Renewal behavior prediction method and device, electronic device and storage medium

Abstract

The present invention provides a renewal behavior prediction method and apparatus, electronic device, and storage medium, belonging to the field of artificial intelligence technology. The method comprises: obtaining insurance speech sentences, clustering the insurance speech sentences to obtain speech expression features and speech categories of the speech expression features, obtaining sample renewal prompt dialogue sentences, extracting features from the sample renewal prompt dialogue sentences to obtain target renewal objection features, predicting renewal behavior based on the target renewal objection features to obtain a target renewal category, screening the dialogue categories based on the speech expression features and the target renewal objection features to obtain a target category, constructing a model based on the target category, the target renewal category, and preset target features to obtain a target renewal prediction model, and predicting renewal behavior for the target renewal prompt dialogue sentences using the target renewal prediction model, thereby improving the accuracy of renewal behavior prediction.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology and is applicable to the field of financial technology, and in particular to a renewal behavior prediction method and device, electronic device and storage medium. Background Art

[0002] In the insurance industry, premium renewal is crucial for maintaining policy validity and providing ongoing coverage. Insurance companies typically remind users to renew their policies via phone calls, and recorded call data contains detailed information about interactions between customers and insurers. Related technologies use keyword matching to extract keywords from recorded call data and predict renewal behavior based on these keywords. However, this approach cannot accurately identify renewal intentions. For example, if a call recording contains the phrase "Are you not paying anymore?", keyword matching would interpret this phrase as a non-renewal, resulting in low renewal prediction accuracy. Summary of the Invention

[0003] The main purpose of the embodiments of the present application is to propose a renewal behavior prediction method and device, electronic device and storage medium, aiming to improve the accuracy of renewal behavior prediction.

[0004] To achieve the above objectives, a first aspect of an embodiment of the present application provides a renewal behavior prediction method, the method comprising:

[0005] Get insurance jargon;

[0006] Clustering the insurance speech sentences to obtain speech expression features and speech categories of the speech expression features;

[0007] Get sample renewal reminder dialogue sentences;

[0008] Extract features from the sample renewal prompt dialogue sentences to obtain renewal objection features;

[0009] Predicting renewal behavior based on the renewal objection characteristics of the object to obtain a renewal category of the object; the renewal category of the object is used to indicate whether to renew the policy or not;

[0010] Filtering the speech categories according to the speech expression characteristics and the renewal objection characteristics of the target to obtain a target category;

[0011] A model is constructed based on the target category, the object renewal category, and preset object features to obtain a target renewal prediction model; the preset object features are pre-acquired features of the preset renewal object;

[0012] The target renewal prediction model is used to predict the renewal behavior of the preset target renewal prompt dialogue sentence.

[0013] In some embodiments, the model is constructed based on the target category, the object renewal category, and preset object features to obtain a target renewal prediction model, including:

[0014] Constructing a decision tree based on the target category, the object renewal category, and the preset object features to obtain a first decision tree;

[0015] Predicting renewal behavior for the target category and the preset object features using the first decision tree to obtain a predicted renewal category;

[0016] Obtaining the renewal prediction weight of the first decision tree;

[0017] Calculating a reference sample gradient according to the object renewal category, the predicted renewal category, and the renewal prediction weight;

[0018] Constructing a decision tree based on the reference sample gradient, the target category, the object renewal category, and the preset object features to obtain a second decision tree;

[0019] The target renewal prediction model is determined based on the first decision tree and the second decision tree.

[0020] In some embodiments, constructing a decision tree based on the target category, the object renewal category, and the preset object features to obtain a first decision tree includes:

[0021] Performing feature screening on the target category and the preset object features to obtain original features;

[0022] Discretizing the original features to obtain discrete sub-features;

[0023] Calculating a feature gain of the discrete sub-feature according to the object renewal category;

[0024] Selecting the discrete sub-feature with the largest feature gain as the reference sub-feature;

[0025] Node construction is performed according to the reference sub-features to obtain a decision node;

[0026] The first decision tree is constructed according to the decision nodes.

[0027] In some embodiments, calculating the feature gain of the discrete sub-feature according to the object renewal category includes:

[0028] Determine the reference renewal category based on the renewal category of the object;

[0029] screening the renewal categories of the object according to the discrete sub-features to obtain candidate renewal categories and the number of candidate renewal categories;

[0030] determining an original sample gradient according to the reference renewal category, the candidate renewal category, and the number of categories of the candidate renewal category;

[0031] The feature gain of the discrete sub-feature is calculated according to the original sample gradient.

[0032] In some embodiments, calculating the feature gain of the discrete sub-feature according to the original sample gradient includes:

[0033] Performing feature splitting on the discrete sub-features to obtain a first sub-feature and a second sub-feature;

[0034] Calculating a first sub-gain of the first sub-feature according to the original sample gradient;

[0035] Calculating a second sub-gain of the second sub-feature according to the original sample gradient;

[0036] The first sub-gain and the second sub-gain are summed to obtain a feature gain of the discrete sub-feature.

[0037] In some embodiments, obtaining the renewal prediction weight of the first decision tree includes:

[0038] Calculate the loss based on the target renewal category and the predicted renewal category to obtain target loss data;

[0039] determining first-order moment data and second-order moment data according to the target loss data;

[0040] The renewal prediction weight of the first decision tree is calculated based on the first-order moment data and the second-order moment data.

