Transaction and evaluation authenticity probability identification method based on behavior link consistency

Abstract

This invention discloses a method, system, electronic device, and storage medium for probabilistic identification of transaction and review authenticity based on behavioral link consistency, belonging to the field of internet platform risk control and data analysis technology. The method acquires user access behavior, interaction behavior, device behavior, network behavior, and review behavior data; performs anonymization and feature extraction on the behavioral data to generate behavioral link feature vectors; and conducts consistency analysis and risk assessment based on an authenticity probability identification model to generate transaction authenticity probability and review authenticity scores, thereby determining the authenticity of target user behavior and outputting a risk level. This method can achieve probabilistic identification of the authenticity of cross-platform transactions, offline transactions, and private domain transactions without directly obtaining third-party order and payment information, and can be used for internet platform risk control and recommendation ranking optimization.

Description

Technical Field

[0001] This invention relates to the fields of risk control on internet platforms, artificial intelligence, data analysis and generative recommendation technology, and in particular to a method, system, electronic device and storage medium for probabilistic identification of the authenticity of transactions and evaluations based on behavioral link consistency. Background Technology

[0002] With the development of internet platforms, e-commerce platforms, content communities, and generative AI recommendation systems, more and more platforms are using methods such as content recommendation, merchant display, transaction recommendation, lead matching, and generative search to guide users to third-party platforms, offline stores, or merchants' private domain systems to complete transactions.

[0003] In cross-platform transactions, offline transactions, and private domain transactions, platforms typically cannot directly obtain complete order information, payment vouchers, or third-party transaction details.

[0004] At the same time, directly obtaining third-party order details, payment records, bank card information and chat logs may pose privacy risks, data compliance risks and cross-platform data sharing restrictions.

[0005] Most existing technologies rely on single-behavior statistics, rule matching, or blacklist mechanisms for risk identification, making it difficult to effectively infer the authenticity of cross-platform transactions and reviews without obtaining real order information. Summary of the Invention

[0006] This invention provides a probabilistic method for identifying the authenticity of transactions and evaluations based on behavioral link consistency, in order to solve the problem that existing technologies cannot effectively identify the authenticity of cross-platform transactions.

[0007] The present invention adopts the following technical solution:

[0008] Obtain target user behavior data;

[0009] This includes access behavior data, interaction behavior data, device behavior data, network behavior data, and evaluation behavior data.

[0010] The behavioral data is anonymized, aggregated, and feature-extracted to generate behavioral link feature vectors.

[0011] The page dwell characteristics include one or more of the following: mean page dwell time, variance of dwell time, and proportion of long-stay time.

[0012] The equipment stability characteristics include one or more of the following: number of equipment switching times, number of days of continuous equipment use, equipment behavior anomaly rate, and number of equipment-related risk events.

[0013] The device-related risk events include one or more of the following: abnormal review complaints associated with the device, account ban history, blacklist hit records, and abnormal login records.

[0014] Without directly obtaining third-party platform order numbers, payment records, bank card information, and complete delivery addresses, the system performs behavioral sequence analysis, behavioral time dependency analysis, behavioral continuity analysis, behavioral transition path analysis, consistency analysis, and risk assessment on behavioral link feature vectors.

[0015] The authenticity probability recognition model includes one or more of the following: rule recognition model, machine learning model, abnormal behavior detection model, and behavior correlation analysis model.

[0016] The machine learning model includes one or more of the following: logistic regression model, random forest model, gradient boosting tree model, and neural network model.

[0017] The authenticity probability recognition model is trained using a weakly supervised training method. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings described below are only a part of the embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 is a flowchart illustrating a probabilistic identification method for the authenticity of transactions and evaluations based on behavioral link consistency, provided by an embodiment of the present invention.

[0020] Figure 2 is a schematic diagram of the transaction and evaluation authenticity probability recognition system provided in an embodiment of the present invention.

[0021] Figures 3 to 5 are schematic diagrams related to behavioral link consistency analysis provided in the embodiments of the present invention, including schematic diagrams of user device, cloud service and database interaction, schematic diagram of AI Agent generative recommendation behavior entry point, and schematic diagram of behavioral link consistency analysis and ranking adjustment. Terminology Definition

[0022] In this invention, "AI Agent" refers to an intelligent dialogue agent deployed on an internet platform, used to answer user questions, recommend merchant content, or guide users to target pages through natural language interaction.

