Product recommendation system based on user portrait and recharging behavior
By integrating user static attributes and dynamic behavior data to generate multi-dimensional user portraits, combined with recharge behavior analysis and reinforcement learning framework, the problem of insufficient granularity of user portraits is solved, and efficient and accurate product recommendations are achieved.
Patent Information
- Application Number
- CN202510566762.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-30
- Publication Date
- 2025-08-15
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the prior art, the granularity of user portraits is insufficient, resulting in product recommendation results that cannot match users' actual needs and potential preferences, resulting in recommendation results lag behind user interest migration.
By integrating the user's static attribute data and dynamic behavior data, multi-dimensional user portrait feature vectors are generated, combined with recharge behavior analysis, and dynamically adjusting recommendation strategies using the reinforcement learning framework, generating a personalized recommendation list, and continuously updating model parameters through a closed-loop feedback mechanism.
It improves the matching degree and granularity of user portraits, accurately reflects users' real needs and potential preferences, maximizes recommendation conversion rate and user satisfaction, and improves the accuracy and efficiency of product recommendations.
Smart Images
Figure CN120494931A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of product recommendation, and specifically to a product recommendation system based on user portraits and recharge behavior. Background Art
[0002] In the current information technology landscape, intelligent recommendation systems have become a key tool for addressing information overload. In advertising and product sales platforms, existing technologies often recommend products based on user profiles. However, these technologies often rely on static attributes or single dimensions, resulting in insufficient granularity in the profiles, making it difficult to fully characterize user characteristics and accurately reflect users' true needs and potential preferences. Furthermore, in product recommendations, the lack of user characteristics results in a poor match with the user's actual needs, causing recommendations to lag behind user interest shifts.
[0003] Therefore, there is an urgent need for a more intelligent and accurate product recommendation technology. Summary of the Invention
[0004] The purpose of this application is to provide a product recommendation system based on user portraits and recharge behavior to solve the technical problems raised in the above background technology.
[0005] To achieve the above objectives, the present application discloses the following technical solution: a product recommendation system based on user profiles and recharge behavior, the method comprising the following steps:
[0006] The user portrait construction module is configured to generate a multi-dimensional user portrait feature vector by integrating the user's static attribute data and dynamic behavior data; the static attribute data includes user identity information and device information; the dynamic behavior data includes user click trajectory, browsing time and interaction frequency;
[0007] The recharge behavior analysis module is configured to: perform time series modeling on the user's historical recharge records to extract the user's recharge cycle, amount distribution, and payment channel preference;
[0008] The target joint optimization module is configured to: dynamically adjust the weight parameters of the recommendation strategy through a reinforcement learning framework based on the user portrait feature vector and recharge behavior analysis results, to maximize the joint optimization goals of recommendation conversion rate and user satisfaction;
[0009] The dynamic product recommendation module is configured to: input the user's current behavioral characteristics into the trained recommendation model, output a personalized recommendation list, and continuously update the model parameters through a closed-loop feedback mechanism; wherein, the generation process of the personalized recommendation list includes multi-objective optimization constraints, and the multi-objective optimization constraints include at least budget control, advertiser demand matching and user interest attenuation factor.
[0010] Preferably, the recharge behavior analysis module is further configured to: predict the user's future recharge cycle through an LSTM-TCN hybrid model, and generate a probability distribution in combination with Monte Carlo simulation.
[0011] As a preferred method, the LSTM-TCN hybrid model for predicting the user's future recharge cycle has a loss function of:
[0012] L=λ×MES(T pred ,T true )+(1-λ)×KL(P pred ||P true )+μ×Uncertainty(T pred )
[0013] Among them, λ is the balance coefficient, ranging from 0 to 1; T pred is the time series of the predicted recharge cycle; T true is the time series of the actual recharge cycle; P pred is the probability distribution value of the predicted recharge cycle; P true is the probability distribution value of the actual recharge cycle; μ is the uncertainty penalty coefficient, which is used to quantify the confidence of the model prediction. pred ) is calculated by entropy.
[0014] Preferably, the recharge behavior analysis module is further configured to: analyze the risk level of the user's current recharge behavior based on the user's credit score and historical fraud records, and trigger a differentiated recommendation strategy.