[0041] In some embodiments, clustering the insurance speech sentences to obtain speech expression features and speech categories of the speech expression features includes:

[0042] Performing sentence vectorization processing on the insurance speech phrase to obtain an insurance speech phrase vector;

[0043] Calculating the distance between the insurance speech vector and a preset cluster center to obtain a speech distance;

[0044] Clustering the insurance speech vectors according to the speech distance to obtain insurance speech clusters;

[0045] Performing an average calculation on the insurance speech clusters to obtain the speech expression features;

[0046] The speech category of the speech expression feature is determined based on the insurance speech cluster.

[0047] To achieve the above-mentioned objectives, a second aspect of an embodiment of the present application provides a renewal behavior prediction device, comprising:

[0048] A speech sentence acquisition module, which acquires insurance speech sentences;

[0049] A clustering module, configured to cluster the insurance speech sentences to obtain speech expression features and speech categories of the speech expression features;

[0050] A dialogue statement acquisition module is used to obtain sample renewal prompt dialogue statements;

[0051] A feature extraction module is used to extract features from the sample renewal prompt dialogue sentences to obtain the renewal objection features of the object;

[0052] an original renewal behavior prediction module, configured to predict renewal behavior based on the renewal objection characteristics of the object, and obtain a renewal category of the object; the renewal category of the object is used to indicate whether to renew or not renew;

[0053] A screening module, configured to screen the speech categories according to the speech expression characteristics and the renewal objection characteristics of the object, to obtain a target category;

[0054] A construction module, configured to construct a model based on the target category, the object renewal category, and preset object features to obtain a target renewal prediction model; the preset object features are pre-acquired features of the preset renewal object;

[0055] The target renewal behavior prediction module is used to predict the renewal behavior of the preset target renewal prompt dialogue sentence through the target renewal prediction model.

[0056] To achieve the above-mentioned purpose, the third aspect of an embodiment of the present application proposes an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the method described in the first aspect when executing the computer program.

[0057] To achieve the above-mentioned purpose, the fourth aspect of the embodiments of the present application proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the method described in the first aspect.

[0058] The renewal behavior prediction method, renewal behavior prediction device, electronic device, and computer-readable storage medium proposed in the embodiments of the present application obtain insurance speech sentences, cluster the insurance speech sentences, classify semantically similar insurance speech sentences into one category, obtain speech expression features and speech categories of speech expression features, and construct labels for insurance speech sentences. Sample renewal prompt dialogue sentences are obtained to predict renewal behavior. In renewal behavior prediction, object objections can reflect the renewal object's renewal tendency, behavior pattern, and potential needs. Feature extraction is performed on sample renewal prompt dialogue sentences to obtain features related to object objections and obtain object renewal objection features. In order to obtain the real renewal label required for model construction, renewal behavior prediction is performed based on the object renewal objection features to obtain the object renewal category. In order to accurately identify renewal intentions, the dialogue categories are screened based on the speech expression features and the object renewal objection features to filter out renewal intentions that match the object renewal objection features from the pre-built label system and obtain the target category. To improve the accuracy of renewal behavior prediction, the target category was incorporated into the model building process as a new feature. A model was constructed based on the target category, the target renewal category, and pre-set object features to obtain a model with superior prediction performance, resulting in the target renewal prediction model. This target renewal prediction model was used to predict renewal behavior for pre-set target renewal prompt dialogue sentences, accurately identifying the renewal intention implicit in the target renewal prompt dialogue sentences and improving the accuracy of renewal behavior prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flowchart of the renewal behavior prediction method provided in an embodiment of the present application;

[0060] Figure 2 yes Figure 1 Flowchart of step S120 in FIG.

[0061] Figure 3 yes Figure 1 Flowchart of step S170 in FIG.

[0062] Figure 4 yes Figure 3 Flowchart of step S310 in FIG.

[0063] Figure 5 yes Figure 4 Flowchart of step S430 in FIG.

[0064] Figure 6 yes Figure 5 Flowchart of step S540 in FIG.

[0065] Figure 7 yes Figure 3 Flowchart of step S330 in FIG.

[0066] Figure 8 This is a schematic diagram of the structure of the renewal behavior prediction device provided in an embodiment of the present application;

[0067] Figure 9 This is a schematic diagram of the hardware structure of the electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0069] It should be noted that although the device schematics illustrate functional module divisions and the flowcharts illustrate logical sequences, in certain circumstances, the steps shown or described may be performed in a sequence that differs from the module divisions in the device or the sequence in the flowcharts. The terms "first," "second," and so on, in the specification, claims, and drawings, are used to distinguish similar items and are not necessarily used to describe a specific sequence or precedence.

[0070] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0071] In the insurance industry, premium renewal is crucial for maintaining policy validity and providing ongoing coverage. Insurance companies often remind users to renew their policies via phone calls, and recorded call data contains detailed information about interactions between customers and insurers. Related technologies use keyword matching to extract keywords from recorded call data and predict renewal behavior based on these keywords. However, this approach cannot accurately predict renewal behavior. For example, if a call recording contains the phrase "Are you not paying?", keyword matching would interpret this phrase as a non-renewal, resulting in low renewal prediction accuracy.

[0072] Based on this, the embodiments of the present application provide a renewal behavior prediction method, a renewal behavior prediction device, an electronic device and a computer-readable storage medium, aiming to improve the accuracy of renewal behavior prediction.

[0073] The renewal behavior prediction method, renewal behavior prediction device, electronic device and computer-readable storage medium provided in the embodiments of the present application are specifically illustrated through the following embodiments. First, the renewal behavior prediction method in the embodiments of the present application is described.

