Customer recommendation method based on artificial intelligence

By constructing customer multi-type feature data graph templates and iterative samples, calculating the recommendation index correction coefficients, and directly outputting the financial product recommendation results, solving the problem of frequent iteration and update of machine learning models, improving recommendation efficiency and accuracy, and reducing costs.

CN119090645APending Publication Date: 2024-12-06HANGZHOU CREDIT INFORMATION CO LTD
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411580299.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-11-07
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Machine learning has frequent iterations and updates of models in financial product recommendations, which leads to high time-consuming and maintenance costs, and it is difficult to quickly identify the impact of customer sample changes on recommendation accuracy.

Method used

By constructing a customer multi-type feature data graph template, constructing iterative samples and calculating the recommended index correction coefficient, directly outputting the corrected recommended results, reducing the number of model iterations, and quickly identifying changes that affect model performance when the customer sample changes.

Benefits of technology

It improves the efficiency and accuracy of financial product recommendations, reduces the frequency of model iteration and updates, enhances model performance and sample diversity, and ensures recommendation accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119090645A_ABST
    Figure CN119090645A_ABST
Patent Text Reader

Abstract

The invention discloses a customer recommendation method based on artificial intelligence, and the method comprises the steps: constructing a customer multi-type feature data graph template and iterative samples, judging an invariant pixel site, capturing the contribution degree of feature elements to the prediction performance of a model, calculating a recommendation index correction coefficient of each iterative sample, and the like. According to the method, when a customer sample obtained at the # imgabs0 # moment and an iterative sample in a correction sample set have feature similarity, iterative updating does not need to be performed on a recommendation index prediction model at the # imgabs1 #, and a corresponding recommendation index predicted and output at the # imgabs2 # moment is corrected and then output by directly extracting a proper recommendation index correction coefficient; the frequency of model iteration updating is reduced, and high correction accuracy is ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and in particular to a customer recommendation method based on artificial intelligence. Background Art

[0002] There are various types of financial products, such as different types of fund products, stock products, bond products, and financial products. Each financial product is usually aimed at different customer purchasing groups. Different customer purchasing groups have different customer characteristics. In order to improve the accuracy of recommendations for different financial products, financial product recommendation methods based on machine learning have been widely used in recent years. Machine learning learns the customer characteristics of the customer purchasing group for each financial product, and establishes a mapping relationship between the group customer characteristics and the specified type of financial products. Compared with the traditional method of matching customer characteristics similarity, it has higher accuracy in recommending financial products.

[0003] However, financial product recommendations through machine learning have the following problems: Although machine learning can explore deeper customer characteristics and financial product characteristics, once the training samples change, in order to ensure the accuracy of the recommendation, the recommendation model usually needs to be iteratively updated and trained. When the model training algorithm is complex or the customer samples are prone to change at different times, each model iteration and update is time-consuming and complicated, and the maintenance cost of the model is high.

[0004] Therefore, the field expects: 1. When recommending financial products, it is possible to take advantage of the accuracy of customer recommendations provided by machine learning and quickly identify whether changes in customer samples have occurred that are sufficient to affect the accuracy of the model's recommendations, and decide whether to update the model. 2. When the changed customer samples are not sufficient to affect the product recommendation performance of the model, the financial product recommendation results for the changed customer samples are directly output without model prediction to improve prediction efficiency.

[0005] However, how to realize the above two expectations is a technical problem that needs to be solved urgently in this field. Summary of the invention

[0006] The present invention aims to reduce the number of model iterations and give full play to the advantages of machine learning in customer recommendation accuracy by directly outputting recommendation results of specified types of financial products to suspected potential customers with the help of the financial product recommendation results historically output by the recommendation model, thereby providing a customer recommendation method based on artificial intelligence.

[0007] To achieve this object, the present invention adopts the following technical solutions: Provided is an artificial intelligence-based customer recommendation method, comprising the steps of: S1, using A customer multi-type feature data graph template that is built or updated for a specified type of financial product at any time. The multi-type feature data of each suspected customer who is suspected of intending to purchase the specified type of financial product is preprocessed to obtain Iterative sample data set at time; S2, according to The iterative sample data set at time t is constructed Iteration samples of the model iteration are performed at each moment, and a recommendation index correction coefficient corresponding to each iteration sample is calculated; S3, according to Each iterative sample constructed at the moment, update The recommendation index prediction model trained at each moment is then updated to verify the correctness of the recommendation index correction coefficient associated with each iterative sample added to the sample set to be corrected, and the iterative samples that pass the verification are added to the In the calibration sample set bound at the moment; S4, yes Calculating feature similarity between the customer sample acquired at each moment and each of the iterative samples in the correction sample set, so as to extract a specified recommendation index correction coefficient associated with the specified iterative sample having the maximum feature similarity with the customer sample; S5, extracting the template sample that has been used to calculate the recommended index correction coefficient for the specified iteration sample. Recommendation index at the moment; S6, correcting the recommendation index with the designated recommendation index correction coefficient, and using the correction result as a recommendation result for recommending the designated type of financial product to the customer represented by the customer sample.

[0008] Preferably, in step S1, the specified type of financial product is constructed or updated. The method for generating a customer multi-type feature data graph template corresponding to a certain time comprises the steps of: A1, filter out Compare model updates at all times Always update the difference samples of the training samples used by the model; A2, converting each feature element in each type of feature array in each of the difference samples into a corresponding color feature, and obtaining the feature element in The first template color comprehensive feature in the first customer multi-type feature data graph template bound at the moment; A3, calculating the average value of the color feature of each feature element in each difference sample and the first template color comprehensive feature, as the average value of the color feature of each feature element in each difference sample The second template color comprehensive features in the second customer multi-type feature data graph template are updated at all times.