[0015] Preferably, the risk level of the user's current recharge behavior is analyzed, specifically including:
[0016] S1-Calculate the risk score of the current recharge behavior through the Bayesian network. The higher the risk score, the greater the risk of the user behavior. The calculation formula of the risk score is:
[0017]
[0018] Among them, f i (x) is the i-th risk factor function, and each factor outputs a risk score in the range of 0 to 1; w i is the weight of the i-th risk factor; ΔT is the time interval between the current recharge behavior and the last recharge; υ is an adjustable parameter for controlling the attenuation intensity, and β is an adjustable parameter for controlling time sensitivity, which is optimized through A / B testing; t pred is the predicted value of the future recharge cycle, is the actual average value of the user's historical recharge period; γ is the penalty coefficient, ranging from 0 to 1;
[0019] S2-Determine the risk level of the user's current recharge behavior based on the calculated risk score and a preset risk level classification mechanism.
[0020] Preferably, the step of obtaining the weight of the i-th risk factor comprises the following specific steps:
[0021] S11-constructing a directed acyclic graph, wherein the nodes of the directed acyclic graph include risk factors and recharge cycles, wherein the risk factors include credit score, fraud record, and payment channel anomaly;
[0022] S12-Learning conditional probability distributions using maximum likelihood estimation or Markov chain Monte Carlo methods;
[0023] S13-Assign weight w to each risk factor based on posterior probability i , and satisfy ∑w i =1.
[0024] Preferably, the objective function of the reinforcement learning framework is:
[0025] J(W t )=ζ c ×CTR t +ζ s ×Satisfaction t -ζ R ×R,
[0026] Among them, CTR t Click-through rate for the advertised product, Satisfaction t For user satisfaction, c ,ζ s ,ζ s is the weight coefficient.
[0027] Preferably, the budget control constraint is specifically: ∑M i ×m i <B, where M i is the cost of advertising product i, m i To recommend this product, m i is 0 or 1, and B is the preset budget cap;
[0028] The constraints on the advertiser demand matching degree are specifically: ∑X i ×m i ≥Du, where X i is the advertiser tag matching degree of advertised product i, and Du is the minimum matching threshold;
[0029] The constraints of the user interest attenuation factor are specifically: m i≤exp(-θ×time i ), where time i is the time interval between the most recent interaction between the advertising product i and the user, and θ is the decay rate parameter.
[0030] Preferably, the user portrait construction module is further configured to: perform correlation analysis on the user's social relationship chain, cross-platform behavior logs and third-party service call records through a graph neural network to generate an enhanced user portrait; and incrementally update the user portrait feature vector based on a sliding time window, and screen key features through an attention mechanism to reduce noise interference.
[0031] Preferably, the target joint optimization module is specifically configured as follows:
[0032] D1-Obtain the multi-dimensional user portrait feature vector S constructed by the user portrait construction model, S = [C, F, H, T pred ], where C is the user's credit score, which is calculated by weighted average based on historical credit records; F is the frequency of historical fraud records, which is extracted through a time series statistical model; H is the payment channel preference, which is calculated based on the proportion of payment behaviors; t pred is the predicted value of the future recharge cycle;
[0033] D2- Obtain the risk score R obtained by the recharge behavior analysis module based on the user's credit score and historical fraud records;
[0034] D3 - Dynamically adjust the weight parameters of the recommendation strategy through the reinforcement learning framework to maximize the joint optimization goal of recommendation conversion rate and user satisfaction; the calculation formula of the weight parameters is:
[0035]
[0036] Among them, W t is the weight parameter at time step t; J(W t ) is the objective function of the reinforcement learning framework; δ is the preset adjustment factor; t ture The actual value of the user's historical recharge cycle; η t is the learning rate, η0 is the initial learning rate; ε is the smoothing term; τ is the risk score adjustment coefficient; G t is the squared gradient accumulation term, g i is the gradient.
[0037] Beneficial effects: The product recommendation system based on user portrait and recharge behavior of this application generates a multi-dimensional user portrait feature vector by integrating the user's static attribute data and dynamic behavior data, thereby improving the matching degree between the user portrait and the user's actual situation, improving the granularity of the user portrait, and accurately reflecting the user's real needs and potential preferences. Combined with the results of the analysis of the user's recharge behavior, it maximizes the joint optimization goal of recommendation conversion rate and user satisfaction, customizes personalized recommendation lists for users, and improves product recommendation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.
