An ai intelligent marketing service management system based on user behavior graph

Abstract

The application discloses an AI intelligent marketing service management system based on a user behavior graph, and particularly relates to the technical field of marketing service management; the user behavior graph is generated by acquiring cross-channel user behavior records, user identification records and marketing event records, and generating user association edges; user clusters are generated according to the connection relationship, portrait features, marketing groupings, touch frequency controls, marketing task states and attribution states are calculated, and user association edge dependency chains are recorded; when the effective time of the user association edge expires, the confidence version decreases or the source state changes, the target bridge edge is identified, the connected components of the user behavior graph after the target bridge edge is inactivated are divided, the split user subgraph and the version mapping relationship are generated; the contaminated derived marketing state is filtered and written into a withdrawal queue, the contaminated derived marketing state is cancelled according to the withdrawal queue, the derived marketing state is recalculated in the split user subgraph, and a marketing service control instruction is generated.

Description

Technical Field

[0001] This invention relates to the field of marketing service management technology, and more specifically, to an AI-powered intelligent marketing service management system based on user behavior graphs. Background Technology

[0002] Existing methods, when forming user behavior graphs based on the relationships between user identifiers, typically focus more on whether user identifiers can be connected, and do not adequately consider the changes in the effectiveness of user identifier relationships over time, or the impact of these changes on downstream marketing service status. When some user identifiers are shared, reallocated, periodically changed, or their relationships become invalid, the user clusters generated based on the user behavior graph may not match the actual marketing service targets. Consequently, the profile features, marketing segmentation, reach frequency control, marketing task status, and attribution status calculated from user clusters continue to use graph association results that no longer match, resulting in inconsistencies between the user association status in the user behavior graph and the status of AI-powered intelligent marketing service management. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an AI-powered intelligent marketing service management system based on user behavior graphs to address the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] An AI-powered intelligent marketing service management system based on user behavior graphs includes:

[0006] The graph construction module acquires cross-channel user behavior records, user identifier records, and marketing event records, generates user association edges according to user identifier relationships, and forms a user behavior graph from the user association edges;

[0007] State Derivation Module: Generates user clusters based on the connectivity of the user behavior graph, calculates derived marketing states for user clusters, and records the user-related edge dependency chains corresponding to the derived marketing states.

[0008] Bridge identification module: When the validity period of a user-related edge expires, the confidence version decreases, or the source status changes, it determines whether the user-related edge is a time-distorted bridge edge connecting different user subgraphs and obtains the target bridge edge.

[0009] Subgraph splitting module: Deactivates the target bridging edge from the user behavior graph, performs connected component partitioning to generate the split user subgraph, and marks the version mapping relationship between the original user cluster version and the split user subgraph version;

[0010] Pollution Filtering Module: Based on the user-related edge dependency chain and version mapping relationship, filter the polluted derived marketing states generated by the dependent target bridging edge or the user subgraph behavior records after the split, and write them to the rollback queue;

[0011] Rollback and Recalculation Module: Undoes the contaminated derived marketing state according to the rollback queue, recalculates the derived marketing state in the split user subgraph, and generates marketing service control instructions.

[0012] In a preferred embodiment, cross-channel user behavior records, user identifier records, and marketing event records are acquired; user association edges are generated according to user identifier association relationships; and user behavior graphs are formed from user association edges, specifically as follows:

[0013] Time normalization and source marking are performed on cross-channel user behavior records, user identification records, and marketing event records. User identification associations are determined based on account correspondence, device identification correspondence, communication identification correspondence, advertising identification correspondence, network co-location relationship, and behavior association.

[0014] Configure the source, validity period, and confidence version for user identification associations to generate user association edges;

[0015] A user behavior graph is formed by connecting the corresponding user identifier records with the user-related edges.

[0016] In a preferred embodiment, user clusters are generated based on the connectivity of the user behavior graph, derived marketing states are calculated for each user cluster, and the user-related edge dependency chains corresponding to the derived marketing states are recorded, specifically as follows:

[0017] User clusters are divided based on user identifier records connected by user association edges in the user behavior graph. Cross-channel user behavior records and marketing event records within the user clusters are aggregated, and derived marketing states, including profile features, marketing segmentation, reach frequency control, marketing task status, and attribution status, are calculated.

[0018] When each derived marketing state is generated, the user identifiers, cross-channel user behavior records, marketing event records, and user association edges involved in the calculation are recorded to form a user association edge dependency chain corresponding to each derived marketing state.