[0009] Preferably, in step S1, the method for preprocessing the multi-type feature data of each of the suspected potential customers comprises the steps of: B1, yes Each newly added sample at a certain moment constructs a corresponding multi-type feature data graph of potential customers. The construction method is: Convert each feature element in each type of feature array in the newly added sample into a corresponding color feature, and then assign it to a specified pixel position of each feature element in the preset empty image; B2, for each pixel site in the multi-type feature data graph of the intended customer belonging to the same type of feature array, determine whether it is an unchanged pixel site one by one, and calculate the ratio of the number of the unchanged pixel sites to the total number of assigned pixel sites in the feature array to which it belongs; B3: When it is determined that the ratio is less than a preset ratio threshold, the feature array is added as a construction unit to the iterative sample data set. Preferably, in step B2, the method for determining the unchanged pixel position comprises the steps of: B21, calculating the color feature differences between two pixel sites in the multi-type feature data map of the intended customer and the multi-type feature data map template belonging to the same type of the feature array and having the same designated position, and determining the pixel sites in the multi-type feature data map of the intended customer whose color feature differences are greater than a difference threshold as suspected changed pixel sites; B22, determining whether the neighboring site of the suspected changed pixel site is a suspected unchanged pixel site or an unchanged pixel site, If yes, the suspected changed pixel position is corrected to an unchanged pixel position; If not, the suspected changed pixel site is determined to be a changed pixel site.

[0010] Preferably, the setting positions of the pixel sites corresponding to the same characteristic element are the same in the multi-type characteristic data map of intended customers and the multi-type characteristic data map template of customers; the color feature difference is the absolute value of the difference in color features between the two pixel sites.

[0011] Preferably, the method of specifying the value assignment position in the client multi-type feature data graph template for each feature element belonging to the same feature array comprises the steps of: C1, from Extracting the feature elements belonging to the same feature array from the sample set acquired at the moment to form a unit sample set, extracting the remaining sample set as a background sample set; and defining the pixel window of the feature array in the image template according to the number of the feature elements in the unit sample set; C2, extracting one unit sample from the unit sample set one by one and the background sample set to form a corresponding valued sample set, and training each valued sample set with the same parameters to obtain a recommendation index prediction unit model corresponding to each unit sample; C3, evaluating the performance of each of the recommendation index prediction unit models, and sorting the corresponding unit samples from low to high according to performance; C4, taking the unit sample with the median performance value as the central pixel position of the pixel window, and specifying the corresponding pixel positions of the unit samples arranged in an outward expansion manner from left to right and from top to bottom relative to the central pixel position in the pixel window.

[0012] Preferably, the number of pixel sites in the pixel window is odd.

[0013] Preferably, in step S2, the method for constructing the iterative sample comprises the steps of: D1, specifying the corresponding template sample for each iterative feature array in the iterative sample data set; D2, replacing the template feature array of the same type in the template sample with the iterative feature array, to obtain the iterative sample in which the template sample is modified to have the iterative feature array.

[0014] Preferably, in step S2, the method for calculating the recommendation index correction coefficient corresponding to the iterative sample is: Determine whether the absolute value of the difference between the sample characteristic value deviation associated with the iterative sample and the recommendation index prediction deviation is less than a preset difference threshold, If yes, then taking the recommendation index prediction deviation corresponding to the iteration sample as the recommendation index correction coefficient corresponding to the iteration sample, and adding the iteration sample to the set of samples to be corrected; If not, it is determined that the calculation of the recommendation index correction coefficient for the iterative sample fails; The sample feature value deviation is: the array feature value deviation between the iteration feature array and the template sample array having a replacement relationship; The recommendation index prediction deviation is: The deviation between the intention recommendation index for the specified type of financial product predicted by the recommendation index prediction model trained at each moment for the iterative sample and the template recommendation index predicted by the same recommendation index prediction model for the template sample with a replacement relationship.

[0015] Preferably, in step S3, the method for verifying the correctness of the recommendation index correction coefficient associated with each of the iterative samples in the sample set to be corrected comprises the steps of: S31, the update model takes the iteration sample as input, predicts and outputs the first recommendation index corresponding to the iteration sample, and then obtains The recommendation index prediction model trained at the moment predicts a second recommendation index outputted by the template sample having a replacement relationship with the iteration sample; S32, calculating a first absolute value of a difference between the first recommendation index and the second recommendation index, and then determining whether a second absolute value of a difference between the first absolute value of the difference and the recommendation index correction coefficient associated with the iterative sample is less than a preset absolute value threshold of the difference, If so, it is determined that the correctness verification of the recommendation index correction coefficient has passed, and when the first recommendation index is greater than or equal to the second recommendation index, the recommendation index correction coefficient associated with the iterative sample is recorded as a positive value, otherwise it is recorded as a negative value; If not, the verification is determined to have failed.

[0016] The present invention has the following beneficial effects: 1. By Build or update customer multi-type feature data graph templates at all times, construct Iteration samples that iterate the model at all times, as well as calculation of the deviation between the iteration samples and the template samples, enable rapid identification Whether the customer samples obtained at any time are sufficient to affect the financial product recommendation performance of the model, and if it is judged that they are not sufficient to affect the model performance, The constantly updated model continues to iterate, reducing the number of model iterations; 2. By calculation The recommended index correction coefficient for each iterative sample constructed at the moment, and The recommended index corresponding to the template sample with the maximum feature similarity obtained at the moment is corrected by the iterative sample with the maximum feature similarity, and the correction result is directly used as the reference. The recommendation index of the specified type of financial products recommended to the specified customers by the customer samples obtained at all times realizes the direct output of the recommendation results of the specified type of financial products for the suspected potential customers, and improves the recommendation efficiency while ensuring the accuracy of the recommendation of the specified type of financial products.

[0017] 3. Through the construction of customer multi-type feature data graph templates and potential customer multi-type feature data graphs, the image segmentation method is used to achieve The iterative samples that iterate the model are constructed at all times; and through the judgment of the unchanged pixel positions, the constructed iterative samples retain the background characteristics of the customer's multi-type data graph template. Comparison Under the condition that the number of new samples at the moment is limited, the number of iterative samples is increased, which improves The diversity of samples used in model update training is beneficial to improving the performance of the updated model.

