Digital operation strategy self-optimization method and system based on multi-scene conjoint analysis

Abstract

The invention discloses a digital operation strategy self-optimization method and system based on multi-scene conjoint analysis, and relates to the technical field of big data, and the method comprises the steps: receiving desensitization data after edge computing nodes deployed in all business scenes are subjected to localization processing, and generating a user feature vector according to the desensitization data; aggregating the user feature vectors, determining aggregated node features, and generating a group behavior map based on the aggregated node features; determining a cross-service interaction sequence based on the group behavior map, converting the cross-service interaction sequence into a hidden space vector, and generating a dynamic feature tag according to the hidden space vector; and performing right matching on the dynamic feature tag based on a right matching model, and outputting a right matching result. Through the mode, the dynamic feature tag is generated based on the edge node data, and the right matching model adjusts the strategy according to the dynamic feature tag, so that a closed-loop optimization system which is self-evolved according to group behavior changes is formed, and the problem that an operation strategy cannot be self-optimized in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data, and particularly relates to a digital operation strategy self-optimization method and system based on multi-scene joint analysis. BACKGROUND

[0002] With the deepening of digital transformation, various entity business scenarios such as local life services, energy supply, supermarket retail and travel services increasingly rely on data-driven intelligent decision-making. In these scenarios, in order to deliver corresponding benefits to users, a preset benefit allocation strategy is usually used to match benefits according to user privacy information, and then corresponding benefits are pushed. However, this approach may result in allocated benefits that are not what users need.

[0003] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0004] The main purpose of the present application is to provide a digital operation strategy self-optimization method and system based on multi-scene joint analysis, aiming to solve the technical problem that the existing technology cannot self-optimize the benefit matching strategy when matching benefits.

[0005] To achieve the above purpose, the present application provides a digital operation strategy self-optimization method based on multi-scene joint analysis, which comprises: receiving desensitized data processed locally by edge computing nodes deployed in each business scenario, and generating a user feature vector according to the desensitized data; aggregating the user feature vector to determine an aggregated node feature, and generating a group behavior graph based on the aggregated node feature; determining a cross-business interaction sequence based on the group behavior graph, converting the cross-business interaction sequence into a latent space vector, and generating a dynamic feature label according to the latent space vector; performing benefit matching on the dynamic feature label based on a benefit matching model, and outputting a benefit matching result.

[0006] In an embodiment, the step of aggregating the user feature vector to determine an aggregated node feature, and generating a group behavior graph based on the aggregated node feature comprises: determining a target scene in the user feature vector, taking the target scene as a target aggregation point, fusing the target aggregation point when the target aggregation point is the same, and determining the target aggregation point as an aggregation center; aggregating the user feature vector based on the aggregation center to obtain a business scenario switching relationship; determining an aggregation number of the service scenario switching relations pointing to the aggregation center, when a proportion of the aggregation number in a total number is greater than a preset proportion, determining the aggregation center as a target aggregation center, and determining the service scenario switching relations pointing to the target aggregation center as an aggregation node feature; generating a group behavior graph with the target aggregation center and the aggregation node feature.

[0007] In an embodiment, after the step of determining the service scenario switching relations pointing to the target aggregation center as an aggregation node feature, the method further comprises: determining remaining service scenario switching relations other than the aggregation node feature as a free feature; when a new user feature vector is received, determining a new target aggregation center based on the free feature, the new user feature vector, and a latest total number; when the new target aggregation center exists, determining a corresponding new aggregation node feature based on the new target aggregation center; updating the group behavior graph based on the new target aggregation center and the new aggregation node feature.

[0008] In an embodiment, the step of determining a cross-service interaction sequence based on the group behavior graph, converting the cross-service interaction sequence into a latent space vector, and generating a dynamic feature label according to the latent space vector comprises: identifying continuous interaction behaviors of a user among multiple service scenarios based on the group behavior graph, and constructing a cross-service interaction sequence based on the continuous interaction behaviors; mapping the cross-service interaction sequence to a low-dimensional latent space based on a service mapping model, and generating a latent space vector representing a user behavior pattern; performing clustering analysis and pattern recognition on the latent space vector, and generating a dynamic feature label reflecting real-time behavior characteristics of the user.

[0009] In an embodiment, the step of identifying continuous interaction behaviors of a user among multiple service scenarios based on the group behavior graph, and constructing a cross-service interaction sequence based on the continuous interaction behaviors comprises: extracting a time sequence behavior pattern in the group behavior graph, and identifying continuous access trajectories of a user among different service scenarios within a preset time window based on the time sequence behavior pattern; sorting the scenario switching behaviors in the continuous access trajectories based on time sequence, and constructing a cross-service interaction sequence with a scene identifier as an element.

[0010] In an embodiment, the step of mapping the cross-service interaction sequence to a low-dimensional latent space based on a service mapping model, and generating a latent space vector representing a user behavior pattern comprises: encoding processing on the cross-business interaction sequence based on a business mapping model to obtain a one-hot encoding based on a business type; mapping the one-hot encoding to a low-dimensional hidden space to generate a numerical sequence; parallel computing on the numerical sequence based on an attention mechanism to determine the correlation between any two elements in the numerical sequence, and generating a short vector based on the correlation; forward propagation on the short vector to generate a hidden space vector representing a user behavior pattern.

[0011] In an embodiment, the step of clustering analysis and pattern recognition on the hidden space vector to generate a dynamic feature label reflecting the real-time behavior characteristics of the user includes: unsupervised clustering on the hidden space vector to identify user groups of target cross-scene behavior patterns; backtracking the cross-business interaction sequence and the de-identified data corresponding to each user group, and performing pattern recognition on the cross-business interaction sequence and the de-identified data to extract group behavior characteristics; based on the group behavior characteristics, generating a dynamic feature label combination for each user group; similarity calculation between the hidden space vector and each user group center to obtain a matching similarity set; determining the target user group center corresponding to the maximum matching similarity in the matching similarity; generating a dynamic feature label corresponding to the target user group center.

