Customer behavior analysis method and system based on deep learning

Abstract

The invention relates to the technical field of deep learning, in particular to a customer behavior analysis method and system based on deep learning and a storage medium. Comprising the following steps: collecting and preprocessing first historical data of a customer, and obtaining action groups and behavior types through adjacent action time differences to form first data; desensitizing the first data to train a first model; acquiring second historical data in a preset time period before the starting time of each action group, extracting a feature vector, and taking the feature vector and the behavior type as training data to train a second model; and inputting the current behavior type of the customer into the second model to obtain a predicted behavior, comparing the predicted behavior with the real-time behavior, and updating the behavior type and the predicted behavior according to a result. The method utilizes deep learning to automatically extract features, adapts to dynamic change of data, protects privacy through desensitization processing, improves efficiency and accuracy of customer behavior analysis, and provides technical support for scenes such as precision marketing and risk early warning.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and particularly to a customer behavior analysis method, system and storage medium based on deep learning. Background Art

[0002] With the advent of the digital age, customer behavior analysis has become a key link in enterprise precision marketing, product optimization and service improvement. Traditional analysis methods are mostly based on statistical models, relying on manual rule setting and feature extraction, and it is difficult to cope with the massive, complex and dynamic nature of customer behavior data. In recent years, machine learning methods have gradually emerged, including decision trees and support vector machines in supervised learning, as well as clustering algorithms in unsupervised learning. Although there is a certain improvement in effect, there are still limitations, such as cumbersome feature engineering and insufficient model generalization ability. Against the backdrop of the booming development of deep learning technology, it has achieved breakthrough results in fields such as image recognition and natural language processing, demonstrating powerful automatic feature extraction and learning capabilities, providing new ideas for solving the problems of customer behavior analysis. Deep learning can automatically mine high-level features from a large amount of raw data, effectively capture the potential laws and complex patterns of customer behavior, and can adapt to the dynamic changes of data and update the analysis results in real time.

[0003] In the existing related technologies, such as the Chinese patent with the publication number CN114880575A, a user behavior prediction method, device, equipment, medium and product are proposed. By obtaining the historical behavior data pairs of users, future time and attribute information of candidate objects, a behavior sequence is determined, and then it is processed by a time-aware user interest extraction network model to obtain the predicted value of the user's execution of a preset behavior on the candidate object at a future time. There is also a US patent with the publication number US11498577B2, which discloses a behavior prediction device, including a moving object behavior detection unit, a behavior prediction model database, a behavior prediction calculation unit, a prediction deviation judgment unit, a deviation cause estimation unit and an update necessity judgment unit, etc., for detecting, predicting the behavior of moving objects and performing deviation processing and model update.

[0004] However, there are still some deficiencies in the existing technologies for customer behavior analysis: First, the data processing ability is limited, and it is difficult to efficiently process and analyze large-scale and multi-dimensional customer behavior data; second, the model adaptability is insufficient, and it cannot well adapt to the changes of different customer groups, business scenarios and market environments; third, the update of behavior types is not timely, and it is unable to quickly capture the dynamic changes of customer behavior types, resulting in lagging analysis results; fourth, the data security and privacy protection are insufficient, and it is easy to cause the leakage of customer sensitive information. Summary of the Invention

[0005] The present invention provides a customer behavior analysis method based on deep learning, which realizes efficient analysis and prediction of customer behavior through deep learning technology, and effectively solves the deficiencies of traditional methods in aspects such as data processing ability, model adaptability, and privacy protection. The method includes: Step S1: Collect the first historical data of customers, perform preprocessing to obtain first target data, obtain action groups and corresponding behavior types through the time difference between every two adjacent actions in the first target data, and obtain first data based on the action groups and the corresponding behavior types. Step S2: Desensitize the first data, and use the desensitized first data to train a first model. Step S3: Obtain the start time of each action group, collect the second historical data within a preset time period before the start time, perform preprocessing on the second historical data to obtain second target data, extract feature vectors related to customer behavior from the second target data, and use the feature vectors and the corresponding behavior types as training data to train a second model. Step S4: Input the current behavior type of the customer into the second model to obtain the predicted behavior of the customer, collect the real-time behavior of the customer, compare the real-time behavior with the predicted behavior, and update the behavior type and predicted behavior of the customer according to the comparison result.