[0018] 4. Through steps C1-C3, a method for capturing the contribution of feature elements to the model prediction performance is provided, so that the neighboring sites assigned to the same feature array in the customer's multi-type feature data graph template have a close contribution to the model performance, and further determine whether the neighboring sites of the suspected changed pixel sites determined in step B21 are unchanged pixel sites. In this way, when correcting the unchanged pixel points, it is not easy to mistakenly correct the suspected changed pixel sites that have a large contribution to the model performance as unchanged pixel sites, thereby improving Compared with the iterative samples obtained at the moment The difference of training samples at each moment can improve The model iteration effect at all times ensures the accuracy of customer recommendations.

[0019] 5. The data richness in the iterative sample data set at the time can be controlled by setting the explicit standard of the neighboring sites in step B22. The higher the explicit standard of the neighboring sites, the worse the data richness. However, the data obtained by steps B1-B3 is The iterative sample data set at time The greater the difference in sample characteristics of the sample set at the moment, the greater the difference in sample characteristics of the sample set at the moment. The higher the performance of the recommendation index prediction model, the higher the training performance of the iterative sample data set obtained at any time. The lower the clear standard of the adjacent sites, the lower the performance of the recommendation index prediction model. The data richness of the iterative sample data set at the moment, which is When the sample data of the sample set obtained at any time is insufficient, it will further affect The performance of the model is updated all the time, so a balance can be found between the data richness of the iterative sample dataset and ensuring the performance of the updated model by setting reasonable criteria for clarifying neighboring sites.

[0020] 6. When When the customer sample obtained at the moment has feature similarity with the iterative sample in the calibration sample set, there is no need to The recommendation index prediction model is iteratively updated at all times, and the appropriate recommendation index correction coefficient is directly extracted to The corresponding recommendation index output by the moment prediction can be output after correction, which reduces the frequency of model iteration and ensures a high correction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 is a diagram of implementation steps of a customer recommendation method based on artificial intelligence provided by an embodiment of the present invention; Figure 2 It is a technical roadmap of the artificial intelligence-based customer recommendation method provided by an embodiment of the present invention; Figure 3 This is an example diagram of a multi-type feature data diagram of potential customers; Figure 4 This is an example diagram of the customer multi-type feature data diagram template; Figure 5 This is an example image of a one-pixel window. DETAILED DESCRIPTION

[0023] The technical solution of the present invention is further described below with reference to the accompanying drawings and through specific implementation methods.

[0024] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on this patent; in order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0025] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "inner", "outer", etc. appear, the orientation or position relationship indicated is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as a limitation on this patent. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0026] In the description of the present invention, unless otherwise clearly specified and limited, if the term "connection" or the like appears to indicate the connection relationship between components, the term should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two components or the interaction relationship between two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0027] An embodiment of the present invention provides a customer recommendation method based on artificial intelligence, such as Figure 1 and Figure 2 As shown, the steps include: S1, using A customer multi-type feature data graph template that is built or updated for a specified type of financial product at any time. The multi-type feature data of each suspected customer who is suspected of intending to purchase the specified type of product is preprocessed to obtain Iterative sample data set at time; S2, according to Iterative sample data set at time, construct Iteration samples of the model iteration are performed at all times, and the recommendation index correction coefficient corresponding to each iteration sample is calculated; S3, according to Each iterative sample constructed at the moment is updated The recommendation index prediction model is trained at all times, and then the correctness of the recommendation index correction coefficient associated with each iterative sample is verified using the updated model, and each iterative sample that passes the verification is added to In the moment-bound correction sample set; S4, yes The customer samples acquired at each moment are subjected to feature similarity calculation with each iteration sample in the correction sample set, so as to extract the specified recommendation index correction coefficient associated with the specified iteration sample having the maximum feature similarity with the customer sample; S5, extracted in The template sample that has been used to calculate the recommended index correction coefficient with the specified iteration sample at the time Recommendation index at the moment; S6, correcting the recommended index with a specified recommended index correction coefficient, and using the correction result as a direct recommendation result for recommending a specified type of financial product to the customer represented by the customer sample.

[0028] The following describes in detail how to implement each of the above steps S1-S6.

[0029] In step S1, the designated type of financial products includes fund products, stock products, financial products, bond products, etc.

[0030] Indicates The moment to build or update the customer multi-type feature data graph template corresponding to the specified type of financial product. express The last moment; Time display The next moment of the moment.

[0031] Customers’ multi-type characteristic data include their historical transaction data, historical consumption data, etc. The historical transaction data includes the customers’ historical transaction data on stocks, funds, bonds, and wealth management in financial institutions; the historical consumption data includes the customers’ historical purchase data on daily necessities and services.

[0032] In step S1, a financial product of a specified type is constructed or updated. The method for generating a customer multi-type feature data graph template corresponding to a certain moment specifically comprises the following steps: A1, filter out Compare model updates at all times Always update the difference samples of the training samples used by the model; For example, suppose The training samples used to update the model at any time include e1, e2, and e3. The training samples to be used for model updating at all times include e1, e2, e3, e4, and e5. Samples e4 and e5 are the difference samples selected in step A1.

[0033] A2, convert each feature element in each type of feature array in each difference sample into the corresponding color feature, and obtain the feature element in The first template color comprehensive feature in the first customer multi-type feature data graph template bound at the moment; The customer's historical transaction data corresponds to the first type of feature array, and the customer's historical consumption data corresponds to the second type of feature array. The customer's historical stock transaction data in the financial institution in the customer's historical transaction data is one of the feature elements in the first type of feature array; the customer's historical fund transaction data in the financial institution in the customer's historical transaction data is another feature element in the first type of feature array.

