A method and system for active intervention decision-making based on recharging entropy

Abstract

The present application relates to a kind of active intervention decision method and system based on recharging entropy in intelligent payment technical field.It is method that includes: establishing the decision information set based on time series, define the action space of user decision, the action space at least includes immediate recharging and non-immediate recharging, and the conditional probability of each action is calculated;Using information entropy theory, calculate recharging entropy;Real-time monitoring the change characteristics of the recharging entropy, when detecting that the change characteristics of recharging entropy satisfy preset trigger condition, reach pre-trigger condition;In preset future time window, find the moment that makes the net intervention value reach maximum as the best intervention opportunity, and generate active intervention instruction.The present application can be based on the accurate perception to user immediate psychological state, ensure that system only intervenes in the moment that user needs most auxiliary decision.

Description

Technical Field

[0001] This invention relates to the field of smart payment technology. More specifically, this invention relates to a proactive intervention decision-making method and system based on recharge entropy. Background Technology

[0002] With the widespread adoption of mobile payment technology and its comprehensive coverage of digital life scenarios, e-wallets and virtual accounts have become core carriers for users' daily economic activities. Whether it's high-frequency transportation (such as subways and ride-hailing services), immediate daily consumption, or continuous digital entertainment subscriptions, the adequacy and availability of account funds directly affect service continuity and user experience quality. To maintain account liquidity, existing asset management systems generally introduce automatic top-up or low-balance alert functions, attempting to provide timely replenishment of funds when users' balances are insufficient.

[0003] Currently, the most mainstream trigger decision technology in the industry is a passive trigger mechanism based on a fixed numerical threshold. The operating logic of this technology is as follows: the system continuously monitors the user's account balance in the background, and only when the balance falls below a certain absolute value preset by the user or defaulted by the system, such as 10 yuan or 50 yuan, does it generate a recharge instruction or reminder notification. This mechanism is simple in logic and easy to implement, but when faced with complex and ever-changing real-world application scenarios, it exposes serious timeliness defects and a disconnect between timeliness and user experience, specifically manifested in the following three aspects: First, in high-frequency or high-pressure scenarios, the fixed threshold mechanism has a significant response lag, which can easily lead to the interruption of critical services.

[0004] For example, existing technologies only focus on the static characteristic of the current balance, ignoring dynamic features such as the rate of fund depletion. In many real-world scenarios, users may make multiple consecutive transactions within a short period, or make a single large transaction. For instance, when a user takes multiple subway rides or a long-distance taxi, although the initial balance may be higher than a preset threshold (e.g., a balance of 20 yuan exceeding the threshold of 10 yuan), the system determines that no intervention is needed. However, once the trip ends or consecutive charges occur, the balance will instantly be depleted and become negative. Because the system failed to intervene in time, the user may be stopped at the gate or unable to pay the fare. Therefore, this triggering logic cannot meet the stringent requirements of fund continuity in real-time service scenarios.

[0005] Secondly, during periods of low anxiety or inactivity, the fixed threshold mechanism lacks context awareness, which can easily lead to ineffective disturbances and over-intervention.

[0006] Existing systems typically decouple their trigger decisions from the user's activity level. Whenever the balance falls below a threshold, regardless of whether the user is resting late at night, in an important meeting, or simply experiencing a natural deduction due to prolonged account inactivity (such as monthly fee deduction), the system indiscriminately pops up a recharge reminder or performs an automatic deduction. In these scenarios, users often don't have an immediate need for funds; the system's intervention not only fails to create value but also intrudes on the user's attention and may even trigger unnecessary anxieties about the safety of their funds. Therefore, this trigger logic leads to an extremely high user closure rate.

[0007] Finally, existing technologies lack quantitative assessment methods for the uncertainty of users' decision-making intentions.

[0008] Specifically, the human payment decision-making process is not a black-and-white binary logic, but a process full of probability and psychological game. When the balance is in an intermediate state (neither too much nor too little), users are often in a period of hesitation, possibly considering whether to switch payment methods, whether to wait for a refund, or whether to top up. Existing rule engines usually use hard if-then rules, which cannot understand this ambiguous psychological state. The system either intervenes too early (forcibly popping up a window when the user is hesitant, causing resentment) or too late (the system still has not responded when the user has already decided to top up and is anxiously looking for the entry point). Therefore, existing systems can never find the optimal golden balance between what the user needs most and what the user least dislikes.