[0074] The renewal behavior prediction method provided in the embodiment of the present application relates to the field of artificial intelligence technology. The renewal behavior prediction method provided in the embodiment of the present application can be applied to a terminal, can be applied to a server side, or can be software running in a terminal or a server side. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc.; the server side can be configured as an independent physical server, or as a server cluster or distributed system composed of multiple physical servers, or as a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application that implements the renewal behavior prediction method, etc., but is not limited to the above forms.

[0075] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and the like. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments in which tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0076] Figure 1 This is an optional flowchart of the renewal behavior prediction method provided in an embodiment of the present application, which may include but is not limited to steps S110 to S180.

[0077] Step S110, obtaining insurance terminology sentences;

[0078] Step S120: Clustering insurance speech sentences to obtain speech expression features and speech categories of the speech expression features;

[0079] Step S130, obtaining a sample renewal prompt dialogue statement;

[0080] Step S140: extracting features from the sample renewal prompt dialogue sentences to obtain renewal objection features;

[0081] Step S150: predicting the renewal behavior of the object based on the renewal objection characteristics to obtain the object renewal category; the object renewal category is used to indicate whether to renew or not renew;

[0082] Step S160 , filtering the dialogue categories based on the dialogue expression features and the renewal objection features of the target to obtain the target category;

[0083] Step S170: constructing a model based on the target category, the target renewal category, and preset target features to obtain a target renewal prediction model; the preset target features are pre-acquired features of the preset renewal target;

[0084] Step S180: predicting the renewal behavior of the preset target renewal prompt dialogue sentence using the target renewal prediction model.

[0085] In steps S110 to S180, as shown in this embodiment of the present application, a pre-constructed tagging system is used to extract customer objection speech. Tags with semantically similar expressions to the customer objection speech are extracted from the tagging system to accurately identify renewal intent. These tags are then used as new features in model construction to improve the accuracy of the model's prediction of renewal behavior.

[0086] In some embodiments, step S110 obtains all possible questions or expressions that a user might raise in an insurance scenario to obtain insurance jargon sentences. Insurance jargon sentences can be used to convey the user's needs for insurance products. Examples of insurance jargon sentences include "This insurance product is useless" and "This insurance product is too expensive."

[0087] See also Figure 2 In some embodiments, step S120 may include but is not limited to steps S210 to S250:

[0088] Step S210, performing sentence vectorization processing on the insurance speech phrases to obtain insurance speech phrase vectors;

[0089] Step S220, calculating the distance between the insurance speech vector and the preset cluster center to obtain the speech distance;

[0090] Step S230: clustering the insurance speech vectors according to speech distance to obtain insurance speech clusters;

[0091] Step S240: averaging the insurance speech clusters to obtain speech expression features;

[0092] Step S250: Determine the speech category of the speech expression feature based on the insurance speech cluster.

[0093] In step S210 of some embodiments, the insurance sales话术语句 is subjected to sentence vectorization processing, converting the insurance sales话术语句 into a fixed-length vector representation, extracting a semantic vector that reflects the sentence semantics and structural information from the insurance sales话术语句, and obtaining an insurance sales vector.

[0094] Specifically, the insurance sales话术语句 is tokenized to obtain insurance sales words. Stop words such as "de" (的) and "le" (了) in the insurance sales words are removed to obtain words that can provide key information, resulting in initial话术词语. The initial话术词语 are subjected to word feature extraction through a pre-trained word embedding model to obtain word semantic features. The word embedding model can be a word2vec model, a glove model, etc. The word semantic features of all initial话术词语 are semantically encoded to obtain an insurance sales vector.

[0095] The word semantic features of all initial话术词语 can be averaged to obtain an insurance sales vector. For example, if the number of initial话术词语 in the insurance sales话术语句 is C, and the word semantic feature of the i-th initial话术词语 is represented as xi, then the insurance sales vector is represented as:

[0096]

[0097] It is also possible to obtain the word weights of the initial话术词语, and perform weighted calculation on the word semantic features of the initial话术词语 according to the word weights to obtain an insurance sales vector. The process of obtaining the word weights is as follows: Obtain the frequency of occurrence of the initial话术词语 in the insurance sales话术语句 to obtain the term frequency. Use the total number of insurance sales话术语句 as the first quantity, and the number of insurance sales话术语句 in which the initial话术词语 appears as the second quantity, and calculate the inverse document frequency according to the first quantity and the second quantity. Multiply the term frequency and the inverse document frequency to obtain the word weight. The calculation formula for the term frequency is expressed as:

[0098]

[0099] Where TF represents the term frequency; t represents the number of times the initial话术词语 appears in the insurance sales话术语句; d represents the total number of words in the insurance sales话术语句.

[0100] The calculation formula for the inverse document frequency is expressed as:

[0101]

[0102] Where IDF represents the inverse document frequency; N represents the first quantity; DF represents the second quantity.

[0103] It is also possible to use an encoding network to perform feature encoding on the word semantic features of all initial话术词语 to obtain an insurance sales vector. The encoding network can be a recurrent neural network, a transformer model, etc.

[0104] In step S220 of some embodiments, K insurance speech vectors are randomly selected as K preset cluster centers. For each insurance speech vector, the Euclidean distance between the insurance speech vector and each preset cluster center is calculated to obtain the speech distance from the insurance speech vector to each preset cluster center. The larger the speech distance, the lower the similarity between the insurance speech vector and the preset cluster center.