[0034] The same feature element in different training samples has the same or different first template color comprehensive feature assignment method in the customer multi-type feature data graph template. For example, for the feature element of stock historical transaction data, the template color comprehensive feature assignment method is: Assume that At this moment, there are three samples, e1, e2, and e3, as the basis for constructing or updating the first customer multi-type feature data graph template. The number of stock historical transactions recorded in the stock historical transaction data in e1, e2, and e3 is, for example, 3, 5, and 8 transactions respectively. Assuming that the initial color value given to the feature element of stock historical transaction data is "003", and each time a historical transaction is added, the color value is deepened, then the color feature of the feature element of the stock historical transaction data of the e1 sample in the first multi-type feature data graph template can be defined as "003-3", the e2 sample is "003-5", and the e3 sample is "003-8". "003-3", "003-5", and "003-8" have corresponding color values ​​respectively. The first template color comprehensive feature of the feature element of the stock historical transaction data in the first customer multi-type feature data graph template is the average value of the color values ​​corresponding to "003-3", "003-5", and "003-8" respectively.

[0035] A3, calculate the average value of the color feature of each feature element in each difference sample and the first template color comprehensive feature, as the average value of the feature element in each difference sample The second template color comprehensive features in the second customer multi-type feature data graph template are updated at all times.

[0036] Taking the characteristic element of stock historical transaction data as an example, assuming that the color features of difference samples e4 and e5 are "003-6" and "003-11" respectively, the second template color comprehensive feature of this characteristic element in the second customer multi-type feature data graph template is the average of "003-6", "003-11" and the first template color comprehensive feature.

[0037] In step S1, The multi-type feature data of each suspected customer who is suspected of intending to purchase the specified type of financial product is preprocessed to obtain Iterative sample data set at time; There are many ways to obtain potential customers at any time. For example, financial products A and B have product similarity. Customers a, b, and c have purchased financial product A in the past. When analyzing suspected potential customers of financial product B at any time, customers a, b, and c can be regarded as suspected potential customers who purchase financial product B.

[0038] The method for preprocessing multi-type feature data of suspected potential customers includes the following steps: B1, yes Each newly added sample at a certain moment constructs a corresponding multi-type feature data graph of potential customers. The construction method is: Convert each feature element in each feature array of the newly added samples into a corresponding color feature, and then assign it to the specified pixel position of each feature element in the preset empty image; The following combination Figure 3 The following is a detailed description of how to construct a multi-type feature data graph of potential customers: Assumptions There are five customer samples at the moment: e1, e2, e3, e4, and e5. There are three customer samples at the moment: e1, e2, and e3. Then e4 and e5 are Compare to time Each new sample carries several types of feature arrays, including the customer's historical transaction data and customer's historical consumption data. The feature array also includes the customer's historical stock transaction data and bond transaction data in financial institutions. Each feature element has a corresponding color feature assignment method. At this time, the customer feature carried by each feature element in the new sample is converted into a corresponding color feature, and then assigned to each feature element in Figure 3 At the specified pixel location in the example preset empty image, the assignment process here is like the process of coloring the customer features carried by the feature elements on the preset empty image.

[0039] For example, suppose the feature array of "customer historical transaction data" carried by the newly added sample e4 includes 12 feature elements. The system generates a feature array of "customer historical transaction data" in the following order: Figure 3 The designated coloring area on the preset empty image shown is, for example, area 100. The coloring position of the feature element "historical stock trading data of the customer in the financial institution" in the feature array in area 100 is designated as pixel position 1, and the color feature of the feature element "historical stock trading data of the customer in the financial institution" of the newly added sample e4 is assigned to pixel position 1.

[0040] B2, for each pixel site in the multi-type feature data graph of the intended customer that belongs to the same type of feature array, determine whether it is an unchanged pixel site one by one, and calculate the ratio of the number of unchanged pixel sites to the total number of assigned pixel sites in the feature array to which it belongs; For example, suppose Figure 3 The region 100 in the preset empty image of the example includes 9 pixel sites, each of which refers to a corresponding feature element. However, after adding the customer feature of the newly added sample e4 to the preset empty image, it is found that only the pixel sites 1-7 in the region 100 are assigned values, so the total number of assigned pixel sites in the region 100 is 7 instead of 9.

[0041] In this embodiment, The newly added samples need to be used to determine the unchanged pixel locations. Therefore, in order to facilitate the machine to quickly identify the image position of the basis area for the judgment of the unchanged pixel position from the template, preferably, the setting position of the pixel position corresponding to the same feature element is the same in the intended customer multi-type feature data map and the customer multi-type feature data map template. For example, the size of the intended customer multi-type feature data map and the customer multi-type feature data map template is exactly the same, and area 100 is Figure 3 The setting position in the multi-type feature data map of potential customers shown in the figure is exactly the same as the setting position in the multi-type feature data map template of customers. The same feature element, such as "historical stock trading data of customers in financial institutions", has the same designated position of the pixel site in the multi-type feature data map of potential customers and the multi-type feature data map template of customers, for example, both are designated as Figure 3 Pixel location 1 shown in .

[0042] The method for determining the unchanged pixel position specifically includes the following steps: B21, calculating the color feature differences between two pixel sites in the potential customer multi-type feature data map and the customer multi-type feature data map template that belong to the same type feature array and have the same designated position, and determining the pixel sites in the potential customer multi-type feature data map whose color feature differences are greater than the difference threshold as suspected changed pixel sites; For example, Figure 4 The feature array belonging to area 100' in the example customer multi-type feature data map template is Figure 3 The type of feature array to which region 100 in the multi-type feature data map of potential customers in the example is assigned is the same, both belonging to the feature array of “customer historical transaction data”, and the setting position and area range of region 100 in the multi-type feature data map of potential customers are the same as the setting position and area range of region 100' in the customer multi-type feature data map template. And the customer multi-type feature data map template and the multi-type feature data map of potential customers have the same size. The feature element corresponding to pixel position 1' in region 100' is, for example, “customer’s historical stock transaction record in a financial institution”, and its setting position in region 100' is the same as that in the multi-type feature data map template. Figure 3 The pixel site 1 in the region 100 is set at the same designated position, and the feature element corresponding to the pixel site 1 is also "the customer's stock transaction history record in the financial institution". Therefore, the pixel site 1' and the pixel site 1 meet the conditions of belonging to the same type of feature array and having the same designated position, and thus the color feature difference (the color feature of the two pixel sites (the color value assigned, where the color feature of the pixel site 1' is the color value assigned to the pixel site 1') is calculated for the pixel site 1' and the pixel site 1. The comprehensive template color feature of each sample obtained at the moment is the mean; and the color feature of pixel position 1 is the color feature of the newly added sample assigned at pixel position 1, which is the absolute value of the difference between the single value).

