Operations back-end management methods and management systems

Abstract

This invention discloses an operation backend management method and system, relating to the field of artificial intelligence technology. The method includes: collecting operation backend data to construct a heterogeneous behavior graph, deploying a graph embedding model, concatenating the output user embedding vector with the time-series state vector encoded by TCN, and inputting it into a pre-constructed risk scoring model to generate a user risk profile. Its key technical points are: using a double difference method to evaluate the net effect of the strategy, combining relevant data of the target class strategy to calculate a health score. When the health score is lower than a defined threshold, not only is the risk control effectiveness score of the relevant actions reduced, but the number of false alarm users is also fed back as a difficult sample and given a higher loss weight in the next round of training of the graph embedding model and the risk scoring model. This solves the problems of rigid traditional operation deployment and lack of effect verification, enabling the system to have continuous learning and self-repair capabilities.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an operational back-end management method and management system. Background Technology

[0002] Artificial intelligence (AI) is the technology that uses computers to simulate human intelligent behavior. Current mainstream technologies include machine learning, deep learning, and reinforcement learning. Machine learning automatically learns patterns from data through algorithms; typical methods include logistic regression, support vector machines, and random forests. Deep learning utilizes neural networks to process image, text, time-series, and graph-structured data. Reinforcement learning optimizes decision-making strategies through agent-environment interaction. Furthermore, interpretable AI is used to improve model transparency, and federated learning supports privacy-preserving collaborative training. These technologies have been widely applied in scenarios such as risk control, recommendation systems, intelligent customer service, and anomaly detection. Traditionally, the construction of backend management systems also requires the use of AI-related technologies, with machine learning and deep learning forming the core foundation. Currently, operational back-end systems generally rely on manually configured static rules or isolated thresholds for risk control and strategy intervention, leading to frequent misjudgments or slow responses in complex scenarios. For example, when faced with fraudulent order practices, rules such as "more than 10 orders per day" are insufficient to identify the use of multiple accounts and low-frequency fraudulent behavior. While blindly issuing coupons during promotional periods may boost GMV in the short term, it could lead to a breakdown in user experience or even compliance risks due to neglecting customer service manpower or user disturbance limits. More importantly, traditionally configured operational management systems often lack a scientific evaluation mechanism for the true effectiveness of strategies, failing to distinguish between effective strategies and natural growth, resulting in ineffective or even harmful strategies remaining online for extended periods. The root causes of these problems lie in data fragmentation, rigid decision-making, and a lack of feedback. Summary of the Invention

[0003] To achieve the above objectives, the present invention provides the following technical solution: Operations backend management method, which includes: Collect operational backend data to construct a heterogeneous behavior graph, deploy a graph embedding model, concatenate the output user embedding vector with the TCN-encoded time-series state vector, and input it into a pre-built risk scoring model to generate a user risk profile; Build a rules engine, define at least M strategy styles and their weight configurations, trigger action scoring and filtering strategies based on the derived business objective weight vector and combined with user risk profiles, select the action with the highest score as the effective action to be executed, and output structured strategy logs; Based on structured policy logs, the DID method is used to evaluate the net effect of the policy, and a strong correlation design model is triggered based on the evaluation results to generate a health score. The health score is then used to guide the adjustment of the feedback policy.

[0004] Furthermore, the operational backend data includes at least: user behavior events, device information, IP addresses, and product interactions; among which, user behavior events include at least: login, order placement, and payment failure; the heterogeneous behavior graph contains four types of nodes: user U, device D, IP address P, and product I, with edges representing interaction relationships.

[0005] Furthermore, the embedded vector e_U and the temporal state vector t_U are concatenated and input into the risk scoring model to form a user risk profile, which includes: the final user state representation s_U=[e_U;t_U]∈R 192 It also includes three quantifiable labels: risk score ri_s, behavioral consistency score be_c, and contextual credibility level cn_tr_level.