[0105] In some embodiments, in step S230, the preset cluster center with the smallest speech distance is selected as the original cluster center, and the insurance speech vectors are clustered to the original cluster center to obtain an original speech cluster. The mean of the insurance speech vectors in the original speech cluster is calculated to obtain a mean speech vector, and the original cluster center is updated based on the mean speech vector to obtain a reference cluster center. If the reference cluster center changes slightly for multiple consecutive times or the number of clusterings reaches a preset threshold, the speech distance between the insurance speech vector and the reference cluster center is calculated, and the reference cluster center with the smallest speech distance is selected as the target cluster center. The insurance speech vectors are clustered to the target cluster center to obtain an insurance speech cluster. Clustering can group semantically similar insurance speech vectors into the same insurance speech cluster. For example, "I don't have money to pay for now" and "I'm a bit tight on cash lately" can be grouped into the same insurance speech cluster, and "This insurance product is useless" and "This insurance product is too expensive" can be grouped into the same insurance speech cluster.

[0106] In step S240 of some embodiments, the insurance speech vectors in the insurance speech cluster are averaged to obtain speech expression features. For example, if the insurance speech cluster includes three insurance speech vectors s1, s2, and s3, the speech expression feature can be expressed as (s1+s2+s3) / 3.

[0107] In some embodiments, step S250 obtains a category label for the insurance speech cluster, obtaining the speech category of the speech expression feature. The speech category indicates the speech intent expressed by the insurance speech phrase, such as financial difficulties or dissatisfaction with the insurance product. The speech expression features and the speech category of the speech expression features are recorded in a hash table to construct a precise speech labeling system.

[0108] Through the above steps S210 to S250, a speech tag system can be constructed to accurately and quickly identify the user's renewal intention based on the speech tag system, thereby improving the accuracy of renewal behavior prediction and performing real-time renewal behavior prediction.

[0109] In step S130 of some embodiments, the insurance company uses a telephone call to remind the renewal subject to renew their premiums. The call recording data contains detailed information about the interaction between the renewal subject and the insurance company. Noise reduction is performed on the audio file containing the call recording data to remove background noise from the audio file, resulting in the original audio file. The original audio file is edited to highlight important audio content, resulting in a target audio file. The audio content of the target audio file is converted into text using an automatic speech recognition system to obtain a sample renewal reminder dialogue sentence. The sample renewal reminder dialogue sentence is used as a sample for training the renewal prediction model and includes the content of the insurance company's reminder and guidance to the renewal subject to renew their premiums, as well as the renewal subject's response to the renewal premium payment.

[0110] In step S140 of some embodiments, the large language model is trained on a large amount of text data and is capable of capturing complex semantic relationships and contextual information in the text. Leveraging the large language model's powerful semantic understanding capabilities, the present embodiment inputs prompt instructions into the large language model, performs feature extraction on sample renewal prompt dialogue sentences based on the prompt instructions, and extracts the content of the renewal subject's objection to the renewal premium from the call recording text to obtain the subject renewal objection feature. The subject renewal objection feature can include the reason why the renewal subject did not renew the premium, such as financial reasons, family reasons, dissatisfaction with the insurance product, or inquiries about other insurance products.

[0111] In step S150 of some embodiments, the large language model can process complex text data and semantic relationships. It uses the large language model to perform deep semantic understanding and contextual awareness of the renewal objection characteristics of the subject. By analyzing the language patterns and sentiment of the renewal object, it identifies key factors influencing renewal behavior, accurately predicts the renewal object's renewal behavior, and determines the object's renewal category. The object's renewal category indicates whether the renewal object will renew or not renew.

[0112] In step S160 of some embodiments, the similarity between the speech expression feature and the object renewal objection feature is calculated, and the speech category corresponding to the speech expression feature with the greatest similarity is used as the target category. This can match the object renewal objection feature with a suitable speech category label to identify the specific reason why the renewal object does not renew the premium, thereby accurately identifying the renewal intention hidden in the sample renewal prompt dialogue sentence.

[0113] See also Figure 3 In some embodiments, step S170 may include but is not limited to steps S310 to S360:

[0114] Step S310: constructing a decision tree based on the target category, the renewal category of the object, and the preset object features to obtain a first decision tree;

[0115] Step S320, predicting renewal behavior based on the target category and preset object features using the first decision tree to obtain a predicted renewal category;

[0116] Step S330, obtaining the renewal prediction weight of the first decision tree;

[0117] Step S340, calculating a reference sample gradient based on the target renewal category, the predicted renewal category, and the renewal prediction weight;

[0118] Step S350: constructing a decision tree based on the reference sample gradient, the target category, the object renewal category, and the preset object features to obtain a second decision tree;

[0119] Step S360: Determine a target renewal prediction model based on the first decision tree and the second decision tree.

[0120] In step S310 of some embodiments, the preset subject features are pre-acquired features of the renewal subject. These preset subject features can influence the renewal subject's renewal behavior, such as the renewal subject's age, renewal term, insurance product to be renewed, insurance company selected by the renewal subject, insurance product price, and coverage. Renewal behavior is influenced by multiple factors, such as renewal term, renewal subject's age, vehicle age, and insurance company. To capture the nonlinear dependencies and complex feature interactions among these multiple factors and to intuitively display features that significantly influence renewal behavior, embodiments of the present application employ a decision tree as the underlying model for predicting renewal behavior.