[0009] In summary, existing asset management technologies are constrained by the rigid logic of static thresholds in selecting the timing of interventions, making them unable to adapt to dynamically changing consumption rates and user psychological intentions. How to construct a method that can perceive multi-dimensional contexts in real time, accurately quantify decision-making intentions, and dynamically find the optimal intervention time based on the principle of value maximization is a key technical problem that urgently needs to be solved to improve the intelligence level of asset management. Summary of the Invention

[0010] To address the aforementioned challenge of constructing a method capable of real-time perception of multidimensional context, accurate quantification of decision-making intent, and dynamic search for the optimal intervention moment based on the principle of maximizing value, this invention discloses a proactive intervention decision-making method and system based on recharge entropy.

[0011] In a first aspect, this invention discloses a proactive intervention decision-making method based on recharge entropy. This method operates on the user terminal device side and includes: establishing a time-series-based decision information set, which includes user account balance change characteristics and intent characteristics reflecting user intent; defining a user decision action space, which includes at least immediate recharge and non-immediate recharge; calculating the conditional probability of each action in the action space based on the decision information set; using information entropy theory, calculating recharge entropy based on the conditional probability of each action in the action space to quantify the uncertainty of the user's current decision intent; monitoring the change characteristics of the recharge entropy in real time, and achieving a pre-trigger condition when the change characteristics of the recharge entropy meet a preset trigger condition; calculating the net intervention value in response to achieving the pre-trigger condition, where the net intervention value is the difference between the marginal benefit of proactive intervention and the user disturbance cost; and finding the moment within a preset future time window where the net intervention value reaches its maximum value as the optimal intervention opportunity and generating a proactive intervention instruction.

[0012] Further, the calculation of the recharge entropy includes: for an action in the action space, summing the product of the conditional probability of the action and the logarithm of the conditional probability, and taking the negative value of the summation result to obtain the current recharge entropy.

[0013] Further, calculating the conditional probability of each action in the action space includes: constructing a comprehensive cost function, which is calculated by weighting at least based on the risk of business failure and the cost of time delay; for an action in the action space, combining the prior probability of the action with the comprehensive cost function, and using a Bayesian network or logistic regression model to calculate the conditional probability of the action.

[0014] Furthermore, the preset triggering conditions include a combination of one or more of the following conditions: the recharge entropy is greater than a set high threshold for recharge entropy; the absolute value of the recharge entropy change rate is greater than a set threshold for the recharge entropy change rate; and the recharge entropy change rate is negative.

[0015] Furthermore, the calculation of the marginal benefit in the net intervention value includes: defining an expected utility function, which is obtained by multiplying the scenario severity coefficient by the success probability of real-time prediction; and calculating the difference between the expected utility generated by immediate intervention at the current moment and the expected utility generated by intervention postponed to the next moment, based on the expected utility function, as the marginal benefit of intervention.

[0016] Furthermore, the scenario severity coefficient is determined based on the current business scenario type. The scenario severity coefficient corresponding to a high-risk scenario is higher than that corresponding to a low-risk scenario. The high-risk scenarios include at least ride-hailing scenarios or gate passage scenarios.

[0017] Furthermore, the user disturbance cost in the calculation of net intervention value includes: the user disturbance cost is related to the current time period and the current attention occupancy, wherein the user disturbance cost is greater during the preset rest period than during the non-rest period; the user disturbance cost is greater when the user is focused on running the application or when the screen is always on than when the user is not focused on running the application or when the screen is not always on.

[0018] Furthermore, within a preset future time window, the optimal intervention time is determined by finding the moment when the net intervention value reaches its maximum value. This includes: traversing each time point within the preset future time window; calculating the net intervention value for each time point; and selecting the time point with the maximum net intervention value as the optimal intervention time.

[0019] Furthermore, the intent features include one or more of the following: the event flow context of the current scenario, the user's historical payment behavior features, the amount of refund pending payment, and the credit limit; correspondingly, the action space also includes one or more of waiting for the refund to be credited and switching to the credit limit.

[0020] On the other hand, the present invention discloses an active intervention decision-making system based on recharge entropy, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the active intervention decision-making method based on recharge entropy as described above.

[0021] The beneficial effects of this invention are as follows: This invention achieves precise intervention timing. Existing technologies typically rely on a fixed absolute balance value to trigger recharges, failing to perceive the user's true level of urgency in different scenarios. This invention innovatively introduces information entropy theory to calculate recharge entropy, which can quantify the uncertainty of the user's decision-making intention. By monitoring changes in entropy values ​​in real time, the system can keenly capture the moment when the user's psychological phase transitions from hesitation to clear intention. This allows intervention to be based on a precise perception of the user's immediate psychological state, ensuring that the system intervenes only when the user most needs assistance in decision-making.

