A report intelligent insight system for banking marketing process data analysis

Abstract

The present application relates to the technical field of data management, in particular to a report intelligent insight system for bank marketing process data analysis, which represents customer behavior, captures minute-level product clicks and other short-term event streams by using time graph network to generate dynamic behavior vectors, and analyzes long-term sequences such as fund flow spanning several months by using Transformer architecture to generate static behavior vectors; the two vectors are fused with customer portrait, input into a multi-layer perception network, and the loan intention strength is calculated by applying semantic projection; at the same time, the system dynamically groups similar customers by using clustering algorithm, and quantifies the group fatigue degree according to the feedback of the group to the recent marketing; the loan intention, the positive trend of the account and the group fatigue degree are integrated as penalty terms to calculate the marketing recommendation coefficient, and the optimal decision covering the reach channel, product and language is generated according to the matching of the business rule library, and the customer feedback will update the fatigue degree in real time to form a dynamic optimization closed loop.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of enterprise data management, in particular to a report intelligent insight system for bank marketing process data analysis. BACKGROUND

[0002] The existing customer marketing analysis system is heavily dependent on static or low-frequency updated data conditions, and cannot effectively capture the real-time demand and intention changes of customers within a specific time window. The understanding of the system to the customers is often one-sided and lagging, and it is difficult to respond to the fleeting financial service opportunities revealed by real-time behavior (such as in-app browsing and searching). Most marketing automation systems work based on pre-set and hard-coded rule engines.

[0003] For example, by setting a simple threshold such as "the number of times of browsing a certain product page exceeds 3 times" or "the search frequency of a certain keyword reaches a set value" to trigger marketing actions, this mechanism lacks flexibility and adaptability, and not only cannot identify users with complex behavior patterns but real intentions, but also often generates a large number of false marketing triggers due to simple behaviors (such as accidental clicks and regular browsing), resulting in waste of marketing resources and user harassment. The current system design is generally oriented towards "maximizing marketing reach", focusing on identifying potential positive signals, but almost completely ignores the quantification and management of the possible negative effects of marketing activities, and cannot automatically reduce marketing frequency or adjust strategies when customers show impatience or resistance. Ultimately, it may lead to a decrease in customer satisfaction and even cause customer loss.

[0004] Therefore, a report intelligent insight system for bank marketing process data analysis is proposed. SUMMARY

[0005] The purpose of the present application is to provide a report intelligent insight system for bank marketing process data analysis, which solves the problem of low marketing efficiency and poor customer experience caused by inaccurate judgment of customer intentions and general neglect of customer marketing fatigue in existing bank marketing systems. In order to achieve the above purpose, the present application provides the following technical solutions:

[0006] A report intelligent insight system for bank marketing process data analysis, comprising:

[0007] A behavior representation module: using a time graph unit to capture a stream of short-term behavior events of a customer and generate a dynamic behavior vector; using a long sequence unit to analyze a long-term behavior sequence of the customer and generate a static behavior vector; inputting the dynamic behavior vector and the static behavior vector into a gated cross-attention fusion process to generate a loan intention vector through semantic projection calculation;

[0008] The loan intention generation module: integrates the basic identity attributes, asset information and credit status to obtain a customer portrait vector, and inputs the customer portrait vector and a loan intention vector into a multi-layer perception network to generate a loan intention strength;

[0009] The fatigue degree calculation module: performs one-hot encoding and embedding processing on the customer portrait vector, splices the customer portrait vector and the loan intention vector to form a customer panoramic state vector, defines a customer group by using a K-Means clustering algorithm, and integrates a group fatigue degree of recent marketing feedback of the customer group according to the customer panoramic state vector;

[0010] The decision recommendation module: obtains a loan trend score from a loan intention survey questionnaire, calculates a marketing recommendation coefficient by combining the loan trend score, the loan intention strength and the group fatigue degree, and generates a marketing decision recommendation according to the marketing recommendation coefficient and a business rule library.

[0011] Preferably, the behavior characterization module comprises: the short-term behavior event stream comprises application-in-page clicks of recommended products, financial tool use records generated by the customer within a minute-level and / or hour-level time window; the time graph network unit is inputted, the customer, the product, the channel and the transaction counterparty are taken as nodes, a heterogeneous graph is constructed, the time sequence dependence relationship and the burst signal between the short-term behavior event streams are captured by internal message passing and node memory updating to obtain a dynamic behavior vector, and the dynamic behavior vector represents the instant state of the customer in the current network; the long-term behavior sequence comprises large-sum fund transactions, credit product application and repayment records and periodic asset allocation patterns of the customer over a history of months or even years; a long sequence unit based on a Transformer architecture is executed, the analysis process dynamically evaluates and weights the importance of each behavior in the historical sequence to the loan intention judgment by using an internal self-attention mechanism to obtain a static behavior vector.

