Business recommendation method, device and system based on big data and customer portrait

Abstract

The present invention discloses a business recommendation method, device and system based on big data and customer portraits. The business recommendation method includes: collecting and processing static data and / or dynamic data of multiple customers to obtain standardized and labeled customer portraits; performing data analysis and mining on the standardized and labeled customer portraits through an algorithm based on a gradient boosting tree to exclude features that lack behavioral regularities and then capture cross-features; formulating a business recommendation strategy based on an explainable algorithm based on the cross-features, including introducing a scenario adaptive mechanism into the algorithm-based business recommendation strategy, wherein the scenario adaptive mechanism learns from the data the importance of each cross-feature in different scenarios; and pushing the business recommendation strategy to the corresponding customers. The present invention combines the algorithm based on the gradient boosting tree and the adaptive mechanism to improve the business recommendation performance while ensuring interpretability.

Description

Technical Field

[0001] The present invention relates to the field of big data and artificial intelligence, and in particular to a business recommendation method, device and system based on big data and customer portraits. Background Art

[0002] In the past, bank marketing mainly relied on traditional physical channels. These traditional marketing methods include:

[0003] Physical branch marketing: relying on a wide range of physical branches, through tellers and sales managers to communicate with customers face to face, provide financial products and services, display promotional materials in the branches, attract customers to pay attention to and understand the bank's products;

[0004] Traditional media advertising and marketing: Use newspapers, magazines, television, radio and other traditional media to place advertisements to expand brand awareness and influence;

[0005] Promotional marketing: regularly hold various promotional activities, such as gifts for account opening and points for deposits, to attract new customers and retain old customers. Through promotional activities, we strengthen interaction with customers and improve customer satisfaction and loyalty;

[0006] Relationship Marketing: By building long-term relationships with customers, understanding their needs and providing customized services, sales managers regularly stay in touch with customers to provide financial advice and solutions to maintain customer relationships and promote business growth.

[0007] Although traditional marketing methods have a wide coverage, they are costly and difficult to accurately measure marketing results. In addition, the interaction and feedback with customers are relatively limited.

[0008] With the rapid development of the Internet and digital technology, banks' marketing methods have also undergone tremendous changes. Modern digital marketing mainly includes the following aspects: Online banking and mobile banking marketing: By developing convenient online banking and mobile banking applications, 24-hour online services are provided to meet customers' financial needs anytime and anywhere; Social media marketing: Using social media platforms (such as WeChat, Weibo, etc.) for brand promotion, product promotion and customer service. Through real-time interaction with customers, banks can more effectively understand customer needs and provide customized services; Content marketing: By writing blog articles, making video tutorials, etc., valuable financial knowledge and information are provided to attract and retain potential customers; Search engine optimization and paid advertising marketing: By optimizing website content and structure, the ranking of bank websites in search engines is improved, and keyword ads are placed to attract potential customers.

[0009] However, current digital marketing generally provides potential customers with a wealth of digital information, requiring customers to independently screen the services they need. Once customers do not have the energy to take the initiative, marketing cannot be achieved.

[0010] The disclosure of the above background technology content is only used to assist in understanding the inventive concept and technical solution of the present invention. It does not necessarily belong to the prior art of the present invention application, nor does it necessarily provide technical guidance. In the absence of clear evidence that the above content has been disclosed before the filing date of the present invention application, the above background technology should not be used to evaluate the novelty and creativity of the present application. Summary of the invention

[0011] The purpose of the present invention is to provide a business recommendation method based on big data and customer portraits, combining an algorithm based on a gradient boosting tree and an adaptive mechanism to improve the business recommendation performance while ensuring interpretability.

[0012] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0013] A business recommendation method based on big data and customer portraits, comprising the following steps:

[0014] Collect and process static data and / or dynamic data of multiple customers to obtain standardized and labeled customer portraits, where the customer portraits include multiple data items;

[0015] Performing data analysis and mining on the standardized and labeled customer portraits through a gradient boosting tree-based algorithm to exclude features that lack behavioral regularities and thereby capture cross-features;

[0016] Formulate a business recommendation strategy based on an explainable algorithm according to the cross-features, including introducing a scenario adaptation mechanism into the algorithm-based business recommendation strategy, wherein the scenario adaptation mechanism learns the importance of each cross-feature in different scenarios from the data;

[0017] The business recommendation strategy is pushed to the corresponding customers.

[0018] Further, based on any one of the above-mentioned technical solutions or a combination of multiple technical solutions, after pushing the service recommendation strategy to the corresponding customer, the method further includes:

[0019] Obtaining the customer's feedback data and evaluating it to obtain the feedback data after the execution of the business recommendation strategy;

[0020] The customer profile of the customer is updated according to the post-execution feedback data so as to be iterated into the algorithm of the business recommendation strategy.