[0022] This invention proposes a net intervention value function, modeling the decision-making process as a game between the marginal benefit of intervention and the cost of user disruption. The system not only determines whether a top-up is needed, but also solves the problem of when to trigger the least annoying pop-up by finding the maximum value within a future time window. This mechanism effectively avoids unnecessary disruptions during late-night rest or when highly focused, while ensuring timely intervention in high-risk scenarios such as taxi rides and turnstiles, significantly improving the user's payment experience and the system's intelligence.

[0023] The method of this invention operates entirely on the user's terminal device. All sensitive data (such as geographic location clustering, calendar events, and payment habits) is collected, processed, and destroyed locally, without needing to be uploaded to the cloud, thus completely eliminating the risk of user privacy leakage. Furthermore, because it does not rely on network communication for decision-making, the system can maintain millisecond-level decision response speeds even in weak network or congested environments, ensuring high availability of asset management services. Attached Figure Description

[0024] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the selection of pre-triggering and active intervention timing according to an embodiment of the present invention; Figure 3 This is a system schematic diagram according to an embodiment of the present invention. Detailed Implementation

[0025] 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, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] This invention provides a proactive intervention decision-making method based on recharge entropy. This method operates entirely on the user's terminal device (such as a smartphone, tablet, or wearable device), aiming to solve the technical problems of existing technologies that rely on fixed balance thresholds, leading to rigid intervention timing and an inability to perceive the user's true urgency. This embodiment quantifies decision uncertainty by introducing information entropy theory and combines it with a marginal benefit model to find the optimal intervention timing, achieving a technological leap from passive response to proactive perception.

[0028] like Figure 1 As shown, the specific embodiments of the present invention will be described in detail below with reference to the method steps described in claim 1.

[0029] This method operates on the user terminal device side and includes: S1. Establish a time-series-based decision information set, which includes the balance change characteristics of user accounts and intent characteristics to reflect user intentions; define the action space for user decisions, which includes at least immediate top-up and non-immediate top-up. S2, based on the decision information set, calculate the conditional probability of each action in the action space; S3. Using information entropy theory, calculate the recharge entropy based on the conditional probability of each action in the action space to quantify the uncertainty of the user's current decision intention; monitor the change characteristics of the recharge entropy in real time, and when the change characteristics of the recharge entropy meet the preset triggering conditions, the pre-triggering condition is achieved. S4, in response to the achievement of the pre-triggering condition, calculate the net intervention value, which is the difference between the marginal benefit of proactive intervention and the user disturbance cost; S5. Within a preset future time window, find the moment when the net intervention value reaches its maximum value as the optimal intervention time, and generate a trigger command to call the payment execution interface.

[0030] The following is a detailed explanation. Step S1 includes: S101, establishing a time-series-based decision information set and defining the user's decision-making action space. The decision information set includes not only the user's account balance change characteristics but also intent characteristics reflecting the user's intentions.

[0031] In one embodiment, intent features include the event flow context of the current scenario and the user's historical payment behavior features.

[0032] In traditional technologies, recharge decisions often rely solely on a single static indicator: the current account balance. However, a single balance value cannot reflect the dynamic process of fund consumption or the environmental pressures faced by the user. The core of this step lies in constructing a decision information set. This decision information set is not a static snapshot of the current moment, but rather a set based on... - A time series set constructed algebraically, containing data from the system's initial time 0 to the current time. All historical memories and contextual accumulation.

[0033] Specifically, intent features include: First, the characteristics of balance changes. The system focuses not only on the balance, but also on the speed and acceleration of fund consumption by analyzing the time series trajectory of the balance.

[0034] Second, the event flow context of the current scene. This refers to the external environment and user behavior state perceived by the device. Based on this, the system can infer where the user is, what they are doing, and what they will do next.

[0035] Third, historical payment behavior characteristics. This is the system's memory of users' personalized habits, serving as prior knowledge for subsequent probability calculations.

[0036] The system constructs a decision information set The definition is as follows:

[0037] In other words, the characteristics of balance changes, the event flow context of the current scenario, and the characteristics of historical payment behavior together constitute the moment. decision information set .