[0012] In an embodiment, the step of matching rights based on the rights matching model and outputting the rights matching result includes: real-time generation of an individualized rights matching strategy based on the group characteristics of the user and the dynamic feature label; evaluation of the individualized rights matching strategy that has been executed, with user conversion rate and input-output ratio as reward signals; optimization of the inner loop weight parameters of the rights matching model based on the reward signals to obtain optimized inner loop weight parameters; updating the inner loop strategy model based on the optimized inner loop weight parameters, and outputting the rights matching result based on the optimized inner loop strategy model.

[0013] In an embodiment, the step of generating a user feature vector based on the de-identified data includes: analyzing the de-identified data to determine the business scene switching sequence in the de-identified data; Determine the occurrence frequency of the service scene switching sequence, take the occurrence frequency as a feature weight, and generate a user feature vector based on the feature weight and the service scene switching sequence.

[0014] In addition, to achieve the above object, the present application also proposes a digital operation strategy self-optimization system based on multi-scene joint analysis, which comprises: A data processing module is configured to receive desensitized data processed by edge computing nodes deployed in each business scene, and generate a user feature vector based on the desensitized data; A feature processing module is configured to aggregate the user feature vector, determine aggregated node features, and generate a group behavior graph based on the aggregated node features; A business processing module is configured to determine a cross-business interaction sequence based on the group behavior graph, convert the cross-business interaction sequence into a latent space vector, and generate a dynamic feature label based on the latent space vector; An equity matching module is configured to match the dynamic feature label based on an equity matching model, and output an equity matching result.

[0015] In addition, to achieve the above object, the present application also proposes a digital operation strategy self-optimization device based on multi-scene joint analysis, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the above-mentioned digital operation strategy self-optimization method based on multi-scene joint analysis.

[0016] In addition, to achieve the above object, the present application also proposes a storage medium, which is a computer readable storage medium, and the storage medium stores a computer program, the computer program being executed by a processor to implement the steps of the above-mentioned digital operation strategy self-optimization method based on multi-scene joint analysis.

[0017] In addition, to achieve the above object, the present application also provides a computer program product, which comprises a computer program, the computer program being executed by a processor to implement the steps of the above-mentioned digital operation strategy self-optimization method based on multi-scene joint analysis.

[0018] The application provides a digital operation strategy self-optimization method based on multi-scene joint analysis, which solves the problem that the equity matching strategy cannot be self-optimized by constructing a closed-loop system based on real-time data flow and machine learning driving: first, the edge computing node continuously provides desensitized data processed locally, so that the system can dynamically generate feature vectors reflecting the real-time behavior of users; then, by aggregating these vectors to form a group behavior map, and extracting cross-business interaction sequences, they are converted into latent space vectors to capture complex behavior pattern changes, thereby generating dynamic feature labels that can adaptively reflect the evolution of user demand; finally, the equity matching model uses these dynamic feature labels for matching, and continuously optimizes the model parameters through an implicit feedback mechanism, so that the matching strategy can be automatically adjusted and improved according to new data, realizing a self-optimization cycle. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings incorporated in and forming a part of the specification illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the application.

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative labor.

[0021] Figure 1 The flowchart of the first embodiment of the digital operation strategy self-optimization method based on multi-scene joint analysis of the present application; Figure 2 The group behavior diagram of the first embodiment of the digital operation strategy self-optimization method based on multi-scene joint analysis of the present application; Figure 3 The module structure diagram of the digital operation strategy self-optimization system based on multi-scene joint analysis of the embodiment of the present application; Figure 4 The device structure diagram of the hardware running environment involved in the digital operation strategy self-optimization method based on multi-scene joint analysis in the embodiment of the present application.

[0022] The implementation of the present application, functional characteristics and advantages will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0023] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application, and are not used to limit the present application.

[0024] For better understanding of the technical solutions of the present application, the following will be described in detail in combination with the drawings of the specification and specific embodiments.

[0025] The main solution of the embodiment of the present application is: receiving desensitized data processed by edge computing nodes deployed in various business scenarios, generating a user feature vector according to the desensitized data; Aggregating the user feature vector to determine an aggregated node feature, and generating a group behavior graph based on the aggregated node feature; Determining a cross-business interaction sequence based on the group behavior graph, converting the cross-business interaction sequence into a latent space vector, and generating a dynamic feature label according to the latent space vector; Matching the dynamic feature label based on an equity matching model, and outputting an equity matching result.

[0026] At present, with the deepening of digital transformation, various entity business scenarios such as local life services, energy supply, supermarket retail and travel services increasingly rely on data-driven intelligent decision-making. In these scenarios, in order to deliver corresponding rights to users, a preset equity allocation strategy is usually used to match the rights according to the user's privacy information, and then the corresponding rights are pushed. However, this approach may result in the allocation of rights that are not needed by the user.

[0027] The present application provides a solution to solve the problem that the equity matching strategy cannot be self-optimized by constructing a closed-loop system based on real-time data flow and machine learning. First, the edge computing node continuously provides desensitized data processed locally, so that the system can dynamically generate a feature vector reflecting the user's real-time behavior. Then, by aggregating these vectors to form a group behavior graph, and extracting cross-business interaction sequences, they are converted into latent space vectors to capture complex behavior pattern changes, thereby generating dynamic feature labels that can adaptively reflect the evolution of user demand. Finally, the equity matching model uses these dynamic feature labels for matching, and continuously optimizes the model parameters through an implicit feedback mechanism, so that the matching strategy can be automatically adjusted and improved according to new data, realizing a self-optimization cycle.

[0028] It should be noted that the execution subject of the present embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, a digital operation strategy self-optimization device based on multi-scene joint analysis, etc. The present embodiment does not make specific limitations on this. The following will take the digital operation strategy self-optimization device based on multi-scene joint analysis as an example to describe the present embodiment and the following embodiments.

[0029] All actions of obtaining signals, information or data in this application are carried out in compliance with the corresponding data protection regulations and policies of the country where the device is located, and with the authorization given by the corresponding device owner.

[0030] The embodiment of the present application provides a digital operation strategy self-optimization method based on multi-scene joint analysis. Figure 1 , Figure 1 FIG. 1 is a flowchart of a first embodiment of the digital operation strategy self-optimization method based on multi-scene joint analysis.