[0019] In a preferred embodiment, when the validity period of a user-associated edge expires, the confidence level decreases, or the source state changes, it is determined whether the user-associated edge is a time-distortion bridging edge connecting different user subgraphs, and the target bridging edge is obtained, specifically as follows:

[0020] When the validity period of a user-related edge expires, the confidence level decreases, or the source status changes, the user-related edge that meets any of these conditions will be used as the user-related edge to be judged.

[0021] Starting from the user identifier records at both ends of the user association edge to be determined, search for the set of reachable user identifier records along the user association edges other than the user association edge to be determined.

[0022] When there is no intersection between the sets of reachable user identifier records, the user association edge to be judged is determined to be a time-distortion bridging edge connecting different user subgraphs, and the time-distortion bridging edge is taken as the target bridging edge.

[0023] In a preferred embodiment, the target bridging edge is deactivated from the user behavior graph, connected component partitioning is performed to generate a split user subgraph, and the version mapping relationship between the original user cluster version and the split user subgraph version is marked, specifically as follows:

[0024] The connection status of the target bridging edge is marked as inactive, and the source, validity period and confidence version of the target bridging edge are retained in the user behavior graph;

[0025] Based on the user identifier records that continue to connect to each other through user-associated edges after the target bridging edge is deactivated, perform connected component partitioning to generate the split user subgraph;

[0026] Record the original user cluster version and the corresponding user subgraph version after splitting to form a version mapping relationship.

[0027] In a preferred embodiment, based on the user-related edge dependency chain and version mapping relationship, the polluted derived marketing state generated from the dependent target bridging edge or the user subgraph behavior record after the split is filtered and written into the rollback queue, specifically:

[0028] Based on the user association edge dependency chain, read the user identifier record, cross-channel user behavior record, marketing event record and user association edge recorded when each derived marketing state is generated;

[0029] Combine the version mapping relationship to determine whether the user association edge recorded when each derived marketing state is generated contains the target bridging edge, or to determine whether the user identification record, cross-channel user behavior record and marketing event record recorded when each derived marketing state is generated belong to different split user subgraphs;

[0030] Mark derived marketing states that contain target bridging edges or belong to different split user subgraphs as contaminated derived marketing states and write contaminated derived marketing states to the rollback queue.

[0031] In a preferred embodiment, the contaminated derived marketing state is revoked according to the rollback queue, the derived marketing state is recalculated within the split user subgraph, and a marketing service control instruction is generated, specifically as follows:

[0032] Read the contaminated derivative marketing status according to the withdrawal queue, and mark the profile features, marketing groups, reach frequency control, marketing task status and attribution status corresponding to the contaminated derivative marketing status as withdrawal status;

[0033] Based on the version mapping relationship, the derived marketing states that have not been written to the rollback queue are retained in the corresponding split user subgraph;

[0034] Within the split user subgraph, recalculate profile features, marketing segmentation, reach frequency control, marketing task status, and attribution status based on user identification records, cross-channel user behavior records, marketing event records, and user association edges.

[0035] The recalculated profile features, marketing segments, reach frequency control, marketing task status, and attribution status are used as the recalculated derived marketing status to generate marketing service control instructions.

[0036] The technical effects and advantages of the AI-powered intelligent marketing service management system based on user behavior graphs proposed in this invention are as follows:

[0037] By configuring the source, validity period, and confidence version of user-related edges in the user behavior graph, the connection relationships between cross-channel user identifiers have traceable time attributes and a reliable state. By synchronously recording the dependency chain of user-related edges when user clusters generate derived marketing states, profile features, marketing segmentation, reach frequency control, marketing task status, and attribution status can be mapped to the user identifier records, cross-channel user behavior records, marketing event records, and user-related edges involved in the calculation. By identifying target bridging edges when user-related edges fail or their reliability decreases, and generating split user subgraphs and version mapping relationships after the target bridging edges become inactive, the correspondence between the original user clusters and the split user subgraphs remains clear. By filtering and reverting contaminated derived marketing states, and then recalculating derived marketing states within the split user subgraphs, it is possible to prevent failed user relationships from continuing to affect marketing service control instructions, ensuring that the connection state of the user behavior graph is consistent with the state of AI-powered intelligent marketing service management. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the structure of an AI-powered intelligent marketing service management system based on user behavior graphs according to the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0040] Example 1

[0041] Figure 1 The present invention discloses an AI-powered intelligent marketing service management system based on user behavior graphs, comprising:

[0042] The graph construction module acquires cross-channel user behavior records, user identifier records, and marketing event records, generates user association edges according to user identifier relationships, and forms a user behavior graph from the user association edges;

[0043] State Derivation Module: Generates user clusters based on the connectivity of the user behavior graph, calculates derived marketing states for user clusters, and records the user-related edge dependency chains corresponding to the derived marketing states.