[0038] In one embodiment, assume a user is using a ride-hailing app. In the balance change characteristics, the system detects that the user's balance dropped from 50 yuan to 3 yuan in the past 10 minutes after two consecutive deductions. At this point, the balance change is -47 yuan; this precipitous drop in balance reflects the urgency more strongly than a simple 3 yuan balance. In the event stream context, the system reads that the underlying location service is frequently requesting GPS, and the location is on a road in a non-residential area. Simultaneously, the calendar shows a meeting in 10 minutes, indicating the user is currently commuting. Historical payment behavior characteristics show that in commuting scenarios, there is a 90% probability that the user will immediately top up their account once the balance falls below 10 yuan. According to this embodiment, the impact of the information set on decision-making can be immediately determined.

[0039] The above embodiments involve relatively little information. To more accurately infer the user's true intent, balance change characteristics, the event flow context of the current scenario, and historical payment behavior characteristics can contain more information. In an engineering implementation application scenario: Indicates from time 0 to The balance trajectory. In engineering implementation, the feature extraction vector can be represented as... The three features represent the balance change over the past 5 minutes, the current balance, and the recent payment frequency, respectively.

[0040] Representing the scene event flow, its feature extraction vector can be expressed as follows in engineering implementation: .

[0041] Here, `event_type` represents the discrete event type encoding, which can be an integer enumeration value used to identify the current macroscopic activity state perceived by the device. The system performs multimodal judgment by reading calendar events, application running status, and motion sensor data. For example: a value of 1 represents commuting (e.g., detecting that a taxi is being hailed or passing through a gate); a value of 2 represents offline consumption (e.g., being in a supermarket or restaurant area); a value of 3 represents digital entertainment (e.g., playing a game or live streaming); and a value of 0 represents no specific event / standby. This feature is used to assess the tolerance of the current scenario for financial interruption.

[0042] `gps_cluster` represents a geographic location clustering index, which can use a discrete index value to identify the current location. In one embodiment, the system can perform clustering based on historical location data using algorithms such as K-Means. For example: 0 represents home / residential area; 1 represents company / work area; 2 represents consumption area; 3 represents unfamiliar area. Unfamiliar areas usually indicate higher risk control attention or special payment needs. To protect privacy, discrete values ​​are used instead of precise latitude and longitude coordinates.

[0043] `time_slot` represents a time-susceptibility label, a discrete encoding of the current time, used to help determine the user's biological clock status and the cost of disturbing the user. For example: 0 represents late-night rest time; 1 represents the morning rush hour; 2 represents working hours.

[0044] Represents payment history, used to calculate prior probabilities. In one embodiment, define Feature extraction vector: The specific definitions of each feature are as follows: freq month Monthly recharge frequency represents the total number of times a user has initiated a recharge within the past 30 calendar days. This value reflects the user's financial activity level. Users with higher recharge frequencies generally have a higher prior probability of making another recharge.

[0045] bal trigger_avg The historical average trigger threshold represents the average remaining balance in a user's account when all past top-up actions occurred. For example, one user might habitually top up when they have 50 yuan left, while another user might only top up after their balance reaches zero. This feature measures a user's financial anxiety threshold. If the current balance is significantly lower than this average, the probability of immediately topping up increases significantly.

[0046] ratio pay_nowThe "Instant Top-Up Conversion Rate" represents the historical probability percentage of users choosing to top up immediately when faced with a low balance warning or similar low balance status. This historical probability percentage can be represented by recording the number of instant top-up actions and the number of non-instant top-up actions, with the ratio of the number of instant top-up actions to the sum of the two types of actions.

[0047] In one embodiment, it can be in ratio pay_now Based on the prior probability, combined with freq month , and bal trigger_avg The difference between the base prior probability and the current balance is used to adjust the base prior probability to obtain the final prior probability. freq month Indicates activity level, bal trigger_avg The difference between the current balance and the amount of money deposited indicates the urgency of the deposit, and both are related to prior probabilities. The correlation is positive. The specific calculation formula can be defined by those skilled in the art based on existing technology and specific needs, and will not be elaborated here.

[0048] Step S1 also includes: S102, based on the decision information set, abstracting the user's decision action space into an action set, and calculating the conditional probability of various actions.

[0049] To achieve timely response with limited computing power on the device side, complex payment decision-making behavior is abstracted into multiple actions: one is immediate top-up, which means the user has a clear intention and wants the system to assist; the other is non-immediate top-up, which includes all situations where the system does not need to intervene, such as top-up later, ignoring, or giving up.

[0050] Based on the information set constructed using S101, the system employs probabilistic models (such as Bayesian networks, logistic regression, or lightweight neural networks) to infer user intent. The essence of this process is calculating the posterior probability, i.e., given the current environment... In this scenario, the system determines the likelihood of a user choosing to top up immediately. During the calculation, a comprehensive cost function is introduced, which weighs the risks and costs of performing this action. This is explained in detail below.