[0006] Furthermore, the risk score ri_s is based on the embedding vector e_U and the temporal state vector t_U. It is input into a lightweight binary classification neural network, trained offline using historical labeled data, and outputs the fraud probability, which is then converted into a risk score through linear mapping. In the historical labeled data, positive samples are confirmed fraud / brushing accounts, and negative samples are normal active users. The behavioral consistency score be_c is obtained by training an autoencoder model using the user's behavioral logs over the past 30 days to learn a low-dimensional representation of their normal behavior; the current 24-hour behavioral sequence is input into the autoencoder to calculate the reconstruction error, and the reconstruction error is mapped to the behavioral consistency score through normalization. The context trust level cn_tr_level is set by the rule engine: if the device / IP has never had a negative record, the context trust level cn_tr_level is high; if there are at most 2 minor anomalies, the context trust level cn_tr_level is medium; if there are more than 2 minor anomalies or other situations, the context trust level cn_tr_level is low.

[0007] Furthermore, when M is 3, three strategy styles are preset, and each strategy style corresponds to a set of business objective weight vectors W = (w1, w2, w3); where w1 represents business revenue, w2 represents risk control, and w3 represents user experience. The corresponding strategy styles include: conservative, balanced, and aggressive.

[0008] Furthermore, the triggered action scoring and filtering strategy process is as follows: A candidate action library A = {a1, a2, ..., a...} is set up and maintained. n}, for each action a i Calculate its action score based on: Score(a) i) = w1×Br_i + w2×Rs_i + w3×Es_i; where Br_i, Rs_i, and Es_i are the action a, respectively. i The expected business revenue score, risk control efficiency score, and user experience score are calculated. i represents the number of the corresponding action, and i = 1, 2, ..., n. The corresponding action includes at least: issuing coupons, limiting traffic, and transferring to human assistance.

[0009] Furthermore, the screening process includes at least a constraint filtering mechanism and a risk adaptation mechanism. The constraint filtering mechanism is based on removing actions that violate the current activation constraints. The types of activation constraints include at least resource constraints, compliance constraints, and experience constraints. The trigger condition for resource constraints is that the number of marketing SMS messages sent today in the current period is ≥100,000. The trigger condition for compliance constraints is that the user's actual age is <18. The trigger condition for experience constraints is that the user has received ≥3 marketing pushes within 7 days. The risk adaptation mechanism is based on the following: when the risk score ri_s exceeds the scoring threshold, a strong intervention action is executed: the weight of the risk control effectiveness score is increased according to the set tier value; when the context credibility level cn_tr_level is low and the risk score ri_s exceeds the scoring threshold, the review action is directly triggered.

[0010] Furthermore, the process of evaluating the net effect of the strategy using the difference-in-differences (DID) method is as follows: For each strategy style, a treatment group and a control group are divided, and the net effect of the strategy on indicator Y is calculated using DID; where, the treatment group consists of users affected by the target strategy; the control group consists of users not affected by the target strategy; the basis for calculating the net effect of indicator Y is: ΔY=(Y_Tafter-Y_Tbefore)-(Y_Cafter-Y_Cbefore); In the formula, Y_Tafter represents the average retention rate of the treatment group after T days of policy implementation, Y_Tbefore represents the average retention rate of the treatment group before T days of policy implementation, and Y_Cafter and Y_Cbefore represent the average retention rates of the control group after C days and before C days of policy implementation, respectively.

[0011] Furthermore, the process of triggering the strong association design model based on the evaluation results is as follows: using the DID result ΔY and target-type strategy-related data as input, the strong association design model is run to generate a health score H; among which, the target-type strategy-related data includes at least: the maximum retention rate of the control group, the total number of executions of the target-type strategy, the value corresponding to whether a compliance complaint occurred, and the number of users who were intervened but confirmed as normal; the basis for guiding the adjustment feedback strategy based on the health score is as follows: when H < H_th, the adjustment feedback strategy is as follows: the Rs_i corresponding to the effective action is attenuated, that is, the original Rs_i is multiplied by the screening factor, and the value range of the screening factor is [0, 1]; the original user behavior data in N_FP is simultaneously fed back to the graph embedding model and the risk scoring model, and used as a hard sample; where N_FP represents the number of false positive users.