[0043] Assuming that the color feature difference between the pixel site 1 and the pixel site 1' is greater than a preset difference threshold, the pixel site 1 is determined as a suspected changed pixel site.

[0044] It should be noted here that the data of customer characteristics (feature elements) of users who purchase different financial products are rich and diverse, and the degree to which each customer characteristic affects the customer's willingness to purchase a specified type of financial product is usually different. In this embodiment, this degree of influence is defined as the contribution of customer characteristics to the model recommendation performance. The greater the degree of influence, the greater the contribution of the customer characteristic. Therefore, in order to capture the contribution of different feature elements to model prediction, in this embodiment, a set of methods for capturing the contribution of feature elements to model prediction performance is provided through steps C1-C3, so that the neighboring sites belonging to the same feature array assigned in the customer multi-type feature data graph template have a close contribution to the model performance, and further judge whether the neighboring sites of the suspected changed pixel sites determined in step B21 are unchanged pixel sites. In this way, when correcting unchanged pixels, it is not easy to mistakenly correct the suspected changed pixel sites that have a large contribution to the model performance as unchanged pixel sites, thereby improving Compared with the iterative samples obtained at the moment The difference of training samples at each moment can improve The model iteration effect at all times ensures the accuracy of customer recommendations.

[0045] Specifically, the following steps C1-C3 are used to capture the contribution of each feature element belonging to the same feature array to the model prediction performance, and to specify the assignment position in the customer multi-type feature data graph template for each feature element: C1, from Extract the characteristic elements belonging to the same characteristic array from the sample set acquired at the moment to form a unit sample set, and extract the remaining sample set as the background sample set; and define the pixel window of the characteristic array in the image template according to the number of characteristic elements in the unit sample set; For example, Assume that we have obtained 3 customer samples, e1, e2, and e3, which constitute Corresponding sample set. When it is necessary to capture the contribution of each feature element in the "customer historical transaction data" to the model prediction performance, extract each feature element belonging to the feature array of "customer historical transaction data" from each customer sample in the sample set. For example, the feature array of "customer historical transaction data" in sample e1 includes the feature element of "customer's historical stock transaction data in financial institutions", and the feature array of "customer historical transaction data" in sample e2 includes the feature element of "customer's historical bond transaction data in financial institutions". Then add the two feature elements under e1 and e2 to the unit sample set. From After extracting these two characteristic elements from the sample set acquired at each moment, the remaining sample set is used as the background sample set.

[0046] An example of a method for defining the pixel window of the feature array in the image template according to the number of feature elements in the unit sample set is as follows: Assuming that the number of characteristic elements in the unit sample set is less than or equal to 9, the size of the corresponding pixel window is set to three rows and three columns with a total of 9 pixel sites. Assuming that the number of characteristic elements in the unit sample set is greater than 9 but less than or equal to 25, the size of the corresponding pixel window is set to five rows and five columns with a total of 25 pixel sites. Assuming that the number of characteristic elements in the unit sample set is greater than 25 but less than or equal to 49, the size of the corresponding pixel window is set to seven rows and seven columns with a total of 49 pixel sites, and so on. When the number of characteristic elements in the unit sample set is less than the number of pixel sites in the pixel window, the color value of the pixel site that is not assigned a color feature is "null value". When judging the unchanged pixel site, the system will automatically filter out these "null value" pixel sites.

[0047] C2, extract one unit sample from the unit sample set and the background sample set one by one to form a corresponding assignment sample set, and use the same parameters to train each assignment sample set to obtain the recommendation index prediction unit model corresponding to each unit sample; For example, the unit sample (feature element) of "the customer's historical stock trading data in a financial institution" in sample e1 in the above example and the background sample set constitute the corresponding e1 value assignment sample set; the unit sample of "the customer's historical bond trading data in a financial institution" in sample e2 and the same background sample set constitute the corresponding e2 value assignment sample set.

[0048] Then, the same training parameters are used to train each assigned sample set to obtain the e1 recommendation index prediction unit model corresponding to the unit sample of "historical stock trading data of customers in financial institutions" in sample e1, and the e2 recommendation index prediction unit model corresponding to the unit sample of "historical bond trading data of customers in financial institutions" in sample e2.

[0049] C3, evaluates the performance of each recommendation index prediction unit model and sorts the corresponding unit samples from low to high according to performance; For example, assuming that the e1 recommendation index prediction unit model has higher model performance than the e2 recommendation index prediction unit model, and the e3 recommendation index prediction unit model has higher model performance than the e1 recommendation index prediction unit model, then the unit sample e2 of the e2 recommendation index prediction unit model is trained. 单元 Sort first, train the e1 recommendation index prediction unit model unit sample e1 单元 Next in ranking, train the e3 recommendation index prediction unit model with unit sample e3 单元 Sort at the end.

[0050] C4, taking the unit sample with the median performance as the central pixel position of the pixel window, and specifying the corresponding pixel positions for each unit sample arranged in an outward expansion manner from left to right and from top to bottom relative to the central pixel position in the pixel window.