[0044] Bridge identification module: When the validity period of a user-related edge expires, the confidence version decreases, or the source status changes, it determines whether the user-related edge is a time-distorted bridge edge connecting different user subgraphs and obtains the target bridge edge.

[0045] Subgraph splitting module: Deactivates the target bridging edge from the user behavior graph, performs connected component partitioning to generate the split user subgraph, and marks the version mapping relationship between the original user cluster version and the split user subgraph version;

[0046] Pollution Filtering Module: Based on the user-related edge dependency chain and version mapping relationship, filter the polluted derived marketing states generated by the dependent target bridging edge or the user subgraph behavior records after the split, and write them to the rollback queue;

[0047] Rollback and Recalculation Module: Undoes the contaminated derived marketing state according to the rollback queue, recalculates the derived marketing state in the split user subgraph, and generates marketing service control instructions.

[0048] Acquire cross-channel user behavior records, user identifier records, and marketing event records; generate user association edges according to user identifier relationships; and form a user behavior graph from these user association edges, including:

[0049] In one implementation, cross-channel user behavior records include browsing, clicking, favorites, adding to cart, placing orders, payments, unsubscribing, complaints, and inquiries from applications, web pages, SMS outreach pages, email outreach pages, offline store POS records, customer service conversation records, and advertising landing pages; user identification records include account identifiers, device identifiers, communication identifiers, advertising identifiers, network addresses, store membership identifiers, and temporary channel identifiers; marketing event records include marketing task identifiers, marketing channels, marketing content types, outreach time, outreach target identifiers, and outreach results; when time-normalizing cross-channel user behavior records, user identification records, and marketing event records, the original occurrence time, channel source time format, and channel source time zone carried in each record are read, and then the original occurrence time is converted into a standard timestamp under a unified time zone. The system generates a normalized time field according to a preset time granularity. The preset time granularity is determined based on the reach frequency of marketing event records and the minimum collection interval of cross-channel user behavior records. For example, when cross-channel user behavior records are collected at the second level and marketing event records are reached at the minute level, the preset time granularity can be set to the minute level. When marking the source of cross-channel user behavior records, user identification records, and marketing event records, the system writes the channel type, channel name, collection entry, record generation method, and source trust level for each record. The source trust level is determined based on whether the record comes from a login account, whether it comes from a verified communication identifier, whether it comes from the same device identifier, whether it comes from a network co-location relationship, and whether it comes from a behavior association relationship. Different source trust levels are used to configure the source and confidence version of the user identifier association relationship in the future.

[0050] After completing time normalization and source tagging, user identifier associations are determined based on account correspondence, device identifier correspondence, communication identifier correspondence, advertising identifier correspondence, network co-addressing, and behavioral association. Account correspondence is determined by the same account identifier or account binding records; device identifier correspondence is determined by the same device identifier corresponding to different user identifier records within a valid time range; communication identifier correspondence is determined by the same mobile phone number, email address, or encrypted communication identifier corresponding to different user identifier records; advertising identifier correspondence is determined by the same advertising identifier corresponding to different user identifier records within a valid time range; network co-addressing is determined by the same network address, similar normalized time fields, and the same data collection entry point corresponding to different user identifier records; and behavioral association is determined by cross-channel user behavior records based on the normalized time field, access object, access order, and marketing data. The matching relationship between the response order of marketing event records is determined; when determining the behavioral association relationship, cross-channel user behavior records belonging to different user identifier records are sorted according to the normalized time field, and then the continuity of the access objects, the continuity of the access order, and whether the corresponding response appears within the limited time window after the marketing event record is reached are compared. The limited time window is set according to the common response cycle of marketing channels and the collection interval of cross-channel user behavior records. For example, the SMS page can be set to within 24 hours after the reach time, and the advertising landing page can be set to within 2 hours after the reach time; when any account correspondence, device identifier correspondence, communication identifier correspondence, advertising identifier correspondence, network co-address relationship, and behavioral association relationship are satisfied, a user identifier association relationship is generated between the corresponding user identifier records, and the basis for formation and participation record are written into the user identifier association relationship.