[0006] As a preferred technical solution of the present invention, the obtaining of the action groups and the corresponding behavior types includes: Extract each minimum action unit from the first target data, use the minimum action unit as an action, obtain the target action in the minimum action unit, calculate the time difference between two adjacent actions before the target action, divide the two adjacent actions with a time difference less than a set threshold into the same action group, obtain the behavior type corresponding to the action group according to the target action, and also use the actions in the action group, the time difference between two adjacent actions, and the behavior type as the first data.

[0007] As a preferred technical solution of the present invention, desensitizing the first data includes: Standardize the format of the action groups in the first data to obtain standard action groups, and extract sensitive information from the standard action groups. Add noise to the sensitive information with a set noise intensity through a noise addition unit, obtain the desensitized information, compare the desensitized information with the value range of the sensitive information, and adjust the noise intensity according to the comparison result. When the comparison result is that the desensitized information is not within the value range corresponding to the sensitive information, reduce the noise intensity. On the contrary, when the comparison result is that the desensitized information is within the value range corresponding to the sensitive information, increase the noise intensity. Repeat this step until the comparison result satisfies that the desensitized information is within the value range corresponding to the sensitive information, and use the maximum noise intensity as the final noise intensity, where the noise is Gaussian noise.

[0008] As a preferred technical solution of the present invention, the training of the second model includes: Based on the start time of each action group in the first data, obtain a preset time period before the start time, collect the second historical data related to the target behavior type corresponding to the action group within the preset time period, preprocess the second historical data to obtain the second target data, extract the feature vectors related to the target behavior type from the second target data, and use the feature vectors and the target behavior type as training data to train the second model.

[0009] As a preferred technical solution of the present invention, obtaining the predicted behavior of the customer includes: Input the current behavior type of the customer into the first model to obtain the predicted behavior of the customer in the future time period. Also, compare the collected real-time behavior of the customer with the predicted behavior and obtain the comparison result. When the comparison result is that the difference value between the real-time behavior and the predicted behavior is less than or equal to the set difference value, the behavior type of the customer remains unchanged; When the comparison result is that the difference value between the real-time behavior and the predicted behavior is greater than the set difference value, collect the second data related to the behavior type of the customer within a preset time period before the current time, extract the feature vectors from the second data, input the feature vectors into the second model to obtain the output result, and obtain the real-time behavior type of the customer based on the output result. According to the real-time behavior type and the first model, re-obtain the predicted behavior of the customer in the future time period.

[0010] As a preferred technical solution of the present invention, when none of the customer's all behavior types are included in the output result, real-time actions of the customer are collected until a target behavior appears. A new action group corresponding to the target behavior and a corresponding new behavior type are obtained from all the collected real-time behaviors, and the first model is trained based on the new action group and the corresponding new behavior type. Third data related to the customer behavior type within a preset time period before the collected new action group is also used, and new feature vectors are extracted from the third data. The second model is trained based on the new feature vectors and the corresponding new behavior type.

[0011] As a preferred technical solution of the present invention, during the training process of the first model, the accuracy of the first model is also verified through verification data. When the accuracy is lower than the preset accuracy, the intensity of the noise in the first data is reduced, and the first model is retrained with the first data after noise adjustment.

[0012] As a preferred technical solution of the present invention, the customer's behavior types at least include the purchase behavior, selling behavior, additional behavior, and pre-purchase behavior of financial products.

[0013] The present invention also provides a customer behavior analysis system based on deep learning for implementing the above method. The system includes: A collection unit for collecting the first historical data of the customer and preprocessing the first historical data to obtain first target data; A grouping unit for obtaining an action group and a corresponding behavior type through the time difference between every two adjacent actions in the first target data, and obtaining first data based on the action group and the corresponding behavior type; A training unit for desensitizing the first data and training a first model with the desensitized first data. It is also used to obtain the start time of each action group, collect the second historical data within a preset time period before the start time, and preprocess it to obtain second target data. Feature vectors related to the customer behavior are extracted from the second target data, and the feature vectors and the behavior type corresponding to the action group are used as training data to train the second model; A prediction unit for inputting the current behavior type of the customer into the second model to obtain the predicted behavior of the customer, collecting the real-time behavior of the customer, comparing the real-time behavior and the predicted behavior, and updating the behavior type and predicted behavior of the customer according to the comparison.