[0012] Preferably, the behavior characterization module further comprises: the static behavior vector and the dynamic behavior vector are taken as inputs, a gated cross-attention fusion process is executed, the static behavior vector is taken as a query to review key information in the dynamic behavior vector, and the most relevant part of the short-term behavior event stream to the long-term pattern is amplified / inhibited by a gating unit to obtain a fused behavior vector; a pre-trained natural language processing model is used to convert text describing a loan business into a semantic vector, an average value of the semantic vector is calculated, a fixed vector representing the meaning of the loan is synthesized, and the fixed vector is taken as a reference direction of the intention judgment; a projection value of the fused behavior vector in the reference direction is calculated as the loan intention vector by using semantic projection.

[0013] Preferably, the customer portrait vector generated by the loan intention generation module includes: from the core transaction system, the credit system, the CRM system, the mobile bank background multiple sources, through ETL extraction, conversion, loading and real-time data synchronization, taking the customer identity as the unique identifier, obtaining and integrating the basic identity attributes, the asset information and the credit status, constructing the customer portrait vector; and splicing the customer portrait vector with the loan intention vector and inputting into the multi-layer perception network to calculate and generate the loan intention strength.

[0014] Preferably, the fatigue degree calculation module includes: performing one-hot encoding and embedding processing on the static portrait of the customer, splicing with the loan intention vector to form a customer panoramic state vector; using a K-Means clustering algorithm, performing real-time clustering operation on the customer panoramic state vectors of the whole line; dynamically dividing the customers with similar current behavior patterns, long-term habits and basic attributes into the same customer group; for the customer group, calculating the fatigue degree score based on the average click rate, application conversion rate and ignore closing rate of the same type of loan marketing information in the preset evaluation period, and setting the fatigue degree score as the group fatigue degree.

[0015] Preferably, the decision recommendation module includes: the loan trend score is calculated based on the analysis of the customer's loan intention survey questionnaire; the loan intention strength is the output value of the loan intention generation module; the group fatigue degree is the output of the fatigue degree calculation module; the marketing recommendation coefficient is calculated through a preset weighting formula, and the weighting formula gives positive weight to the loan intention strength and the loan trend score, and gives negative weight to the group fatigue degree.

[0016] The marketing recommendation coefficient value is compared with the multi-layer threshold value, and the specific conditions in the customer portrait vector are combined to retrieve and match the optimal action strategy in the business rule library to obtain a structured instruction set, and the instruction set defines the recommended touch channel, the recommended specific product, the recommended marketing speech template ID and the configurable preferential strategy; the touch channel includes a call from a customer manager, an APP pop-up window and a short message.

[0017] Preferably, the system further comprises a feedback module for incorporating the marketing acceptance degree into the group fatigue degree, comprising: adopting an event response type dynamic scoring model, according to the feedback behavior of the customer to the marketing action, the feedback behavior including clicking, applying, ignoring / closing, calculating an individual dynamic score, if the customer has no feedback within a preset time window, the individual dynamic score will decay at a corresponding rate according to the tolerance of the latest marketing; adopting a weighted average algorithm, applying the weight corresponding to the individual dynamic score to the marketing feedback data of all members in the customer group, establishing a feedback value matrix; when receiving the feedback behavior, the adjustment amount of the individual dynamic score is determined by the base score multiplied by the reference weight corresponding to the current situation in the feedback value matrix, thereby quantitatively calculating the marketing acceptance degree, and incorporating the marketing acceptance degree into the group fatigue degree to complete the closed-loop feedback.

[0018] Compared with the prior art, the application has the following beneficial effects:

[0019] 1. The application fuses the long-term and short-term behaviors of customers, constructs a comprehensive and dynamic behavior representation, and converts fuzzy intentions into accurate quantitative scores by projecting complex behavior vectors to a loan semantic benchmark; this method fundamentally surpasses traditional rule-based judgments, greatly improves the accuracy and reliability of intention recognition, enables marketing resources to accurately lock high-willingness customers, and significantly improves the conversion rate;

[0020] 2. The application introduces and quantifies the group fatigue degree as a decision variable, calculates the fatigue score through dynamic clustering and marketing feedback, and uses it as a negative weight in decision-making to form an intelligent braking mechanism, effectively suppressing excessive marketing; this avoids customer aversion caused by information bombardment, greatly improves customer experience while pursuing performance, and improves customer long-term loyalty;

[0021] 3. The system constructs an intelligent closed loop from insight to decision-making to feedback, and the system can automatically generate specific executable marketing instructions by comprehensively analyzing multi-dimensional information; the core is that the real feedback of customers dynamically updates the group fatigue degree, forming a self-optimizing learning loop; this closed loop mechanism enables the marketing strategy to continuously and adaptively evolve, significantly improves the automation and scientificity of decision-making, and reduces the cost of manual calculation. BRIEF DESCRIPTION OF DRAWINGS

[0022] Fig. 1 A step flow chart of the report intelligent insight system for bank marketing process data analysis of the application;

[0023] Fig. 2 A structural schematic diagram of the report intelligent insight system for bank marketing process data analysis of the application;