[0031] In the embodiment, the digital operation strategy self-optimization method based on multi-scene joint analysis comprises steps S10-S40: Step S10, receiving desensitized data processed locally by edge computing nodes deployed in each business scene, and generating a user feature vector according to the desensitized data.

[0032] It should be noted that the business scene refers to local services in life, such as energy supply, supermarket retail, and travel services. The edge computing node refers to a computing device in a business site such as a gas station, a chain store, and a charging station, which is responsible for preliminary processing and data desensitization at the data source. The desensitized data is non-sensitive data after local processing and removal of personal identification information, which only retains abstract features that can be used for analysis. The user feature vector refers to a structured numerical vector extracted from the desensitized data for representing user behavior preferences or states.

[0033] It can be understood that each edge node locally cleanses, normalizes, and desensitizes original business data (such as transaction records, behavior logs, etc.), and sends the desensitized data to a central aggregation node through a secure transmission protocol. After receiving the data, the central system fuses and vectorizes the multi-source data, and finally generates a user feature vector in a unified format.

[0034] In a feasible implementation, the step of generating a user feature vector according to the desensitized data comprises: analyzing the desensitized data to determine a business scene switching sequence in the desensitized data; determining a frequency of occurrence of the business scene switching sequence, taking the frequency of occurrence as a feature weight, and generating a user feature vector based on the feature weight and the business scene switching sequence.

[0035] It should be noted that the business scene switching sequence refers to a sequential expression of switching from one business scene to another, for example, a user refuels at a gas station and then consumes in a supermarket, so the business scene switching sequence of the user in this process can be determined as from the gas station to the supermarket. The feature weight refers to the probability of the user switching from one scene to another.

[0036] In a specific implementation, the received desensitized user behavior logs are first sorted by time, the transition relationship between adjacent business scenarios is identified, and all possible scenario switching sequences are extracted. Then the occurrence frequency of each sequence, i.e., the number of occurrences of the sequence , is counted. The sequence and its weight are integrated into a user feature vector. Let the set of all possible switching sequences be , then the user feature vector is represented as:

[0037] Each dimension corresponds to the occurrence frequency of a sequence. The vector not only retains the time sequence characteristics of the user's cross-scene behavior, but also highlights its main behavior patterns through frequency weights.

[0038] In step S20, the user feature vectors are aggregated to determine aggregated node features, and a group behavior graph is generated based on the aggregated node features.

[0039] It should be noted that the aggregated node features refer to the feature vectors obtained by clustering the feature vectors of all users in a node, such as a specific business district, time period, or user group, representing the common patterns or trends of the group in business scenario switching.

[0040] It should be understood that the group behavior graph refers to a model that describes the flow relationship and intensity between different business scenarios in the form of a graph structure, as shown in Figure 2 , Figure 2 , which is a group behavior diagram. In the diagram, the nodes represent different business scenarios, the edges represent the switching relationship between the scenarios, and the weight of the edge reflects the frequency of switching. For example, the group behavior graph of the lunch and lunch break behavior patterns of the white-collar group in a central business district of a city during the noon period includes office building A, chain fast food store B, coffee shop C, convenience store D, and city park E. The edges formed between the nodes can be represented as office building A→chain fast food store B (weight: 0.65), office building A→coffee shop C (weight: 0.25), chain fast food store B→convenience store D (weight: 0.15), chain fast food store B→office building A (weight: 0.80), coffee shop C→city park E (weight: 0.40), coffee shop C→office building A (weight: 0.55), and city park E→office building A (weight: 0.90). This graph shows how the group behavior graph can convert abstract data into intuitive and actionable business insights, clearly revealing the movement rules and preferences of specific groups in physical space.

[0041] It can be understood that the user feature vectors are grouped according to dimensions such as regions and time, an aggregation operation is performed on the vectors in each group, and thus aggregated node features are obtained. Then, a directed graph with business scenarios as nodes and inter-scene switching probabilities or frequencies as edge weights, i.e., a group behavior graph, is constructed according to the weights of different scene switching sequences in the aggregated features. The graph can be further used for community discovery, key path analysis and other tasks, so as to reveal the internal structure and rules of group behavior.

[0042] In a feasible implementation, the step of aggregating the user feature vectors, determining aggregated node features, and generating a group behavior graph based on the aggregated node features comprises: determining a target scene in the user feature vectors, taking the target scene as a target aggregation point, fusing the target aggregation points when the target aggregation points are the same, and determining the target aggregation points as aggregation centers; aggregating the user feature vectors based on the aggregation centers to obtain business scene switching relationships; determining the number of aggregations of the business scene switching relationships pointing to the aggregation centers, and determining the aggregation centers as target aggregation centers when the proportion of the number of aggregations in the total number is greater than a preset proportion, and determining the business scene switching relationships pointing to the target aggregation centers as aggregated node features; generating a group behavior graph based on the target aggregation centers and the aggregated node features.

[0043] It should be noted that the target scene refers to a business scene that is the end point of user behavior in the user feature vector, the target aggregation point is a temporary grouping identifier for classifying user behaviors with the same target scene, and the aggregation center is a representative point formed by fusing all user feature vectors with the same target scene, which represents a specific destination scene and all user behaviors attracted by the destination scene. The business scene switching relationship refers to the complete path sequence of a user from a starting scene to a target aggregation center.

[0044] In a specific implementation, first, each user feature vector is parsed, the end point of the behavior sequence thereof is extracted as a target scene, and the target scene is used as a target aggregation point to preliminarily group users. The user feature vectors with the same target aggregation point are fused to form aggregation centers. Then, the number of scene switching paths with each aggregation center as the end point is counted, and the proportion thereof is calculated. The core destinations with a proportion exceeding a preset proportion are selected as target aggregation centers. Finally, all paths pointing to the target aggregation centers and the weights (frequencies) thereof are extracted as aggregated node features, which are used to generate a group behavior graph with the target aggregation centers as nodes and the inflow paths as edges.

[0045] Let the set of all user feature vectors be where each vector contains a switching sequence, the end of which is the target scene. All unique target scene sets are For a certain target scene its aggregate center is obtained by fusing all user feature vectors.

[0046]

[0047] where, is the clustering function.