[0014] The present invention also provides a computer-readable storage medium, on which instructions are stored. When the instructions are executed by a processor, the above method is implemented.

[0015] The beneficial effects of the present invention are as follows: The present invention realizes the efficient analysis and prediction of customer behavior through deep learning technology, effectively solving the deficiencies of traditional methods in aspects such as data processing capabilities, model adaptability, and privacy protection. First, the action grouping mechanism based on time difference can accurately identify customer behavior types. By extracting the time difference between adjacent actions and dividing action groups, and combining with target actions to determine behavior types, the model can automatically capture complex behavior patterns from massive data, avoiding the cumbersome process of traditional methods relying on manual feature engineering. The dual-model collaborative working mechanism significantly improves the real-time performance and accuracy of prediction: the first model is trained based on historical behavior data and is used to predict future behavior trends; the second model learns the data before and after the behavior type conversion to capture the dynamic incentives that trigger behavior changes. The combination of the two realizes the dynamic tracking and adaptive update of customer behavior. When the deviation between real-time behavior and the prediction result is large, the system quickly identifies the new behavior type through the second model and retrains the model to ensure that the prediction result always fits the actual changes. In terms of data security, the dynamic adjustment Gaussian noise desensitization technology is adopted. By comparing the value range of desensitized information with the original data, the noise intensity is adaptively optimized. While maximizing the protection of customer sensitive information, it ensures that the desensitized data can still effectively support model training, taking into account both privacy security and model performance. In addition, the system supports the automatic discovery of new behavior types and model updates. By collecting new behavior data in real time and generating new action groups, the recognition ability of the model is dynamically expanded, enhancing the adaptability of the system to diverse business scenarios and customer groups. Through the mutual cooperation of the above technical solutions, compared with the prior art, the present invention not only improves the processing efficiency of large-scale and high-dimensional customer behavior data, but also significantly improves the prediction accuracy and real-time performance through the dual-model linkage and dynamic update mechanism. At the same time, the data security is guaranteed through innovative desensitization methods, providing reliable technical support for financial service platforms in scenarios such as precision marketing and risk warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0017] Figure 1 It is a flowchart of the method for analyzing customer behavior based on deep learning in the embodiments of the present invention; Figure 2 It is a flowchart of the method for obtaining action groups and corresponding behavior types in the embodiments of the present invention; Figure 3It is a flowchart of a method for obtaining the predicted behavior of customers in an embodiment of the present invention; Figure 4 It is a structural diagram of a customer behavior analysis system based on deep learning in an embodiment of the present invention. Detailed implementation manners

[0018] Embodiments of the present invention provide a customer behavior analysis method, system and medium based on deep learning. Terms such as "first", "second", "third", "fourth", etc. (if any) in the description and claims of the present invention and the above drawings are used to distinguish similar objects and do not have to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described herein can be implemented in an order other than those illustrated or described herein. In addition, the term "comprising" or "having" and any variation thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0019] For ease of understanding, the specific process of the embodiments of the present invention is described below. As Figure 1 shown, an embodiment of a customer behavior analysis method based on deep learning in an embodiment of the present invention includes: Step S1: Collect the first historical data of the customer, preprocess the first historical data, obtain and obtain the action group and the corresponding behavior type through the time difference between every two adjacent actions in the first target data, and obtain the first data based on the action group and the corresponding behavior type; Specifically, by collecting the historical behavior data of the customer, such as actions of browsing, collecting, purchasing financial products, etc., through preprocessing operations such as data cleaning, denoising, and formatting, structured data is formed, the minimum action unit is extracted, such as a single click, browsing, and the time difference between adjacent actions is calculated. If the time difference between adjacent actions is less than a set threshold, such as 24 hours, it is classified into the same action group. For example, continuously browsing the same product multiple times, according to the target action within the group, such as the final purchase behavior, the behavior type is determined, and the action sequence, time difference, and behavior type within the action group are integrated into a structured data set, that is, the first data. Through the above technical solution, structured first data can be obtained, laying a foundation for subsequent model training, especially suitable for capturing the temporal correlation of behaviors.