[0039] Figure 1 This is a structural block diagram of the product recommendation system based on user portraits and recharge behavior provided in an embodiment of the present application. DETAILED DESCRIPTION
[0040] The following is a clear and complete description of the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.
[0041] In this document, the term "comprising" is intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising..." does not preclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.
[0042] This embodiment discloses Figure 1 The product recommendation system based on user portrait and recharge behavior shown in the figure includes a user portrait construction module, a recharge behavior analysis module, a target joint optimization module and a product dynamic recommendation module.
[0043] Below, each module is introduced in detail.
[0044] The user portrait construction module is configured to generate a multi-dimensional user portrait feature vector by integrating static attribute data reflecting long-term stable characteristics and dynamic behavior data reflecting short-term interests and habits; the static attribute data includes: user identity information and device information; the dynamic behavior data includes: user click trajectory, browsing time and interaction frequency.
[0045] The recharge behavior analysis module is configured to: perform time series modeling on the user's historical recharge records, and extract the user's recharge cycle, amount distribution, and payment channel preference.
[0046] The target joint optimization module is configured to: based on the user portrait feature vector and recharge behavior analysis results, dynamically adjust the weight parameters of the recommendation strategy through the reinforcement learning framework to maximize the joint optimization goals of recommendation conversion rate and user satisfaction.
[0047] The dynamic product recommendation module is configured to: input the user's current behavioral characteristics into the trained recommendation model, output a personalized recommendation list, and continuously update the model parameters through a closed-loop feedback mechanism; wherein, the generation process of the personalized recommendation list includes multi-objective optimization constraints, and the multi-objective optimization constraints include at least budget control, advertiser demand matching and user interest attenuation factor.
[0048] In a feasible embodiment, in order to improve the timeliness of the recommendation system, the recharge behavior analysis module is further configured to: predict the user's future recharge cycle through the LSTM-TCN hybrid model, and generate a probability distribution in combination with Monte Carlo simulation. The purpose of generating the probability distribution through Monte Carlo simulation is to quantify the uncertainty of the model prediction and avoid the deviation caused by the prediction of a single time point.
[0049] Preferably, the loss function of the LSTM-TCN hybrid model for predicting the user's future recharge cycle is:
[0050] L=λ×MES(T pred ,T true )+(1-λ)×KL(P pred ||P true )+μ×Uncertainty(T pred )
[0051] Among them, MES(T pred ,T true ) is the mean square error, which is used to measure the deviation between the time point of the recharge cycle and the time point of the actual recharge cycle to ensure the accuracy of the model's prediction of the time point; λ is the balance coefficient, which ranges from 0 to 1 and is determined by the stability of the user's historical data; KL calculation is used to measure the difference in the predicted probability distribution and optimize the model's overall fitting ability to the probability distribution; T predis the time series of the predicted recharge cycle; T true is the time series of the actual recharge cycle; P pred is the probability distribution value of the predicted recharge cycle; P true is the probability distribution value of the actual recharge cycle; μ is the uncertainty penalty coefficient, which is used to quantify the confidence of the model prediction and avoid overconfident prediction. pred ) is calculated by entropy.
[0052] In a feasible embodiment, the recharge behavior analysis module is further configured to: analyze the risk level of the user's current recharge behavior based on the user's credit score and historical fraud records, and trigger a differentiated recommendation strategy.