[0051] After generating user identifier associations, the source, validity period, and confidence level are configured for each association to generate user association edges. The source is taken from the basis for forming the user identifier association. When the basis is account correspondence, the source record is the account correspondence; when the basis is device identifier correspondence, the source record is the device identifier correspondence; when the basis is communication identifier correspondence, the source record is the communication identifier correspondence; when the basis is advertising identifier correspondence, the source record is the advertising identifier correspondence; when the basis is network co-addressing, the source record is the network co-addressing; and when the basis is behavioral association, the source record is the behavioral association. The validity period is determined based on the normalized time field of the participating records in the user identifier association and the basis for its formation. The validity period for account correspondence can extend from the account binding start time to the account unbinding time; the validity period for device identifier correspondence can extend from the normalized time field of the first occurrence of the same device identifier to the device identifier reset time; and the validity period for communication identifier correspondence can extend from the communication identifier verification time. The validity period of the relationship between advertising identifiers extends from the time the advertising identifier first appears to the time the advertising identifier is reset. The validity period of the relationship between network co-locations can be determined according to the continuous co-occurrence time period of network addresses. The validity period of the relationship between behaviors can be determined according to the continuous behavior time period formed by the access order and response order. The confidence version is generated based on the source trust level, validity period update, and changes in participation records. An initial confidence version is configured when the user identifier relationship is first generated. An increased confidence version is generated when a new participation record supports the user identifier relationship. A decreased confidence version is generated when the validity period expires, the number of participation records decreases, or the source status changes. Each change in confidence version is recorded with a corresponding change time and reason. The user association edge is generated by the user identifier relationship, source, validity period, and confidence version. The two endpoints of the user association edge are the user identifier records participating in the user identifier relationship. The edge attributes of the user association edge include source, validity period, confidence version, formation basis, and participation record.

[0052] After generating user association edges, user behavior graphs are formed by connecting corresponding user identifier records according to these edges. When forming the user behavior graph, each user identifier record is treated as a graph node, and each user association edge is used as an edge connecting two user identifier records. Cross-channel user behavior records and marketing event records are then linked to their corresponding user identifier records according to the user identifier values ​​carried within them. When the same cross-channel user behavior record carries multiple user identifier values, the cross-channel user behavior record establishes record attribution relationships with each of the multiple corresponding user identifier records, retaining the normalized time field and source marker in the record attribution relationships. Similarly, when the same marketing event record carries multiple reach object identifiers, the marketing event record establishes event attribution relationships with each of the multiple corresponding user identifier records, retaining the reach time, marketing channel, and marketing content type in the event attribution relationships. The user behavior graph includes user identifier records, user association edges, record attribution relationships between cross-channel user behavior records and user identifier records, and event attribution relationships between marketing event records and user identifier records. The user association edges in the user behavior graph retain the source, validity time, and confidence version, used to segment user clusters based on the user identifier records connected by the user association edges.

[0053] User clusters are generated based on the connectivity relationships in the user behavior graph. Derived marketing states are calculated for each user cluster, and the user-related edge dependency chains corresponding to these derived marketing states are recorded, including:

[0054] When segmenting user clusters, the two endpoints of each user-related edge in the user behavior graph are read, and the user identifier records corresponding to the two endpoints are used as connectable objects. Then, user-related edges participating in the segmentation are filtered according to their effective time. The filtering method is as follows: the reach time of the marketing event record that needs to generate derived marketing status is compared with the effective time of the user-related edge. When the reach time of the marketing event record is within the effective time of the user-related edge, the user-related edge participates in the user cluster segmentation. When the reach time of the marketing event record is not within the effective time of the user-related edge, the user-related edge does not participate in the user cluster segmentation. When the marketing event record does not have a reach time, the planned reach time corresponding to the marketing event record is compared with the effective time of the user-related edge.

[0055] When dividing user clusters based on the user identifier records connected by the filtered user association edges, starting from any user identifier record not yet assigned to a user cluster, the search continues along the filtered user association edges to find other reachable user identifier records. The found user identifier records are then assigned to the same user cluster as the user identifier record corresponding to the search starting point. When no new reachable user identifier records exist within the same user cluster, the current user cluster division ends, and the search continues from the remaining user identifier records not yet assigned to a user cluster, until all user identifier records in the user behavior graph are assigned to their corresponding user clusters. Each user cluster records the user cluster identifier, user cluster generation time, user identifier records participating in the user cluster division, user association edges participating in the user cluster division, and the source, validity time, and confidence version of the corresponding user association edges. The user cluster identifier can be generated from the string obtained by sorting the user cluster generation time and the user identifier records within the user cluster.