[0048] Let be the total number of business scene switching relations, the number of switching relations pointing to the aggregate center is When , is determined as the target aggregate center, where is a preset proportion threshold. Finally, the group behavior graph is generated based on the target aggregate center and the aggregate node features.

[0049] In a possible implementation, after the step of determining the aggregate node features pointing to the target aggregate center, the method further includes: determining the remaining business scene switching relations other than the aggregate node features as free features; when a new user feature vector is received, determining a new target aggregate center based on the free features, the new user feature vector and the latest total number; when the new target aggregate center exists, determining corresponding new aggregate node features based on the new target aggregate center; updating the group behavior graph based on the new target aggregate center and the new aggregate node features.

[0050] It can be understood that, in the aggregation process, there are some business scene switching relations that are not included in the target aggregate center, which are referred to as free features, and correspond to extremely rare behavior patterns, which are temporarily non-mainstream, low-frequency or emerging user behavior patterns.

[0051] It should be understood that when new user feature vectors are received, the new data will be merged with the historically accumulated free features, and the proportion of each potential aggregation center will be recalculated based on the latest total number at that time. If the proportion of a feature exceeds the threshold, it will be identified as a new target aggregation center, and the corresponding new aggregation node feature will be extracted. Finally, the existing group behavior graph is dynamically updated by adding nodes and edges, so that the graph can adaptively reflect the evolution of user behavior patterns.

[0052] Step S30, determining a cross-business interaction sequence based on the group behavior graph, converting the cross-business interaction sequence into a latent space vector, and generating a dynamic feature label according to the latent space vector.

[0053] It should be noted that the cross-business interaction sequence refers to an ordered path extracted from the group behavior graph, representing the transition of a typical user group between different business scenarios, such as "office building → convenience store → coffee shop". The latent space vector is a numerical representation of a low-dimensional, dense continuous vector space obtained by embedding technology to map discrete sequences. This vector can capture deep semantic information and behavior patterns. The dynamic feature label is a label generated by a clustering or classification model to describe the dynamic behavior characteristics of the group, such as leisure socialization, midday efficient shopping, etc. The label can be dynamically updated with the graph and data.

[0054] It can be understood that a large number of node sequences are generated on the group behavior graph to simulate user cross-scenario interaction behavior, thereby obtaining a set of cross-business interaction sequences. Then, each scene or the entire sequence in these sequences is encoded into a fixed-dimensional vector in the latent space. Finally, based on these vectors, unsupervised clustering or classification combined with a small amount of labeled data is performed to convert each vector or a group of vectors into readable and usable dynamic feature labels, thereby completing the conversion from graph structure data to semantic labels.

[0055] In one possible implementation, the step of determining a cross-business interaction sequence based on the group behavior graph, converting the cross-business interaction sequence into a latent space vector, and generating a dynamic feature label according to the latent space vector includes: identifying continuous interaction behavior of a user between multiple business scenarios based on the group behavior graph, and constructing a cross-business interaction sequence based on the continuous interaction behavior; mapping the cross-business interaction sequence to a low-dimensional latent space based on a business mapping model to generate a latent space vector representing a user behavior pattern; performing clustering analysis and pattern recognition on the latent space vector to generate a dynamic feature label reflecting real-time behavior characteristics of the user.

[0056] In a specific implementation, when constructing the dynamic feature label, sampling is performed on the group behavior graph, a user behavior path is simulated to perform multiple walks from each node in the graph, and a large number of fixed-length sequences are generated. These sequences are cross-business interaction sequences. Then, the discrete sequences are mapped to a low-dimensional latent space to generate latent space vectors. The obtained set of latent space vectors is subjected to unsupervised clustering analysis, and k cluster centers are determined by minimizing the within-cluster sum of squares, each cluster representing a group of users with similar behavior patterns. Then, a semantic label is assigned to each cluster, such as "commuting energy supplement type". These labels are derived from the latest behavior data and are therefore dynamic, reflecting real-time or near real-time changes in user behavior characteristics.

[0057] In a feasible implementation, the step of identifying continuous interaction behaviors of users between multiple business scenarios based on the group behavior graph and constructing cross-business interaction sequences based on the continuous interaction behaviors includes: extracting a time sequence behavior pattern in the group behavior graph, and identifying continuous access trajectories of users between different business scenarios within a preset time window based on the time sequence behavior pattern; sorting the scene switching behaviors in the continuous access trajectories based on time sequence, and constructing a cross-business interaction sequence with scene identifiers as elements.

[0058] In a specific implementation, when constructing cross-business interaction sequences, the time sequence behavior pattern can be extracted and the continuous access trajectory can be identified first. The edges in the group behavior graph not only contain switching frequencies, but also usually imply time sequence information. By analyzing the high-frequency paths and node transition probabilities in the graph, a general time sequence behavior pattern is extracted. For example, refueling a new energy vehicle after work is a pattern. Based on these patterns, a preset time window, such as 2 hours, can be set. Then, in the original user behavior log, a series of continuous business scenario access events of each user within the window are filtered out to form the continuous access trajectory of the user. The trajectory is a set of events arranged in ascending order of timestamp. Then, for each identified continuous access trajectory, the system ignores the specific timestamp and strictly sorts the scene access events in time sequence to extract an ordered list of pure scene identifiers. This list is the final generated cross-business interaction sequence.

[0059] In a feasible implementation, the step of mapping the cross-business interaction sequence to a low-dimensional latent space based on the business mapping model to generate a latent space vector representing a user behavior pattern includes: encoding the cross-business interaction sequence based on the business mapping model to obtain a one-hot encoding based on business types; mapping the one-hot encoding to a low-dimensional latent space to generate a numerical sequence; perform parallel calculation on the numerical sequence based on an attention mechanism, determine the correlation between any two elements in the numerical sequence, and generate a short vector based on the correlation; perform forward propagation on the short vector to generate a latent space vector representing a user behavior pattern.