[0023] In this invention, "Content Optimization Technology for Generative AI Search Results (GEO)" is used. Generative Engine Optimization"" refers to content optimization methods for generative artificial intelligence search systems, AI question-answering systems, and generative recommendation systems, used to increase the exposure probability, citation probability, or ranking weight of target content in generative search results, AI recommendation results, and generative answers. Detailed Implementation

[0024] In one implementation, when a target user simultaneously meets at least three of the following conditions within a preset time window of 1 to 72 hours, the system will label the corresponding behavioral sample as a positive weak label:

[0025] (1) There was a click on the contact information;

[0026] (2) Navigation call behavior exists;

[0027] (3) The page dwell time exceeds the preset threshold;

[0028] (4) There was a secondary visit to the merchant's details page;

[0029] (5) There was coupon redemption behavior;

[0030] (6) There is evaluation and publication behavior.

[0031] In one implementation, the page dwell time threshold is 30 seconds to 5 minutes.

[0032] When a target user only exhibits short-term browsing behavior, does not exhibit redirection behavior, does not exhibit navigation behavior, does not exhibit contact information click behavior, and the page dwell time is less than a preset threshold, the system will mark the corresponding behavior sample as a negative weak label.

[0033] To reduce the impact of weak label noise on model training, the system uses a confidence weighting mechanism to screen samples that have been labeled as positive weak labels.

[0034] For samples that simultaneously meet multiple behavioral conditions, increase the confidence weight of the corresponding sample.

[0035] For samples that meet fewer behavioral conditions, reduce the confidence weight of the corresponding samples.

[0036] The system further combines historical manual review results, historical anomaly handling results, and user complaint results to iteratively correct weak labels.

[0037] The weakly supervised training method employs one or more of the following: PU learning, self-training learning, semi-supervised learning, or multi-instance learning.

[0038] The consistency analysis includes one or more of the following: time consistency analysis, device consistency analysis, network consistency analysis, and content semantic consistency analysis.

[0039] The time-related score is generated based on the time interval of user behavior.

[0040] In one implementation, the time-related score is generated as follows:

[0041]

[0042] in:

[0043] S_t represents the time-related score;

[0044] Δt is the time interval between actions;

[0045] T is the time decay coefficient.

[0046] When the time interval between multiple actions is less than a preset time window, the corresponding time correlation score is increased.

[0047] The device association score is generated based on the device's unique identifier, device switching frequency, device historical stability, and device-related risk events.

[0048] When multiple behaviors use the same device identifier and the device switching frequency is lower than a preset threshold, the device association score is increased.

[0049] The network association score is generated based on IP address, network operator, geographical location, and network switching frequency.

[0050] When multiple behaviors have the same or similar network environment within a preset time window, the network correlation score is improved.

[0051] The content semantic consistency analysis employs one or more of the following: word vector model, TF-IDF model, text similarity model, DistilBERT model, or lightweight semantic coding model.

[0052] In one implementation, the DistilBERT model can be used for lightweight semantic coding analysis to reduce computational resource consumption and response latency in real-time risk control scenarios.

[0053] The system generates semantic matching scores using text cosine similarity.

[0054] The system generates a behavioral link consistency score based on time association score, device association score, network association score, and semantic matching score.

[0055] The preset consistency threshold is generated through historical sample statistics, grid search, model validation set results, or quantile statistics.

[0056] In one implementation, the preset consistency threshold range is 0.65 to 0.90.

[0057] The system dynamically adjusts the time consistency weight, device consistency weight, network consistency weight, and semantic consistency weight based on historical identification results, the accuracy of abnormal behavior identification, the contribution of behavioral features, and the trend of platform risk changes.

[0058] In one implementation, the system adaptively updates the time consistency weight, device consistency weight, network consistency weight, and semantic consistency weight based on an online learning framework.

[0059] The system dynamically adjusts the weights of each dimension based on historical identification results, manual review results, and the accuracy of abnormal behavior identification.

[0060] There is a complementary relationship between time consistency, device consistency, network consistency, and content semantic consistency.

[0061] When a single dimension is noisy or missing, the remaining dimensions still maintain the ability to judge the integrity of the behavioral chain, thereby improving the accuracy of cross-platform transaction authenticity identification.

[0062] For example, when the target user's use of a proxy network leads to a decrease in network consistency, if device consistency, behavior sequence consistency, and content semantic consistency are high, the system can still maintain a high authenticity score, thereby reducing the false positive rate caused by the proxy network.

[0063] The system analyzes not only the behavioral characteristics themselves, but also the sequence of behavior occurrence, the temporal dependence of behavior, the continuity of behavior, and the behavior transfer path.

[0064] In one implementation, the sequence of user actions includes the following steps:

[0065] Step 1 (Content Reach): Target users enter the merchant's post page or product display page through the generative search sorting in the AI ​​Agent's question-and-answer results.