[0053] Preferably, the risk level of the user's current recharge behavior analysis specifically includes:
[0054] S1-Calculate the risk score of the current recharge behavior through the Bayesian network. The higher the risk score, the greater the risk of the user behavior. The calculation formula of the risk score is:
[0055]
[0056] Among them, f i (x) is the i-th risk factor function, and each factor outputs a risk score in the range of 0 to 1 (such as credit score C, fraud record F, payment channel anomaly P, etc.); w i is the weight of the i-th risk factor, which is learned from historical data through the Bayesian network; ΔT is the time interval between the current recharge behavior and the last recharge; υ is an adjustable parameter that controls the attenuation intensity, and β is an adjustable parameter that controls time sensitivity, which is optimized through A / B testing; t pred The value of the predicted future recharge cycle is a specific value in the time series of the predicted recharge cycle. is the actual average value of the user's historical recharge cycle; γ is the penalty coefficient, ranging from 0 to 1, indicating the degree of suppression of the cycle prediction deviation on the risk score; if the predicted future recharge cycle t pred Significant deviation from the actual average value of the user's historical recharge cycle The risk score is lowered (i.e. the user behavior is considered unstable and caution is required). Combining the prediction of future recharge cycles with Bayesian risk assessment can achieve dynamic acquisition of risk scores and avoid the limitations of static features. Combined with the periodic attenuation correction term and the prediction deviation penalty The risk score can be adjusted in real time as user behavior changes.
[0057] S2-Determine the risk level of the user's current recharge behavior based on the calculated risk score and a preset risk level classification mechanism.
[0058] According to the numerical range of the risk score \(R\in[0,1]\), the risk level can be divided into four levels according to the risk level division mechanism. When \(R > 0.8\), it indicates that the user has an extremely high fraud or default risk, and restrictive measures or denial of service should be taken immediately; when \(0.6 < R\leq0.8\), it indicates that the user's behavior is abnormal, and monitoring needs to be strengthened and the recommendation of high-value products should be restricted; when \(0.4 < R\leq0.6\), it indicates that the user has a certain risk, and the recommendation strategy needs to be optimized to reduce the recommendation intensity; when \(R\leq0.4\), it indicates that the user has good credit, and high-value products can be normally recommended and incentives can be provided. For the differentiated recommendation strategy, when \(R > 0.8\), a strict restriction and active intervention strategy is executed, specifically: freezing the user account, suspending all recommendation functions, triggering the intervention of the risk control team, setting hard intercept rules in the recommendation engine, calling external anti-fraud interfaces (such as third-party blacklist libraries) for cross-verification, verifying the user's historical behavior, sending a high-risk warning to the user, and requiring the supplement of identity verification materials (such as bank card binding, face recognition). When \(0.6 < R\leq0.8\), a cautious recommendation and enhanced monitoring strategy is executed, specifically: only recommending advertising products with lower prices, reducing the recommendation frequency (such as no more than 3 times a day), avoiding excessive stimulation of consumption, recording the user's click and payment paths in real time, and triggering secondary verification if abnormalities occur (such as high-frequency operations in a short period of time). When \(0.4 < R\leq0.6\), an optimized experience and promotion strategy is executed, specifically: recommending medium-value advertising products in combination with the user's historical preferences. When \(R\leq0.4\), a free recommendation and deep incentive strategy is executed, specifically: recommending high-value advertising products based on the user portrait, issuing targeted coupons, and enhancing user stickiness through the membership level system (such as VIP rights). It should be noted that an advertising product can be, but is not limited to, an advertising channel product of an advertising platform. After purchasing an advertising product, the operator will投放 the advertising content required by the user according to the specific content of the advertising product (such as the placement location, placement time, placement market, etc.).
[0059] Thereby, the system can achieve refined operation while ensuring security, and at the same time significantly improve user satisfaction and business revenue.
[0060] Preferably, the steps for obtaining the weight of the \(i\)-th risk factor include the following specific steps:
[0061] S11 - Construct a directed acyclic graph, the nodes of which include risk factors and recharge cycles, and the risk factors include credit scores, fraud records, and abnormal payment channels;
[0062] S12 - Use the maximum likelihood estimation or Markov chain Monte Carlo method to learn the conditional probability distribution;
[0063] S13 - Allocate the weight \(w\) of each risk factor according to the posterior probability i, and satisfy ∑w i =1.
[0064] The aforementioned recharge behavior analysis module, based on a combined LSTM-TCN hybrid model and Bayesian networks, achieves accurate prediction of user recharge cycles and dynamic risk assessment. A multi-objective loss function balances time point accuracy with probability distribution fitting to adapt to diverse business scenarios. Uncertainty quantification improves model robustness by generating probability distributions through Monte Carlo simulation. Risk-driven recommendations significantly reduce fraud risk and improve conversion rates through differentiated strategies.