[0056] After generating user clusters, cross-channel user behavior records and marketing event records within each cluster are aggregated. When aggregating cross-channel user behavior records, the record attribution relationships in the user behavior graph are read. Cross-channel user behavior records whose user identifiers are connected by these relationships and belong to the same user cluster are grouped into the same user cluster. The aggregation result retains the normalized time field, source marker, channel type, access target, access order, and behavior type of the cross-channel user behavior records. When aggregating marketing event records, the event attribution relationships in the user behavior graph are read. Marketing event records whose user identifiers are connected by these relationships and belong to the same user cluster are grouped into the same user cluster. The aggregation result retains the marketing task identifier, marketing channel, marketing content type, reach time, reach target identifier, and reach result of the marketing event records. When a cross-channel user behavior record belongs to multiple user identifier records simultaneously, the user cluster to which the cross-channel user behavior record belongs is determined by all user identifier records stored in the record attribution relationships. When a marketing event record belongs to multiple user identifier records simultaneously, the user cluster to which the marketing event record belongs is determined by all user identifier records stored in the event attribution relationships.

[0057] When calculating user profile features, cross-channel user behavior records and marketing event records collected within the user cluster are used as input. Statistical items are formed according to behavior type, channel type, target audience, and normalized time field. These statistical items include the number of views, clicks, favorites, add-to-cart, orders, payments, unsubscribes, complaints, inquiries, the time of the most recent behavior, the time of the most recent contact, and the number of responses after contact. The time of the most recent behavior is obtained by comparing the normalized time field of cross-channel user behavior records within the user cluster. The time of the most recent contact is obtained by comparing the contact time of marketing event records within the user cluster. The number of responses after contact is obtained by determining whether there are cross-channel user behavior records corresponding to the marketing content type or target audience after the contact time of the marketing event record. The response time window is used to determine the response after contact. The response time window is set according to the marketing channel, marketing content type, and the collection interval of cross-channel user behavior records. For example, the response time window for SMS contact pages can be set to 24 hours, and the response time window for advertising landing pages can be set to 2 hours.

[0058] When calculating marketing segments, user profile features and marketing event records within user clusters are used as input. User clusters are matched against pre-configured marketing segmentation conditions. These conditions consist of a combination of behavior types, a time interval between behavior occurrences, marketing channels, marketing content types, number of responses after outreach, and outreach results. The threshold for marketing segmentation conditions is determined based on the distribution of user profile features within a continuous historical marketing period. The length of this continuous historical marketing period is set based on the marketing campaign settlement cycle and the cross-channel user behavior record accumulation cycle; for example, the most recent 30 days or the most recent 90 days can be used as the continuous historical marketing period. When a user profile feature meets a certain marketing segmentation condition, a corresponding marketing segment is generated. When a user profile feature meets multiple marketing segmentation conditions simultaneously, marketing segments are generated according to the pre-set segmentation priority in the marketing segmentation conditions. The segmentation priority is determined based on the marketing task execution order, marketing content type, and outreach channel configuration.

[0059] When calculating reach frequency control, marketing event records within a user cluster are used as input. The system calculates the number of times a user cluster is reached within a specified time window, the most recent reach time, and the time interval between adjacent reach visits, based on the marketing channel, marketing content type, and reach time. The specified time window is set according to the reach frequency limit of the marketing channel and the duration of the marketing content type. For example, a calendar day can be used for SMS channels, and a calendar week can be used for email channels. Reach frequency control includes the number of channel reach visits, the number of content reach visits, the most recent reach time, and the next allowed reach time. The next allowed reach time is obtained by adding the interval duration configured for the corresponding marketing channel or marketing content type to the most recent reach time. During the generation of reach frequency control, the system records the marketing event records involved in the statistics, the user identifier records to which the marketing event records belong, and the user association edges connecting the user identifier records to form the user cluster.

[0060] When calculating the marketing task status, marketing event records within user clusters, reach frequency control, and marketing segmentation are used as inputs. Marketing event records are aggregated according to marketing task identifiers, and the marketing task status is determined based on the reach results, reach time, reach target identifier, and reach frequency control of the marketing event records. Marketing task status includes pending reach, reached, responded to, suppressed, and completed states. The pending reach state corresponds to marketing event records with a marketing task identifier but no reach time; the reached state corresponds to marketing event records with a reach time but no response after reach; the responded state corresponds to marketing event records with corresponding cross-channel user behavior records within the response time window after the reach time; the suppressed state corresponds to marketing event records where the reach frequency control shows the next allowed reach time is later than the planned reach time of the marketing task; and the completed state corresponds to marketing event records where all reach results corresponding to the marketing task identifier have been recorded as complete records.