[0066] Step 2 (External Redirect): Users are redirected to external product links customized by the merchant.

[0067] Step 3 (Behavioral Analysis): Based on user behavior links, device behavior characteristics, network behavior characteristics, page dwell behavior, redirection behavior, and historical behavior characteristics, the system performs a probability analysis of the authenticity of target user behavior and generates a comprehensive risk score.

[0068] Step 4 (Follow-up Verification): When the comprehensive risk score meets the preset conditions, the system initiates the collection of follow-up confirmation signals to the target user and generates follow-up confirmation features based on the user's response behavior.

[0069] Step 5 (Cross-validation): The system further performs multi-dimensional cross-validation on the target order based on the order placement time, estimated logistics cycle, user return visit time, evaluation posting time, and behavioral consistency results.

[0070] Step Six (Evaluation Permission Allocation): The system dynamically allocates corresponding evaluation feedback permissions based on the cross-consistency verification results.

[0071] Step 7 (Evaluation Re-verification): After the target user submits a product usage experience evaluation, the system further conducts a cross-consistency analysis on the evaluation content, evaluation behavior, device behavior, and network behavior.

[0072] Step 8 (Data Storage and Sorting Adjustment): The verified evaluation data is entered into the corresponding product evaluation database and used to adjust the generative search ranking, AI recommendation ranking, and corresponding content exposure weight.

[0073] When the behavioral chain satisfies the continuous behavioral sequence described in steps one to four above, the system increases the probability of completion and the authenticity score of the corresponding transaction.

[0074] When the evaluation behavior and behavior link have a consistency score above a preset consistency threshold in terms of time, device, network, and content semantics, the system determines that the corresponding evaluation has a high authenticity score.

[0075] When evaluation behavior involves text duplication, abnormal concentrated release, abnormal device association, or abnormal network association, the system determines that the corresponding evaluation has a high risk of being abnormal.

[0076] The authenticity score results are used for ranking adjustments in AI recommendation models and generative search optimization models.

[0077] For merchants, content, or service providers with high transaction completion probability, high payment completion probability, and high ratings for authenticity of reviews, the system will increase their recommendation weight.

[0078] For merchants, content, or service recipients with abnormal reviews, abnormal behavior associations, paid online commentators, or abnormal conversion behavior, the system will reduce their recommendation weight, restrict their display, or trigger manual review.

[0079] Compared to single-behavior statistics solutions, this invention can improve the identification rate of batch abnormal behaviors, the identification rate of false reviews, and the identification rate of cross-platform abnormal conversions.

[0080] This invention can also reduce the misjudgment rate caused by device spoofing, network switching, short-term abnormal evaluation, and troll gang behavior.

[0081] Compared with existing technologies, this invention does not require direct acquisition of third-party order and payment information. It can make probabilistic inferences about the authenticity of transactions based on the consistency of behavioral links, and can improve the accuracy of identifying abnormal transactions, false reviews and batch abnormal behaviors.

Claims

1. A method for probabilistic identification of the authenticity of transactions and evaluations based on behavioral link consistency, characterized in that, Includes the following steps: S1. Obtain behavioral data of target users on internet platforms; The behavioral data includes access behavior data, interaction behavior data, device behavior data, network behavior data, and evaluation behavior data. S2. Without directly obtaining the order number, payment record, bank card information, chat log, and complete shipping address from the third-party platform, the behavioral data is anonymized, aggregated, and feature-extracted to generate a behavioral link feature vector. The behavioral link feature vector includes one or more of the following: page dwell features, jump behavior features, navigation behavior features, contact information click features, return visit behavior features, device stability features, network stability features, and evaluation behavior features; S3. Based on the authenticity probability identification model, perform behavior sequence analysis, behavior time dependency analysis, behavior continuity analysis, behavior transfer path analysis, consistency analysis and risk assessment on the behavior link feature vector to generate transaction completion probability, payment completion probability and evaluation authenticity score; The authenticity probability recognition model includes one or more of the following: rule recognition model, machine learning model, abnormal behavior detection model, and behavior correlation analysis model. S4. Based on the transaction completion probability, payment completion probability, and evaluation authenticity score, determine the authenticity of the target user's behavior and output the corresponding risk level. S5. Implement the corresponding platform risk control strategy according to the risk level.

2. The method according to claim 1, characterized in that: The behavior link feature vector is generated by numerical processing of user behavior data; The numerical processing includes one or more of the following: behavior frequency statistics, behavior time interval statistics, behavior path statistics, behavior stability statistics, and abnormal behavior statistics.