[0020] Step S2: Perform desensitization processing on the first data, and train the first model with the desensitized first data; Specifically, by unifying the format of action groups, identifying sensitive fields such as transaction amount and account ID, and generating desensitized data under initial noise intensity, the value range of the desensitized data is compared with that of the original data. If the desensitized data exceeds the original range, the noise intensity is reduced; if it is within the range, the intensity is increased, and the process is iterated until the maximum available noise intensity is found to ensure that the desensitized data can protect privacy while retaining statistical characteristics, such as amount distribution. The desensitized first data is used to train deep learning models, such as LSTM and Transformer, to learn the timing pattern of customer behavior. The model input is the action sequence and time difference, and the output is the probability distribution of future behavior. Through the above technical solution, based on dynamic noise adjustment, while protecting sensitive information, the data validity is retained to the maximum extent. The first model can capture the long-term patterns of customer behavior, such as cyclical purchasing habits.

[0021] Step S3: obtaining the start time of each action group, and collecting preprocessed second target data within a preset time period before the start time, extracting feature vectors related to customer behavior from the second target data, and using the feature vectors and the behavior types corresponding to the action groups as training data to train the second model; Specifically, for the start time of each action group, extract the second historical data of its leading time period, such as industry news, price fluctuations, and account balance changes 24 hours before purchase. After preprocessing the second historical data, extract multidimensional feature vectors, such as price change rate and browsing frequency. Use the feature vectors and corresponding behavior types, such as from "browsing" to "buying", as training data to train the second model and learn the driving factors of behavior change. Through the above technical solution, identify changes in external environment or internal state based on the second model, for example, a sudden drop in price triggers a purchase, thereby enhancing the ability to explain sudden changes in behavior. The first model predicts trends, and the second model analyzes inducements. The dual models are linked to improve the robustness of the system.

[0022] Step S4: Input the customer's current behavior type into the second model, obtain the customer's predicted behavior, collect the customer's real-time behavior, compare the real-time behavior with the predicted behavior, and update the customer's behavior type and predicted behavior based on the comparison.

[0023] Specifically, the current behavior type is input into the second model, and the short-term predicted behavior is output. By collecting the actual behaviors of customers in real time and comparing them with the prediction results, the difference value is calculated. When the difference value is less than or equal to the set difference value, the current behavior type is maintained. When the difference value is greater than the set difference value, the second model is triggered to re-analyze the recent data to identify a new behavior type, such as "urgent sell", and update the above-mentioned first model and second model based on the new behavior type. Through the feedback loop, the above technical solution can quickly respond to behavior changes, reduce prediction lag. For example, it can capture sudden selling behaviors in a timely manner, obtain the new behavior types of customers in a timely manner, and update the above-mentioned first model and second model in a timely manner, improving the prediction accuracy of customer behaviors and laying a foundation for the marketing or promotion of financial products.

[0024] Further, the acquisition of the action groups and the corresponding behavior types, as Figure 2 shown, includes: Extracting each minimum action unit from the first target data, taking the minimum action unit as an action, obtaining the target action in the minimum action unit, calculating the time difference between two adjacent actions before the target action, dividing the two adjacent actions with the time difference less than the set threshold into the same action group, obtaining the behavior type corresponding to the action group according to the target action, and taking the actions in the action group, the time difference between two adjacent actions, and the behavior type as the first data.

[0025] Specifically, since there are various customer actions related to finance, but it is impossible to specifically determine which ones are related to the above target actions. In order to obtain the regular behavior types of customers, each of the above minimum action units, i.e., the minimum customer behavior units, is extracted from the above first target data, such as: browsing financial products, etc. The target actions in the above minimum action units are also obtained, such as: buying or selling actions of financial products. The time difference between any two adjacent actions before the above target action is calculated. For example: the first action, the second action, the third action, the target action, the third time difference between the target action and the third action, the second time difference between the third action and the second action, and the first time difference between the second action and the first action are calculated respectively. When the third time difference and the second time difference are less than the above set threshold, and the first time difference is greater than or equal to the above set threshold, the value of the above set threshold is 24 hours. The above second action, third action, and the above target action form an action group, and the above first action does not belong to the above action group. The behavior type to which the above action group belongs is also determined according to the above target action, such as: buying behavior or selling behavior. Each action in the above action group, the time difference between adjacent two actions, and the above behavior type are used as the above first data. Through the above technical solution, the behavior type corresponding to each above target action can be obtained, and a foundation is laid for training the first model based on the first data corresponding to the above behavior type.