[0051] In one embodiment, a binary action set can be defined. ,in To recharge immediately, For non-immediate top-up.

[0052] Calculate the action to take conditional probability The formula is as follows:

[0053] in: It can be based on historical data The prior probability is obtained from statistics. k represents the action number; since there are two actions, k = 0 or 1.

[0054] It is a comprehensive cost function. For immediate top-up actions... Its cost mainly comes from the complexity of the operation; for non-immediate top-up actions The costs mainly come from the risk of business interruption and the cost of delays. It is an adjustment coefficient used to control the model's sensitivity to cost differences.

[0055] In one embodiment, The fail_risk function is for assessing the risk of service interruption. The cost function of delay_cost We obtain the result by weighted summation.

[0056] Risk function fail_risk Definition: Execution of an action This could lead to a higher probability that critical services (such as hailing a taxi, passing through turnstiles, and deducting fees) will be interrupted due to insufficient balance.

[0057] Risk function fail_risk Primarily relies on balance trajectory features and scene event flow features Cross-calculation. For immediate top-up. fail_risk For non-immediate top-ups Based on the following factors, a probability value can be obtained for prediction.

[0058] For example, according to Given the current balance and its change, calculate how long the current balance can sustain operations at the current consumption rate. If the calculated result is less than the minimum completion time required by the business, then `fail_risk` is applied. Larger.

[0059] For example, if the scene event flow characteristics If `event_type=1` indicates commuting, then service interruption means being stuck on the road. `fail_risk` Larger. If event_type=3, it means that in the game, service interruption only means purchase failure, and fail_risk is relatively high. Smaller.

[0060] For example, if the scene event flow characteristics In this context, `gps_cluster` points to a remote / unknown region. If a payment fails, the user faces significant difficulties in resolving the issue, resulting in a high failure risk. Larger.

[0061] The specific risk function is fail_risk The calculation method can refer to mature existing technologies, and will not be elaborated here.

[0062] The cost function for delaying disturbances: delay_cost It is the execution of actions. The degree to which the user's current action is interrupted, and the user's level of aversion to it. For non-immediate top-ups, delay_cost is 0 because the user's current operation will not be interrupted by choosing to ignore or wait. For immediate top-ups, delay_cost is high and can be calculated based on the following factors.

[0063] For example, scene event stream features If the event_type indicates that the user is playing a full-screen game or making a video call, the pop-up window will be extremely disruptive. delay_cost Extremely high. If event_type indicates the user is browsing a news feed or in standby mode, the interruption is less noticeable, and delay_cost is low. Lower.

[0064] For example, if the scene event flow characteristics If the time_slot is set to late at night, waking the user to perform an operation would cause extreme user annoyance, and delay_cost would be too high. Extremely high. If the time slot is for commuting, users have a higher tolerance for pop-ups, and delay_cost is low. Lower.

[0065] The cost function for delaying disturbances: delay_cost The calculation method can refer to mature existing technologies, and will not be elaborated here.

[0066] It should be noted that the above embodiments are only one implementation method. Therefore, other implementation methods can be used to measure the overall cost. For example, the complexity of user operations and the user's emphasis on funds can also be taken into account.

[0067] Taking the above ride-hailing scenario as an example, for non-immediate top-up actions... Since the user is currently on their journey and has only 3 yuan remaining, a failed payment could result in them being unable to disembark or suffering credit damage, posing a very high risk of business interruption and leading to substantial overall costs. Regarding the action... Although it requires payment, it avoids the aforementioned significant risks and has a relatively low overall cost. After substituting the values ​​into the conditional probability formula, the system yields... That means there is a 95% chance that users will want to top up immediately.

[0068] In summary, step S102 further calculates the conditional probabilities of various actions based on step S101. It should be noted that in the above embodiments, the action set is binary, while in other embodiments, non-immediate top-up can include more options, such as credit payment, waiting for a refund, etc. That is, the intent features include one or more of the following: the event flow context of the current scenario, the user's historical payment behavior characteristics, the amount of refund pending, and the credit limit; of course, to support these more actions, the decision information set also needs to include more information corresponding to these actions. For example, the action space also includes one or more of waiting for a refund and switching to credit. Specific details will not be elaborated here.

[0069] Step S2 involves using information entropy theory to calculate the real-time recharge entropy based on the conditional probability, thereby quantifying the uncertainty of the user's current decision-making intention. This is one of the core innovations of this invention. Traditional threshold triggering mechanisms are deterministic, while human decision-making processes are often fraught with uncertainty. This invention introduces Shannon entropy to measure this psychological state.