[0024] Fig. 3 The flowchart of the marketing recommendation decision of the present application is shown. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0026] Please refer to Figs. 1 to 3 The present application provides a report intelligent insight system for bank marketing process data analysis, referring to Fig. 1 The step flowchart of the present system is shown, Fig. 2 The structural schematic diagram of the present system is shown, and the technical solution is as follows: a report intelligent insight system for bank marketing process data analysis comprises:

[0027] The behavior representation module: a time graph unit is used to capture a short-term behavior event stream of a customer and generate a dynamic behavior vector; a long sequence unit is used to analyze a long-term behavior sequence of the customer and generate a static behavior vector; the dynamic behavior vector and the static behavior vector are input into a gated cross-attention fusion process to generate a loan intention vector through semantic projection calculation;

[0028] The loan intention generation module: a customer portrait vector is obtained by integrating basic identity attributes, asset information and credit status, and input into a multilayer perception network together with the loan intention vector to generate a loan intention strength;

[0029] The fatigue degree calculation module: the customer portrait vector is one-hot encoded and embedded, spliced with the loan intention vector to form a customer panoramic state vector, a K-Means clustering algorithm is used to define a customer group, and a group fatigue degree of recent marketing feedback of the customer group is integrated according to the customer panoramic state vector;

[0030] The decision recommendation module: a loan trend score is obtained from a loan intention survey questionnaire, a marketing recommendation coefficient is calculated by combining the loan trend score, the loan intention strength and the group fatigue degree, and a marketing decision recommendation is generated according to the marketing recommendation coefficient and a business rule library.

[0031] Embodiment one:

[0032] This embodiment mainly introduces that the behavior representation module of the system comprehensively captures short-term real-time behavior and long-term historical behavior of a customer to generate a loan intention vector.

[0033] Firstly, the data sources for short-term behavior event stream capture and dynamic behavior vector generation include:

[0034] Mobile banking app logs, online banking system logs, counter transaction records, ATM transaction records, SMS interaction records, and customer service call recordings transcribed into text;

[0035] Time Graph Unit: Employing a 2-layer GAT network with 4 attention heads, outputting 256 dimensions, and using customers, products, channels, and counterparties (merchants) as nodes to construct a heterogeneous graph; Customer A performed the following actions within 10 minutes: logged into the mobile banking app, clicked on the "Deposit Products" page, browsed the details page of the "Current Deposit" product pushed to them, clicked to learn about the details page of the "Fixed Deposit" product; and consulted customer service about "Wealth Management Products" before closing the app.

[0036] Heterogeneous graph construction: Nodes include: Customer A, Mobile Banking App, Deposit Products, Current Deposits, Time Deposits, Wealth Management Products, Fund Zone, Money Market Fund, and a specific money market fund; Edges represent behavioral sequences and timestamps;

[0037] Furthermore, the temporal graph network unit updates the state of each node (e.g., customer A's interest in "current account deposits") by aggregating neighbor node information; for example, if customer A clicks on "current account deposits" and then clicks on "time deposits," the system will transmit information to update customer A's overall interest in "deposit products"; and perform node memory updates: the temporal graph network unit maintains the memory state of each node as it changes over time; for example, customer A's active behavior on deposits and wealth management products will reflect a high immediate interest in "wealth management" in their dynamic behavior vector;

[0038] The dynamic behavior vector generation process involves processing the data through a time-map network unit to generate a low-dimensional dense vector (e.g., 256-dimensional) that represents customer A's immediate behavioral patterns and potential intentions within a time window. This vector captures the temporal dependencies and sudden signals between events; for example, continuously viewing deposits and funds may indicate that the customer has a need for fund allocation.

[0039] Long-term behavioral sequence analysis and static behavioral vector generation: The main data sources include core transaction systems (large-amount transfers, salary deposits), credit systems (loan applications, approvals, repayment records), CRM systems (customer complaints, suggestions), and asset management systems (investment purchases, redemptions).

[0040] Based on the long sequence unit of the Transformer architecture, a 4-layer encoder and 8 attention heads are used to output a 512-dimensional sequence of long-term behaviors of customer A in the past three years: March 2021: applied for and approved a housing mortgage loan, paid on time every month; July 2021: opened a fund account and regularly invested in an index fund; January 2022: received a large salary, part of which was transferred to a current account and part of which was used for consumption; May 2022: applied for a credit card, actively used and fully repaid; February 2023: redeemed part of the fund for family education expenses; August 2023: consulted several small consumer loan products, but did not apply.

[0041] In the Transformer architecture, each event in the long-term behavior sequence (such as "apply for a mortgage" and "fund investment") is considered a node, and the importance of each behavior in the historical sequence to the loan intention judgment is dynamically evaluated and weighted by calculating the attention weight of each event to other events; for example, when judging the loan intention of customer A, the event of applying for a mortgage and paying on time may have a higher weight because it is directly related to credit behavior; and salary entry may also have a higher weight because it reflects the customer's repayment ability; redeeming funds for education expenses may imply future financial needs;

[0042] Extending the "loan intention" to a comprehensive prediction of the customer's "key life cycle events", this is not limited to loans, but also covers a series of major life events such as marriage, childbirth, car purchase, education, retirement, etc.