[0026] Further, desensitizing the first data includes: Standardizing the format of the action groups in the first data to obtain standard action groups, and extracting sensitive information from the standard action groups; Adding noise to the sensitive information by a noise addition unit with a set noise intensity to obtain desensitized information, comparing the desensitized information with the value range of the sensitive information, and adjusting the noise intensity according to the comparison result. When the comparison result is that the desensitized information is not within the value range of the corresponding sensitive information, the noise intensity is reduced. On the contrary, when the comparison result is that the desensitized information is within the value range of the corresponding sensitive information, the noise intensity is increased. Repeat this step until the comparison result satisfies that the desensitized information is within the value range of the corresponding sensitive information, and take the maximum noise intensity as the final noise intensity, where the noise is Gaussian noise.

[0027] Specifically, since the actions of the customers on the financial service platform at least include the purchase information, selling information, browsing information, consulting information, etc. of financial products. Among them, the purchase information and selling information will inevitably involve account information and transaction amount, that is, sensitive information. If the first model is directly trained with the above first data, once it is obtained by network attackers, it may cause inestimable losses to the customers. Therefore, it is very necessary to desensitize the above first data. Also, since Gaussian noise has a good local desensitization effect, Gaussian noise is used to desensitize the above sensitive information. In order to facilitate setting the noise intensity, each of the above actions in the first data is grouped for format standardization, and the standard action groups are obtained. The sensitive information is extracted from the action groups according to the set sensitive information type, and the noise adding unit adds noise to the sensitive information with the set noise intensity, and the desensitized information corresponding to the sensitive information is obtained. The desensitized information is compared with the value range corresponding to the sensitive information. Since the data will change after adding noise to the sensitive information, but in order to enable the desensitized first data to still obtain an accurate first model through training, it is necessary to make the desensitized information still have a high similarity with the sensitive information although it has added noise. Therefore, the desensitized information is compared with the value range of the corresponding sensitive information. When the comparison result is that the desensitized information is not within the value range of the corresponding sensitive information, it means that the noise intensity is too large, resulting in a large difference between the desensitized information and the sensitive information. For example, if the customer's purchase information is negative, at this time, the noise intensity should be reduced. On the contrary, when the comparison result is that the desensitized information is within the value range of the corresponding sensitive information, it may be that the noise intensity is too low to fully protect the sensitive information. Therefore, it is necessary to increase the noise intensity, and the maximum noise intensity that can satisfy the desensitized information within the above value range is used as the final noise intensity. Through the above technical solution, the desensitized first data can not only meet the requirements of training a high-precision first model, but also maximize the protection of the customers' sensitive information from being leaked.

[0028] Further, the training of the second model includes: Based on the start time of each action group in the first data, a preset time period before the start time is obtained, the second historical data related to the target behavior type corresponding to the action group within the preset time period is collected, and the second historical data is preprocessed to obtain the second target data. Feature vectors related to the target behavior type are extracted from the second target data, and the feature vectors and the target behavior type are used as training data to train the second model.

[0029] Specifically, the behavior trend of the customer in the future time period can be accurately predicted through the customer's behavior type and the above-mentioned first model. However, with the changes in the external environment and information, the customer's behavior type will also change accordingly. Therefore, it is necessary to obtain a preset time period before the start time of the above-mentioned action group and collect the above-mentioned second historical data related to the change in the customer's behavior. The above-mentioned second historical data is the inducement for triggering the above-mentioned target behavior type, and the above-mentioned second historical data is also preprocessed. The above-mentioned second historical data at least includes the industry dynamics, price changes, account fund changes of the customer, and browsing times information of the financial product concerned. The above-mentioned preprocessing includes the elimination of redundant values and outliers to obtain the above-mentioned second target data. It is also necessary to extract the feature vectors related to the above-mentioned customer behavior type from the above-mentioned second target data, and use the above-mentioned feature vectors and the corresponding above-mentioned target behavior type as the above-mentioned training data, and train the above-mentioned second model based on the above-mentioned training data. Through the above technical solution, the above-mentioned second model with higher accuracy can be obtained, laying a foundation for accurately predicting the above-mentioned customer's financial behavior through the mutual cooperation of the above-mentioned second model and the above-mentioned first model.