[0021] Furthermore, based on any one of the technical solutions or a combination of multiple technical solutions mentioned above, the static data includes one or more of personal information, purchase preferences, historical transaction records, and historical marketing records; the dynamic data includes one or more of real-time transaction data, real-time page access event data, and product browsing event data.

[0022] Furthermore, based on any one of the technical solutions or a combination of multiple technical solutions mentioned above, the collected static data and dynamic data are cleaned, deduplicated, format converted and standardized, and key features are extracted and labeled to form a customer profile.

[0023] Further, based on any one of the above-mentioned technical solutions or a combination of multiple technical solutions, a set of customer profiles of multiple customers is defined as D = {x i ,y i}, where |D| = N, N is the number of customers, i is an integer from 1 to N, R is a real number set, x i ∈R M ,y i ∈R, M is the number of data items for each customer profile;

[0024] The calculation formula based on the gradient boosting tree is as follows:

[0025] Γ={f(x)=w t(x)}(t:R M →S,w∈R S ), where Γ is the formal definition of the gradient boosting tree algorithm, each f(x) is a base learner and has a corresponding weight w for learning its importance score, S is the number of base learners, and the base learner is used to independently learn the weight of each sample, w t(x) is the leaf node weight, M is the number of data items for each customer portrait, and t(x) is the characterization function of the leaf node;

[0026] f s ∈Γ, where is the formal definition of the algorithm prediction, which calculates the final prediction result by adding the weights of each base learner; f s (x i ) is the sth item in the data set Γ;

[0027] Where p is the pth iteration in the process of recursively solving the model of each decision tree through the training method of additive training, L (p) is the formal definition of the objective function, Ω(f p ) is the regularization term, l() is the loss function, y i For real results, is the prediction result of the pth iteration, that is, the prediction value of the i-th instance Equal to the predicted value of the p-1th iteration Add the predicted value f of the pth tree p (x i ), the predicted value f p (x i ) is the leaf node weight w t(x) ;

[0028] in, is the prediction function of the second-order Taylor expansion of the loss function, g i is the first-order derivative of the loss function l, h i is the second-order derivative of the loss function l.

[0029] Further, based on any one of the above-mentioned technical solutions or a combination of multiple technical solutions, the regularization term Ω(f p ) is:

[0030] Among them, γ and λ are regularization factors;

[0031] According to the prediction function of the second-order Taylor expansion of the objective function, the set of all leaf nodes obtained in the decision tree is converted into a multi-hot vector t=G(x|Tr)=[Tr1(x),…,Tr S (x)], where Tr is a decision tree, G is a gradient boosting tree function, the vector t is the concatenation of all leaf nodes in the decision tree, the number of elements with a value of 1 in t is the number of decision trees, each decision tree has only one active leaf node, Tr1(x) is the leaf node obtained by the first decision tree with x as input, and Tr S (x) is the leaf node obtained by the Sth decision tree with x as input.

[0032] Further, based on any one of the above-mentioned technical solutions or a combination of multiple technical solutions, the contribution of each data item in the customer portrait to the model prediction is calculated based on the gradient boosting tree algorithm, and the cross-feature v is captured in the following way: l :v l ={x∈x i |Tr i (x) = 1};

[0033] The scene adaptation mechanism calculates the personalized weight of each cross feature in different scenes as follows:

[0034]

[0035] Among them, h is the weight vector, h∈R a, Relu() is the linear rectification function, is the concatenation operation, W is the weight matrix of the hidden layer, W∈R a×2k , b is the bias vector of the hidden layer, b∈R a , a is the size of the hidden layer, k is the size of the embedding layer, u i is the customer ID vector, e j is the scene ID vector, v l is the cross feature, exp() is the natural exponential function, a ijl for u i 、e j 、v l The personalized weight value of .

[0036] Further, based on any one of the above-mentioned technical solutions or a combination of multiple technical solutions, the service recommendation method provided by the present invention also includes calculating the probability of recommending a service to a specific customer in a specific scenario based on the calculation result of the personalized weight by the following formula:

[0037]

[0038] Among them, max_pool is the maximum pooling, V is the cross feature set, w0 is the initial leaf node weight, w d is the weight of the dth leaf node, x d For the d-th customer portrait, P∈R k and Q∈R k is the weight of the linear regression layer, and ⊙ is the dot product operation.

[0039] Furthermore, based on any one of the technical solutions or a combination of multiple technical solutions described above, after training the gradient boosting tree-based algorithm to obtain the minimized loss value, the cross-features are input into the model based on the scene adaptation mechanism for secondary training to obtain interpretable weights for prediction.

