CN-122022909-A - Multi-mode full-link collaboration method for rights supermarket

CN122022909ACN 122022909 ACN122022909 ACN 122022909ACN-122022909-A

Abstract

The application provides a multi-mode full-link collaboration method for a rights supermarket, belongs to the field of rights personalized service, and is used for solving the problems of rights rule fragmentation, slow multi-scene adaptation and inaccurate user grouping matching in the related technology. The method collects multi-mode data such as behaviors and environments, generates user clusters through the self-adaptive density clustering by integrating the characteristics of a rights-business correlation matrix and an attention mechanism, adopts a mixed pre-judgment model to pre-judge a matching result by combining Bayesian weight inference and LTV prediction modeling influence degree, generates a cluster exclusive rule through multi-objective optimization, and realizes full-link dynamic coordination by matching with a conflict regulation scheme of intention perception, thereby improving service accuracy and operation efficiency.

Inventors

GAO DEYANG
WANG RUIYING
Li Zhangti
WU RUIQI
WANG PENG
Han Jinshuang

Assignees

联通在线信息科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260126

Claims (10)

1. A multi-mode full-link collaboration method for a rights supermarket is characterized by aiming at rights service rule fragmentation, multi-scene high-frequency update and user grouping value difference characteristics, Collecting multimodal data comprising equity exclusive dimensions, wherein equity exclusive dimensions are characteristic dimensions which are designed for equity supermarket business and are different from general service dimensions, and comprise equity acquisition sequences, integral consumption characteristics and other dimensions directly related to equity business, the data comprise a behavior mode, an environment mode, an emotion intention mode and a business rule mode, Behavior mode contains equity acquisition sequence and integral consumption characteristics, environment mode contains provincial region labels and terminal types, emotion intention mode contains equity consultation semantic characteristics, business rule mode contains provincial compliance labels and integral thresholds, Based on the data, constructing an end-to-end link of data fusion, modeling, prejudging, rule optimizing, conflict adjusting and feedback iteration, and realizing the dynamic adaptation of the fragmentation rule.
2. The equity-oriented supermarket multi-mode full-link collaboration method according to claim 1, wherein the multi-mode data fusion process includes constructing equity exclusive mode system, The system comprises a equity combination selection sequence of a behavior mode, a provincial region code of an environment mode, a consultation text emotion score of an emotion intention mode and an equity validity period label of a business rule mode, Based on the system, a rights-service correlation matrix is constructed, attention mechanism fusion characteristics regulated by provincial correlation degree are adopted, and fusion characteristics of fitting provincial rules are output.
3. The equity supermarket oriented multi-mode full-link collaboration method of claim 2, wherein equity influence level modeling process includes constructing equity exclusive dimension set, The set contains a base dimension, a scene-specific dimension and a grouping-specific dimension, the base dimension contains membership grade and equity combination preference, the scene-specific dimension is a characteristic dimension preset for different equity scenes, the grouping-specific dimension is a characteristic dimension customized for different user groupings, And determining dimension weights by adopting rights-oriented Bayesian weight inference, determining dynamic thresholds based on user grouping values, scene importance and number of rights, and adjusting influence scores by combining the tolerance of the integral gaps.
4. The equity-oriented supermarket multi-mode full-link collaborative method according to claim 3, wherein the pre-judging module adopts a mixed pre-judging model of fusion gradient lifting model, meta learning and antagonism learning, The generator generates simulation matching data based on the provincial regional rule labels and integral constraints, the discriminator distinguishes between true and generated data, The method comprises the steps that a small number of sample scenes adopt a meta-gradient accumulation mechanism to accelerate adaptation, network positioning failure reasons are attributed through rights and special errors, and a calibration strategy is triggered, wherein the small number of sample scenes are rights and popularization scenes with the number of existing samples not exceeding a preset sample number threshold value.
5. The equity supermarket oriented multi-mode full-link collaboration method of claim 3, wherein the rule optimizing process includes constructing equity compliance attribution network, Solving the pareto optimal rule through a multi-objective optimization module, optimizing the objective conversion efficiency, the compliance rate and the benefit cost, Generating exclusive rules for different user groups, wherein the high-value member rules relax the constraint of the point gap, the new guest rules simplify the constraint conditions, the high-value member rules correspond to the high-value groups, the high-value groups are the user groups of which the long-term value predicted value of the user is not lower than the preset user value, And after the candidate rule is qualified through compliance verification and matching verification, synchronizing to the corresponding provincial rule subset.
6. The equity oriented supermarket multi-mode full-link collaboration method of claim 3, wherein the conflict reconciliation process includes calculating conflict priorities based on equity core priorities, The priority resets the rights effective option to the highest, which is a member class, a provincial domain and an integral in turn, Constructing a grouping customized scheme library, wherein the high-value member scheme comprises a point coupon and a equity delay, the new guest scheme comprises a point acceleration package and a simplification rule, the high-value member scheme corresponds to high-value grouping, the high-value grouping is user grouping in which the predicted value of the long-term value of the user is not lower than the preset user value, And establishing a mediation-guidance-upgrading closed loop by adopting a constraint reinforcement learning optimization strategy, automatically upgrading and grouping after the user task reaches the standard, and downregulating a rule threshold.
7. The equity-oriented supermarket multi-mode full-link collaboration method of claim 2, wherein the user grouping process includes extracting equity-specific behavior feature sets, The collection contains the rights acquisition frequency, the integral consumption duty ratio, the provincial rights click rate and the long-term value predicted value, Generating core clustering by adopting a rights self-adaptive density clustering algorithm, triggering temporary clustering by an active scene, merging according to long-term value after the scene is ended, And (3) implementing resource tilting based on the grouping value, wherein the high-value grouping matches with a high-quality equity rule, and the high-value grouping is the grouping of users with the long-term value predicted value of the users not lower than the preset user value.
8. The equity-oriented multi-modal all-link collaboration method of claim 3 further comprising the step of equity mathematical model customization, Applying mathematical constraint to the weight of the exclusive dimension of the rights, strengthening the weight ratio of the core dimension through prior distribution, Constructing a rights and interests subsection rewarding mechanism, setting rewarding values according to standard conditions by rigid constraint, setting gradient rewarding according to point gaps by flexible constraint, And constructing a long-term value model based on the rights consumption time sequence, and converting the grouping cost upper limit into a constraint embedded objective function.
9. The equity-oriented supermarket multi-mode full-link collaboration method of claim 4, further comprising a scene feedback iteration step, When the user triggers the rights core behavior, the influence degree weight and the multi-mode characteristic weight are finely adjusted in real time, The peak scene shortens the fine adjustment period of the mixed pre-judgment model, improves the learning rate, updates data and optimizes driving rules in stages in the off-peak period, And updating the modal matrix and the grouping model every week, and triggering emergency feedback when the pre-judgment error exceeds the standard.
10. The equity-oriented supermarket multi-mode full-link collaboration method of claim 5, further comprising: and step of provincial rule subset synchronization and verification, namely synchronizing the provincial rule subset to the rights data cache nodes of the corresponding region after updating, verifying the suitability of the updated rule and the user grouping through a feature comparison mechanism, and triggering the adjustment of dimension weight when the adaptation deviation exceeds a set range.