[0061] When calculating attribution status, marketing event records within the user cluster, cross-channel user behavior records, and marketing task status are used as inputs. The reach time of the marketing event record is used as the attribution starting point. Cross-channel user behavior records corresponding to orders, payments, cancellations, complaints, and inquiries that occur after the reach time are used as attributable behavior records. It is determined whether the normalized time field of the attributable behavior record is within the attribution time window. The attribution time window is set according to the average conversion cycle of the marketing channel, marketing content type, and target audience. The average conversion cycle can be determined by the distribution of time difference between the reach time and the order time or payment time within a continuous historical marketing cycle. For example, the length of time that covers most historical conversion records in the time difference distribution can be set as the attribution time window. Attribution status includes marketing task identifier, reach time, attributable behavior record, attribution time window, and attribution result. The attribution result is generated based on the matching relationship between the attributable behavior record and the marketing event record in terms of target audience identifier, target audience, and marketing content type.

[0062] When any derived marketing state is generated from profile features, marketing segmentation, reach frequency control, marketing task status, and attribution status, the user association edge dependency chain corresponding to each derived marketing state is simultaneously recorded. The user association edge dependency chain includes the derived marketing state identifier, derived marketing state type, user cluster identifier, generation time, user identifier records involved in the calculation, cross-channel user behavior records involved in the calculation, marketing event records involved in the calculation, and user association edges involved in the calculation. The user association edges involved in the calculation are determined based on the connection paths between the user identifier records involved in the calculation within the user cluster. The connection paths consist of user association edges that have been filtered and divided into user clusters. When a cross-channel user behavior record or a marketing event record involved in the calculation belongs to multiple user identifier records, the user association edge dependency chain records multiple user identifier records and all user association edges connecting multiple user identifier records.

[0063] When the validity period of a user-related edge expires, the confidence level decreases, or the source state changes, it is determined whether the user-related edge is a time-distortion bridging edge connecting different user subgraphs, and the target bridging edge is obtained, including:

[0064] This process involves reading user association edges, user identifier records connected by these edges, the source of the user association edges, their validity period, confidence level, formation basis, and participation records from the user behavior graph. The validity period indicates the time range within which the user association edge participates in the connection judgment of user identifier records in the user behavior graph. The confidence level indicates the credibility status of the user association edge at different change times. The source status indicates whether the formation basis and participation records still support the user identifier association relationship. A state judgment time is determined for each user association edge, using the reach time of the marketing event record that needs to generate a derived marketing state. When the marketing event record does not have a reach time, the planned reach time corresponding to the marketing event record is used. When the derived marketing state does not correspond to a specific marketing event record, the generation time of the derived marketing state is used.

[0065] When determining whether the validity period of a user-associated edge has expired, the edge state judgment time is compared with the validity period of the user-associated edge. The validity period of a user-associated edge includes the valid start time and the valid end time. If the edge state judgment time is earlier than the valid start time or later than the valid end time, the user-associated edge is determined to have expired. When determining whether the confidence version of a user-associated edge has decreased, the most recently recorded confidence version of the user-associated edge before the edge state judgment time and the confidence version of the user-associated edge at the edge state judgment time are read and compared in order of confidence version level. The confidence version level order is generated based on the source trust level, validity period update, and changes in participation records. For example, account correspondence and verified communication identifier correspondence can be configured as higher level, device identifier correspondence and advertising identifier correspondence can be configured as intermediate level, and network co-address relationship and behavior association relationship can be configured as lower level. When the confidence version of the user-associated edge at the edge state judgment time is lower than the most recently recorded confidence version of the user-associated edge before the edge state judgment time, the user-associated edge is determined to have decreased confidence version.

[0066] When determining whether the source status of a user-related edge has changed, the source, formation basis, and participation records of the user-related edge are read, and the source, formation basis, and participation records of the user-related edge are compared before and after the edge status judgment time to see if they have changed. Changes in source status include account correspondence changing from account binding to account unbinding, device identifier correspondence showing device identifier reset records, communication identifier correspondence showing communication identifier unbinding records or communication identifier re-verification records, advertising identifier correspondence showing advertising identifier reset records, network co-addressing relationships no longer meeting the continuous co-occurrence conditions of the same network address and similar normalized time fields, and behavioral association relationships no longer meeting the conditions of continuous access objects, sequential access order, and marketing event record response order matching. When a user-related edge experiences any of the following situations: expiration of the effective time, decrease in confidence version, or change in source status, the user-related edge is treated as a user-related edge to be judged, and the judgment triggering reason, edge status judgment time, and participation records corresponding to the triggering reason are recorded.