[0043] By integrating the customer's financial behavior (short-term dynamics and long-term statics) and authorized external data (such as public information on social media and consumption trends at partner merchants), the system can predict the customer's upcoming major life changes; for example, if the system monitors that the user has recently frequently browsed baby products, education savings products, and has made large family-related purchases, it predicts not "loan intention", but "newborn family" event; the bank can then push a complete set of services such as family insurance, child-specific savings accounts, and mortgage counseling for larger space housing, upgrading marketing from "single-point sales" to "scenario-based, consultant-style comprehensive services."

[0044] Static behavior vector generation includes a fixed-length vector (e.g. 512-dimensional) generated after processing by the Transformer encoder, capturing customer A's long-term behavior patterns, risk preferences, consumption habits, and fund management style. This vector is relatively stable and represents the customer's static portrait.

[0045] The raw data is specially processed to be converted into a vector that can be understood by the machine. The time graph network technology is used to capture the instant behaviors of the customers in a short period of time (such as a few hours), such as clicking and browsing, to generate a dynamic behavior vector. At the same time, the Transformer model is used to analyze the historical records of the customers for a long period of time, such as transactions and repayments, to generate a static behavior vector, which provides a core data basis for accurately judging the intention of the customers.

[0046] Further, the gated cross-attention fusion takes the dynamic behavior vector (the instant state of the short-term behavior) and the static behavior vector (the stable pattern of the long-term behavior) as inputs; the static behavior vector is taken as a query (Query) to examine the key information in the dynamic behavior vector, and the part most related to the long-term pattern in the stream of short-term behavior events is amplified / inhibited by the gating unit;

[0047] The customer C recently frequently browses the car loan page on the mobile bank (the dynamic behavior vector shows a strong car loan intention), but the static behavior vector shows that he has never had a record of large consumption loan in the past, and his asset allocation is conservative; the gating unit will appropriately inhibit the loan intention that is too "aggressive" in the dynamic behavior vector according to the "conservative" characteristics of the static behavior vector, or amplify the signals related to "cautious assessment" therein; conversely, if the static behavior vector shows that customer C is a high net worth customer with multiple successful loan records, the gating unit will amplify the loan intention in the dynamic behavior vector;

[0048] The semantic projection calculates the loan intention intensity, including "loan" semantic vector synthesis: with the help of a pre-trained natural language processing model (such as BERT, Word2Vec), texts describing "loan" business, such as "personal consumption loan", "housing mortgage loan", "credit loan", and "enterprise operation loan", are converted into semantic vectors; the average value of these semantic vectors is calculated to synthesize a fixed vector (for example, 300 dimensions) representing the meaning of "loan", which serves as the reference direction for intention judgment;

[0049] The semantic projection uses cosine similarity or dot product to calculate the projection value of the loan intention vector on the reference direction as the loan intention intensity; the higher the projection value, the closer the loan intention vector of the customer to the "loan" semantic vector, and the greater the loan intention intensity; the L2 norm of the loan intention vector is 1, and the L2 norm of the standardized "loan" semantic vector is also 1; if the cosine similarity of the loan intention vector and the "loan" semantic vector is 0.85, the loan intention intensity is 0.85.

[0050] The dynamic vector representing the immediate interest and the static vector representing the long-term habit are fused through a gated cross-attention mechanism; then, through a semantic projection technology, the projection value of the fusion vector in the business direction of "loan" is calculated, and finally a precise quantitative score is output, which can intuitively reflect the willingness intensity of the customer.

[0051] Embodiment two:

[0052] This embodiment mainly introduces the loan intention generation module, which is responsible for integrating the customer's basic identity attributes, asset information and credit status, generating a customer portrait vector, and fusing it with the loan intention vector, and finally generating the loan intention intensity;

[0053] The data sources for constructing the customer portrait vector include: core transaction system (deposit, transfer record), credit system (historical loan, credit card, overdue record), CRM system (occupation, education, family information), mobile bank background (device information, login frequency);

[0054] ETL extraction, conversion, loading and real-time data synchronization: extraction is to extract data from the above multi-source systems regularly or in real time; conversion is to clean, de-duplicate and standardize the original data, for example, unify the "gender" field in different systems as "male / female", and unify the "occupation"; loading is to load the processed data to a unified customer data platform (such as data warehouse or lake); real-time data synchronization uses message queue such as Kafka to synchronize the key operations of customers in mobile bank or online bank in real time, ensuring the real-time update of customer portrait;

[0055] The customer portrait vector generation takes the customer's identity as the unique identifier, integrates the basic identity attributes (such as age, gender, occupation, marital status, region), asset information (such as deposit balance, financial product holding, fund market value), credit status (such as total debt, overdue record, credit score); these features are processed through one-hot encoding, embedding, etc. After processing, it is spliced into a customer portrait vector (such as 128 dimensions);