[0121] The embodiment of the present application constructs a new feature, namely the target category, through a sample renewal prompt dialogue statement. The target category indicates the objection reason for the renewal subject's non-renewal of premiums and the needs of the renewal subject. The target category is related to the renewal behavior of the renewal subject. In order to improve the accuracy of the renewal behavior prediction, the target category is used as a new feature in the training process of the renewal behavior prediction model. Specifically, the target category and the preset object feature are used as samples, the object renewal category is used as the true label of the sample, and a first decision tree is constructed based on the sample and the true label. A single decision tree is prone to overfitting, resulting in the model being unable to accurately predict the renewal behavior of unknown samples. In order to improve the generalization ability of the model, the embodiment of the present application uses multiple decision trees to learn different features and patterns from the samples to provide comprehensive and accurate renewal behavior prediction.

[0122] In step S320 of some embodiments, the target category and the preset object features are input into the first decision tree to predict the renewal behavior and obtain a predicted renewal category. The predicted renewal category is used to indicate whether the renewal object renews or does not renew.

[0123] In some embodiments, step S330 sets weights for each decision tree to control its contribution to the renewal prediction results, prioritizing decision trees that contribute significantly to the prediction during model training. This balances the contributions of each decision tree, rather than allowing a single decision tree to be so powerful as to dominate the entire model's predictions. The weight of the first decision tree is obtained to obtain a renewal prediction weight. The renewal prediction weight reflects the contribution of the first decision tree to the renewal prediction; a higher renewal prediction weight indicates a greater contribution of the first decision tree.

[0124] In step S340 of some embodiments, multiple samples are used in the construction of the first decision tree. Each sample includes a target category, an object renewal category, and a preset object feature. The object renewal category has an object renewal value. If the object renewal value is 1, the object renewal category is renewal; if the object renewal value is 0, the object renewal category is non-renewal. The object renewal values ​​of all samples are averaged to obtain a renewal mean. An initial renewal label is determined based on the renewal mean. If the renewal mean is greater than a preset renewal threshold, the initial renewal label is determined to be renewal based on the renewal mean. The preset renewal threshold can be 0.5. If the renewal mean is less than or equal to the preset renewal threshold, the initial renewal label is determined to be non-renewal based on the renewal mean.

[0125] Multiply the renewal prediction weight by the predicted renewal category, and add the result to the initial renewal label to obtain the pseudo-renewal category. Calculate the gradient of the first decision tree based on the pseudo-renewal category and the target renewal category to obtain the reference sample gradient. The formula for gradient calculation is:

[0126]

[0127] Where g represents the reference sample gradient; represents the pseudo renewal category; y represents the object renewal category.

[0128] In some embodiments, in step S350, to improve the accuracy of renewal behavior prediction, it is necessary to focus on samples for which the first decision tree failed to accurately predict renewal. Referring to step S310, the reference sample gradient is used as a new feature. A decision tree is constructed based on the reference sample gradient, the target category, the object renewal category, and the preset object features. This fits the reference sample gradient and corrects the prediction errors generated by the first decision tree, resulting in a second decision tree. It should be noted that the reference sample gradient, the target category, and the preset object features are all sample features, and the object renewal category is the true label of the sample features.

[0129] In step S360 of some embodiments, the first decision tree and the second decision tree are used as target renewal prediction models.

[0130] In the above steps S310 to S360, different feature representations can be learned from different perspectives through multiple decision trees, and the multiple decision trees can be integrated into a target renewal prediction model to combine the prediction results of multiple decision trees, thereby reducing prediction errors and improving the accuracy of renewal behavior prediction.

[0131] See also Figure 4 In some embodiments, step S310 may include but is not limited to steps S410 to S460:

[0132] Step S410, performing feature screening on the target category and the preset object features to obtain original features;

[0133] Step S420, discretizing the original features to obtain discrete sub-features;

[0134] Step S430, calculating the feature gain of the discrete sub-feature according to the policy renewal category of the object;

[0135] Step S440, selecting the discrete sub-feature with the largest feature gain as the reference sub-feature;

[0136] Step S450, constructing nodes based on the reference sub-features to obtain decision nodes;

[0137] Step S460: construct a first decision tree based on the decision nodes.

[0138] In step S410 of some embodiments, characteristic attributes of the target category and the preset object feature are obtained, the Gini impurity of the characteristic attributes is calculated, and the characteristic attribute with the largest Gini impurity is selected as the original feature. The target category and the preset object feature are both values ​​of the characteristic attributes. For example, if the characteristic attribute of the target category is speech skills, the target category is financially difficult, the characteristic attribute of the preset object feature is age, and the preset object feature is 18.

[0139] In step S420 of some embodiments, if the original feature is a continuous feature, the feature value range and discrete interval of the original feature are obtained, and the feature value range of the original feature is discretized according to the discrete interval to obtain discrete sub-features, and the feature value range of the discrete sub-features is within the feature value range of the original feature. For example, the original feature is age, and its feature value range is 0-100, and the discrete interval is 20. The original feature can be discretized into 5 discrete sub-features according to the discrete interval, and each discrete sub-feature has a different value range, which is divided into 0-20, 20-40, 40-60, 60-80, and 80-100. If the original feature is a discrete feature, there is no need to discretize the original feature, and the node is directly constructed according to the original feature to obtain a decision node.

[0140] In some embodiments, in step S430, to evaluate the feature importance of the discrete sub-feature, a feature gain of the discrete sub-feature is calculated based on the policy renewal category. A larger feature gain indicates a more important role of the discrete sub-feature in predicting policy renewal behavior.