[0051] For example, the unit sample e1 with the median performance 单元 exist Figure 3 The assignment position in the pixel window (area 100) shown is designated as pixel position 5 (pixel position 5 is the center pixel position of the pixel window). Then, in the pixel window, the unit samples e2 are arranged in order from left to right and from top to bottom relative to the center pixel position. 单元 、e3 单元 Specify the corresponding pixel locations, i.e. unit sample e2 单元 The pixel locations specified are Figure 3 Pixel location 1 in region 100, unit sample e3 单元 The pixel locations specified are Figure 3 The pixel site 2 in the area 100 in the pixel window. The "external expansion method" described here means: assuming that there are 25 pixel sites in five rows and five columns in the pixel window, the pixel sites are numbered in order from left to right and from top to bottom, and the central pixel site is pixel site 13. Assuming that there are only 6 unit samples in the unit sample set, the pixel sites specified by the 6 unit samples in the pixel window are pixel sites 7-9 and 12-14 respectively. Assuming that there are 10 unit samples in the unit sample set, the pixel sites specified by the 10 unit samples in the pixel window are pixel sites 7-9, 12-14, 17-19, and 1 respectively.

[0052] The method for specifying the assignment position of each feature element in the multi-type feature data graph of potential customers is the same as the method for specifying the assignment position of the feature element in the customer multi-type feature data graph template, and will not be repeated here.

[0053] After identifying each suspected changed pixel position belonging to the same type of feature array through step B21, the determination method of the unchanged pixel position in step B2 proceeds to step: B22, determining whether the neighboring site of the suspected changed pixel site is a suspected unchanged pixel site or an unchanged pixel site, If so, the suspected changed pixel position is corrected to an unchanged pixel position; If not, the suspected changed pixel site is determined to be a changed pixel site.

[0054] For example, suppose Figure 5 The pixel site 7 in step B21 is determined as a suspected changed pixel site. In this embodiment, the neighboring sites refer to eight neighboring sites, namely, the top, bottom, left, right, upper left, lower left, upper right, and lower right of the pixel site. For example, Figure 5 In the figure, the neighboring pixels of pixel 13 are pixel 7-9, 12, 14, 17-19. The color features of the top, left, bottom left, and top right of the suspected changed pixel 7 are assigned "null value". Therefore, the final judgment of whether it is an unchanged pixel or a changed pixel is based on its neighboring pixel 1, 8, 12, and 13.

[0055] It should be noted here that The data richness in the iterative sample data set at the time can be controlled by setting the explicit standard of the neighboring sites in step B22. The higher the explicit standard of the neighboring sites, the worse the data richness. However, the data obtained by steps B1-B3 is The iterative sample data set at time The greater the difference in sample characteristics of the sample set at the moment, the greater the difference in sample characteristics of the sample set at the moment. The higher the performance of the recommendation index prediction model, the higher the training performance of the iterative sample data set obtained at any time. The lower the clear standard of the adjacent sites, the lower the performance of the recommendation index prediction model. The data richness of the iterative sample data set at the moment, which is When the sample data of the sample set obtained at any time is insufficient, it will further affect The performance of the model is updated all the time, so a balance can be found between the data richness of the iterative sample dataset and ensuring the performance of the updated model by setting reasonable criteria for clarifying neighboring sites.

[0056] In this embodiment, in step B22, preferably, when any one of the neighboring sites of the suspected changed pixel site is determined to be a suspected unchanged pixel site or an unchanged pixel site, the suspected changed pixel site is corrected to an unchanged pixel site. Alternatively, when all the neighboring sites of the suspected changed pixel site are determined to be suspected unchanged pixel sites or unchanged pixel sites, the suspected changed pixel site is corrected to an unchanged pixel site. Alternatively, when the neighboring sites of the suspected changed pixel site are determined to be suspected unchanged pixel sites or the ratio of the number of neighboring sites of the unchanged pixel site to the total number of neighboring sites is greater than a preset ratio threshold, the suspected changed site is corrected to an unchanged pixel site.

[0057] In step B2, the method for calculating the ratio of the number of unchanged pixel sites to the total number of assigned pixel sites in the attribute array is exemplified as follows: Assumptions Figure 5 In the figure, pixel positions 1, 7, and 8 are determined to be unchanged pixel positions. Figure 5 The total number of pixel sites assigned values ​​is 10, and the ratio calculated in step B2 is three tenths.

[0058] After step B2, in step S1, the method of preprocessing the multi-type feature data of each suspected potential customer proceeds to step: B3, when the ratio calculated in step B2 is determined to be less than a preset ratio threshold, the feature array is added as a construction unit to the iterative sample data set.

[0059] The implementation of step S2 is described in detail below: In step S2, according to The iterative sample data set at time t is constructed The method of iterating samples at each moment of model iteration comprises the steps of: D1, specifies the corresponding template sample for each iterative feature array in the iterative sample data set; For example, suppose you update The customer samples of the customer multi-type feature data graph template corresponding to the moment include template samples m1, m2, m3, and m4. The iteration sample data set includes iteration feature arrays n1, n2, and n3. Then, the template sample corresponding to iteration feature array n1 can be specified as m2, the template sample corresponding to n2 as m1, and the template sample corresponding to n3 as m4.

[0060] D2, replace the template feature array of the same type in the template sample with the iterative feature array, and obtain the template sample modified to have the iterative sample with the iterative feature array.

[0061] For example, the template feature array of the same type as the iteration feature array n3 in the template sample m4 is n3', then n3 is replaced with n3' in m4, and the template sample m4 is modified to be an iteration sample with the iteration feature array n3.

[0062] In step S2, the method for calculating the recommendation index correction coefficient corresponding to each iterative sample is: Determine whether the absolute value of the difference between the sample characteristic value deviation of the associated iterative sample and the recommendation index prediction deviation is less than the preset difference threshold, If yes, then taking the recommendation index prediction deviation corresponding to the iteration sample as the recommendation index correction coefficient corresponding to the iteration sample; If not, it is determined that the calculation of the recommended index correction coefficient for the iterative sample has failed.

[0063] In this embodiment, the sample feature value deviation of the associated iterative sample is: the array feature value deviation between the iterative feature array and the template feature array having a replacement relationship.

[0064] For example, in the above example, the iterative feature array n3 replaces the template feature array n3' in the template sample m4, so n3 and n3' have a replacement relationship.