[0040] According to another aspect of the present invention, the present invention provides a business recommendation device based on big data and customer portraits, comprising the following modules:

[0041] A data collection and processing module is configured to collect and process static data and / or dynamic data of multiple customers to obtain a standardized and labeled customer portrait, wherein the customer portrait includes multiple data items;

[0042] A data analysis and mining module is configured to perform data analysis and mining on the standardized and labeled customer portraits through a gradient boosting tree-based algorithm to exclude features that lack behavioral regularities and thereby capture cross-features;

[0043] A recommendation strategy formulation module, which is configured to formulate a business recommendation strategy based on an explainable algorithm according to the cross-features, including introducing a scenario adaptation mechanism into the algorithm-based business recommendation strategy, wherein the scenario adaptation mechanism learns the importance of each cross-feature in different scenarios from data;

[0044] A strategy push module is configured to push the business recommendation strategy to the corresponding customer.

[0045] According to another aspect of the present invention, the present invention provides a business recommendation system based on big data and customer portraits, including a management platform, a dynamic big data analysis module, a customer portrait module, an intelligent strategy module, and a recommendation task management module, wherein:

[0046] The big data analysis module is configured to collect dynamic data of customers, and update customer labels into the customer portrait module after standardization and labeling;

[0047] The customer portrait module is configured to load the customer portrait into the smart strategy module;

[0048] The intelligent strategy module performs data analysis and mining on the standardized and labeled customer portraits through an algorithm based on a gradient boosting tree to exclude features that lack behavioral regularities, thereby capturing cross-features, and formulating a business recommendation strategy based on an explainable algorithm based on the cross-features;

[0049] The recommendation task management module pushes the service recommendation strategy to the corresponding customer in response to the triggering task drive of the intelligent strategy module;

[0050] The management platform is configured to receive feedback information after customer contact, load data into the dynamic big data analysis module, and update customer information in the customer portrait module.

[0051] The beneficial effects brought by the technical solution provided by the present invention are as follows:

[0052] a. Multi-dimensional data collection: Most current technologies mainly collect static data, and do not fully consider and utilize multi-dimensional dynamic data such as customers' historical behavior, current point-in-time information, historical marketing records, and real-time events;

[0053] b. Generation of explainable intelligent marketing strategies: Existing marketing strategies are mostly based on simple rules or difficult-to-explain machine learning models, and fail to customize strategies for specific marketing scenarios. Based on basic feature analysis and high-order feature construction, by combining algorithms based on gradient boosting trees and adaptive mechanisms, we eventually constructed a variety of rule-based and algorithm-based explainable intelligent marketing strategies.

[0054] c. Closed-loop marketing management: Existing technologies have failed to effectively integrate customer relationship management into the marketing system and lack the ability to manage customers throughout their life cycle, making it difficult to achieve closed-loop operations for intelligent marketing; closed-loop marketing management for the entire life cycle of multi-mode customers has been achieved, integrating the customer relationship management system with the traditional marketing platform functions to achieve closed-loop operations for intelligent marketing. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0056] Figure 1 A flowchart of a business recommendation method based on big data and customer portraits provided for an exemplary embodiment of the present invention;

[0057] Figure 2 A schematic diagram of the framework structure of a business recommendation system based on big data and customer portraits provided for an exemplary embodiment of the present invention;

[0058] Figure 3 A schematic diagram of a hybrid model structure of a gradient boosting tree-based algorithm and an adaptive mechanism is provided for an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0060] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, device, product or equipment that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0061] The intelligent marketing based on dynamic big data and customer portrait provided by the embodiment of the present invention mainly belongs to the technical fields of big data, artificial intelligence and financial technology. Big data technology can obtain in-depth insights into customer behavior, preferences, credit status and other aspects by collecting, processing and analyzing massive and dynamic data. These insights provide an important basis for banks to formulate precise marketing strategies. Secondly, customer portrait technology enables banks to understand their customer groups more deeply. By building multi-dimensional customer portraits, banks can grasp information such as customer consumption habits, risk preferences, financial needs, etc., so as to provide customers with more personalized financial products and services. Thirdly, artificial intelligence technology makes the entire marketing process more intelligent and automated. Through machine learning, deep learning and other technologies, the intelligent marketing platform can automatically analyze customer data, predict customer behavior, and provide customized marketing suggestions for banks. Finally, as an application scenario of the entire solution, the field of financial technology also provides the necessary technical support and market environment. With the help of such intelligent marketing methods and platforms, small and medium-sized banks can not only improve marketing effects, but also optimize customer service and enhance market competitiveness.