[0060] In a specific implementation, a cross-business interaction sequence is encoded based on a business mapping model, a unique code is assigned to each cross-business interaction sequence, and then the unique code is all-hot encoded and mapped to a low-dimensional latent space to obtain a numerical sequence. The numerical sequence is then input into an encoding layer based on a self-attention mechanism through an attention-generated context vector. The self-attention mechanism captures long-distance dependencies by calculating the correlation score between any two elements in the sequence. The calculation formula can be expressed as:

[0061] wherein, and are trainable query and key weight matrices, is the dimension of the key vector, is the correlation score, and are any two elements.

[0062] After the correlation score is determined, the correlation score can be normalized and converted into an attention weight. The numerical sequence is weighted and summed based on the attention weight to generate a new vector containing global context information for each cross-business interaction sequence. Finally, a fixed short vector is obtained by pooling the new vector of the sequence, which comprehensively represents the information of the entire sequence. Finally, a latent space vector is generated through forward propagation. The short vector is input into a feedforward neural network, which usually includes one or more fully connected layers and activation functions for nonlinear transformation and dimension adjustment. The output is the final latent space vector representing the user behavior pattern.

[0063] In a feasible implementation, the step of performing clustering analysis and pattern recognition on the latent space vector to generate a dynamic feature label reflecting the real-time behavior characteristics of the user includes: performing unsupervised clustering on the latent space vector to identify user groups of target cross-scenario behavior patterns; backtracking the cross-business interaction sequences and the de-identified data corresponding to each of the user groups, performing pattern recognition on the cross-business interaction sequences and the de-identified data, and extracting group behavior characteristics; generating a dynamic feature label combination for each of the user groups based on the group behavior characteristics; The similarity between the latent space vector and each user group center is calculated to obtain a matching similarity set; The target user group center corresponding to the maximum matching similarity in the matching similarity is determined; The dynamic feature label corresponding to the target user group center is combined to generate a dynamic feature label.

[0064] In a specific implementation, the user groups are identified by unsupervised clustering. The clustering algorithm K-Means is used to divide the entire set of latent space vectors The goal is to find K clusters that minimize the intra-cluster variance:

[0065] wherein, represents the kth cluster, is the user group center. Each cluster represents a group of users with similar target cross-scenario behavior patterns.

[0066] Then the data is traced back and the group behavior features are extracted. The original cross-business interaction sequences and de-identified data corresponding to all latent space vectors in each cluster are traced back. By statistically analyzing these sequences and data, group behavior features such as "most frequently visited scenario sequence", "average stay duration", "consumption time preference", etc. are extracted. A dynamic feature label combination is generated. Based on the extracted group behavior features, a dynamic feature label combination that is easy to understand is generated for each user group For example, if the behavior features of a group show "morning rush hour on weekdays, from residential area to commercial building, fixed path", the label combination may be generated. Subsequently, similarity calculation and label matching are performed. When a new latent space vector needs to be labeled, its similarity with all existing group centers is calculated to form a matching similarity set. The maximum similarity is determined, and the corresponding user group center is determined as the target user group center. Finally, the target user group center is used as the final dynamic feature label for the new user vector.

[0067]

[0068] It should be noted that the benefit matching result refers to one or more optimal benefit schemes selected and recommended for a user group with a specific dynamic feature label. This result usually includes specific benefit identifiers, push strategies, and expected evaluations, such as pushing a combination of "cooperative gas station discount coupons" and "highway service area consumption discount coupons" to users labeled as "high-frequency cross-city commuters".

[0069] It can be understood that the process of matching the dynamic feature label with the benefit based on the benefit matching model and outputting the benefit matching result depends on the pre-trained benefit matching model. The model can be a rule-based expert system or a machine learning model. The input is the dynamic feature label, and the output is a list of benefits sorted by matching degree. The model first performs multidimensional matching calculation between the label and each benefit in the benefit library, considering factors such as target audience definition, scene relevance, budget limit, and historical conversion rate. Then, the model calculates a comprehensive matching score for each benefit, filters out the highest score or a group of benefits, and forms the final benefit matching result. This result can be directly provided to the business system to trigger actual operation actions such as coupon issuance and message pushing.

[0070] In a feasible implementation, the step of matching the dynamic feature label with the benefit based on the benefit matching model and outputting the benefit matching result includes: generating a personalized benefit matching strategy in real time based on the group characteristics of the user and the dynamic feature label; evaluating the executed personalized benefit matching strategy, taking user conversion rate and input-output ratio as reward signals; optimizing the inner loop weight parameters of the benefit matching model based on the reward signals to obtain optimized inner loop weight parameters; updating the inner loop strategy model based on the optimized inner loop weight parameters, and outputting the benefit matching result based on the optimized inner loop strategy model.

[0071] In specific implementation, a personalized benefit matching strategy is generated in real time by an inner loop strategy model according to the group characteristics of the user and the specific dynamic feature label of the user. The strategy specifies what kind of benefit (such as a specific fuel coupon) and push parameters (such as denomination and validity period) should be pushed to the user. Then, the executed strategy is continuously evaluated. The evaluation is based on actual business data, and user conversion rate (such as coupon cancellation rate) and input-output ratio are taken as core reward signals. These quantitative reward signals reflect the effectiveness of the strategy. Then, the reward signals are used to optimize the inner loop weight parameters of the benefit matching model through reinforcement learning algorithms such as policy gradient. The goal is to adjust the weights of the neural network so that the model is more likely to make decisions that can obtain higher cumulative rewards in the future. This optimization process can be formalized as maximizing expected rewards. Finally, the optimized weight parameters are updated to the inner loop strategy model. This new model, which has been learned online and enhanced in performance, is used to process subsequent matching requests, thereby outputting more accurate and higher return-on-investment benefit matching results, forming an automatic closed-loop optimization system of “decision-evaluation-optimization-re-decision”.

[0072] The embodiment provides a digital operation strategy self-optimization method based on multi-scene joint analysis, which solves the problem that an equity matching strategy cannot be self-optimized by constructing a closed-loop system based on real-time data flow and machine learning driving: first, an edge computing node continuously provides desensitized data processed locally, so that the system can dynamically generate a feature vector reflecting real-time behavior of a user; then, a group behavior graph is formed by aggregating the vectors, and a cross-business interaction sequence is extracted and converted into a latent space vector to capture complex behavior pattern changes, so as to generate dynamic feature labels that can adaptively reflect evolution of user demand; finally, an equity matching model uses the dynamic feature labels for matching, and continuously optimizes model parameters through an implicit feedback mechanism, so that the matching strategy can be automatically adjusted and improved according to new data, and a self-optimization cycle is realized.