[0065] The array feature deviation with a replacement relationship refers to: the weighted average of the absolute values ​​of the differences between the element eigenvalues ​​of the feature elements in n3 and the eigenvalues ​​of the same type of feature elements in n3'. For example, n3 includes three feature elements, x1, x2, and x3, and the corresponding element eigenvalues ​​are y1, y2, and y3, respectively. The eigenvalue of the feature element x1' of the same type as x1 in n3' is, for example, y1', the eigenvalue of the feature element x2' of the same type as x2 is, for example, y2', and the eigenvalue of the feature element x3' of the same type as x3 is, for example, y3'. Then the array feature deviation of n3 and n3' is the weighted average of the absolute values ​​of the first difference between y1 and y1', the absolute value of the second difference between y2 and y2', and the absolute value of the third difference between y3 and y3'.

[0066] The recommendation index prediction deviation with substitution relationship refers to: using The deviation between the intention recommendation index for a specified type of financial product output by the iterative sample prediction model trained at each moment and the template recommendation index output by the same recommendation index prediction model for a template sample with a replacement relationship (also taken as the absolute value of the difference). The implementation of step S3 is described in detail below: In step S3, the method for verifying the correctness of the recommendation index correction coefficient associated with each iterative sample in the sample set to be corrected includes the following steps: S31, update the model to take the iteration sample as input, predict the first recommendation index corresponding to the output iteration sample, and then obtain The recommendation index prediction model trained at the moment predicts a second recommendation index outputted by the template sample having a replacement relationship with the iteration sample; S32, calculating a first absolute value of a difference between the first recommendation index and the second recommendation index, and then determining whether a second absolute value of a difference between the first absolute value of the difference and a recommendation index correction coefficient associated with the iteration sample is less than a preset absolute value threshold of the difference, If yes, it is determined that the correctness verification of the recommendation index correction coefficient has passed, and when the first recommendation index is greater than or equal to the second recommendation index, the recommendation index correction coefficient associated with the iterative sample is recorded as a positive value, otherwise it is recorded as a negative value; If not, the verification is determined to have failed.

[0067] The implementation of step S4 is described in detail below: In step S4, There are many existing methods for calculating feature similarity between customer samples acquired at any time and each iterative sample in the correction sample set. Since the feature similarity calculation method is not within the scope of rights to be protected by this application, it will not be explained in detail.

[0068] The purpose of the feature similarity calculation in step S4 is to: extract the recommendation index correction coefficient (defined as the specified recommendation index correction coefficient) associated with the iteration sample (defined as the specified iteration sample) having the maximum feature similarity with the customer sample; It should be noted here that, through step S32, each recommendation index correction coefficient in the correction sample set has been assigned a positive value or a negative value.

[0069] The implementation of step S5 is described in detail below: In step S5, the template sample for which the recommended index correction coefficient has been calculated with the specified iteration sample is extracted. The recommendation index at the moment. That is: extract The recommendation index prediction model trained at the moment predicts the recommendation index output by the template sample having a replacement relationship with the specified iteration sample; The implementation of step S6 is described in detail below: In step S6, the recommended index extracted in step S5 is corrected by the correction coefficient of the specified recommended index extracted in step S4, and the correction method is to find the sum of the two. It should be noted here that the correction coefficient of the specified recommended index extracted in step S4 has been assigned a positive value or a negative value through steps S31-S32. Therefore, the sum of the two may be the sum of two positive numbers, or the difference between one positive and the other negative. The correction result is directly used as the correction coefficient of the recommended index extracted in step S4. The customer sample obtained at the moment refers to the recommendation result of the specified type of financial product. When the customer sample obtained at the moment has feature similarity with the iterative sample in the calibration sample set, there is no need to The recommendation index prediction model is iteratively updated again, and the appropriate recommendation index correction coefficient is directly extracted to The corresponding recommendation index output by the moment prediction can be output after correction, which reduces the frequency of model iteration and ensures a high correction accuracy.

[0070] It should be noted that the above specific implementations are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art should understand that various modifications, equivalent substitutions, changes, etc. can be made to the present invention. However, as long as these changes do not deviate from the spirit of the present invention, they should be within the scope of protection of the present invention. In addition, some terms used in the specification and claims of this application are not restrictive, but are only for the convenience of description.

Claims

1. A customer recommendation method based on artificial intelligence, characterized in that: Includes steps: S1, using A customer multi-type feature data graph template that is built or updated for a specified type of financial product at any time. The multi-type feature data of each suspected customer who is suspected of intending to purchase the specified type of financial product is preprocessed to obtain Iterative sample data set at time; S2, according to The iterative sample data set at time t is constructed Iteration samples of the model iteration are performed at each moment, and a recommendation index correction coefficient corresponding to each iteration sample is calculated; S3, according to Each iterative sample constructed at the moment, update The recommendation index prediction model trained at each moment is then updated to verify the correctness of the recommendation index correction coefficient associated with each iterative sample added to the sample set to be corrected, and the iterative samples that pass the verification are added to the In the calibration sample set bound at the moment; S4, yes Calculating feature similarity between the customer sample acquired at each moment and each of the iterative samples in the correction sample set, so as to extract a specified recommendation index correction coefficient associated with the specified iterative sample having the maximum feature similarity with the customer sample; S5, extracting the template sample that has been used to calculate the recommended index correction coefficient for the specified iteration sample. Recommendation index at the moment; S6, correcting the recommendation index with the designated recommendation index correction coefficient, and using the correction result as a recommendation result for recommending the designated type of financial product to the customer represented by the customer sample.

2. The artificial intelligence-based customer recommendation method according to claim 1, characterized in that: In step S1, the specified type of financial product is constructed or updated The method for obtaining the customer multi-type feature data graph template corresponding to each time comprises the steps of: A1, filter out Compare model updates at all times Always update the difference samples of the training samples used by the model; A2, converting each feature element in each type of feature array in each of the difference samples into a corresponding color feature, and obtaining the feature element in The first template color comprehensive feature in the first customer multi-type feature data graph template bound at the moment; A3, calculating the average value of the color feature of each feature element in each difference sample and the first template color comprehensive feature, as the average value of the color feature of each feature element in each difference sample The second template color comprehensive features in the second customer multi-type feature data graph template are updated at all times.