[0065] In a feasible embodiment, the objective function of the reinforcement learning framework is specifically:
[0066] J(W t )=ζ c ×CTR t +ζ s ×Satisfaction t -ζ R ×R,
[0067] Among them, CTR t Satisfaction is the ratio of clicks to impressions of the advertised product, which measures the attractiveness of the recommended content to users; t User satisfaction reflects the user's acceptance of the recommendation results, which can be obtained through any method in the existing technology (such as user feedback, etc.); c ,ζ s ,ζ s The weight coefficient is customized according to business goals, and the impact of different objective functions on actual indicators is verified through A / B testing; the weight coefficient ζ is adjusted based on domain knowledge (such as advertiser ROI and user retention rate). c ,ζ s ,ζ s .
[0068] Through the design of the above objective function, the global optimization of the recommendation system is achieved, taking into account click-through conversion rate, user satisfaction and risk, avoiding the imbalance caused by a single objective, and ensuring the reliability of the joint optimization goal of maximizing recommendation conversion rate and user satisfaction.
[0069] In a feasible embodiment, the budget control constraint is specifically: ∑M i ×m i <B, where M i is the cost of advertising product i, m i To recommend this product, m i is 0 or 1, and B is the preset budget cap;
[0070] The constraints on the advertiser demand matching degree are specifically: ∑X i ×m i ≥Du, where X i is the advertiser tag matching degree of advertised product i, and Du is the minimum matching threshold;
[0071] The constraints of the user interest attenuation factor are specifically: m i ≤exp(-θ×time i ), where time i is the time interval between the most recent interaction between the advertising product i and the user, and θ is the decay rate parameter.
[0072] In a feasible embodiment, the user portrait construction module is further configured to: perform correlation analysis on the user's social relationship chain, cross-platform behavior logs and third-party service call records through the graph neural network in the prior art to generate an enhanced user portrait and capture the user's multi-dimensional interests and behavior patterns; and incrementally update the user portrait feature vector based on the sliding time window in the prior art, and screen key features through the attention mechanism to reduce noise interference.
[0073] By optimizing the user portrait construction module, we have achieved deep correlation of cross-domain data, significantly improving the comprehensiveness of user portraits. By combining the sliding window and attention mechanism, we have solved the problem of insufficient timeliness of traditional portraits.
[0074] In a feasible embodiment, the target joint optimization module is specifically configured as follows:
[0075] D1-Obtain the multi-dimensional user portrait feature vector S constructed by the user portrait construction model, S = [C, F, H, T pred ], where C is the user's credit score, which is calculated by weighted average based on historical credit records; F is the frequency of historical fraud records, which is extracted through a time series statistical model; H is the payment channel preference, which is calculated based on the proportion of payment behaviors; t pred is the predicted value of the future recharge cycle;
[0076] D2- Obtain the risk score R obtained by the recharge behavior analysis module based on the user's credit score and historical fraud records;
[0077] D3 - Dynamically adjust the weight parameters of the recommendation strategy through the reinforcement learning framework to maximize the joint optimization goal of recommendation conversion rate and user satisfaction; the calculation formula of the weight parameters is:
[0078]
[0079] Among them, W t is the weight parameter at time step t; is the gradient of the objective function J with respect to the parameter W, which measures the optimization direction of the current strategy; J(W t ) is the objective function of the reinforcement learning framework, which is used to measure the weight parameter W of the recommendation strategy t The effect at time step t; δ is a preset adjustment factor used to control the impact of period deviation on weight update, which is 0 or 1; t ture is the actual value of the user's historical recharge cycle, which is a specific value in the time series of the actual recharge cycle; η t is the learning rate, η0 is the initial learning rate, which is used to control the step size of parameter update and determine the amplitude of weight adjustment; ε is the smoothing term, which is a very small constant to prevent the denominator from being zero to ensure numerical stability; τ is the risk score adjustment coefficient, which is used to control the inhibitory strength of the risk score on the learning rate; G t is the squared gradient accumulation term, g i is the gradient, which represents the loss function J(W t ) for the weight parameter W t The partial derivative (gradient direction) of . A correction factor is formed. When the predicted future recharge cycle exceeds the actual value of the user's historical recharge cycle, the weight update amplitude is reduced to prevent overfitting high-latency scenarios, and the prediction time deviation is quantified to avoid strategy failure caused by delay.