[0062] Intelligent recommendation is a service developed based on big data and artificial intelligence technology. Early recommendation systems were mainly based on simple rules and filtering algorithms. With the evolution of technology, collaborative filtering algorithms, machine learning and deep learning technologies have been widely used in intelligent recommendation, enabling recommendation systems to understand customer needs more accurately. However, most current intelligent recommendation methods study features as independent factors, ignoring the interaction between features and lacking interpretability. In addition, in the process of evaluating marketing effectiveness, they focus on the statistics of the effect, directly adjust the strategy based on the statistical results, and do not consider the marketing history data of the customer throughout the life cycle from the customer dimension.

[0063] In one embodiment of the present invention, a business recommendation method based on big data and customer portraits is provided, which combines an algorithm based on a gradient boosting tree and an adaptive mechanism to improve business recommendation performance while ensuring interpretability.

[0064] In one embodiment of the present invention, a business recommendation system based on big data and customer portraits is provided. Figure 2 As shown, the system includes a management platform, a dynamic big data analysis module, a customer portrait module, an intelligent strategy module, and a recommendation task management module, among which:

[0065] The big data analysis module is configured to collect dynamic data of customers, and update customer labels into the customer portrait module after standardization and labeling;

[0066] The customer portrait module is configured to load the customer portrait into the smart strategy module;

[0067] The intelligent strategy module performs data analysis and mining on the standardized and labeled customer portraits through an algorithm based on a gradient boosting tree to exclude features that lack behavioral regularities, thereby capturing cross-features, and formulating a business recommendation strategy based on an explainable algorithm based on the cross-features;

[0068] The recommendation task management module pushes the service recommendation strategy to the corresponding customer in response to the triggering task drive of the intelligent strategy module;

[0069] The management platform is configured to receive feedback information after customer contact, load data into the dynamic big data analysis module, and update customer information in the customer portrait module.

[0070] like Figure 1 As shown, the business recommendation method based on big data and customer portraits of this embodiment includes the following steps:

[0071] The first step is to collect and process static data and / or dynamic data of multiple customers to obtain a standardized and labeled customer portrait, wherein the customer portrait includes multiple data items;

[0072] Based on the high-quality basic data provided by the self-built data warehouse, the variable center and event center are combined to collect multi-dimensional static and dynamic data. Static data collection includes historical customer data such as personal information, transaction records, purchase preferences, and historical marketing records such as past marketing activity participation and marketing effectiveness. Dynamic data collection includes current point-in-time data such as real-time transaction data and online behavior data, as well as real-time event data such as customers visiting specific pages and browsing specific products.

[0073] The collected data is cleaned, deduplicated, format converted and other basic operations are standardized to ensure data quality and uniform format. Based on a series of dynamic and static data such as basic customer information and behavioral data, key features are extracted and labeled to form customer portraits, which facilitates subsequent data analysis and marketing strategy formulation.

[0074] Specifically, the static data includes one or more of personal information, purchase preferences, historical transaction records, and historical marketing records; the dynamic data includes one or more of real-time transaction data, real-time page access event data, and product browsing event data.

[0075] Furthermore, based on any one of the technical solutions or a combination of multiple technical solutions mentioned above, the collected static data and dynamic data are cleaned, deduplicated, format converted and standardized, and key features are extracted and labeled to form a customer profile.

[0076] Further, based on any one of the above-mentioned technical solutions or a combination of multiple technical solutions, a set of customer profiles of multiple customers is defined as D = {x i ,y i}, where |D| = N, N is the number of customers, i is an integer from 1 to N, R is a real number set, x i ∈R M ,y i ∈R, M is the number of data items for each customer profile.

[0077] Step 2: Perform data analysis and mining on the standardized and labeled customer portraits through a gradient boosting tree-based algorithm to exclude features that lack behavioral regularities and capture cross-features.

[0078] Data analysis and mining include basic feature analysis and high-order feature construction. The basic feature analysis uses algorithms based on gradient boosting trees (such as XGBoost, LightGBM, etc.) to preliminarily screen standardized static data and dynamic data that are effective for marketing, and can analyze the basic characteristics and behavioral patterns of customers. By analyzing important features such as "age" and "income", the purchasing behavior and preferences of customers of different ages and income segments can be analyzed. This will provide a deeper understanding of customer needs and provide a basis for precision marketing and personalized recommendations. The algorithm based on the gradient boosting tree calculates the contribution of each feature to the model prediction based on the marketing data and obtains its importance score. The algorithm based on the gradient boosting tree calculates the contribution of each data item in the customer portrait to the model prediction to capture cross-features; the calculation formula based on the gradient boosting tree is as follows:

