prompt word and the plurality of conversation feedback data; Use the preset large model to recognize each input text separately and output the topic category corresponding to each conversation feedback data.

4. The method according to claim 3, characterized in that The conversation feedback data is voice data; the step of constructing a plurality of input texts according to the target prompt word and the plurality of conversation feedback data includes: Performing text conversion processing on the plurality of pieces of conversation feedback data to obtain a plurality of pieces of text data; Constructing a plurality of input texts according to the target prompt word and the plurality of text data.

5. The method according to any one of claims 1 to 4, characterized in that: The step of screening the plurality of feature fields based on the reinforcement learning model to determine the input fields corresponding to the plurality of prediction dimensions respectively includes: According to the plurality of characteristic fields, a plurality of different subsets to be selected are constructed; the subsets to be selected include at least one characteristic field; For each prediction dimension, a reward value for transferring from a current state to a next state is calculated; the current state corresponds to a first subset, the next state corresponds to a second subset, and the first subset and the second subset are any two different subsets to be selected; According to the reward values ​​of multiple state transitions, determine the target subset corresponding to the prediction dimension; The characteristic field corresponding to the target subset is determined as the input field corresponding to the predicted dimension.

6. The method according to claim 5, characterized in that The calculation of the reward value for transferring from the current state to the next state includes: Obtain training samples corresponding to the predicted dimension; Extracting first data from the training sample according to the feature field corresponding to the first subset, and extracting second data from the training sample according to the feature field corresponding to the second subset; Using the first data to train a preset model to determine a first accuracy rate, and using the second data to train the preset model to determine a second accuracy rate; A reward value for transitioning from a current state to a next state is determined according to the first accuracy rate and the second accuracy rate.

7. The method according to any one of claims 1 to 4, characterized in that The step of identifying whether the target user is a derogatory user according to the satisfaction rating results of the target user in the multiple prediction dimensions includes: Determining a comprehensive satisfaction score of the target user according to the satisfaction rating results of the target user in the multiple prediction dimensions; According to the comprehensive satisfaction score, it is identified whether the target user is a derogatory user.

8. A derogatory user identification device, characterized in that The device comprises: An acquisition module is used to acquire several pieces of conversation feedback data; A determination module, configured to determine a feedback category feature corresponding to each user according to the plurality of pieces of conversation feedback data; The acquisition module is also used to acquire several characteristic fields for satisfaction rating; The determination module is further used to screen the plurality of feature fields based on the reinforcement learning model to determine the input fields corresponding to the plurality of prediction dimensions respectively; The acquisition module is further used to acquire input data corresponding to the target user according to the input field corresponding to the target prediction dimension, and acquire target feedback category features corresponding to the target user; the target prediction dimension is any one of the multiple prediction dimensions; A scoring module, used to input the input data and the target feedback category features into a preset satisfaction scoring model to obtain a satisfaction scoring result of the target user in the target prediction dimension; An identification module is used to identify whether the target user is a derogatory user based on the satisfaction rating results of the target user in the multiple prediction dimensions.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images