[0012] The operation back-end management system includes: User profile generation module: Collects operational backend data to construct a heterogeneous behavior graph, deploys a graph embedding model, concatenates the output user embedding vector with the TCN-encoded time-series state vector, and inputs it into a pre-built risk scoring model to generate a user risk profile; Strategy generation module: Build a rule engine, define at least M strategy styles and their weight configurations, trigger action scoring and filtering strategies based on the derived business objective weight vector and combined with user risk profiles, select the action with the highest score as the effective action to be executed, and output structured strategy logs; Feedback Adjustment Module: Based on structured strategy logs, the DID method is used to evaluate the net effect of the strategy, and a strong correlation design model is triggered based on the evaluation results to generate a health score. The health score is then used to guide the adjustment of the feedback strategy.

[0013] This invention provides an operation backend management method and management system, which have the following beneficial effects: (1) This solution constructs a heterogeneous behavior graph and deploys a graph embedding model to aggregate the interaction relationships of four types of nodes: users, devices, IPs, and products, and generates embedding vectors containing contextual semantics. At the same time, it combines TCN to perform time-series encoding on the 24-hour behavior sequence to capture dynamic risk signals and finally obtain a multi-angle user risk profile, which improves the accuracy and interpretability of anomaly identification. This profile is not only the basic input for subsequent strategy decisions, but also provides a standardized benchmark for effect backtracking, solving the problems of high misjudgment rate and delayed response caused by traditional risk control relying on static rules or single indicators. (2) This solution sets a strategy style, with each style corresponding to a clear business objective weight configuration. Operations personnel can switch as needed. At the same time, the action selection is not only based on the weighted calculation of revenue, risk control, and experience scores obtained from historical A / B testing, but also forcibly embeds three types of constraint filtering mechanisms: resources, compliance, and experience. In high-risk scenarios, the risk control weight is strengthened to ensure that the strategy is both flexible and adaptable to the business stage and safe and controllable. This overcomes the shortcomings of traditional rule engines that cannot balance revenue, risk control, and experience and are prone to violating resource or compliance restrictions. (3) This scheme uses the double difference method to evaluate the net effect of the strategy. Combined with the relevant data of the target class strategy, the health score is calculated. When the health score is lower than the defined threshold, not only is the risk control effectiveness score of the relevant action reduced, but the number of false alarm users is also sent back and used as a difficult sample. In the next round of training of the graph embedding model and the risk scoring model, a higher loss weight is given. This solves the problem of the traditional operation deployment being fixed and lacking effect verification, and enables the system to have continuous learning and self-repair capabilities. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the overall steps of the operation backend management method in this invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0016] Example 1: Please see Figure 1 This embodiment provides an operation backend management method, which includes the following specific steps: S1. Collect operational backend data to construct a heterogeneous behavior graph, deploy a graph embedding model, concatenate the output user embedding vector with the TCN-encoded time-series state vector, and input it into a pre-built risk scoring model to generate a user risk profile. The operational backend data includes user behavior events, device information, IP addresses, and product interactions. User behavior events include login, order placement, and payment failure. The heterogeneous behavior graph contains four types of nodes: user U, device D, IP address P, and product I, with edges representing interaction relationships. For example, UD represents a user logging in using a certain device. The graph embedding model deployed in this embodiment can employ a lightweight graph neural network, such as GraphSAGE, to generate an embedding vector e_U∈R for each user node. 128This embedding vector not only includes the user's own history, but also aggregates the characteristics of their associated devices, IPs, and products, thereby automatically identifying abnormal patterns. Where R, as well as R mentioned below, represents the set of real numbers; The temporal state vector representation using TCN encoding: For the user's behavioral event stream within the last 24 hours, ordered by time, a temporal convolutional network (TCN) is used to encode it into a fixed-length temporal state vector t_U∈R. 64 TCN can effectively capture risk signals, which include behavioral jumps and frequency abrupt changes. Behavioral jumps are represented by, for example, suddenly attempting to make a large payment after browsing low-priced goods. Frequency abrupt changes are represented by, for example, multiple failed card binding attempts within 1 minute. A specific example is: if a user has no activity in the past 23 hours, and then fails to bind 8 bank cards in a row from a new device and a new IP in the 24th hour, the t_U output by the TCN will deviate significantly from the normal distribution, thus triggering a risk signal. In this embodiment, the risk scoring model can employ a lightweight multilayer perceptron (MLP), and its operation is based on: The embedded vector e_U and the temporal state vector t_U are concatenated and input into a lightweight multilayer perceptron (MLP). The resulting user risk profile includes: the final user state representation s_U = [e_U; t_U] ∈ R. 192 It also includes three quantifiable labels: risk score ri_s, behavioral consistency score be_c, and contextual credibility level cn_tr_level. Wherein, the risk score ri_s represents: This comprehensively reflects the degree of suspicion of a user's current behavior in terms of fraud, abuse, and abnormal operations; the higher the corresponding value, the greater the risk. Based on the embedding vector e_U and the temporal state vector t_U, a lightweight binary classification neural network, such as a two-layer fully connected network, is input. This network is trained offline using historical labeled data, and the output fraud probability is converted into a risk score through linear mapping. Among them, in the historical labeled data, positive samples are confirmed fraud / brushing accounts, and negative samples are normal active users. The behavioral consistency score be_c represents: The system measures the degree of deviation between a user's current behavior sequence and its historical behavior patterns. The closer the value is to 0, the more inconsistent and suspicious the behavior is. An autoencoder model is trained using the user's behavior logs over the past 30 days to learn a low-dimensional representation of their normal behavior. The current 24-hour behavior sequence is input into the autoencoder to calculate the reconstruction error ε. The error is then normalized and mapped to a behavior consistency score. Context trust level cn_tr_level: Build a device / IP reputation database: Record the number of negative events associated with each device ID or IP address in the past 90 days, such as being reported, triggering risk control, and being associated with black market activities; Configure the rule engine: If a device / IP has never had a negative record, the context trust level cn_tr_level is set to high; if there are at most two minor anomalies, such as logging in from a different location but the transaction is successful, the context trust level cn_tr_level is set to medium; if there are more than two minor anomalies or other situations, the context trust level cn_tr_level is set to low; It should be noted that the above high, medium, and low are corresponding labels and do not require specific numerical settings; In summary, the three tags described above characterize the user's status from three dimensions: overall risk probability, degree of behavioral deviation, and environmental credibility. They complement each other, avoid bias from a single indicator, require no manual intervention during generation, and have a standardized format. They can be directly parsed or traced back by the supporting engine in subsequent solutions, achieving the effect of data reusability.