[0079] Γ={f(x)=w t(x)}(t:R M →S,w∈R S), where Γ is the formal definition of the gradient boosting tree algorithm, each f(x) is a base learner and has a corresponding learning weight w, i.e., importance score, S is the number of base learners, and the base learner is used to independently learn the weight of each sample, w t(x) is the leaf node weight, i.e., the contribution degree, M is the number of data items of each customer portrait, and t(x) is the characterization function of the leaf node;

[0080] f s ∈Γ, where is the formal definition of algorithm prediction, that is, the model prediction, which calculates the final prediction result by adding the weights of each base learner; f s (x i ) is the sth item in the data set Γ;

[0081] Where p is the pth iteration in the process of recursively solving the model of each decision tree through the training method of additive training, L (p) is the formal definition of the objective function, Ω(f p ) is the regularization term, and its expression is Among them, γ and λ are regularization factors; l() is the loss function, y i For real results, is the prediction result of the pth iteration, that is, the prediction value of the i-th instance Equal to the predicted value of the p-1th iteration Add the predicted value f of the pth tree p (x i ), the predicted value f p (x i ) is the leaf node weight w t(x) ; Each iteration of the gradient boosting tree will get a new decision tree in the direction of reducing the residual, so the p in the above p-th iteration and the p-th tree is the same.

[0082] in, is the prediction function of the second-order Taylor expansion of the loss function, g i is the first-order derivative of the loss function l, h i is the second-order derivative of the loss function l.

[0083] The algorithm based on the second-order Taylor expansion can better fit complex nonlinear data sets and can be used to obtain the regularization term Ω(f p ) to reduce model complexity and prevent overfitting.

[0084] High-order feature construction uses historical data and algorithms based on gradient boosting trees to construct second-order and higher features to reveal customers' potential needs and preferences in more depth. In the industry, systems use a large number of features to improve the accuracy of the model. These features can show different information from different dimensions, and the cross-features of these feature combinations are also of great significance. Traditional second-order cross-features are usually designed manually, and the complexity will increase exponentially for high-order cross-features. Such features have extremely high requirements for designers with domain knowledge, and this labor-intensive work is costly, restrictive, and has poor scalability. In order to avoid these shortcomings, some systems have studied the use of deep models to automatically learn high-order cross-features. However, due to the black box nature of the deep model, the cross-features obtained are implicit and difficult to interpret. Therefore, the present invention chooses to use an algorithm based on gradient boosting trees to capture effective cross-features. Although the algorithm is not specifically designed for extracting cross-features, its leaf nodes can be used to represent cross-features. The algorithm optimizes the prediction of cross-features in iterations to obtain leaf nodes with higher importance scores and form reasonable cross-features.

[0085] For a set of decision trees Tr={Tr1,…,Tr S}, each decision tree has leaf nodes that map feature vectors x to weights w. Different from the original feature analysis that sums the weights of active leaf nodes as the predicted value, the high-order feature construction extracts active leaf nodes as valid cross-features and provides them to subsequent modules for processing. The prediction function of the second-order Taylor expansion of the objective function is used to convert the set of all leaf nodes obtained in the decision tree into a multi-hot vector t = G(x|Tr) = [Tr1(x),…,Tr S (x)] where Tr is the leaf node of the decision tree, G is the gradient boosting tree function, vector t is the concatenation of all leaf nodes in the decision tree, the number of elements with value 1 in t is the number of decision trees, each decision tree has only one active leaf node, Tr1(x) is the leaf node obtained by the first decision tree with x as input, Tr S (x) is the leaf node obtained by the Sth decision tree with x as input; vector t is the concatenation of all leaf nodes in the decision tree, and the number of elements with a value of 1 in t is the number of decision trees. Each decision tree has only one active leaf node, so the multi-hot vector can be replaced with an effective cross feature. As the depth of the decision tree continues to increase, the algorithm can construct more complex high-order features.

[0086] The gradient boosting tree-based algorithm calculates the contribution of each data item in the customer profile to the model prediction, capturing the cross-feature v in the following way l :v l ={x∈x i |Tr i(x)=1}, that is, the cross feature is the combination of features that are 1 in vector t. For example, a series of features [x1,x2,x3,x4] corresponds to a series of values ​​[1,0,1,0], then the final cross feature is [x1,x3].

[0087] Step 3: Formulate a business recommendation strategy based on an explainable algorithm based on the cross-features, including introducing a scenario adaptation mechanism into the algorithm-based business recommendation strategy, wherein the scenario adaptation mechanism learns the importance of each cross-feature in different scenarios from the data;

[0088] Utilize the data analysis and mining methods to formulate marketing strategies, which are divided into rule-based marketing recommendations and algorithm-based marketing recommendations.

[0089] Rule-based marketing recommendation: Based on expert experience, the basic features and high-level features constructed in the second step are used as the basis, and rules and conditions are preset to issue marketing tasks to account managers and recommend corresponding products and services. Experts adjust marketing rules based on marketing feedback and results, so as to timely adjust and optimize marketing strategies and achieve basic marketing goals.