[0030] Further, obtain the predicted behavior of the customer, as Figure 3 shown, including: Input the current behavior type of the customer into the first model to obtain the predicted behavior of the customer in the future time period. Also, compare the real-time behavior of the customer collected with the predicted behavior and obtain the comparison result. When the comparison result shows that the difference value between the real-time behavior and the predicted behavior is less than or equal to the set difference value, the behavior type of the customer remains unchanged; When the comparison result shows that the difference value between the real-time behavior and the predicted behavior is greater than the set difference value, collect the second data related to the behavior type of the customer within a preset time period before the current time, extract the feature vectors from the second data, input the feature vectors into the second model to obtain the output result, and obtain the real-time behavior type of the customer based on the output result. According to the real-time behavior type and the first model, re-obtain the predicted behavior of the customer in the future time period.

[0031] Specifically, input the current behavior type of the above-mentioned customer into the above-mentioned first model to obtain the predicted behavior of the above-mentioned customer within a future time period. Since the behavior type of the customer may change due to changes in the external environment and information, the real-time behavior of the above-mentioned customer is also obtained in real time and compared with the above-mentioned predicted behavior, and the comparison result is obtained. When the comparison result shows that the difference value between the two is small, that is, when the difference value is less than or equal to the set difference value, it is considered that the customer's behavior is consistent with the expectation, indicating that the customer's behavior type has not changed. When the difference value between the two is large, that is, when the difference value is greater than the set difference value, it is considered that the customer's behavior is inconsistent with the expectation, indicating that the customer's behavior type has changed. Therefore, by collecting the above-mentioned second data related to the customer's behavior type within a preset time period before the current time, and extracting the above-mentioned feature vector from the above-mentioned second data, that is, the feature vector that causes the change in the customer's behavior type, and inputting the above-mentioned feature vector into the above-mentioned second model, and obtaining the output result, and identifying the current real behavior type of the customer, that is, the above-mentioned real-time behavior type, based on the above-mentioned output result, and inputting the above-mentioned real-time behavior type into the above-mentioned first model to re-obtain the predicted behavior of the above-mentioned customer within a future preset time period, and repeating the method of this step to obtain the real-time predicted behavior of the above-mentioned customer. Through the above technical solution, the real-time predicted behavior of the above-mentioned customer can be obtained, and the above-mentioned predicted behavior can be updated in real time as the customer's behavior type changes.

[0032] Furthermore, when the output result does not include any one of all the behavior types of the customer, the real-time actions of the customer are collected in real time until the target behavior appears. The new action group corresponding to the target behavior and the corresponding new behavior type are obtained from all the collected real-time behaviors, and the first model is trained based on the new action group and the corresponding new behavior type. Also, based on the third data related to the customer's behavior type within a preset time period before the collected new action group, and extracting new feature vectors from the third data, the second model is trained based on the new feature vectors and the corresponding new behavior types.

[0033] Specifically, when any one of all the behavior types of the customer is not included in the above output result, the real-time behavior of the above customer may be a random behavior or a new behavior type. Therefore, by collecting the real-time behavior of the above customer in real time until the target behavior appears, and obtaining the above new action grouping through the time difference between two adjacent real-time behaviors in time, this method is the same as the method of dividing action groups in step S2. Add the above new action grouping and the corresponding new behavior type to the training data of the above first model, and train the above first model. Similarly, based on the above new feature vectors and new behavior types extracted from the above third data, add them to the training data of the above second model and retrain the above second model, so as to dynamically adapt to the behavior changes of customers. Through the above technical solutions, not only can the prediction accuracy of customer behavior be improved, but also the dynamic changes of customer behavior can be adapted, providing a basis for accurately providing corresponding financial product marketing strategies according to the predicted behavior of customers.

[0034] Further, during the training process of the first model, the accuracy of the first model is also verified through verification data. When the accuracy is lower than the preset accuracy, reduce the intensity of the noise in the first data, and retrain the first model with the first data after noise adjustment.

[0035] Specifically, since the noise added to the above action grouping may affect the training accuracy of the first model, when the accuracy of the above first model is low after training, that is, when the accuracy is lower than the above preset accuracy, reduce the noise intensity in the action grouping of the above first data, and retrain the above first model based on the first data after adjusting the noise. Through the above technical solutions, while protecting the sensitive information of customers, the accuracy of the first model can be ensured.

[0036] Further, the behavior types of the customer at least include the purchase behavior, sell behavior, additional purchase behavior, and pre-purchase behavior of financial products.