[0017] This solution constructs a heterogeneous behavior graph and deploys a graph embedding model to aggregate the interaction relationships of four types of nodes: users, devices, IPs, and products, generating embedding vectors containing contextual semantics. Simultaneously, it combines TCN to perform time-series encoding on 24-hour behavior sequences, capturing dynamic risk signals, and ultimately derives a multi-faceted user risk profile, improving the accuracy and interpretability of anomaly identification. This profile not only serves as the basic input for subsequent strategy decisions but also provides a standardized benchmark for effect backtesting, solving the problems of high misjudgment rates and delayed responses caused by traditional risk control relying on static rules or single indicators.

[0018] S2. Build a rules engine, define at least M strategy styles and their weight configurations, and based on the obtained business target weight vector and combined with the user risk profile output in S1, trigger the action scoring and filtering strategy, select the action with the highest score as the effective action to be executed, and output the structured strategy log. In this embodiment, the value of M is 3; Three strategy styles are preset, each corresponding to a set of business objective weight vectors W = (w1, w2, w3), which represent the degree of importance attached to the following objectives, where w1 represents business revenue, such as GMV and conversion rate; w2 represents risk control, such as fraud prevention and loss avoidance; and w3 represents user experience, such as interruption frequency and operation smoothness. Therefore, in this embodiment, the strategy styles include: conservative, balanced, and aggressive. The weight configuration can be found in Table 1 below: Table 1: Correspondence between strategy style and business objective weight vector:

[0019] The specific mechanism is as follows: Operations personnel can select a style from the drop-down menu on the strategy center page in the backend, and the system will load the corresponding weight in real time; The triggered action scoring and filtering strategy process is as follows: Define and maintain a candidate action library A = {a1, a2, ..., a...} n}, for each action a i Calculate its action score based on: Score(a) i ) = w1×Br_i + w2×Rs_i + w3×Es_i; where Br_i, Rs_i, and Es_i are the action a, respectively. i The expected business revenue score, risk control efficiency score, and user experience score are calculated. Specifically, Br_i is derived by normalizing the historical A / B test average to [0, 1]. Rs_i is 0.9 if "freeze" is displayed, 0.3 if "reminder" is displayed, and 0.6 otherwise. Es_i is 0.9 if "do not disturb" is displayed, 0.3 if "force verification" is displayed, and 0.6 otherwise. It should be noted that the values ​​are based on historical results and business experience. Rs_i is high because of strong risk control during freezing, and Es_i is high because of the superior experience of "do not disturb". For others, the middle value of 0.6 is used to balance the conservative estimate of unknown actions. i represents the number of the corresponding action, and i = 1, 2, ..., n. The corresponding actions include: issuing coupons, limiting traffic, and transferring to human assistance, etc. The screening process should include at least a constraint filtering mechanism and a risk adaptation mechanism. The constraint filtering mechanism is based on the following: Actions that violate the current activation constraints are removed. These constraints include at least three types: resource constraints, compliance constraints, and experience constraints. Specifically, the trigger condition for resource constraints is: the number of marketing SMS messages sent today within the current period is ≥100,000, which prevents resource overspending. The trigger condition for compliance constraints is: the user's actual age is <18. The trigger condition for experience constraints is: the user has received ≥3 marketing push notifications within 7 days. The trigger conditions can be set according to actual needs; the above are the corresponding constraints used in this embodiment. The risk adaptation mechanism is based on the following: When the risk score ri_s exceeds the scoring threshold, a strong intervention action is executed: the weight of the risk control effectiveness score is increased according to the set tier value; even if the conservative type is selected, risk control is given priority; the set tier value is selected as 0.1 or 0.2, etc. If it is actually increased, the value of other weights needs to be reduced by the same amount to ensure that the sum of the weights is 1; for example: if w1 is increased by 0.1, then w2 and w3 both need to be reduced by 0.05; when the context trust level cn_tr_level is low and the risk score ri_s exceeds the scoring threshold, the review action is directly triggered and sent for manual review. The output structured policy log should include at least: user_id: A unique identifier for the user; profile_snapshot: User risk profile; selected_action: The final action to be executed; strategy_style: The selected corresponding strategy style; weights_used: The actual target weight vector used, i.e., w1, w2, w3; constraint_violations_checked: List of constraints checked and results.

[0020] This solution sets out strategy styles, each with a clearly defined weighted business objective. Operations personnel can switch between them as needed. Action selection is not only based on a weighted calculation of revenue, risk control, and user experience scores derived from historical A / B testing, but also forcibly embeds resource, compliance, and user experience constraint filtering mechanisms. Furthermore, it strengthens risk control weights in high-risk scenarios, ensuring that the strategy is both flexible and adaptable to different business stages while remaining secure and controllable. The input to this mechanism directly relies on the user risk profile in S1, and the output forms a structured log, providing a unique and reliable data source for S3 assessment. This overcomes the shortcomings of traditional rule engines, which cannot balance revenue, risk control, and user experience, and are prone to violating resource or compliance restrictions.