3. The method according to claim 1, characterized in that: The page dwell characteristics include one or more of the following: mean page dwell time, variance of dwell time, and proportion of long-duration dwell time; The equipment stability characteristics include one or more of the following: number of equipment switching times, number of days of continuous equipment use, equipment behavior anomaly rate, and number of equipment-related risk events.

4. The method according to claim 1, characterized in that: The authenticity probability recognition model adopts one or more of the following: logistic regression model, random forest model, gradient boosting tree model, and neural network model.

5. The method according to claim 1, characterized in that: The authenticity probability recognition model is trained using a weakly supervised training method; Specifically, when a target user simultaneously meets at least three of the following conditions within a preset time window of 1 to 72 hours, the system will label the corresponding behavioral sample as a positive weak label: (1) There was a click on the contact information; (2) Navigation call behavior exists; (3) The page dwell time exceeds the preset threshold; (4) There was a secondary visit to the merchant's details page; (5) There was coupon redemption behavior; (6) There is evaluation and publication behavior.

6. The method according to claim 5, characterized in that: The threshold for page dwell time is 30 seconds to 5 minutes.

7. The method according to claim 5, characterized in that: When a target user only exhibits short-term browsing behavior, does not exhibit redirection behavior, does not exhibit navigation behavior, does not exhibit contact information click behavior, and the page dwell time is less than a preset threshold, the system will mark the corresponding behavior sample as a negative weak label.

8. The method according to claim 5, characterized in that: The system uses a confidence-weighted mechanism to filter samples that have been labeled as positive weak labels; For samples that simultaneously meet multiple behavioral conditions, increase the confidence weight of the corresponding sample. For samples that meet fewer behavioral conditions, reduce the confidence weight of the corresponding samples.

9. The method according to claim 5, characterized in that: The weakly supervised training method employs one or more of the following: PU learning, self-training learning, semi-supervised learning, or multi-instance learning.

10. The method according to claim 1, characterized in that: The consistency analysis includes one or more of the following: time consistency analysis, device consistency analysis, network consistency analysis, and content semantic consistency analysis.

11. The method according to claim 10, characterized in that: The time-related score is generated based on the interval between user behaviors. The time-related score is generated in the following way: Where S_t is the time-related score, Δt is the behavior time interval, and T is the time decay coefficient.

12. The method according to claim 10, characterized in that: The device association score is generated based on the device's unique identifier, device switching frequency, device historical stability, and device-related risk events.

13. The method according to claim 10, characterized in that: The network association score is generated based on IP address, network operator, geographical location, and network switching frequency.

14. The method according to claim 10, characterized in that: The content semantic consistency analysis employs one or more of the following: word vector model, TF-IDF model, text similarity model, DistilBERT model, or lightweight semantic coding model.

15. The method according to claim 1, characterized in that: The system generates a behavioral link consistency score based on time association score, device association score, network association score, and semantic matching score.

16. The method according to claim 15, characterized in that: The preset consistency threshold is generated through historical sample statistical results, grid search method, model validation set results, or quantile statistical method. The preset consistency threshold ranges from 0.65 to 0.

90.

17. The method according to claim 15, characterized in that: The system dynamically adjusts the time consistency weight, device consistency weight, network consistency weight, and semantic consistency weight based on historical identification results, the accuracy of abnormal behavior identification, the contribution of behavioral features, and the trend of platform risk changes.

18. The method according to claim 1, characterized in that: Time consistency, device consistency, network consistency, and content semantic consistency are complementary; When a single dimension is noisy or missing, the remaining dimensions maintain the ability to judge the integrity of the behavioral chain, thereby improving the accuracy of cross-platform transaction authenticity identification.

19. The method according to claim 1, characterized in that: The system identifies multiple target accounts based on device identification, network environment, behavior time distribution, and access path to identify suspected batch abnormal behavior, online troll behavior, or abnormal conversion behavior.

20. The method according to claim 1, characterized in that: The authenticity score results are used for ranking adjustments in AI recommendation models and generative search optimization models.

21. A transaction and evaluation authenticity probability recognition system, characterized in that, include: The data acquisition module is used to acquire target user behavior data; The feature analysis module is used for de-identification processing of behavioral data, feature extraction, and generation of behavioral link feature vectors; The authenticity probability identification module is used to generate the transaction completion probability, payment completion probability, and authenticity score based on the authenticity probability identification model. The consistency analysis module is used to generate behavioral link consistency scores. The risk assessment module is used to output the corresponding risk level; The execution module is used to execute the corresponding platform risk control strategies.

22. An electronic device, characterized in that, include: A processor, a memory, and a computer program stored in the memory, wherein the computer program, when executed by the processor, implements the steps of the method according to any one of claims 1 to 20.

23. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1 to 20.

Images