[0090] Algorithm-based marketing recommendations: In actual marketing, there are usually multiple marketing scenarios. Although manually constructed rules are effective, they are too costly and heavily dependent on expert experience. For this reason, the present invention proposes algorithm-based marketing recommendations. If the same weight is assigned to the cross-features of all customers for different scenarios, this treatment will limit the prediction effect due to the different importance of different features in the scenarios. Some less used feature interactions may even introduce noise. Therefore, a marketing scenario adaptive mechanism is introduced in the algorithm-based marketing recommendation, which learns from the data the importance of each cross-feature in different marketing scenarios. The modeling process of personalized weights on cross-features improves the performance of the model and improves the interpretability of the model through attention weights. The calculation process of the personalized weights of each cross-feature in different scenarios by the scenario adaptive mechanism is as follows:

[0091] T is the transposed matrix;

[0092] Among them, h is the weight vector, h∈R a , Relu() is the linear rectification function, is the concatenation operation, W is the weight matrix of the hidden layer, W∈R a×2k , b is the bias vector of the hidden layer, b∈R a , a is the size of the hidden layer, k is the size of the embedding layer, u i is the customer ID vector, e j is the scene ID vector, v lis the cross feature, exp() is the natural exponential function, a ijl for u i 、e j 、v l The personalized weight value of .

[0093] When making predictions in different marketing scenarios, the marketing degree of various cross-features is different. The marketing scenario adaptive mechanism converts the customer ID vector u i , scene ID vector e j and the cross feature v l The embeddings are combined to capture the complex interactions between customers, scenarios, and cross-features, while avoiding the use of hidden layers to ensure model transparency.

[0094] The hybrid model structure based on the gradient boosting tree algorithm and the adaptive mechanism is as follows Figure 3 As shown. The overall recommendation model is a shallow model. Due to the additivity of the shallow model, the importance contribution of each component can be provided for the prediction. At the same time, even without a fully connected layer, the embedding mechanism and the adaptive mechanism still make the model have strong representation ability and effectiveness. Based on the calculation results of the personalized weights, the probability of recommending a business to a specific customer in a specific scenario is calculated by the following formula:

[0095] It is an aggregation process of cross-features according to personalized weights, and each sample obtains a unified representation of the cross-features;

[0096] In order to calculate the probability of recommending to a specific customer in a specific scenario, the customer ID, marketing scenario ID and the embedding vectors x, u, e, V of the feature set are integrated to obtain a unified embedding vector, and linear regression is used to make the final prediction. Where max_pool is the maximum pooling, V is the cross feature set, w0 is the initial leaf node weight, w d is the weight of the dth leaf node, x d For the d-th customer portrait, P∈R k and Q∈R k is the weight of the linear regression layer, and ⊙ is the dot product operation.

[0097] In addition, this embodiment regards whether to recommend a certain customer in a certain scenario as a binary classification problem, uses logarithmic loss as the loss function, and uses L2 regularization to prevent overfitting. Since the model consists of two series models, the two models are trained separately. After training the algorithm based on the gradient boosting tree to minimize the loss value, the cross-features are input into the model based on the adaptive mechanism for secondary training, so as to obtain the final interpretable weights for prediction.

[0098] Step 4: Use multi-channel docking to push the business recommendation strategy to the corresponding customers;

[0099] After obtaining the final marketing list under the above different marketing scenarios, the platform will automatically distribute it. Through multi-channel marketing to target customers, it can achieve docking with multiple channels such as SMS, phone calls, outbound call robots, WeChat public accounts, offline visits, mobile applications, etc., to ensure that marketing information can reach customers accurately and efficiently. At the same time, based on customer feedback and latest behavior, the customer portrait is further updated to optimize the marketing recommendation strategy in a cycle, and finally achieve iterative marketing recommendations and efficient marketing execution.

[0100] Step 5: Data statistics and effect evaluation;

[0101] According to the above steps, after the marketing execution of the intelligent marketing method and platform based on dynamic big data and customer portraits, data statistics and effect evaluation are carried out. In data statistics, data is recorded and counted from multiple dimensions such as customers, marketing scenarios, account managers, and branches to fully understand the execution of marketing scenarios. For customers, in-depth analysis of customer purchase behavior, preferences, activity and other information; at the same time, detailed data records are made for different marketing scenarios, such as online activities and offline promotions, in order to understand the marketing effects in various scenarios; for account managers and branches, performance statistics are carried out to better evaluate the team's execution and cooperation effects. Based on statistical data, an in-depth evaluation of the effects of marketing scenarios is carried out, and the advantages and disadvantages of the scenarios are analyzed to provide new ideas and basis for subsequent scenario strategy adjustments.