[0073] It should be noted that the above examples are only used for understanding the present application and do not constitute a limitation on the digital operation strategy self-optimization method based on multi-scene joint analysis of the present application. More forms of simple transformation based on the technical concept are within the protection scope of the present application.

[0074] The present application also provides a digital operation strategy self-optimization system based on multi-scene joint analysis, please refer to Figure 3 The digital operation strategy self-optimization system based on multi-scene joint analysis comprises: A data processing module 10 is configured to receive desensitized data processed locally by edge computing nodes deployed in each business scene, and generate a user feature vector according to the desensitized data; A feature processing module 20 is configured to aggregate the user feature vector, determine an aggregated node feature, and generate a group behavior graph based on the aggregated node feature; A business processing module 30 is configured to determine a cross-business interaction sequence based on the group behavior graph, convert the cross-business interaction sequence into a latent space vector, and generate a dynamic feature label according to the latent space vector; An equity matching module 40 is configured to perform equity matching on the dynamic feature label based on an equity matching model, and output an equity matching result.

[0075] In an implementation, the feature processing module 20 is further configured to determine a target scenario in the user feature vector, take the target scenario as a target aggregation point, fuse the target aggregation points when the target aggregation points are the same, determine the target aggregation points as an aggregation center, aggregate the user feature vector based on the aggregation center to obtain a service scenario switching relationship, determine an aggregation number of the service scenario switching relationship pointing to the aggregation center, and determine the aggregation center as a target aggregation center when a proportion of the aggregation number in a total number is greater than a preset proportion, take the service scenario switching relationship pointing to the target aggregation center as an aggregation node feature, and generate a group behavior graph based on the target aggregation center and the aggregation node feature.

[0076] In an implementation, the feature processing module 20 is further configured to determine a remaining service scenario switching relationship other than the aggregation node feature as a free feature, determine a new target aggregation center based on the free feature, a new user feature vector, and a latest total number when a new user feature vector is received, determine a corresponding new aggregation node feature based on the new target aggregation center when the new target aggregation center exists, and update the group behavior graph based on the new target aggregation center and the new aggregation node feature.

[0077] In an implementation, the service processing module 30 is further configured to identify a continuous interaction behavior of a user between multiple service scenarios based on the group behavior graph, construct a cross-service interaction sequence based on the continuous interaction behavior, map the cross-service interaction sequence to a low-dimensional hidden space based on a service mapping model to generate a hidden space vector representing a user behavior pattern, and generate a dynamic feature label reflecting a real-time behavior feature of the user by performing clustering analysis and pattern recognition on the hidden space vector.

[0078] In an implementation, the service processing module 30 is further configured to extract a time sequence behavior pattern in the group behavior graph, identify a continuous access trajectory of a user between different service scenarios within a preset time window based on the time sequence behavior pattern, sort scene switching behaviors in the continuous access trajectory based on a time sequence, and construct a cross-service interaction sequence taking a scene identifier as an element.

[0079] In an implementable embodiment, the business processing module 30 is further configured to encode the cross-business interaction sequence based on a business mapping model to obtain a one-hot encoding based on a business type, map the one-hot encoding to a low-dimensional latent space to generate a numerical sequence, perform parallel calculation on the numerical sequence based on an attention mechanism to determine a correlation between any two elements in the numerical sequence, generate a short vector based on the correlation, and perform forward propagation on the short vector to generate a latent space vector representing a user behavior pattern.

[0080] In an implementable embodiment, the business processing module 30 is further configured to perform unsupervised clustering on the latent space vector to identify a user group of a target cross-scene behavior pattern, backtrack the cross-business interaction sequence and the de-identified data corresponding to each user group to perform pattern recognition on the cross-business interaction sequence and the de-identified data and extract group behavior features, generate a dynamic feature label combination for each user group based on the group behavior features, perform similarity calculation on the latent space vector and a user group center to obtain a matching similarity set, determine a target user group center corresponding to a maximum matching similarity in the matching similarity, and generate a dynamic feature label based on the dynamic feature label combination corresponding to the target user group center.

[0081] In an implementable embodiment, the benefit matching module 40 is further configured to generate a personalized benefit matching strategy in real time based on the group features of the user and the dynamic feature label, evaluate the personalized benefit matching strategy that has been executed, take a user conversion rate and an input-output ratio as a reward signal, optimize an inner loop weight parameter of a benefit matching model based on the reward signal to obtain an optimized inner loop weight parameter, update an inner loop strategy model based on the optimized inner loop weight parameter, and output a benefit matching result based on the optimized inner loop strategy model.

[0082] In an implementable embodiment, the data processing module 10 is further configured to analyze the de-identified data to determine a business scene switching sequence in the de-identified data, determine a frequency of occurrence of the business scene switching sequence, take the frequency of occurrence as a feature weight, and generate a user feature vector based on the feature weight and the business scene switching sequence.

[0083] The digital operation strategy self-optimization system based on multi-scenario joint analysis provided in the application adopts the digital operation strategy self-optimization method based on multi-scenario joint analysis in the above embodiment, and can solve the technical problem that the equity matching strategy cannot be self-optimized in equity matching. Compared with the prior art, the digital operation strategy self-optimization system based on multi-scenario joint analysis provided in the application has the same beneficial effects as the digital operation strategy self-optimization method based on multi-scenario joint analysis provided in the above embodiment, and other technical features in the digital operation strategy self-optimization system based on multi-scenario joint analysis are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0084] The application provides a digital operation strategy self-optimization device based on multi-scenario joint analysis. The digital operation strategy self-optimization device based on multi-scenario joint analysis comprises at least one processor and a memory in communication connection with the at least one processor. The memory stores instructions executable by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to execute the digital operation strategy self-optimization method based on multi-scenario joint analysis in the above embodiment one.