[0037] Specifically, the above additional purchase behavior is an additional purchase behavior for the already purchased financial product, and the above pre-purchase behavior is a reserved purchase behavior for the pre-sold financial product.

[0038] The present invention also provides a customer behavior analysis system based on deep learning for implementing the above method, as Figure 4 shown, the system includes: A collection unit for collecting the first historical data of the customer and preprocessing the first historical data to obtain first target data; A grouping unit, configured to obtain action groups and corresponding behavior types by the time differences between every two adjacent actions in the first target data, and obtain first data based on the action groups and the corresponding behavior types; A training unit, configured to desensitize the first data, train a first model with the desensitized first data, further configured to obtain the start time of each action group, collect second historical data within a preset time period before the start time, perform preprocessing to obtain second target data, extract feature vectors related to customer behaviors from the second target data, and use the feature vectors and the behavior types corresponding to the action groups as training data to train a second model; A prediction unit, configured to input the current behavior type of a customer into the second model to obtain a predicted behavior of the customer, collect the real-time behavior of the customer, compare the real-time behavior with the predicted behavior, and update the behavior type and predicted behavior of the customer according to the comparison.

[0039] The present invention further provides a computer-readable storage medium, on which instructions are stored, and when the instructions are executed by a processor, the above method is implemented.

[0040] In summary, the present invention realizes the efficient analysis and prediction of customer behavior through deep learning technology, effectively solving the deficiencies of traditional methods in aspects such as data processing capabilities, model adaptability, and privacy protection. First, the action grouping mechanism based on time difference can accurately identify customer behavior types. By extracting the time difference between adjacent actions and dividing action groups, and combining with the target action to determine the behavior type, the model can automatically capture complex behavior patterns from massive data, avoiding the cumbersome process of traditional methods relying on manual feature engineering. The dual-model collaborative working mechanism significantly improves the real-time performance and accuracy of prediction: the first model is trained based on historical behavior data and is used to predict future behavior trends; the second model learns from the data before and after the behavior type conversion to capture the dynamic incentives that trigger behavior changes. The combination of the two realizes the dynamic tracking and adaptive update of customer behavior. When the deviation between the real-time behavior and the prediction result is large, the system quickly identifies the new behavior type through the second model and retrains the model to ensure that the prediction result always conforms to the actual changes. In terms of data security, the Gaussian noise desensitization technology with dynamic adjustment is adopted. By comparing the value ranges of the desensitized information and the original data, the noise intensity is adaptively optimized. While maximizing the protection of customer sensitive information, it ensures that the desensitized data can still effectively support model training, taking into account both privacy security and model performance. In addition, the system supports the automatic discovery of new behavior types and model updates. By real-time collecting new behavior data and generating new action groups, the recognition ability of the model is dynamically expanded, enhancing the adaptability of the system to diverse business scenarios and customer groups. Through the mutual cooperation of the above technical solutions, compared with the prior art, the present invention not only improves the processing efficiency of large-scale and high-dimensional customer behavior data, but also significantly improves the prediction accuracy and real-time performance through the dual-model linkage and dynamic update mechanism. At the same time, the data security is guaranteed through the innovative desensitization method, providing reliable technical support for financial service platforms in scenarios such as precision marketing and risk warning.

[0041] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the above-described systems, systems, and units can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0042] When 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 invention, in essence, 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. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0043] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of various embodiments of the present invention.

Claims

1. A customer behavior analysis method based on deep learning, characterized in that: The method comprises: Step S1: collecting first historical data of the customer and performing preprocessing to obtain first target data, obtaining action groups and corresponding behavior types through the time difference between each two adjacent actions in the first target data, and obtaining first data based on the action groups and the corresponding behavior types; Step S2: desensitizing the first data, and training a first model with the desensitized first data; Step S3: obtaining the start time of each action group, and collecting second historical data within a preset time period before the start time, preprocessing the second historical data, obtaining second target data, extracting feature vectors related to customer behavior from the second target data, using the feature vectors and corresponding behavior types as training data, and training the second model; Step S4: Input the customer's current behavior type into the second model, obtain the customer's predicted behavior, collect the customer's real-time behavior, compare the real-time behavior with the predicted behavior, and update the customer's behavior type and predicted behavior based on the comparison result.