3. The customer recommendation method based on artificial intelligence according to claim 1, characterized in that: In step S1, the method for preprocessing the multi-type feature data of each of the suspected potential customers comprises the following steps: B1, yes Each newly added sample at a certain moment constructs a corresponding multi-type feature data graph of potential customers. The construction method is: Convert each feature element in each type of feature array in the newly added sample into a corresponding color feature, and then assign it to a specified pixel position of each feature element in the preset empty image; B2, for each pixel site in the multi-type feature data graph of the intended customer belonging to the same type of feature array, determine whether it is an unchanged pixel site one by one, and calculate the ratio of the number of the unchanged pixel sites to the total number of assigned pixel sites in the feature array to which it belongs; B3: When it is determined that the ratio is less than a preset ratio threshold, the feature array is added as a construction unit to the iterative sample data set.

4. The artificial intelligence-based customer recommendation method according to claim 3, characterized in that: In step B2, the method for determining the unchanged pixel position includes the steps of: B21, calculating the color feature differences between two pixel sites in the multi-type feature data map of the intended customer and the multi-type feature data map template belonging to the same type of the feature array and having the same designated position, and determining the pixel sites in the multi-type feature data map of the intended customer whose color feature differences are greater than a difference threshold as suspected changed pixel sites; B22, determining whether the neighboring site of the suspected changed pixel site is a suspected unchanged pixel site or an unchanged pixel site, If yes, the suspected changed pixel position is corrected to an unchanged pixel position; If not, the suspected changed pixel site is determined to be a changed pixel site.

5. The artificial intelligence-based customer recommendation method according to claim 4, characterized in that: The setting positions of the pixel sites corresponding to the same characteristic elements are the same in the multi-type characteristic data map of the intended customers and the multi-type characteristic data map template of the customers; the color feature difference is the absolute value of the difference between the color features of the two pixel sites.

6. The artificial intelligence-based customer recommendation method according to claim 3, characterized in that: The method for specifying a value assignment position in the client multi-type feature data graph template for each feature element belonging to the same feature array comprises the steps of: C1, from Extracting the feature elements belonging to the same feature array from the sample set acquired at the moment to form a unit sample set, extracting the remaining sample set as a background sample set; and defining the pixel window of the feature array in the image template according to the number of the feature elements in the unit sample set; C2, extracting one unit sample from the unit sample set one by one and the background sample set to form a corresponding valued sample set, and training each valued sample set with the same parameters to obtain a recommendation index prediction unit model corresponding to each unit sample; C3, evaluating the performance of each of the recommendation index prediction unit models, and sorting the corresponding unit samples from low to high according to performance; C4, taking the unit sample with the median performance value as the central pixel position of the pixel window, and specifying the corresponding pixel positions of the unit samples arranged in an outward expansion manner from left to right and from top to bottom relative to the central pixel position in the pixel window.

7. The artificial intelligence-based customer recommendation method according to claim 6, characterized in that: The number of pixel sites in the pixel window is odd.

8. The artificial intelligence-based customer recommendation method according to any one of claims 1 to 7, characterized in that: In step S2, the method for constructing the iterative sample includes the steps of: D1, specifying the corresponding template sample for each iterative feature array in the iterative sample data set; D2, replacing the template feature array of the same type in the template sample with the iterative feature array, to obtain the iterative sample in which the template sample is modified to have the iterative feature array.

9. The artificial intelligence-based customer recommendation method according to claim 8, characterized in that: In step S2, the calculation method of the recommendation index correction coefficient corresponding to the iterative sample is: Determine whether the absolute value of the difference between the sample characteristic value deviation associated with the iterative sample and the recommendation index prediction deviation is less than a preset difference threshold, If yes, then taking the recommendation index prediction deviation corresponding to the iteration sample as the recommendation index correction coefficient corresponding to the iteration sample, and adding the iteration sample to the set of samples to be corrected; If not, it is determined that the calculation of the recommendation index correction coefficient for the iterative sample fails; The sample feature value deviation is: the array feature value deviation between the iteration feature array and the template sample array having a replacement relationship; The recommendation index prediction deviation is: The deviation between the intention recommendation index for the specified type of financial product predicted by the recommendation index prediction model trained at each moment for the iterative sample and the template recommendation index predicted by the same recommendation index prediction model for the template sample with a replacement relationship.

10. The artificial intelligence-based customer recommendation method according to claim 8, characterized in that: In step S3, the method for verifying the correctness of the recommendation index correction coefficient associated with each of the iterative samples in the sample set to be corrected comprises the following steps: S31, the update model takes the iteration sample as input, predicts and outputs the first recommendation index corresponding to the iteration sample, and then obtains The recommendation index prediction model trained at the moment predicts a second recommendation index outputted for the template sample having a replacement relationship with the iteration sample; S32, calculating a first absolute value of a difference between the first recommendation index and the second recommendation index, and then determining whether a second absolute value of a difference between the first absolute value of the difference and the recommendation index correction coefficient associated with the iterative sample is less than a preset absolute value threshold of the difference, If so, it is determined that the correctness verification of the recommendation index correction coefficient has passed, and when the first recommendation index is greater than or equal to the second recommendation index, the recommendation index correction coefficient associated with the iterative sample is recorded as a positive value, otherwise it is recorded as a negative value; If not, the verification is determined to have failed.

Citation Information

Patent Citations

  • User prediction interpretation Treecap method based on financial product recommendation scene

    CN111882441A

  • Product recommendation method and device, equipment and storage medium

    CN114581177A

  • Message recommendation method, electronic equipment, storage medium and program product

    CN115659037A

  • Customer data processing method and device based on artificial intelligence and computer equipment

    CN116883048A

  • Financial product recommendation method, system, equipment and medium

    CN118505348A