[0056] The loan intention strength generation fuses the customer portrait vector and the loan intention vector (for example, 128+256+512=896 dimensions) and inputs them into a multi-layer perception network MLP; the MLP network includes multiple fully connected layers and an activation function ReLU, the network structure is [896, 256, 64, 1], the activation function is ReLU and Sigmoid; the training parameters use an Adam optimizer with a learning rate of 0.001 to learn the complex nonlinear relationship between different vectors, and finally output a scalar value between 0 and 1, representing the loan intention strength; the training data uses historical marketing data, and customers who successfully apply for loans are marked as positive samples (intention strength close to 1), and customers who do not apply or are rejected are marked as negative samples (intention strength close to 0);

[0057] The customer portrait vector of customer D shows that he is 35 years old, married, a company executive, with an annual income of 500,000, a savings of 5 million, and no bad credit record; the dynamic behavior vector shows that he has recently frequently viewed overseas housing information; the static behavior vector shows that he has a history of multiple successful large investment behaviors; after splicing the three vectors, input into MLP; if the output of MLP is 0.92, it means that the loan intention strength of customer D is very high.

[0058] The scattered data is integrated into a structured customer portrait vector, which provides key background information for subsequent intention judgment, ensuring that the decision is based not only on customer behavior but also on their basic situation.

[0059] Embodiment three:

[0060] This embodiment mainly introduces the fatigue degree calculation module, which defines customer groups according to customer portraits, dynamic behaviors and static behaviors, and integrates recent marketing feedback to calculate group fatigue;

[0061] The customer panoramic state vector generation is to perform one-hot encoding and embedding processing on the customer's static portrait conditions (for example, high net worth, white collar, student), and splice it with the loan intention vector to form a customer panoramic state vector (for example, 128-dimensional portrait conditions + 256-dimensional dynamic behavior + 512-dimensional static behavior = 896-dimensional);

[0062] K-Means clustering algorithm: K value is determined by elbow rule, typical range is 10-100, real-time clustering operation is performed on the panoramic state vector of the full line customer; Clustering process: the system presets K cluster centers, calculates the distance from the panoramic state vector of each customer to the K cluster centers through iteration, assigns the customer to the nearest cluster, and updates the cluster center until the cluster center no longer changes significantly; Dynamic division: the clustering algorithm dynamically divides the customers into different customer groups according to the current behavior mode, long-term habit and basic attribute similarity of the customers; Assuming K=5; After clustering, customer A is divided into the "high-frequency active investor" group, customer B is divided into the "stable deposit customer" group, customer C is divided into the "potential housing demand customer" group, and customer D is divided into the "small and micro enterprise owner" group;

[0063] Further, group fatigue degree calculation is performed, for each customer group, the average click rate, application conversion rate and ignore close rate of the same type of loan marketing information in the preset evaluation period (for example, the past 7 days) are counted;

[0064] Fatigue score calculation formula:

[0065] Group fatigue degree = 1-(w1*average click rate+w2*average application conversion rate-w3*average ignore close rate)

[0066] Wherein, w1, w2, w3 are weight coefficients, which are set according to business experience, for example, w1=0.4, w2=0.5, w3=0.1; The higher the average click rate and application conversion rate, the lower the fatigue degree; The higher the ignore close rate, the higher the fatigue degree; And in a customer experience priority strategy, w1=0.3, w2=0.3, w3=0.4 can be set, so as to increase the influence of negative feedback on fatigue degree calculation. These weights can be dynamically adjusted according to the marketing goals of different business stages,

[0067] The "high-frequency active investor" group received 500 financial product marketing pushes in the past 7 days, with an average click rate of 5%, an average application conversion rate of 2%, and an average ignore close rate of 20%; Group fatigue = 1- (0.3*0.05+0.3*0.02-0.4*0.20) = 1.059 (extremely high fatigue, the negative impact of high ignore rate is significantly amplified). Group fatigue = 1- (0.4*0.05+0.5*0.02-0.1*0.20) = 0.99 (high fatigue, because the ignore close rate is high, although the click rate and conversion rate are low); The "potential home buying demand customer" group received 200 mortgage product marketing pushes in the past 7 days, with an average click rate of 15%, an average application conversion rate of 8%, and an average ignore close rate of 5%; Group fatigue = 1- (0.3*0.15+0.3*0.08-0.4*0.05) = 0.951 (moderately high fatigue, even though the click and conversion rates are high, a certain ignore rate keeps it at a high level). Therefore, group fatigue = 1- (0.4*0.15+0.5*0.08-0.1*0.05) = 0.905 (lower fatigue, because the click rate and conversion rate are higher);

[0068] Quantify the negative impact of marketing, avoid harassing customers, use K-Means clustering algorithm to divide the current customers into the same group as people with similar behavior and attributes; Then the system will calculate the average click rate, ignore rate and other feedback data of this specific group to the same marketing in the near future, and calculate a group fatigue score, which will play a braking role in the final decision.