[0102] Step 6: Customer management and relationship maintenance;

[0103] Through the above process, we can eventually build an intelligent marketing method and platform based on dynamic big data and customer portraits, forming a closed-loop marketing management of the multi-mode customer life cycle process, including customer acquisition, conversion, retention and upgrade stages, such as Figure 2 As shown. The platform realizes personalized and intelligent recommendations for marketing scenarios through the dynamic big data analysis, customer portrait, and intelligent strategy center modules corresponding to steps 2 and 3 above, realizes multi-channel efficient marketing execution through the customer contact module corresponding to step 4 above, and obtains customer manager contact feedback data and scenario results statistics through step 5. In addition, the platform integrates traditional marketing platform functions with customer relationship management systems to provide one-stop customer management services for customer managers, further improving marketing efficiency.

[0104] The fifth and sixth steps are to obtain the feedback data of the customer and evaluate it to obtain the post-execution feedback data of the business recommendation strategy; and update the customer profile of the customer based on the post-execution feedback data to iterate into the algorithm of the business recommendation strategy.

[0105] In one embodiment of the present invention, a business recommendation device based on big data and customer portraits is provided, comprising the following modules:

[0106] A data collection and processing module is configured to collect and process static data and / or dynamic data of multiple customers to obtain a standardized and labeled customer portrait, wherein the customer portrait includes multiple data items;

[0107] A data analysis and mining module is configured to perform data analysis and mining on the standardized and labeled customer portraits through a gradient boosting tree-based algorithm to exclude features that lack behavioral regularities and thereby capture cross-features;

[0108] A recommendation strategy formulation module, which is configured to formulate a business recommendation strategy based on an explainable algorithm according to the cross-features, including introducing a scenario adaptation mechanism into the algorithm-based business recommendation strategy, wherein the scenario adaptation mechanism learns the importance of each cross-feature in different scenarios from data;

[0109] A strategy push module is configured to push the business recommendation strategy to the corresponding customer.

[0110] The inventive concept of the present business recommendation device is the same as the inventive concept of the above-mentioned business recommendation method and business recommendation system. The contents of the above-mentioned business recommendation method embodiment and business recommendation system embodiment are introduced into the present business recommendation device embodiment by full reference.

[0111] The present invention collects and analyzes customer data to more accurately understand customer needs and behavior patterns, thereby formulating more accurate marketing strategies, and using artificial intelligence technology to provide personalized financial advice and services. Through digital technology, banks can more accurately target customer groups and achieve higher marketing results at a lower cost. At the same time, digital marketing also provides more customer feedback channels, which helps banks continuously improve and optimize services.

[0112] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0113] The above is only a specific implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A business recommendation method based on big data and customer portraits, characterized in that: The following steps are involved: Static data and / or dynamic data of multiple customers are collected and processed to obtain standardized and labeled customer portraits, wherein the customer portraits include multiple data items, and the set of customer portraits of multiple customers is defined as D = {x i ,y i }, where |D| = N, N is the number of customers, i is an integer from 1 to N, R is a real number set, x i ∈R M ,y i ∈R, M is the number of data items for each customer profile; The standardized and labeled customer portraits are analyzed and mined by using a gradient boosting tree-based algorithm to exclude features that lack behavioral regularities and capture cross-features v l :v l ={x∈x i |Tr i (x) = 1}; where Tr i (x) is defined as the leaf node of the ith decision tree with x as input, which is obtained as follows: The calculation formula based on the gradient boosting tree is as follows: Γ={f(x)=w t(x) }, t:R M →S,w∈R S , where Γ is the formal definition of the gradient boosting tree algorithm, each f(x) is a base learner and has a corresponding learning weight w, S is the number of base learners, and the base learner is used to independently learn the weight of each sample, w t(x) is the leaf node weight, M is the number of data items for each customer portrait, and t(x) is the characterization function of the leaf node; in, is the formal definition of the algorithm prediction, which calculates the final prediction result by adding the weights of each base learner; f s (x i ) is the sth item in the data set Γ; Where p is the pth iteration in the process of recursively solving the model of each decision tree through the training method of additive training, L (p) is the formal definition of the objective function, Ω(f p ) is the regularization term, l() is the loss function, y i For real results, is the prediction result of the pth iteration; in, is the prediction function of the second-order Taylor expansion of the objective function, g i is the first-order derivative of the loss function l, h i is the second-order derivative of the loss function l; according to the prediction function of the second-order Taylor expansion of the objective function, the set of all leaf nodes obtained in the decision tree is converted into a multi-hot vector t=G(x|Tr)=[Tr1(x),…,Tr S (x)], where Tr is a decision tree, G is a gradient boosting tree function, the vector t is the concatenation of all leaf nodes in the decision tree, the number of elements with a value of 1 in t is the number of decision trees, each decision tree has only one active leaf node, Tr1(x) is the leaf node obtained by the first decision tree with x as input, and Tr S (x) is the leaf node obtained by the Sth decision tree with x as input; According to the cross-features, a business recommendation strategy based on an explainable algorithm is formulated, including introducing a scenario adaptation mechanism into the algorithm-based business recommendation strategy, wherein the scenario adaptation mechanism learns the importance of each cross-feature in different scenarios from data, including: the scenario adaptation mechanism calculates the personalized weight a of each cross-feature in different scenarios ijl The calculation process is: Among them, h is the weight vector, h∈R a , Relu() is the linear rectification function, is the concatenation operation, W is the weight matrix of the hidden layer, W∈R a×2k , b is the bias vector of the hidden layer, b∈R a , a is the size of the hidden layer, k is the size of the embedding layer, u i is the customer ID vector, e j is the scene ID vector, v l is the cross feature, exp() is the natural exponential function, a ijl for u i 、e j 、v l The personalized weight value of Push the service recommendation strategy to the corresponding customer: Based on the calculation result of the personalized weight, calculate the probability of recommending the service to the specific customer in a specific scenario: Among them, max_pool is the maximum pooling, V is the cross feature set, w0 is the initial leaf node weight, w d is the weight of the dth leaf node, x d For the d-th customer portrait, P∈R k and Q∈R k is the weight of the linear regression layer, and ⊙ is the dot product operation.