[0070] Recharge Entropy This reflects the degree of confusion in the user's intent. When the entropy value is high: This indicates... and The situation is largely consistent; the user is in a state of hesitation, and the system should not interrupt them rashly. When the entropy value is very low, it indicates that the probability distribution is extreme (e.g., or The user's intent is very clear. The system focuses on the intent to recharge immediately.

[0071] The formula for calculating recharge entropy is as follows:

[0072] For example, in one scenario, a user is browsing products at home with a balance of 50 yuan and the product costing 55 yuan. The user might then consider whether to top up their balance or switch to a different card to pay, and the calculations would... Substituting into the formula, the entropy value... The value of 1 bit is close to the maximum value of 1 bit, indicating high uncertainty.

[0073] In another scenario, a user is hailing a ride but has insufficient funds. The calculation shows... Substituting into the formula, the entropy value... The bit has an extremely low entropy value, indicating a clear intention to recharge immediately.

[0074] Step S3: Monitor the change characteristics of the recharge entropy in real time. When the change of the recharge entropy meets the preset triggering conditions, the pre-triggering condition is achieved.

[0075] A low recharge entropy value alone may not indicate the optimal timing, for example, if the user consistently maintains a clear intention not to recharge. This invention focuses on the process of decreasing recharge entropy, that is, the moment of psychological phase transition from when the user is unsure whether to recharge to when they decide to recharge. This moment is when the user most needs help and is least averse to system intervention.

[0076] To accurately capture this moment, this embodiment designs flexible trigger detection logic, covering single, dual, and triple trigger modes to adapt to different business sensitivity requirements. The rate of change estimator for recharge entropy is defined as follows: . for The derivative of can be calculated using discrete values, which will not be elaborated here.

[0077] Preset trigger conditions One of the following three logics can be used: Single trigger mode:

[0078] Only the rate of change of recharge entropy is monitored. When the rate of change of recharge entropy is less than a certain negative threshold... At that time, it is triggered directly. This indicates that the user's uncertainty is rapidly being eliminated and intent is being formed. This is an indicator function.

[0079] Dual trigger mode:

[0080] Simultaneously monitor the recharge entropy and its rate of change. When the current recharge entropy... It is at a high level (indicating it is in the decision-making period), and at the same time, it is increasing the rate of change of entropy. Significantly negative, i.e., recharge entropy. If the price is dropping rapidly, it will be triggered immediately. This method is used to capture the moment when hesitation ends, effectively filtering out noise during the user's decision-making process.

[0081] Triple trigger mode:

[0082] Simultaneously monitor the recharge entropy and the rate of change of the recharge entropy. Trigger when the following conditions are met simultaneously: (1) The recharge entropy is high, greater than the high threshold of the recharge entropy. (2) The absolute value of the recharge entropy change rate is greater than the recharge entropy change rate threshold. , such as 0.05bit / s; (3) the direction of change is negative, and sign() is the sign function. On the basis of double triggering, the constraint of the change amplitude is added. All three conditions are met at the same time to ensure that the system intervenes only at the moment when the user's psychological intention changes.

[0083] Step S4: In response to the achievement of the pre-triggered condition, calculate the net intervention value, which is the difference between the marginal benefit of the intervention and the user disturbance cost.

[0084] Meeting the aforementioned pre-triggering condition only indicates that the user wants to recharge immediately, but it does not mean that now is the most appropriate time for the pop-up window to appear. Therefore, the net intervention value is used to measure whether the timing is appropriate, as explained in detail below. It should be noted that the following embodiments are merely one measurement method. In other embodiments, those skilled in the art can flexibly adjust the calculation method, add or subtract various factors involved, and adjust and replace specific formulas.

[0085] In one embodiment, to address the question of whether it is worthwhile to disturb the user now, the system introduces a utility function from economics to calculate the net value of intervention. This value is determined by two parts of a game: first, the marginal benefit of intervention, that is, how much benefit is gained if we intervene now compared to delaying intervention, which reflects the value of timeliness; second, the cost of user disruption, that is, the degree of disruption to users caused by proactive intervention now, which reflects the protection of user experience.

[0086] 1) Specifically, the net intervention value function is defined as:

[0087] The marginal benefit of intervention is expressed as The expected utility is expressed as:

[0088] in, It is a real-time predicted failure probability. Probability of success; This is the scenario severity coefficient, which in one embodiment can be a discrete mapping function. It does not rely on complex real-time calculations, but rather is derived directly from a predefined business rule table based on the currently identified scenario type (event_type). The table below shows one such business rule table.