[0069] Embodiment four:

[0070] This embodiment mainly introduces the calculation of marketing recommendation coefficient: refer to Fig. 3 The flowchart of marketing recommendation decision proposed by the present application; The loan trend score is calculated by real-time analysis of customer transaction flow, and the positive financial change indicators are calculated; The indicator examples are wage growth rate (for example, this month's wage is 10% higher than last month) and continuous growth days of current deposits (for example, current deposits have been growing for 30 consecutive days);

[0071] The calculation method is to normalize these indicators and then weight and sum to obtain the loan trend score (for example, 0-100 points); Loan trend score = 0.6*standardized (wage growth) + 0.4*standardized (deposit growth days);

[0072] Customer E, this month's wage growth rate is 20%, and current deposit has been growing for 45 consecutive days; The loan trend score is calculated as 90 points (high score); Marketing recommendation coefficient formula:

[0073] Marketing recommendation coefficient = w_intent*loan intent intensity + w_trend*loan trend score - w_fatigue*group fatigue

[0074] Wherein, w_intention, w_trend, w_fatigue are weight coefficients, which are set according to business objectives; for example, w_intention = 0.5, w_trend = 0.3, w_fatigue = 0.2;

[0075] Customer F, loan intention intensity 0.8, loan trend score 80 points (0.8 normalized), group fatigue 0.7; marketing recommendation coefficient = 0.5*0.8 + 0.3*0.8 - 0.2*0.7 = 0.5; marketing decision recommendation generation is compared with multi-layer threshold, the marketing recommendation coefficient value is compared with the preset multi-layer threshold (for example, 0.8 or more is "strong recommendation", 0.5-0.8 is "moderate recommendation", 0.3-0.5 is "weak recommendation", and 0.3 or less is not recommended); as shown in Table 2, three example rules of the business rule base, the business rule base matching is combined with the specific conditions in the customer portrait vector (for example, customer F is "small and micro enterprise owner"), and the optimal action strategy is searched and matched in the business rule base.

[0076] Table 2: Business rule base example

[0077]

[0078] The structured instruction set generation is to generate a structured instruction set according to the matched rule, which defines the recommended touch channel, specific product, suggested marketing speech template ID and configurable preferential policy; the marketing recommendation coefficient of customer F is 0.85, and its portrait is "small and micro enterprise owner"; according to the rule base, "small and micro enterprise short-term turnover loan" may be matched, the touch channel is "customer manager call", the marketing speech template ID is "SME_Loan_C", and the preferential policy is "free guarantee".

[0079] For a customer with potential intention but moderate acceptance, the system will not directly recommend "customer manager call", but will automatically arrange a step-by-step warming contact sequence:

[0080] First step (cultivation): push an article about "how to plan family finance" in the APP.

[0081] Second step (trigger): if the customer reads the article, a low-disturbance financial product recommendation banner will be displayed when he logs in a few days later.

[0082] Third step (conversion): if the customer clicks the banner, the system will push this high-potential clue to the customer manager, and attach the complete interaction path of the customer in advance, to assist artificial precise communication.

[0083] This "journey" will be adjusted in real time according to the feedback of customers at each step, realizing the mode transition from "one-off transaction" to "fine cultivation and conversion", and upgrading the single and static instruction of "generating marketing decision recommendation" to "automatic arrangement of adaptive marketing journey".

[0084] The marketing recommendation coefficient is calculated by calculating the previously calculated intention intensity and group fatigue and other indicators. The system automatically matches and generates a specific instruction according to the high and low value of this coefficient, and clearly defines which channel should be used, which product should be recommended, and even what kind of discount should be attached.

[0085] Embodiment five:

[0086] This embodiment mainly introduces the feedback module. According to the feedback of customers on marketing decisions, the marketing acceptance degree is quantitatively calculated and is included in the group fatigue. The individual dynamic score calculation adopts an event response type dynamic scoring model. Specifically, the feedback behavior is defined as clicking (positive feedback), applying (strong positive feedback), ignoring (negative feedback), and closing (strong negative feedback).

[0087] The basic score is set as follows: for example, application = +10 points, click = +2 points, ignore = -5 points, and close = -10 points.

[0088] Individual dynamic score calculation: customer G has accepted 5 times of marketing in the past 30 days:

[0089] First time (loan A): click (+2 points);

[0090] Second time (finance B): ignore (-5 points);

[0091] Third time (loan C): application success (+10 points);

[0092] Fourth time (credit card D): close (-10 points);

[0093] Fifth time (loan E): no feedback (decay);

[0094] The decay mechanism is that if the customer has no feedback within the preset time window (for example, 72 hours), the individual dynamic score will be decayed according to the tolerance of the latest marketing (for example, the customer manager telephone tolerance is higher than the APP pop-up window) at the corresponding rate;

[0095] Customer G has no feedback within 72 hours after the fifth marketing; if this marketing is an APP pop-up window (low tolerance), the individual score decay rate is 10% per day; if it is a call from a customer manager (high tolerance), the decay rate is 20% per day.