[0067] After obtaining the user association edge to be judged, read the user identifier records at both ends of the user association edge to be judged, and use the user identifier records at both ends of the user association edge to be judged as the search starting point respectively; before performing the search, user association edges in the user behavior graph other than the user association edge to be judged are selected as candidate search edges, and user association edges whose edge state judgment time is not within the valid time are excluded from the candidate search edges; the candidate search edges are used to determine whether the user identifier records at both ends of the user association edge to be judged can still maintain the connection through other user association edges without depending on the user association edge to be judged; the search method can adopt breadth-first search or depth-first search. Breadth-first search traverses the user identifier records connected by the candidate search edges in order of distance from the search starting point from near to far. Depth-first search continuously traverses the reachable user identifier records along the connection path formed by the candidate search edges; one search method is used in the same search, and the searched user identifier records are written into the reachable user identifier record set to avoid the same user identifier record being searched repeatedly.

[0068] A first set of reachable user identifier records is obtained by using the user identifier record at one end of the edge connecting the user to be determined as the search starting point, and a second set of reachable user identifier records is obtained by using the user identifier record at the other end of the edge connecting the user to be determined as the search starting point. The first set of reachable user identifier records includes the user identifier record at one end of the edge connecting the user to be determined, and the user identifier records that can be reached from the user identifier record at one end of the edge connecting the user to be determined along the candidate search edge. The second set of reachable user identifier records includes the user identifier record at the other end of the edge connecting the user to be determined, and the user identifier records that can be reached from the user identifier record at the other end of the edge connecting the user to be determined along the candidate search edge. The output format of the reachable user identifier record set includes the search starting point, the edge state determination time, the candidate search edges participating in the search, the searched user identifier records, and the connection path, where the connection path is composed of the candidate search edges participating in the search in sequence.

[0069] When comparing the first set of reachable user identifier records and the second set of reachable user identifier records, it is compared whether the user identifier records in the first set of reachable user identifier records are the same as those in the second set of reachable user identifier records. When there are identical user identifier records in the first set of reachable user identifier records and the second set of reachable user identifier records, the user identifier records at both ends of the user association edge to be judged are still connected through the candidate search edge without depending on the user association edge to be judged, and the user association edge to be judged is not used as a time-distortion bridging edge connecting different user subgraphs. When there is no intersection between the first set of reachable user identifier records and the second set of reachable user identifier records, the user identifier records at both ends of the user association edge to be judged belong to different reachable ranges without depending on the user association edge to be judged, the user association edge to be judged is determined to be a time-distortion bridging edge connecting different user subgraphs, and the time-distortion bridging edge is used as the target bridging edge.

[0070] When using a time-distorted bridging edge as the target bridging edge, the following are retained: the user-associated edge identifier corresponding to the target bridging edge, the user identifier records at both ends of the target bridging edge, the source of the target bridging edge, the effective time of the target bridging edge, the confidence version of the target bridging edge, the judgment triggering reason of the target bridging edge, the first reachable user identifier record set, and the second reachable user identifier record set. The target bridging edge is used to perform inactivation processing in the user behavior graph, and the first reachable user identifier record set and the second reachable user identifier record set are used to characterize the range of user identifier records corresponding to both ends of the target bridging edge.

[0071] The target bridging edge is deactivated from the user behavior graph, connected component partitioning is performed to generate a split user subgraph, and the version mapping relationship between the original user cluster version and the split user subgraph version is marked, including:

[0072] The connection status of the target bridging edge is used to indicate whether the target bridging edge participates in the connection judgment in the user behavior graph. The connection status includes at least a valid state and an inactive state. The valid state indicates that the target bridging edge can connect two user identification records as a user association edge. The inactive state indicates that the target bridging edge is retained in the user behavior graph but is no longer used as the basis for connection between user identification records. When the connection status of the target bridging edge is marked as inactive, the connection status of the target bridging edge is updated in the user behavior graph, and the connection status change time and connection status change reason are written. The connection status change time is the edge status judgment time, and the connection status change reason is the judgment trigger reason of the target bridging edge.