Description

Multi-mode full-link collaboration method for rights supermarket Technical Field The application relates to the field of rights and interests personalized service, in particular to a multi-mode full-link collaboration method for rights and interests supermarkets. Background With the popularization of digital services, rights and interests supermarkets are rapidly developed as a platform for aggregating rights and interests of various users. As used herein, "equity" refers broadly to various types of digital benefits or privileged services that a user may acquire or redeem, typical forms of which include, but are not limited to, proprietary traffic packages for communications or Internet platforms, membership coupons, off-line merchant coupons, video entertainment membership durations, point redemption physical goods, exemption session qualifications, proprietary customer service channels, and the like. The core requirement of the platform is to provide accurate personalized equity matching service for users so as to promote user viscosity and business transformation. Currently, rights service covers multiple regions and multiple scenes, the user scale and the demand complexity continuously increase, and higher requirements are put on the accuracy and response speed of the service. In implementing personalized matching, a number of key business parameters are typically involved, such as: The points are virtual points accumulated by the user through consumption or active behaviors and are used for equity exchange or rating; membership grade, namely obtaining qualification and priority according to different grades corresponding to different rights and interests according to the grades divided by the historical behaviors of the user, consumption capability and other factors; the region, namely the province region where the user belongs to or is currently located, directly influences compliance and applicable rules of the disciplinable rights and interests. In the prior art, the universal AI personalized methods are mostly applied to the fields of e-commerce, content recommendation and the like, and although the methods can realize basic user preference matching, the methods are not customized for the rights and interests business characteristics. Most schemes adopt fixed dimension data modeling, rely on general behavior characteristics, and do not fully integrate core elements such as exclusive integral consumption mode of rights and interests business, membership grade gradient rights and interests, regional difference rule combination strategies and the like. The prior art has the obvious defects that the system cannot adapt to the characteristics of 'provincial rule fragmentation, multi-scene high-frequency update and remarkable user grouping value difference' of the rights and interests service due to the fact that the recognition and matching model fused with the rights and interests parameters deeply is not constructed. The method is characterized in that the problems of poor module cooperativity, lag in regional rule adaptation, high-value user matching deviation and the like are solved, the service requirements of rights and interests in the aspects of accuracy and high efficiency are difficult to meet, and the large-scale development of rights and interests service is restricted. Disclosure of Invention The application provides a multi-mode full-link collaboration method for a rights supermarket, which can accurately adapt to the rights service characteristics, solve the problems of rule fragmentation and personalized matching, and improve the service accuracy and service efficiency. In a first aspect, the application provides a multi-mode full-link collaboration method for a rights supermarket. The multi-mode full-link collaboration method for the equity supermarket is characterized in that multi-mode data comprising equity exclusive dimensions are collected aiming at equity service rule fragmentation, multi-scene high-frequency updating and user grouping value difference characteristics, the equity exclusive dimensions are feature dimensions which are used for indicating equity supermarket service design and are different from general service dimensions, the feature dimensions comprise equity acquisition sequences, integral consumption features and the like and are directly related to equity service, the data comprise behavior modes, environment modes, emotion intention modes and service rule modes, the behavior modes comprise equity acquisition sequences and integral consumption features, the environment modes comprise provincial region labels and terminal types, the emotion intention modes comprise equity consultation semantic features, the service rule modes comprise provincial compliance labels and integral thresholds, and an end-to-end link of 'data fusion-modeling-prejudgment-rule optimization-conflict mediation-feedback iteration' is constructed based on the data, so that the dynamic adaptati