2. The service recommendation method according to claim 1, characterized in that: After pushing the service recommendation strategy to the corresponding customer, the method further includes: Obtaining the customer's feedback data and evaluating it to obtain the feedback data after the execution of the business recommendation strategy; The customer profile of the customer is updated according to the post-execution feedback data so as to be iterated into the algorithm of the business recommendation strategy.

3. The service recommendation method according to claim 1, characterized in that: The static data includes one or more of personal information, purchase preferences, historical transaction records, and historical marketing records; the dynamic data includes one or more of real-time transaction data, real-time page access event data, and product browsing event data.

4. The service recommendation method according to claim 1, characterized in that: The collected static and dynamic data are cleaned, deduplicated, and format converted in a standardized manner, and key features are extracted and labeled to form customer portraits.

5. The service recommendation method according to claim 1, characterized in that: Regularization term Ω(f p ) is: Among them, γ and λ are regularization factors.

6. The service recommendation method according to claim 1, characterized in that: After training the gradient boosting tree-based algorithm to minimize the loss value, the cross-features are input into the model based on the scene adaptation mechanism for secondary training to obtain interpretable weights for prediction.

7. A business recommendation device based on big data and customer portraits, characterized in that: The service recommendation method according to any one of claims 1 to 6 is performed, wherein the service recommendation device comprises the following modules: A data collection and processing module is configured to collect and process static data and / or dynamic data of multiple customers to obtain a standardized and labeled customer portrait, wherein the customer portrait includes multiple data items; A data analysis and mining module is configured to perform data analysis and mining on the standardized and labeled customer portraits through a gradient boosting tree-based algorithm to exclude features that lack behavioral regularities and thereby capture cross-features; A recommendation strategy formulation module, which is configured to formulate a business recommendation strategy based on an explainable algorithm according to the cross-features, including introducing a scenario adaptation mechanism into the algorithm-based business recommendation strategy, wherein the scenario adaptation mechanism learns the importance of each cross-feature in different scenarios from data; A strategy push module is configured to push the business recommendation strategy to the corresponding customer.

8. A business recommendation system based on big data and customer portraits, characterized in that: Execute the business recommendation method according to any one of claims 1 to 6, wherein the business recommendation system includes a management platform, a dynamic big data analysis module, a customer portrait module, an intelligent strategy module, and a recommendation task management module, wherein: The big data analysis module is configured to collect dynamic data of customers, and update customer labels into the customer portrait module after standardization and labeling; The customer portrait module is configured to load the customer portrait into the smart strategy module; The intelligent strategy module performs data analysis and mining on the standardized and labeled customer portraits through an algorithm based on a gradient boosting tree to exclude features that lack behavioral regularities, thereby capturing cross-features, and formulating a business recommendation strategy based on an explainable algorithm based on the cross-features; The recommendation task management module pushes the service recommendation strategy to the corresponding customer in response to the triggering task drive of the intelligent strategy module; The management platform is configured to receive feedback information after customer contact, load data into the dynamic big data analysis module, and update customer information in the customer portrait module.

Images