[0089]

[0090] It is the real-time prediction failure probability, that is, the predicted consumption amount is greater than the current balance. The probability of [the outcome]. Specifically, in one embodiment, the calculation can be performed according to the following steps: The system constructs a probability distribution of consumption amount under different scenarios based on historical data, i.e., a probability distribution curve of consumption amount. This is equal to the area to the right of the consumption distribution curve on the current balance. Here, it's necessary to assume that consumption in a specific scenario follows a normal distribution. For ease of calculation, the Z-score method can be used to simplify the calculation of the standard score Z. , It is the cumulative distribution function of the standard normal distribution.

[0091] For example, in a ride-hailing scenario, the current balance The figure is 25 yuan. The average historical ride-hailing cost for users is 35 yuan, with a predicted standard deviation of 5 yuan. Look up the normal distribution table: .but 978. That is, there is a 97.8% probability that the balance is insufficient.

[0092] Based on the above examples, if the severity coefficient is obtained by looking up the table in the aforementioned ride-hailing scenario... So, expected utility =2.2.

[0093] The following calculations For ease of calculation, assume =1 minute. Within this 1 minute, as the vehicle travels closer to the destination, the final price fluctuation range will rapidly narrow, causing the forecast standard deviation to decrease, for example, from 5 yuan to 2 yuan. Calculated... Approximately equal to 0. Marginal revenue. =2.2. The greater the marginal benefit, the greater the loss of utility from delayed top-ups, and the greater the value recovered by immediate top-ups.

[0094] 2) The cost of user interruption is related to the current time of day and the user's current attention span. Specifically, the cost of interruption is higher during preset rest periods compared to non-rest periods; and the cost of interruption is higher when the user is focused on running the application or keeping the screen on compared to when the user is not focused on running the application or keeping the screen off. Specifically, the cost of user interruption... The calculation formula is:

[0095] For time sensitivity, a lookup table function can be used. For example, according to the time-slot sensitivity label time_slot introduced above, it can be set to 1.0 in the late night (23:00-07:00) and 0.1 in other times. To determine attention occupancy, a lookup table function can be used, for example, based on the discrete event type encoding described above. If it is confirmed that a game or entertainment activity is in progress, then The corresponding value is set to 1. Weight and It can be configured according to specific needs.

[0096] For example, in the scenario of taking a taxi late at night, although the cost of disturbing others at night is high... Very high (because it can be) (Large), but because failing to hail a ride can lead to being stranded, the scenario coefficient... Extremely high, resulting in marginal revenue items Extremely large, far exceeding .therefore The system determines that this situation warrants intervention and immediately displays a pop-up window.

[0097] For example, the cost of disturbing others, such as playing games late at night. Extremely high. And the consequences of failing to buy a skin are very minor. (Very small), marginal returns are negligible. At this time... The system determines that it is not worthwhile to pop up a window at this time, and keeps it silent or only displays it in the notification bar.

[0098] Step S5: Within a preset future time window, find the moment when the net intervention value reaches its maximum value as the optimal intervention time, and generate an active intervention instruction, which can then call the payment execution interface.

[0099] This is the final step in decision-making. The system not only considers whether intervention is worthwhile now, but also makes short-term future predictions. The system operates within a tiny time window. Inside, prediction The curve of change is used to try to find the golden ratio point that maximizes benefits and minimizes disturbance. .

[0100] Within a preset future time window, the optimal intervention time is determined by finding the moment when the net intervention value reaches its maximum value. This includes: iterating through each time point within the preset future time window; calculating the net intervention value for each time point; and selecting the time point with the highest net intervention value as the optimal intervention time. Figure 2 As shown, the system will not pop up a window as soon as the pre-trigger condition is met, but may do so at the trigger time t. now Then wait a few seconds, until the user has just finished a game or just picked up their phone, so that... It has reached its peak.

[0101] Optimal intervention timing and target optimization:

[0102] The function is the function with maximum value as the independent variable. This indicates that the pre-triggering condition has been met. That is, once the system calculates the current time... That is (or very close to) the optimal moment. The agent immediately generates proactive intervention instructions (such as pop-up instructions). Figure 2 As shown, in this embodiment, Ω=200s.

[0103] In one embodiment, the proactive intervention command is sent to the App's UI layer or a third-party payment SDK via the device's internal bus or API, activating the cashier and completing the entire proactive intervention process. The technical details following the sending of the proactive intervention command are prior art and unrelated to this invention, and therefore will not be elaborated upon.