[0141] In step S440 of some embodiments, the discrete sub-feature with the largest feature gain is selected as the reference sub-feature to select features that play an important role in the model prediction performance, thereby improving the accuracy of renewal behavior prediction.

[0142] In step S450 of some embodiments, a tree node is created according to the reference sub-feature to obtain a decision node.

[0143] In step S460 of some embodiments, steps S410 to S450 are repeatedly performed to construct child nodes of the decision node until an iteration stopping condition is met, thereby obtaining a first decision tree. The iteration stopping condition may be that the depth of the decision tree reaches a preset depth threshold, all samples in the decision node belong to the same category, or the number of samples in the decision node is less than a preset minimum number of samples.

[0144] In the above steps S410 to S460, a decision tree is constructed to capture the complex nonlinear dependencies between features and the correlation between features and renewal behavior, so as to accurately predict renewal behavior.

[0145] See also Figure 5 In some embodiments, step S430 may include but is not limited to steps S510 to S540:

[0146] Step S510, determining a reference renewal category according to the target renewal category;

[0147] Step S520: Screen the renewal categories of the object according to the discrete sub-features to obtain candidate renewal categories and the number of candidate renewal categories;

[0148] Step S530, determining an original sample gradient based on the reference renewal category, the candidate renewal category, and the number of candidate renewal categories;

[0149] Step S540: Calculate the feature gain of the discrete sub-feature according to the original sample gradient.

[0150] In step S510 of some embodiments, referring to step S340, the object renewal values ​​of the object renewal categories of all samples are averaged to obtain a renewal average, and a reference renewal category is determined based on the renewal average.

[0151] In step S520 of some embodiments, the target category, the object renewal category, and the preset object features are used as the original samples. To speed up the decision tree construction, the original samples are filtered based on the discrete sub-features to obtain the target samples. The object renewal category of the target samples is used as the candidate renewal category, and the number of target samples is used as the number of candidate renewal categories. For example, if the discrete sub-feature is age 20-40, the original samples meeting the age range of 20-40 are filtered based on the discrete sub-features to obtain the target samples.

[0152] In step S530 of some embodiments, a difference is calculated based on the reference renewal category and the candidate renewal category to obtain a sample difference value. The sample difference value and the number of candidate renewal categories are used as the original sample gradient.

[0153] In step S540 of some embodiments, the feature gain of the discrete sub-feature is calculated based on the sample difference value and the number of candidate renewal categories to evaluate the feature importance of each discrete sub-feature.

[0154] Through the above steps S510 to S540, the feature importance of different features can be evaluated, so that features that are important to the model prediction performance are selected based on the feature importance, thereby improving the accuracy of renewal behavior prediction.

[0155] See also Figure 6 In some embodiments, step S540 may include but is not limited to steps S610 to S640:

[0156] Step S610, performing feature splitting on the discrete sub-feature to obtain a first sub-feature and a second sub-feature;

[0157] Step S620, calculating a first sub-gain of the first sub-feature according to the original sample gradient;

[0158] Step S630, calculating a second sub-gain of the second sub-feature according to the original sample gradient;

[0159] Step S640 : performing gain summation on the first sub-gain and the second sub-gain to obtain a feature gain of the discrete sub-feature.

[0160] In step S610 of some embodiments, the target category, the object renewal category, and the preset object features are used as original samples, the original samples whose feature values ​​are less than or equal to the discrete sub-features are used as the first sub-features, and the original samples whose feature values ​​are greater than the discrete sub-features are used as the second sub-features.

[0161] In some embodiments, in step S620, a sample difference value and the number of candidate renewal categories are obtained from the original sample gradient based on the first sub-feature. The sample difference values ​​are summed to obtain a first total gradient, and the number of candidate renewal categories is summed to obtain a first total number. The first total gradient is squared, and the squared result is divided by the first total number to obtain a first sub-gain.

[0162] In some embodiments, in step S630, a sample difference value and the number of candidate renewal categories are obtained from the original sample gradient based on the second sub-feature. The sample difference values ​​are summed to obtain a second total gradient, and the number of candidate renewal categories is summed to obtain a second total number. The second total gradient is squared, and the squared result is divided by the second total number to obtain a second sub-gain.

[0163] In step S640 of some embodiments, the first sub-gain and the second sub-gain are added to obtain a feature gain of the discrete sub-feature.

[0164] Through the above steps S610 to S640, the feature gains of the discrete sub-features can be obtained, so as to screen the discrete sub-features based on the feature gains and construct a decision tree.

[0165] See also Figure 7 In some embodiments, step S330 may include but is not limited to steps S710 to S730:

[0166] Step S710, performing loss calculation based on the target renewal category and the predicted renewal category to obtain target loss data;

[0167] Step S720, determining first-order moment data and second-order moment data according to the target loss data;

[0168] Step S730: Calculate the renewal prediction weight of the first decision tree based on the first-order moment data and the second-order moment data.

[0169] In step S710 of some embodiments, based on a cross entropy loss function, a cross entropy loss is calculated for the target renewal category and the predicted renewal category to obtain target loss data.