[0021] S3. Based on structured strategy logs, the net effect of the strategy is evaluated using the difference-in-differences (DID) method. The evaluation results trigger a strong correlation design model to generate a health score, which guides the adjustment of the feedback strategy. The process of evaluating the net effect of the strategy using the difference-in-differences (DID) method is as follows: For each strategy style, a treatment group and a control group are divided, and the net impact on indicator Y is calculated using the DID strategy; where, in this embodiment, indicator Y represents the 7-day retention rate. Processing group: Users affected by the target class strategy; Control group: Users not affected by the target strategy; It should be noted that users affected by the target-type policy are the set of users who meet the condition ri_s≥Q1 and actually perform actions in S2 within a preset time period T, denoted as T_group; where actions include traffic limiting, etc.; users not affected by the target-type policy are users who meet the condition ri_s≥Q1 but did not perform actions due to random traffic allocation or resource constraints within the same time period. They are selected from all users through propensity score matching (PSM) to ensure that their distribution is consistent with T_group in terms of age, region, historical behavior, etc., and are denoted as C_group. The basis for calculating the net impact of indicator Y is as follows: ΔY=(Y_Tafter-Y_Tbefore)-(Y_Cafter-Y_Cbefore); In the formula, Y_Tafter represents the average retention rate of the treatment group T days after the strategy is implemented, and Y_Tbefore represents the average retention rate of the treatment group T days before the strategy is implemented; Y_Cafter and Y_Cbefore represent the average retention rates of the control group C days after the strategy is implemented and C days before the strategy is implemented, respectively; it should be noted that DID eliminates time trends and inherent user differences, and only retains the incremental effect brought by the strategy; the values ​​of C and T can be set according to actual needs; The process of triggering a strongly correlated design model based on the evaluation results is as follows: Using the DID result ΔY and target class policy-related data as input, a strongly correlated design model is run, based on the following: ; In the formula, H represents the generated health score, Y_ba is the maximum retention rate of the control group, used for normalization; Fpc is the number of users who were intervened but confirmed as normal, obtained from customer service complaints or behavior retrospectives; Tac is the total number of times the target strategy was executed; I_cf indicates whether a compliance complaint occurred, and the value must satisfy: if = 1, no = 0; q1, q2 and q3 are the corresponding weight coefficients, and the value range is [0, 1]; it should be noted that: (ΔY / Y_ba) is the benefit gain, (1-Fpc / Tac) is the false positive control, and the required health score H can be obtained under their combined effect; Therefore, the relevant data for the target strategy are: the maximum retention rate of the control group, the total number of times the target strategy was executed, the value corresponding to whether a compliance complaint occurred, and the number of users who were intervened but confirmed as normal users; The basis for using health scores to guide adjustments to feedback strategies is as follows: When H < H_th, the adjustment feedback strategy is as follows: the Rs_i corresponding to the effective action in S2 is decayed, that is, the original Rs_i is multiplied by the screening factor, the value range of which is [0, 1], and in this embodiment it is 0.85, with a lower limit of 0.2 to avoid zeroing; otherwise, no response is made; the original user behavior data in N_FP is synchronously transmitted back to the graph embedding model and risk scoring model in S1 and used as hard samples; where N_FP represents the number of false positive users, specifically referring to: the set of individuals who were intervened by the S2 strategy, such as traffic limiting or freezing, but were later confirmed as normal users, and these users were wrongly marked as high risk due to model misjudgment; in S1, when the risk scoring model MLP and graph embedding model GNN are trained next time, these samples are marked as hard samples and given higher weights in the loss function, such as 2.0, to strengthen the model's ability to learn easily confused cases; This solution uses a difference-in-differences method to evaluate the net effect of the strategy. It combines relevant data of the target class strategy to calculate a health score. When the health score is lower than the defined threshold, not only is the risk control effectiveness score of the relevant actions in S2 reduced, but the number of false alarm users is also sent back to S1 as a hard sample. In the next round of training of the graph embedding model and the risk scoring model, it is given a higher loss weight. As a result, the evaluation results of S3 optimize the feature extraction capability of S1 and the strategy library of S2, which solves the problems of rigid deployment and lack of effect verification in traditional operation and makes the system have continuous learning and self-repair capabilities.

[0022] Additionally, H_th represents the threshold value, which is 0.4 in this embodiment; In summary, the profile of S1 is not only used for decision-making but also serves as the benchmark for S3 evaluation. The constraint mechanism of S2 ensures policy security, and its logs provide analytical basis for S3. The evaluation results of S3 optimize the S1 model and S2 policy library in reverse, forming a closed-loop solution of data, decision-making, evaluation, and evolutionary feedback, achieving a synergistic effect of 1+1>2. The operation of the overall solution is not only more intelligent but also has the effect of being interpretable and sustainably evolving.

[0023] Example 2: Based on Example 1, this embodiment also provides an operation back-end management system, which includes: User profile generation module: Collects operational backend data to construct a heterogeneous behavior graph, deploys a graph embedding model, concatenates the output user embedding vector with the TCN-encoded time-series state vector, and inputs it into a pre-built risk scoring model to generate a user risk profile; Strategy generation module: Build a rule engine, define at least M strategy styles and their weight configurations, trigger action scoring and filtering strategies based on the derived business objective weight vector and combined with user risk profiles, select the action with the highest score as the effective action to be executed, and output structured strategy logs; Feedback Adjustment Module: Based on structured strategy logs, the DID method is used to evaluate the net effect of the strategy, and a strong correlation design model is triggered based on the evaluation results to generate a health score. The health score is then used to guide the adjustment of the feedback strategy.