[0104] like Figure 3 As shown, on the other hand, this invention also provides an asset data management and proactive intervention system, which is integrated into a user terminal device. From a hardware architecture perspective, the system includes: a memory for storing computer program code that executes the aforementioned algorithms, and locally cached decision information set data (such as balance history and payment habit models); and a processor (such as a mobile phone's CPU or NPU) for executing the program in the memory. The processor is configured to run the aforementioned steps of recharge entropy calculation, net intervention value calculation, and proactive intervention command control, realizing the entire process logic from data perception, probability inference, recharge entropy value monitoring to command generation. Through hardware and software collaboration, this system achieves highly timely asset management with privacy protection capabilities without relying on a continuous cloud connection.

[0105] The system in this embodiment also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art, and therefore will not be described in detail here.

[0106] In this invention, the aforementioned memory can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0107] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A proactive intervention decision-making method based on recharge entropy, characterized in that, This method operates on the user terminal device side and includes: Establish a time-series-based decision information set, which includes user account balance change characteristics and intent characteristics reflecting user intent; define a user decision action space, which includes at least immediate top-up and non-immediate top-up; and calculate the conditional probability of each action in the action space based on the decision information set. Using information entropy theory, the recharge entropy is calculated based on the conditional probability of each action in the action space to quantify the uncertainty of the user's current decision intention; The change characteristics of the recharge entropy are monitored in real time. When the change characteristics of the recharge entropy meet the preset trigger conditions, the pre-trigger condition is achieved. In response to the achievement of the pre-triggering condition, the net intervention value is calculated, which is the difference between the marginal benefit of proactive intervention and the user disturbance cost; Within a preset future time window, the optimal intervention time is identified as the moment when the net intervention value reaches its maximum, and an active intervention instruction is generated; wherein, For an action in the action space, sum the conditional probability of the action with the product of the logarithm of the conditional probability, and take the negative value of the summation result to obtain the current charging entropy.

2. The method according to claim 1, characterized in that, Calculating the conditional probability of each action in the action space includes: Construct a comprehensive cost function, which is calculated at least based on the business failure risk and the time delay cost; For an action in the action space, the conditional probability of the action is calculated using a Bayesian network or logistic regression model by combining the prior probability of the action with the comprehensive cost function.

3. The method according to claim 1, characterized in that, The preset triggering conditions include a combination of one or more of the following conditions: The recharge entropy is greater than the set high-order threshold of the recharge entropy; The absolute value of the recharge entropy change rate is greater than the set recharge entropy change rate threshold; and The rate of change of recharge entropy is negative.

4. The method according to claim 1, characterized in that, The marginal benefit in calculating the net intervention value includes: Define the expected utility function, which is obtained by multiplying the scenario severity coefficient by the probability of success of real-time prediction; The difference between the expected utility of immediate intervention at the current moment and the expected utility of intervention postponed to the next moment is calculated based on the expected utility function, and is taken as the marginal benefit of intervention.

5. The method according to claim 4, characterized in that, The scenario severity coefficient is determined based on the current business scenario type. The scenario severity coefficient corresponding to a high-risk scenario is higher than that corresponding to a low-risk scenario. The high-risk scenarios include at least ride-hailing scenarios or gate passage scenarios.

6. The method according to claim 1, characterized in that, The user disturbance cost in calculating the net intervention value includes: The cost of disturbing users is related to the current time of day and the current level of attention. Specifically, the cost of disturbing users is greater during preset rest periods compared to non-rest periods; the cost of disturbing users is also greater when they are focused on running applications or keeping the screen on compared to when they are not focused on running applications or keeping the screen off.

7. The method according to claim 1, characterized in that, Within a pre-defined future time window, the optimal intervention time is identified as the moment when the net intervention value reaches its maximum value, including: Iterate through every point in time within the preset future time window; Calculate the net intervention value at each time point; The optimal intervention time is selected based on the point at which the net intervention value is greatest.

8. The method according to any one of claims 1-7, characterized in that, The intent features include one or more of the following: the event flow context of the current scenario, the user's historical payment behavior features, the amount of refunds pending payment, and the credit limit; Correspondingly, the action space also includes one or more of waiting for a refund to be credited and switching to a credit limit.

9. A proactive intervention decision-making system based on recharge entropy, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of a proactive intervention decision-making method based on recharge entropy as described in any one of claims 1 to 8 when executing the computer program.

Images