[0170] In step S720 of some embodiments, first-order moment data and second-order moment data are calculated based on the target loss data. The first-order moment data is used to indicate the gradient direction, and the second-order moment data is used to indicate the gradient change rate. The calculation method of the first moment data is expressed as:

[0171] m1=β1*m01+(1-β1)*L,

[0172] m2=β2*m02+(1-β2)*L 2 ,

[0173] Among them, m1 represents the first-order moment data; m2 represents the second-order moment data; β1 and β2 are both attenuation coefficients; m01 and m02 represent the initial value of the first-order moment and the initial value of the second-order moment, respectively; L represents the target loss data.

[0174] In step S730 of some embodiments, the square root of the second-order moment data is obtained, and the first-order moment data is divided by the square root of the second-order moment data to obtain the renewal prediction weight of the first decision tree.

[0175] Through the above steps S710 to S730, the weights of the decision trees can be obtained, so as to balance the contributions of the decision trees in the renewal behavior prediction based on the weights of the decision trees and improve the generalization ability of the model.

[0176] In some embodiments, in step S180, after training the target renewal prediction model, a target renewal prompt dialogue sentence is obtained. The target renewal prompt dialogue sentence is the call recording text obtained during the model inference phase. The target renewal prediction model includes a first decision tree and a second decision tree. The first decision tree predicts renewal behavior based on the target renewal prompt dialogue sentence, determines the probability that the renewal subject will renew the premium, and obtains a first renewal probability. The second decision tree predicts renewal behavior based on the target renewal prompt dialogue sentence, determines the probability that the renewal subject will renew the premium, and obtains a second renewal probability. The renewal prediction weight of the first decision tree is multiplied by the first renewal probability. The tree weight of the second decision tree is obtained by referring to steps S710 to S730. The tree weight is multiplied by the second renewal probability, and the two multiplication results are added together to obtain the target renewal probability. If the target renewal probability is greater than a preset probability threshold, renewal is indicated. If the target renewal probability is less than or equal to the preset probability threshold, non-renewal is indicated.

[0177] See also Figure 8 The present application also provides a renewal behavior prediction device that can implement the above-mentioned renewal behavior prediction method. The renewal behavior prediction device includes:

[0178] A speech sentence acquisition module 810 acquires insurance speech sentences;

[0179] Clustering module 820, for clustering insurance speech sentences to obtain speech expression features and speech categories of speech expression features;

[0180] A dialogue statement acquisition module 830 is used to acquire sample renewal prompt dialogue statements;

[0181] Feature extraction module 840 is used to extract features from the sample renewal prompt dialogue sentences to obtain the renewal objection features of the object;

[0182] The original renewal behavior prediction module 850 is used to predict the renewal behavior of the target according to the renewal objection characteristics and obtain the target renewal category; the target renewal category is used to indicate whether to renew or not renew;

[0183] A screening module 860 is used to screen the dialogue categories based on the dialogue expression characteristics and the renewal objection characteristics of the target to obtain the target category;

[0184] A construction module 870 is configured to construct a model based on the target category, the target renewal category, and preset target features to obtain a target renewal prediction model; the preset target features are pre-acquired features of the preset renewal target;

[0185] The target renewal behavior prediction module 880 is used to predict the renewal behavior of the preset target renewal prompt dialogue sentence through the target renewal prediction model.

[0186] The specific implementation of the renewal behavior prediction device is basically the same as the specific embodiment of the renewal behavior prediction method described above, and will not be repeated here.

[0187] The present application also provides an electronic device comprising a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described renewal behavior prediction method. The electronic device can be any smart terminal, including a tablet computer and an in-vehicle computer.

[0188] See also Figure 9 , Figure 9 The hardware structure of an electronic device according to another embodiment is shown. The electronic device includes:

[0189] The processor 910 may be implemented using a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is configured to execute relevant programs to implement the technical solutions provided in the embodiments of the present application.

[0190] The memory 920 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 920 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 920 and is called by the processor 910 to execute the renewal behavior prediction method of the embodiments of this application.

[0191] Input / output interface 930, used to implement information input and output;

[0192] Communication interface 940, used to implement communication interaction between this device and other devices, which can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, Wi-Fi, Bluetooth, etc.);

[0193] bus 950 , which transmits information between various components of the device (e.g., processor 910 , memory 920 , input / output interface 930 , and communication interface 940 );

[0194] The processor 910 , the memory 920 , the input / output interface 930 , and the communication interface 940 are connected to each other in communication within the device via a bus 950 .

[0195] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned renewal behavior prediction method.

[0196] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0197] The renewal behavior prediction method, renewal behavior prediction device, electronic device, and computer storage medium provided in the embodiments of this application utilize a pre-constructed labeling system to extract customer objection speech. Labels with semantically similar speech to the customer objection speech are extracted from the labeling system to accurately identify renewal intent. These labels are then incorporated into the model as new features to improve the accuracy of the model's renewal behavior prediction.

[0198] The embodiments described in the embodiments of this application are intended to more clearly illustrate the technical solutions of the embodiments of this application and do not constitute a limitation on the technical solutions provided by the embodiments of this application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0199] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0200] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, i.e., they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0201] Those skilled in the art will appreciate that all or some of the steps in the methods, systems, and functional modules / units in the devices disclosed above may be implemented as software, firmware, hardware, or appropriate combinations thereof.

[0202] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0203] It should be understood that in this application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0204] In the several embodiments provided in this 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 schematic. For example, the division of the above-mentioned units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0205] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0206] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0207] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store programs.

[0208] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and essence of the present invention should be within the scope of the present invention.