2. The method according to claim 1, characterized in that: The acquisition of the action group and the corresponding behavior type includes: Each minimum action unit is extracted from the first target data, and the minimum action unit is taken as the action, the target action in the minimum action unit is obtained, and the time difference between two adjacent actions before the target action is calculated, and the two adjacent actions whose time difference is less than a set threshold are divided into the same action group, and the behavior type corresponding to the action group is obtained according to the target action, and the actions in the action group, the time difference between two adjacent actions and the behavior type are also taken as the first data.

3. The method according to claim 1, characterized in that Performing desensitization processing on the first data includes: Standardizing the format of the action groups in the first data, obtaining standard action groups, and extracting sensitive information from the standard action groups; Noise is added to the sensitive information with a set noise intensity through a noise adding unit to obtain desensitized information, and the value ranges of the desensitized information and the sensitive information are compared, and the noise intensity is adjusted according to the comparison result. When the comparison result is that the desensitized information is not within the value range corresponding to the sensitive information, the noise intensity is reduced. Conversely, when the comparison result is that the desensitized information is within the value range corresponding to the sensitive information, the noise intensity is increased. This step is repeated until the comparison result satisfies that the desensitized information is within the value range corresponding to the sensitive information, and the maximum noise intensity is used as the final noise intensity, wherein the noise is Gaussian noise.

4. The method according to claim 1, characterized in that The training of the second model comprises: Based on the start time of each action group in the first data, a preset time period before the start time is obtained, the second historical data related to the target behavior type corresponding to the action group within the preset time period is collected, and the second historical data is preprocessed to obtain the second target data, and a feature vector related to the target behavior type is extracted from the second target data, and the feature vector and the target behavior type are used as training data to train the second model.

5. The method according to claim 1, characterized in that Obtain the predicted behavior of the customer, including: Inputting the current behavior type of the customer into the first model, obtaining the predicted behavior of the customer in a future time period, and comparing the collected real-time behavior of the customer with the predicted behavior to obtain a comparison result. When the comparison result shows that the difference between the real-time behavior and the predicted behavior is less than or equal to a set difference, the behavior type of the customer remains unchanged; When the comparison result is that the difference between the real-time behavior and the predicted behavior is greater than the set difference, second data related to the customer's behavior type in a preset time period before the current time is collected, and a feature vector is extracted from the second data, and the feature vector is input into the second model to obtain an output result, and the customer's real-time behavior type is obtained based on the output result, and the customer's predicted behavior in the future time period is re-obtained according to the real-time behavior type and the first model.

6. The method according to claim 5, characterized in that When the output result does not include any one of all the behavior types of the customer, the real-time actions of the customer are collected in real time until the target behavior appears, and new action groups and corresponding new behavior types corresponding to the target behavior are obtained from all the collected real-time behaviors, and the first model is trained based on the new action groups and the corresponding new behavior types. In addition, based on third data related to the customer's behavior type within a preset time period before the new action group is collected, a new feature vector is extracted from the third data, and the second model is trained based on the new feature vector and the corresponding new behavior type.

7. The method according to claim 1, characterized in that During the training process of the first model, the accuracy of the first model is also verified by verification data. When the accuracy is lower than the preset accuracy, the intensity of the noise in the first data is reduced and the first model is retrained using the noise-adjusted first data.

8. The method according to claim 1, characterized in that: The customer's behavior types include at least purchase behavior, sale behavior, additional behavior and pre-purchase behavior of financial products.

9. A customer behavior analysis system based on deep learning, used to implement the method according to any one of claims 1 to 8, characterized in that: The system comprises: A collecting unit, used for collecting first historical data of a customer, and preprocessing the first historical data to obtain first target data; A grouping unit, configured to obtain an action group and a corresponding behavior type according to a time difference between each two adjacent actions in the first target data, and obtain first data based on the action group and the corresponding behavior type; a training unit, configured to perform desensitization processing on the first data, and train the first model with the desensitized first data, and further configured to obtain a start time of each of the action groups, collect second historical data within a preset time period before the start time, and perform preprocessing to obtain second target data, extract a feature vector related to the customer behavior from the second target data, and use the feature vector and the behavior type corresponding to the action group as training data to train the second model; The prediction unit is used to input the customer's current behavior type into the second model, obtain the customer's predicted behavior, collect the customer's real-time behavior, compare the real-time behavior with the predicted behavior, and update the customer's behavior type and predicted behavior based on the comparison.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images