[0080] By designing the above-mentioned target joint optimization module, the negative impact of high-latency strategies is suppressed through time-sensitive corrections, the user experience is improved, and it has high robustness.
[0081] Furthermore, the generation of advertising recharge service content in the personalized recommendation list is also combined with eCPM optimization. The specific eCPM optimization is as follows:
[0082] Q1-Calculate the conversion effect of each advertised product using the following formula:
[0083]
[0084] CPC is the cost per click of the advertised product i, LTV is i The total expected benefits generated by users during the use of the product / service, reflecting long-term value, calculated by fitting historical consumption data; CPA i is the cost per conversion; ρ is LTV i The amplification factor of eCPM ranges from 0 to 1;
[0085] Q2- Sort the recommendation list in descending order based on the eCPM calculation results of each advertising product, and give priority to advertising products that meet the budget and advertiser demand constraints.
[0086] pass The introduction of the correction factor gives priority to recommending high-yield advertising products, thereby improving user satisfaction with the recommendation results.
[0087] In the embodiments provided herein, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any appropriate combination thereof. For hardware implementation, the processor can be implemented in one or more of the following units: an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a processor, a controller, a microcontroller, a microprocessor, other electronic units designed to implement the functions described herein, or a combination thereof. For software implementation, part or all of the processes of the embodiments can be completed by instructing the relevant hardware through a computer program. When implemented, the above program can be stored in a computer-readable storage medium or transmitted as one or more instructions or codes on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein the communication media include any medium that facilitates the transmission of a computer program from one place to another. The storage medium can be any available medium that a computer can access. The computer-readable storage medium can include, but is not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer.
[0088] Finally, it should be noted that the above is only a preferred embodiment of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent replacements for some of the technical features therein. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.
Claims
1. A product recommendation system based on user profile and recharge behavior, characterized by: include: The user portrait construction module is configured to: generate a multi-dimensional user portrait feature vector by integrating the user's static attribute data and dynamic behavior data; The static attribute data includes: user identity information and device information; the dynamic behavior data includes: user click trajectory, browsing time and interaction frequency; The recharge behavior analysis module is configured to: perform time series modeling on the user's historical recharge records to extract the user's recharge cycle, amount distribution, and payment channel preference; The target joint optimization module is configured to: dynamically adjust the weight parameters of the recommendation strategy through a reinforcement learning framework based on the user portrait feature vector and recharge behavior analysis results, to maximize the joint optimization goals of recommendation conversion rate and user satisfaction; The dynamic product recommendation module is configured to: input the user's current behavioral characteristics into the trained recommendation model, output a personalized recommendation list, and continuously update the model parameters through a closed-loop feedback mechanism; wherein, the generation process of the personalized recommendation list includes multi-objective optimization constraints, and the multi-objective optimization constraints include at least budget control, advertiser demand matching and user interest attenuation factor.
2. The product recommendation system based on user profile and recharge behavior according to claim 1 is characterized in that: The recharge behavior analysis module is further configured to predict the user's future recharge cycle through an LSTM-TCN hybrid model and generate a probability distribution in combination with Monte Carlo simulation.
3. The product recommendation system based on user profile and recharge behavior according to claim 2 is characterized in that: The loss function of the LSTM-TCN hybrid model for predicting a user's future recharge cycle is: L=λ×MES(T pred ,T true )+(1-λ)×KL(P pred ||P true )+μ×Uncertainty(T pred ) Among them, λ is the balance coefficient, ranging from 0 to 1; T pred is the time series of the predicted recharge cycle; T true is the time series of the actual recharge cycle; P pred is the probability distribution value of the predicted recharge cycle; P true is the probability distribution value of the actual recharge cycle; μ is the uncertainty penalty coefficient, which is used to quantify the confidence of the model prediction. pred ) is calculated by entropy.
4. The product recommendation system based on user profile and recharge behavior according to claim 2, characterized in that: The recharge behavior analysis module is further configured to analyze the risk level of the user's current recharge behavior based on the user's credit score and historical fraud records, and trigger a differentiated recommendation strategy.