Claims

1. A renewal behavior prediction method, characterized in that: The method comprises: Get insurance jargon; Clustering the insurance speech sentences to obtain speech expression features and speech categories of the speech expression features; Get sample renewal reminder dialogue sentences; Extract features from the sample renewal prompt dialogue sentences to obtain renewal objection features; Predicting renewal behavior based on the renewal objection characteristics of the object to obtain a renewal category of the object; the renewal category of the object is used to indicate whether to renew the policy or not; Filtering the speech categories according to the speech expression characteristics and the renewal objection characteristics of the target to obtain a target category; A target renewal prediction model is obtained by constructing a model based on the target category, the target renewal category, and preset target characteristics; the preset target characteristics are pre-acquired characteristics of the preset renewal target, and the preset target characteristics can affect the renewal behavior of the preset renewal target; Predicting renewal behavior for a preset target renewal prompt dialogue sentence using the target renewal prediction model; The model is constructed based on the target category, the object renewal category, and preset object features to obtain a target renewal prediction model, including: Constructing a decision tree based on the target category, the object renewal category, and the preset object features to obtain a first decision tree; Predicting renewal behavior for the target category and the preset object features using the first decision tree to obtain a predicted renewal category; Obtaining the renewal prediction weight of the first decision tree; Calculating a reference sample gradient according to the object renewal category, the predicted renewal category, and the renewal prediction weight; Constructing a decision tree based on the reference sample gradient, the target category, the object renewal category, and the preset object features to obtain a second decision tree; The target renewal prediction model is determined based on the first decision tree and the second decision tree.

2. The method according to claim 1, characterized in that The step of constructing a decision tree according to the target category, the object renewal category, and the preset object features to obtain a first decision tree includes: Performing feature screening on the target category and the preset object features to obtain original features; Discretizing the original features to obtain discrete sub-features; Calculating a feature gain of the discrete sub-feature according to the object renewal category; Selecting the discrete sub-feature with the largest feature gain as the reference sub-feature; Node construction is performed according to the reference sub-features to obtain a decision node; The first decision tree is constructed according to the decision nodes.

3. The method according to claim 2, characterized in that The calculating the feature gain of the discrete sub-feature according to the object renewal category includes: Determine the reference renewal category based on the renewal category of the object; screening the renewal categories of the object according to the discrete sub-features to obtain candidate renewal categories and the number of candidate renewal categories; determining an original sample gradient according to the reference renewal category, the candidate renewal category, and the number of categories of the candidate renewal category; The feature gain of the discrete sub-feature is calculated according to the original sample gradient.

4. The method according to claim 3, characterized in that The calculating the feature gain of the discrete sub-feature according to the original sample gradient includes: Performing feature splitting on the discrete sub-features to obtain a first sub-feature and a second sub-feature; Calculating a first sub-gain of the first sub-feature according to the original sample gradient; Calculating a second sub-gain of the second sub-feature according to the original sample gradient; The first sub-gain and the second sub-gain are summed to obtain a feature gain of the discrete sub-feature.

5. The method according to claim 1, wherein The obtaining of the renewal prediction weight of the first decision tree includes: Calculate the loss based on the target renewal category and the predicted renewal category to obtain target loss data; determining first-order moment data and second-order moment data according to the target loss data; The renewal prediction weight of the first decision tree is calculated based on the first-order moment data and the second-order moment data.

6. The method according to any one of claims 1 to 5, characterized in that Clustering the insurance speech sentences to obtain speech expression features and speech categories of the speech expression features includes: Performing sentence vectorization processing on the insurance speech phrase to obtain an insurance speech phrase vector; Calculating the distance between the insurance speech vector and a preset cluster center to obtain a speech distance; Clustering the insurance speech vectors according to the speech distance to obtain insurance speech clusters; Performing an average calculation on the insurance speech clusters to obtain the speech expression features; The speech category of the speech expression feature is determined based on the insurance speech cluster.

7. A renewal behavior prediction device, characterized in that: The device comprises: A speech sentence acquisition module, which acquires insurance speech sentences; A clustering module, configured to cluster the insurance speech sentences to obtain speech expression features and speech categories of the speech expression features; A dialogue statement acquisition module is used to obtain sample renewal prompt dialogue statements; A feature extraction module is used to extract features from the sample renewal prompt dialogue sentences to obtain the renewal objection features of the object; an original renewal behavior prediction module, configured to predict renewal behavior based on the renewal objection characteristics of the object, and obtain a renewal category of the object; the renewal category of the object is used to indicate whether to renew or not renew; A screening module, configured to screen the speech categories according to the speech expression characteristics and the renewal objection characteristics of the object, to obtain a target category; a construction module, configured to construct a model based on the target category, the object renewal category, and preset object features to obtain a target renewal prediction model; the preset object features are pre-acquired features of the preset renewal object, and the preset object features can affect the renewal behavior of the preset renewal object; A target renewal behavior prediction module is used to predict the renewal behavior of a preset target renewal prompt dialogue sentence using the target renewal prediction model; The device is also used for: A decision tree is constructed based on the target category, the object renewal category, and the preset object characteristics to obtain a first decision tree; renewal behavior is predicted for the target category and the preset object characteristics using the first decision tree to obtain a predicted renewal category; a renewal prediction weight of the first decision tree is obtained; a reference sample gradient is calculated based on the object renewal category, the predicted renewal category, and the renewal prediction weight; a decision tree is constructed based on the reference sample gradient, the target category, the object renewal category, and the preset object characteristics to obtain a second decision tree; and the target renewal prediction model is determined based on the first decision tree and the second decision tree.

8. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images