[0024] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0025] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0026] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An operational backend management method, characterized in that, The method includes: Data from the operational backend is collected to construct a heterogeneous behavior graph. A graph embedding model is deployed, and the output user embedding vector is concatenated with the TCN-encoded temporal state vector. This concatenation is then input into a pre-built risk scoring model to generate a user risk profile. The embedding vector e_U is concatenated with the temporal state vector t_U and input into the risk scoring model. The resulting user risk profile includes: the final user state representation s_U = [e_U; t_U] ∈ R. 192 R represents the set of real numbers, and is accompanied by three quantifiable labels: risk score ri_s, behavioral consistency score be_c, and contextual credibility level cn_tr_level. Build a rules engine, define at least M strategy styles and their weight configurations, trigger action scoring and filtering strategies based on the derived business objective weight vector and combined with user risk profiles, select the action with the highest score as the effective action to be executed, and output structured strategy logs; Based on structured policy logs, the DID method is used to evaluate the net effect of the policy, and a strong correlation design model is triggered based on the evaluation results to generate a health score. The health score is then used to guide the adjustment of the feedback policy. The process of evaluating the net effect of a strategy using the difference-in-differences (DID) method is as follows: For each strategy style, a treatment group and a control group are divided, and the net effect of the strategy on indicator Y is calculated using DID; where, the treatment group consists of users affected by the target strategy; and the control group consists of users not affected by the target strategy; the basis for calculating the net effect of indicator Y is: ΔY=(Y_Tafter-Y_Tbefore)-(Y_Cafter-Y_Cbefore); In the formula, Y_Tafter represents the average retention rate of the treatment group after T days of policy implementation, Y_Tbefore represents the average retention rate of the treatment group before T days of policy implementation; Y_Cafter and Y_Cbefore represent the average retention rates of the control group after C days and before C days of policy implementation, respectively. The process of triggering the strong association design model based on the evaluation results is as follows: Using the DID result ΔY and target-type strategy-related data as input, the strong association design model is run to generate a health score H. The target-type strategy-related data includes at least: the maximum retention rate of the control group, the total number of executions of the target-type strategy, the value corresponding to whether a compliance complaint occurred, and the number of users who were intervened but confirmed as normal. The basis for guiding the adjustment feedback strategy based on the health score is as follows: when H < H_th, the adjustment feedback strategy is: the Rs_i corresponding to the effective action is attenuated, i.e., the original Rs_i is multiplied by a screening factor, and the screening factor's value range is [0, 1]. Simultaneously, the original user behavior data in N_FP is fed back to the graph embedding model and the risk scoring model, and used as a hard sample. N_FP represents the number of false positive users.

2. The operation backend management method according to claim 1, characterized in that: The operational backend data includes at least: user behavior events, device information, IP address, and product interaction; among which, user behavior events include at least: login, order placement, and payment failure; the heterogeneous behavior graph contains four types of nodes: user U, device D, IP address P, and product I, with edges representing interaction relationships.

3. The operation backend management method according to claim 2, characterized in that: The risk score ri_s is based on the embedding vector e_U and the temporal state vector t_U. It is input into a lightweight binary classification neural network, trained offline using historical labeled data, and outputs the fraud probability, which is then converted into a risk score through linear mapping. In the historical labeled data, positive samples are confirmed fraud / brushing accounts, and negative samples are normal active users. The behavioral consistency score be_c is obtained by training an autoencoder model using the user's behavioral logs over the past 30 days to learn a low-dimensional representation of their normal behavior; the current 24-hour behavioral sequence is input into the autoencoder to calculate the reconstruction error, and the reconstruction error is mapped to the behavioral consistency score through normalization. The context trust level cn_tr_level is set by the rule engine: if the device / IP has never had a negative record, the context trust level cn_tr_level is high; if there are at most 2 minor anomalies, the context trust level cn_tr_level is medium; if there are more than 2 minor anomalies or other situations, the context trust level cn_tr_level is low.

4. The operation backend management method according to claim 1, characterized in that: When M is 3, three strategy styles are preset, and each strategy style corresponds to a set of business objective weight vectors W = (w1, w2, w3); where w1 represents business revenue, w2 represents risk control, and w3 represents user experience. The corresponding strategy styles include: conservative, balanced, and aggressive.