[0096] The group fatigue index is updated to establish a feedback value matrix, each element of which represents the benchmark weight of different feedback behaviors in a specific marketing context, as shown in Table 1 below:

[0097] Table 1: Feedback value matrix in this embodiment

[0098]

[0099] Individual dynamic score adjustment: when receiving feedback behavior, the adjustment amount of the individual dynamic score is determined by multiplying the base score by the benchmark weight corresponding to the current context in the feedback value matrix; customer G applies for loan C; if the base score is +10 and the benchmark weight of "loan marketing-application" is 0.9, then the individual dynamic score adjustment amount is 10*0.9=9 points; weighted average algorithm and aggregation statistics: using the weighted average algorithm, the weight corresponding to the individual dynamic score is applied to the marketing feedback data of all members in the customer group, and the group fatigue index is calculated by aggregation statistics;

[0100] Further, the weight function uses a reverse variant of the Sigmoid function, which reduces the negative impact on group fatigue when the individual dynamic score is higher than a certain threshold, and vice versa; the update process is as follows:

[0101] (1) Calculate the latest individual dynamic score of each active customer in the group;

[0102] (2) According to the Sigmoid reverse variant function, assign a weight to the individual dynamic score of the customer; the weight of high positive score (positive feedback) is low, and the impact on fatigue is small; the weight of low negative score (negative feedback) is high, and the impact on fatigue is large;

[0103] (3) Weighted average of these weighted individual feedback data, combined with historical group fatigue, to recalculate the updated group fatigue;

[0104] A "potential home buying demand customer" group has an initial fatigue of 0.905; if a large number of customers in this group have positive feedback (click or apply) on mortgage marketing, then the individual dynamic score of these positive feedback will have a lower weight, and after overall weighted average, the group fatigue will decrease, for example, to 0.85; on the contrary, if a large number of customers have negative feedback, the group fatigue will increase.

[0105] From the "feedback module" and "rule base" to the "reinforcement learning optimizer of marketing strategy", the "feedback module" for updating fatigue and the static "business rule base" for matching decisions are integrated and subverted to build a self-optimizing engine of marketing strategy based on reinforcement learning. The state is defined as the complete vector of the customer (including behavior, portrait, emotional acceptance, and financial health); the action is the set of marketing actions that the system can take (such as pushing a specific product combination through a specific channel at a specific time); the reward is the customer's feedback, a successful application is a huge positive reward, a click is a small positive reward, and an ignore or "no longer remind" is a negative reward;

[0106] The core of the system is no longer to execute fixed If-Then rules, but to autonomously explore what actions to take in what state to maximize the long-term cumulative reward (i.e., the sum of the bank's long-term business value and customer satisfaction) through continuous "trial-feedback-learning"; this method can discover complex strategies that human experts cannot summarize, such as "for a certain class of high-knowledge customers, directly recommending high-value products is not as effective as providing a deep industry analysis report first"; this allows the system to evolve from a passive rule execution tool to a brain that can autonomously learn and create optimal marketing strategies;

[0107] The above steps enable the entire system to learn and evolve. When the customer provides specific feedback on marketing, the system quantifies the value of this interaction, updates the individual dynamic score of the customer, and adjusts the overall fatigue index of the group to which the customer belongs through a weighting algorithm. Each successful marketing or failed attempt becomes an experience for the next decision.

[0108] The system integrates behavior representation, loan intent generation, fatigue calculation, decision recommendation, and feedback modules to build a comprehensive intelligent insight system for the banking industry. It not only deeply understands the real-time and long-term behavior patterns of customers, accurately identifies potential loan intentions, but also effectively manages customer marketing fatigue, thereby achieving personalized and efficient marketing recommendations. Through continuous feedback learning, the system can continuously optimize marketing strategies, improve customer experience and marketing conversion rates, and bring significant economic and social benefits to banks.

[0109] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A report-based intelligent insight system for data analysis in the banking marketing process, characterized in that: include: Behavioral representation module: Uses time-map units to capture short-term customer behavioral event streams and generate dynamic behavioral vectors; Long sequence units are used to analyze long-term customer behavior sequences and generate static behavior vectors. The dynamic behavior vector and the static behavior vector are input into a gated cross-attention fusion process, and a loan intent vector is generated by semantic projection. Loan Intent Generation Module: Integrates basic identity attributes, asset information and credit status to obtain a customer profile vector, and inputs it along with the loan intent vector into a multilayer perceptron network to generate loan intent strength; Fatigue calculation module: The customer profile vector is one-hot encoded and embedded, and then concatenated with the loan intent vector to form a customer panoramic state vector. The customer group is defined by K-Means clustering algorithm, and the group fatigue of the customer group is integrated based on the recent marketing feedback of the customer group according to the customer panoramic state vector. Decision recommendation module: Loan trend score is obtained from loan intention survey questionnaire. The marketing recommendation coefficient is calculated by combining the loan trend score, loan intention strength and group fatigue. Based on marketing recommendation coefficients and business rule base, marketing decision recommendations are generated.