5. The product recommendation system based on user profile and recharge behavior according to claim 4 is characterized in that: The risk level of the user's current recharge behavior analysis specifically includes: S1-Calculate the risk score of the current recharge behavior through the Bayesian network. The higher the risk score, the greater the risk of the user behavior. The calculation formula of the risk score is: Among them, f i (x) is the i-th risk factor function, and each factor outputs a risk score in the range of 0 to 1; w i is the weight of the i-th risk factor; ΔT is the time interval between the current recharge behavior and the last recharge; υ is an adjustable parameter for controlling the attenuation intensity, and β is an adjustable parameter for controlling time sensitivity, which is optimized through A / B testing; t pred is the predicted value of the future recharge cycle, is the actual average value of the user's historical recharge period; γ is the penalty coefficient, ranging from 0 to 1; S2-Determine the risk level of the user's current recharge behavior based on the calculated risk score and a preset risk level classification mechanism.
6. The product recommendation system based on user profile and recharge behavior according to claim 5, characterized in that: The step of obtaining the weight of the i-th risk factor includes the following specific steps: S11-constructing a directed acyclic graph, wherein the nodes of the directed acyclic graph include risk factors and recharge cycles, wherein the risk factors include credit score, fraud record, and payment channel anomaly; S12-Learning conditional probability distributions using maximum likelihood estimation or Markov chain Monte Carlo methods; S13-Assign weight w to each risk factor based on posterior probability i , and satisfy ∑w i =1.
7. The product recommendation system based on user profile and recharge behavior according to claim 1 is characterized in that: The objective function of the reinforcement learning framework is specifically: J(W t )=ζ c ×CTR t +g s ×Satisfaction t -g R ×R, Among them, CTR t Click-through conversion rate for the advertised product, Satisfaction t For user satisfaction, c ,ζ s ,ζ s is the weight coefficient.
8. The product recommendation system based on user profile and recharge behavior according to claim 1 is characterized in that: The budget control constraints are specifically: ∑M i ×m i <B, where M i is the cost of advertising product i, m i To recommend this product, m i is 0 or 1, and B is the preset budget cap; The constraints on the advertiser demand matching degree are specifically: ∑X i ×m i ≥Du, where X i is the advertiser tag matching degree of advertised product i, and Du is the minimum matching threshold; The constraints of the user interest attenuation factor are specifically: m i ≤exp(-θ×time i ), where time i is the time interval between the most recent interaction between the advertising product i and the user, and θ is the decay rate parameter.
9. The product recommendation system based on user profile and recharge behavior according to claim 1, characterized in that: The user portrait construction module is also configured to: perform correlation analysis on the user's social relationship chain, cross-platform behavior logs and third-party service call records through a graph neural network to generate an enhanced user portrait; and incrementally update the user portrait feature vector based on a sliding time window, and screen key features through an attention mechanism to reduce noise interference.
10. The product recommendation system based on user profile and recharge behavior according to claim 1, characterized in that: The target joint optimization module is specifically configured as follows: D1-Obtain the multi-dimensional user portrait feature vector S constructed by the user portrait construction model, S = [C, F, H, T pred ], where C is the user's credit score, which is calculated by weighted average based on historical credit records; F is the frequency of historical fraud records, which is extracted through a time series statistical model; H is the payment channel preference, which is calculated based on the proportion of payment behaviors; t pred is the predicted value of the future recharge cycle; D2- Obtain the risk score R obtained by the recharge behavior analysis module based on the user's credit score and historical fraud records; D3 - Dynamically adjust the weight parameters of the recommendation strategy through the reinforcement learning framework to maximize the joint optimization goal of recommendation conversion rate and user satisfaction; the calculation formula of the weight parameters is: Among them, W t is the weight parameter at time step t; J(W t ) is the objective function of the reinforcement learning framework; δ is the preset adjustment factor; t ture The actual value of the user's historical recharge cycle; η t is the learning rate, η0 is the initial learning rate; ε is the smoothing term; τ is the risk score adjustment coefficient; G t is the squared gradient accumulation term, g i is the gradient.