[0085] Reference will be made to the following description of the embodiments of the application with reference to the drawings. Figure 4 The drawings show a structure diagram of the digital operation strategy self-optimization device based on multi-scenario joint analysis suitable for implementing the embodiments of the application. The digital operation strategy self-optimization device based on multi-scenario joint analysis in the embodiments of the application can include but is not limited to mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (Personal Digital Assistant), PADs (Portable Application Description), PMPs (Portable Media Player), vehicle terminals (such as vehicle navigation terminals), and the like, and fixed terminals such as digital TVs, desktop computers, and the like. Figure 4 The digital operation strategy self-optimization device based on multi-scenario joint analysis shown is only an example and should not impose any limitation on the functions and use range of the embodiments of the application.

[0086] As Figure 4As shown, the multi-scenario joint analysis based digital operation strategy self-optimization device can include a processing apparatus 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to programs stored in a ROM (Read Only Memory) 1002 or programs loaded from a storage apparatus 1003 into a RAM (Random Access Memory) 1004. In the RAM 1004, various programs and data required for operation of the multi-scenario joint analysis based digital operation strategy self-optimization device are also stored. The processing apparatus 1001, the ROM 1002, and the RAM 1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: input apparatuses 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output apparatuses 1008 including, for example, an LCD (Liquid Crystal Display), a speaker, a vibrator, etc.; the storage apparatus 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication apparatus 1009. The communication apparatus 1009 can allow the multi-scenario joint analysis based digital operation strategy self-optimization device to communicate wirelessly or by wire with other devices to exchange data. Although the multi-scenario joint analysis based digital operation strategy self-optimization device having various systems is shown in the figure, it should be understood that all the shown systems are not required to be implemented or possessed. More or less systems can be alternatively implemented or possessed.

[0087] In particular, according to embodiments of the present disclosure, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, embodiments of the present disclosure include a computer program product comprising a computer program carrying out the program codes for performing the methods shown in the flowcharts carried on a computer readable medium. In such embodiments, the computer program can be downloaded and installed from a network through the communication apparatus, or installed from the storage apparatus 1003, or installed from the ROM 1002. When the computer program is executed by the processing apparatus 1001, the above-mentioned functions defined in the methods of the embodiments of the present disclosure are performed.

[0088] The device for self-optimization of digital operation strategy based on multi-scenario joint analysis provided in the application adopts the method for self-optimization of digital operation strategy based on multi-scenario joint analysis in the above embodiment, and can solve the technical problem of self-optimization of digital operation strategy based on multi-scenario joint analysis. Compared with the prior art, the device for self-optimization of digital operation strategy based on multi-scenario joint analysis provided in the application has the same beneficial effects as the method for self-optimization of digital operation strategy based on multi-scenario joint analysis provided in the above embodiment, and other technical features in the device for self-optimization of digital operation strategy based on multi-scenario joint analysis are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0089] It should be understood that various parts of the present application can be realized by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0090] The above is merely specific implementation of the present application, but the protection scope of the present application is not limited to this, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0091] The present application provides a computer readable storage medium having computer readable program instructions (i.e. computer programs) stored thereon, the computer readable program instructions being used to execute the method for self-optimization of digital operation strategy based on multi-scenario joint analysis in the above embodiment.

[0092] The computer readable storage medium provided in the application may be, for example, a U disk, but is not limited to an electric, magnetic, optical, electromagnetic, infrared, or semiconductor system, system or device, or any combination of the above. More specific examples of the computer readable storage medium may include, but are not limited to, an electric connection with one or more conductive wires, a portable computer disk, a hard disk, a RAM (Random Access Memory), a ROM (Read Only Memory), an EPROM (Erasable Programmable Read Only Memory or flash memory), an optical fiber, a CD-ROM (CD-Read Only Memory), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the embodiment, the computer readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer readable storage medium can be transmitted by any suitable medium, including but not limited to: electric wire, optical cable, RF (Radio Frequency), etc., or any suitable combination of the above.

[0093] The computer readable storage medium described above can be contained in the digital operation strategy self-optimization device based on multi-scene joint analysis; or can exist independently without being assembled into the digital operation strategy self-optimization device based on multi-scene joint analysis.

[0094] The computer readable storage medium described above carries one or more programs, when the one or more programs are executed by the digital operation strategy self-optimization device based on multi-scene joint analysis, the digital operation strategy self-optimization device based on multi-scene joint analysis: receives desensitization data processed by edge computing nodes deployed in each business scene, generates a user feature vector according to the desensitization data; The user feature vector is aggregated to determine an aggregated node feature, and a group behavior graph is generated based on the aggregated node feature; A cross-business interaction sequence is determined based on the group behavior graph, the cross-business interaction sequence is converted into a latent space vector, and a dynamic feature label is generated according to the latent space vector; The dynamic feature label is matched with rights and interests based on a rights and interests matching model, and a rights and interests matching result is output.

[0095] Computer program code for carrying out operations of the present application can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0096] The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0097] The modules involved in the embodiments of the present application can be implemented in software or hardware. In some cases, the names of the modules do not constitute a limitation on the modules themselves.

[0098] The readable storage medium provided by the present application is a computer readable storage medium, which stores computer readable program instructions (i.e., computer programs) for executing the above-mentioned digital operation strategy self-optimization method based on multi-scene joint analysis, and can solve the technical problem of digital operation strategy self-optimization based on multi-scene joint analysis. Compared with the prior art, the computer readable storage medium provided by the present application has the same beneficial effects as the digital operation strategy self-optimization method based on multi-scene joint analysis provided by the above-mentioned embodiments, and will not be described here.

[0099] The application also provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the method for self-optimization of digital operation strategy based on multi-scenario joint analysis as described above.

[0100] The computer program product provided by the application can solve the technical problem of self-optimization of digital operation strategy based on multi-scenario joint analysis. Compared with the prior art, the beneficial effects of the computer program product provided by the application are the same as those of the method for self-optimization of digital operation strategy based on multi-scenario joint analysis provided by the above-mentioned embodiments, and are not described here.

[0101] The above is only some embodiments of the application, and does not limit the patent scope of the application. Any equivalent structural transformation made by using the content of the application specification and drawings, or direct / indirect application in other related technical fields is included in the patent protection scope of the application.