5. The operation back-end management method according to claim 4, characterized in that: The triggered action scoring and filtering strategy process is as follows: A candidate action library A = {a1, a2, ..., a...} is set up and maintained. n }, for each action a i Calculate its action score based on: Score(a) i ) = w1×Br_i + w2×Rs_i + w3×Es_i; where Br_i, Rs_i, and Es_i are the action a, respectively. i The expected business revenue score, risk control efficiency score, and user experience score are calculated. i represents the number of the corresponding action, and i = 1, 2, ..., n. The corresponding action includes at least: issuing coupons, limiting traffic, and transferring to human assistance.

6. The operation back-end management method according to claim 3, characterized in that: The screening process should include at least a constraint filtering mechanism and a risk adaptation mechanism. The constraint filtering mechanism is based on removing actions that violate the current activation constraints. The types of activation constraints include at least resource constraints, compliance constraints, and experience constraints. The trigger condition for resource constraints is that the number of marketing SMS messages sent today in the current period is ≥100,000. The trigger condition for compliance constraints is that the user's actual age is <18. The trigger condition for experience constraints is that the user has received ≥3 marketing pushes within 7 days. The risk adaptation mechanism is based on the following: when the risk score ri_s exceeds the scoring threshold, a strong intervention action is executed: the weight of the risk control effectiveness score is increased according to the set tier value; when the context credibility level cn_tr_level is low and the risk score ri_s exceeds the scoring threshold, the review action is directly triggered.

7. An operation back-end management system, characterized in that, The system includes: User profile generation module: Collects operational backend data to construct a heterogeneous behavior graph, deploys a graph embedding model, concatenates the output user embedding vector with the TCN-encoded temporal state vector, and inputs it into a pre-built risk scoring model to generate a user risk profile; concatenates the embedding vector e_U with the temporal state vector t_U and inputs it into the risk scoring model, forming a user risk profile that includes: the final user state representation s_U=[e_U;t_U]∈R 192 R represents the set of real numbers, and is accompanied by three quantifiable labels: risk score ri_s, behavioral consistency score be_c, and contextual credibility level cn_tr_level. Strategy generation module: Build a rule engine, define at least M strategy styles and their weight configurations, trigger action scoring and filtering strategies based on the derived business objective weight vector and combined with user risk profiles, select the action with the highest score as the effective action to be executed, and output structured strategy logs; Feedback Adjustment Module: Based on structured strategy logs, the DID method is used to evaluate the net effect of the strategy, and a strong correlation design model is triggered based on the evaluation results to generate a health score, which is then used to guide the adjustment of the feedback strategy. The process of evaluating the net effect of a strategy using the difference-in-differences (DID) method is as follows: For each strategy style, a treatment group and a control group are divided, and the net effect of the strategy on indicator Y is calculated using DID; where, the treatment group consists of users affected by the target strategy; and the control group consists of users not affected by the target strategy; the basis for calculating the net effect of indicator Y is: ΔY=(Y_Tafter-Y_Tbefore)-(Y_Cafter-Y_Cbefore); In the formula, Y_Tafter represents the average retention rate of the treatment group after T days of policy implementation, Y_Tbefore represents the average retention rate of the treatment group before T days of policy implementation; Y_Cafter and Y_Cbefore represent the average retention rates of the control group after C days and before C days of policy implementation, respectively. The process of triggering the strong association design model based on the evaluation results is as follows: Using the DID result ΔY and target-type strategy-related data as input, the strong association design model is run to generate a health score H. The target-type strategy-related data includes at least: the maximum retention rate of the control group, the total number of executions of the target-type strategy, the value corresponding to whether a compliance complaint occurred, and the number of users who were intervened but confirmed as normal. The basis for guiding the adjustment feedback strategy based on the health score is as follows: when H < H_th, the adjustment feedback strategy is: the Rs_i corresponding to the effective action is attenuated, i.e., the original Rs_i is multiplied by a screening factor, and the screening factor's value range is [0, 1]. Simultaneously, the original user behavior data in N_FP is fed back to the graph embedding model and the risk scoring model, and used as a hard sample. N_FP represents the number of false positive users.

Images