2. The intelligent report insight system for data analysis in the banking marketing process according to claim 1, characterized in that, The behavior representation module includes: The short-term behavioral event stream includes clicks on recommended product pages within the application and records of financial instrument usage generated by the customer within minute-level and / or hour-level time windows. These are input into a time-graph network unit, where customers, products, channels, and counterparties are used as nodes to construct a heterogeneous graph. Through internal message passing and node memory updates, the temporal dependencies and sudden signals between short-term behavioral event streams are captured to obtain dynamic behavioral vectors. These dynamic behavioral vectors represent the customer's immediate state in the current network. The long-term behavioral sequence includes large-amount financial transactions spanning several months or even years in the customer's history, loan application and repayment records, and periodic asset allocation patterns. An analysis process based on the Transformer architecture is executed. This analysis process dynamically evaluates and weights the importance of each behavior in the historical sequence to the loan intent judgment through an internal self-attention mechanism, resulting in a static behavioral vector.

3. The intelligent report insight system for data analysis of the banking marketing process according to claim 1, characterized in that, The behavior representation module also includes: Using static and dynamic behavior vectors as input, a gated cross-attention fusion process is executed. The static behavior vector is used as the query to examine key information in the dynamic behavior vector. The gating unit amplifies / suppresses the part of the short-term behavior event stream most relevant to the long-term pattern to obtain the fused behavior vector. With the help of a pre-trained natural language processing model, the text describing the loan business is converted into semantic vectors. The average value of the semantic vectors is calculated to synthesize a fixed vector representing the meaning of the loan, and the fixed vector is used as the reference direction for intent judgment. Semantic projection is used to calculate the projection value of the fused behavior vector on the reference direction as the loan intent vector.

4. The intelligent report insight system for data analysis in the banking marketing process according to claim 1, characterized in that, The customer profile vector generated by the loan intent generation module includes: From multiple sources including the core transaction system, credit system, CRM system, and mobile banking backend, data is extracted, transformed, loaded, and synchronized in real time via ETL. Using the customer's identity as a unique identifier, the basic identity attributes, asset information, and credit status are obtained and integrated to construct a customer profile vector. The customer profile vector is then concatenated with the loan intent vector and input into a multilayer perceptron network to calculate and generate the loan intent strength.

5. The intelligent report insight system for data analysis in the banking marketing process according to claim 1, characterized in that, The fatigue calculation module includes: The static profile of the customer is one-hot encoded and embedded, and then concatenated with the loan intent vector to form a customer panoramic state vector. The K-Means clustering algorithm is used to perform real-time clustering operations on the panoramic state vectors of all customers in the bank. Customers with similar current behavior patterns, long-term habits, and basic attributes are dynamically grouped into the same customer group. For the customer group, the average click-through rate, application conversion rate, and ignore / close rate of similar loan marketing information are statistically analyzed within a preset evaluation period to calculate a fatigue score, which is then set as the group fatigue level.

6. The intelligent report insight system for data analysis of the banking marketing process according to claim 1, characterized in that, The decision recommendation module includes: The loan trend score is calculated based on the analysis of customer loan intention survey questionnaires; the loan intention strength is the output value of the loan intention generation module; and the group fatigue level is the output of the fatigue level calculation module. The marketing recommendation coefficient is calculated using a preset weighting formula, which assigns positive weights to the intensity of loan intent and loan trend score, and negative weights to group fatigue. The marketing recommendation coefficient value is compared with multi-level thresholds and combined with specific conditions in the customer profile vector. The optimal action strategy is retrieved and matched in the business rule base to obtain a structured instruction set. The instruction set defines the recommended reach channels, the specific recommended products, the suggested marketing script template ID, and the configurable discount strategies. The reach channels include customer manager calls, APP pop-ups, and SMS.

7. The intelligent report insight system for data analysis of the banking marketing process according to claim 1, characterized in that, The system also includes a feedback module for incorporating marketing acceptance levels into group fatigue levels, specifically including: An event-responsive dynamic scoring model is adopted, which calculates an individual dynamic score based on the customer's feedback behavior to marketing actions, including clicking, applying, ignoring / closing. If the customer does not provide feedback within a preset time window, the individual dynamic score will decay at a corresponding rate based on the tolerance of the most recent marketing. A weighted average algorithm is used to apply the weights corresponding to the individual dynamic scores, which are inverse variants of the Sigmoid function, to the marketing feedback data of all members within the customer group to establish a feedback value matrix. When feedback behavior is received, the adjustment amount of the individual dynamic score is determined by multiplying the base score by the benchmark weight in the feedback value matrix corresponding to the current situation. This quantitatively calculates the degree of marketing acceptance and incorporates the degree of marketing acceptance into the group fatigue level to complete the closed-loop feedback.

Images