Claims

1. A self-optimization method for digital operation strategies based on multi-scenario joint analysis, characterized in that, The self-optimization method for digital operation strategies based on multi-scenario joint analysis includes: Receive de-identified data after local processing by edge computing nodes deployed in various business scenarios, and generate user feature vectors based on the de-identified data; The user feature vectors are aggregated to determine the features of the aggregated nodes, and a group behavior map is generated based on the features of the aggregated nodes. Based on the group behavior map, a cross-business interaction sequence is determined, the cross-business interaction sequence is transformed into a latent space vector, and dynamic feature labels are generated based on the latent space vector. The dynamic feature labels are matched for interest based on the interest matching model, and the interest matching results are output.

2. The method as described in claim 1, characterized in that, The step of aggregating the user feature vectors, determining the aggregated node features, and generating a group behavior graph based on the aggregated node features includes: The target scene in the user feature vector is determined, and the target scene is used as the target aggregation point. When the target aggregation points are the same, the target aggregation points are merged, and the target aggregation point is determined as the aggregation center. The user feature vectors are aggregated based on the aggregation center to obtain the business scenario switching relationship; Determine the number of aggregations of the business scenario switching relationship pointing to the aggregation center. When the proportion of the number of aggregations to the total number is greater than a preset proportion, determine the aggregation center as the target aggregation center and use the business scenario switching relationship pointing to the target aggregation center as the aggregation node feature. A group behavior map is generated using the target aggregation center and the features of the aggregation nodes.

3. The method as described in claim 2, characterized in that, After using the business scenario switching relationship pointing to the target aggregation center as the aggregation node feature step, the method further includes: The remaining business scenario switching relationships, excluding the aggregated node features, are defined as detached features; Upon receiving a new user feature vector, a new target aggregation center is determined based on the free features, the new user feature vector, and the latest total number. When the new target aggregation center exists, the corresponding new aggregation node features are determined based on the new target aggregation center; The group behavior map is updated based on the new target aggregation center and the features of the new aggregation nodes.

4. The method as described in claim 1, characterized in that, The steps of determining cross-business interaction sequences based on the group behavior map, converting the cross-business interaction sequences into latent space vectors, and generating dynamic feature labels based on the latent space vectors include: Based on the group behavior graph, identify the continuous interaction behavior of users in multiple business scenarios, and construct a cross-business interaction sequence based on the continuous interaction behavior; Based on the business mapping model, the cross-business interaction sequence is mapped to a low-dimensional latent space to generate a latent space vector representing the user behavior pattern. Cluster analysis and pattern recognition are performed on the latent space vectors to generate dynamic feature labels that reflect the real-time behavioral characteristics of users.

5. The method as described in claim 4, characterized in that, The step of identifying continuous user interaction behaviors across multiple business scenarios based on the group behavior graph, and constructing cross-business interaction sequences based on the continuous interaction behaviors, includes: Extract the temporal behavior patterns from the group behavior graph, and identify the continuous access trajectory of users in different business scenarios within a preset time window based on the temporal behavior patterns. The scene switching behaviors in the continuous access trajectory are sorted according to time order to construct a cross-business interaction sequence with scene identifiers as elements.

6. The method as described in claim 4, characterized in that, The step of mapping the cross-business interaction sequence to a low-dimensional latent space based on the business mapping model to generate latent space vectors representing user behavior patterns includes: The cross-business interaction sequence is encoded based on the business mapping model to obtain a one-hot encoding based on the business type. The one-hot encoding is mapped to a low-dimensional latent space to generate a numerical sequence. The numerical sequence is computed in parallel based on an attention mechanism to determine the correlation between any two elements in the numerical sequence, and a short vector is generated based on the correlation. The short vector is forward-propagated to generate a latent space vector representing the user's behavior pattern.

7. The method as described in claim 4, characterized in that, The step of performing clustering analysis and pattern recognition on the latent space vectors to generate dynamic feature labels reflecting the user's real-time behavior characteristics includes: Unsupervised clustering is performed on the latent space vectors to identify user groups with target cross-scenario behavior patterns; Backtrack the cross-business interaction sequences and the de-identified data corresponding to each user group, perform pattern recognition on the cross-business interaction sequences and the de-identified data, and extract group behavior features; Based on the group behavior characteristics, a dynamic feature label combination is generated for each user group; The latent space vector is compared with the center of each user group to obtain a matching similarity set; Determine the target user group center corresponding to the highest matching similarity among the matching similarities; The dynamic feature tags are generated by combining the dynamic feature tags corresponding to the target user group center.

8. The method as described in claim 1, characterized in that, The step of performing interest matching on the dynamic feature labels based on the interest matching model and outputting the interest matching result includes: Personalized rights matching strategies are generated in real time based on the user's group characteristics and dynamic feature tags. The implemented personalized benefits matching strategy is evaluated, with user conversion rate and return on investment serving as reward signals; Based on the reward signal, the inner loop weight parameters of the equity matching model are optimized to obtain the optimized inner loop weight parameters. The inner loop strategy model is updated based on the optimized inner loop weight parameters, and the equity matching result is output based on the optimized inner loop strategy model.

9. The method as described in claim 1, characterized in that, The steps for generating user feature vectors based on the de-identified data include: The anonymized data is analyzed to determine the business scenario switching sequence within the anonymized data; The frequency of occurrence of the business scenario switching sequence is determined, and the frequency of occurrence is used as a feature weight. A user feature vector is generated based on the feature weight and the business scenario switching sequence.

10. A self-optimizing system for digital operation strategies based on multi-scenario joint analysis, characterized in that, The digital operation strategy self-optimization system based on multi-scenario joint analysis includes: The data processing module is used to receive de-identified data after local processing by edge computing nodes deployed in various business scenarios, and generate user feature vectors based on the de-identified data; The feature processing module is used to aggregate the user feature vectors, determine the aggregate node features, and generate a group behavior map based on the aggregate node features. The business processing module is used to determine cross-business interaction sequences based on the group behavior graph, convert the cross-business interaction sequences into latent space vectors, and generate dynamic feature labels based on the latent space vectors. The rights matching module is used to perform rights matching on the dynamic feature labels based on the rights matching model and output the rights matching results.

Images