[0073] After marking the connection status of the target bridging edge as inactive, the source, validity period, and confidence version of the target bridging edge are still retained in the user behavior graph. When retaining the source of the target bridging edge, the account correspondence, device identifier correspondence, communication identifier correspondence, advertising identifier correspondence, network co-address relationship, or behavior association relationship corresponding to the target bridging edge are also retained. When retaining the validity period of the target bridging edge, the validity start time and validity end time are retained, as well as the edge status judgment time used to determine the expiration of the validity period are retained. When retaining the confidence version of the target bridging edge, the most recently recorded confidence version of the target bridging edge before the edge status judgment time and the confidence version of the target bridging edge corresponding to the edge status judgment time are retained. After the target bridging edge is saved in an inactive state in the user behavior graph, the target bridging edge does not participate in the connected component partitioning, but the source, validity period, and confidence version of the target bridging edge continue to serve as the data source for version mapping relationships and the screening of contaminated derived marketing states.

[0074] When performing connected component partitioning, the user behavior graph after the target bridging edge is deactivated is used as the processing object. The user identifier record and user associated edge in the original user cluster corresponding to the target bridging edge are read. The original user cluster is the user cluster containing the user identifier records at both ends of the target bridging edge. The original user cluster version is the version record corresponding to the original user cluster before the target bridging edge is deactivated. The original user cluster version includes the user cluster identifier, user cluster generation time, user identifier record participating in user cluster partitioning, user associated edge participating in user cluster partitioning, and the source, validity time, and confidence version of the user associated edge. The user associated edge participating in connected component partitioning is the user associated edge in the original user cluster version that is still in a valid state except for the target bridging edge. The validity time of the user associated edge participating in connected component partitioning needs to cover the edge status judgment time. The target bridging edge with a deactivated connection state and the user associated edge whose validity time does not cover the edge status judgment time are not used as edges connecting two user identifier records.

[0075] When performing connected component partitioning based on user identifier records that continue to be interconnected through user association edges after the target bridging edge becomes inactive, the process starts with any user identifier record in the original user cluster version that has not yet been assigned to the split user subgraph. This user identifier record is written into the search queue, and the search continues along the user association edges participating in the connected component partitioning to find reachable user identifier records. Each time a new user identifier record is found, it is written into the same split user subgraph, and the new user identifier record continues to be used as the starting point for searching adjacent user association edges. When there are no unprocessed user identifier records in the search queue, the current split user subgraph is generated. The same search continues, starting with user identifier records in the original user cluster version that have not yet been assigned to any split user subgraph, until all user identifier records in the original user cluster version are assigned to the corresponding split user subgraphs. The output of connected component partitioning includes the user identifier records contained in each split user subgraph, the user association edges contained in each split user subgraph, the cross-channel user behavior records corresponding to each split user subgraph, and the marketing event records corresponding to each split user subgraph.

[0076] When generating split user subgraphs, according to the record affiliation relationships in the user behavior graph, cross-channel user behavior records connected to user identifier records within the split user subgraphs are assigned to the corresponding split user subgraphs; according to the event affiliation relationships in the user behavior graph, marketing event records connected to user identifier records within the split user subgraphs are assigned to the corresponding split user subgraphs; when a cross-channel user behavior record is connected to multiple user identifier records and these records are located in different split user subgraphs, the cross-channel user behavior record simultaneously records multiple corresponding split user subgraph versions to determine cross-split user subgraph behavior records; when a marketing event record is connected to multiple user identifier records and these records are located in different split user subgraphs, the marketing event record simultaneously records multiple corresponding split user subgraph versions to determine whether the marketing event record belongs to a different split user subgraph.

[0077] When recording the correspondence between the original user cluster version and the split user subgraph version, a split user subgraph version is generated for each split user subgraph. The split user subgraph version includes the split user subgraph identifier, subgraph generation time, source original user cluster version, target bridging edge that triggered the split, user identifier records contained in the split user subgraph, user association edges contained in the split user subgraph, cross-channel user behavior records corresponding to the split user subgraph, and marketing event records corresponding to the split user subgraph. The split user subgraph identifier can be derived from the source original user cluster version, subgraph generation time, and the split user subgraph. The string is generated after sorting the internal user identifier records; a version mapping relationship is formed after corresponding records of the original user cluster version and the split user subgraph version. The version mapping relationship includes the original user cluster version, target bridging edge, connection state change time, connection state change reason, split user subgraph version, and the attribution relationship of the user identifier records in the original user cluster version to the split user subgraph version; the version mapping relationship is used to determine whether the user identifier records, cross-channel user behavior records, marketing event records, and user association edges recorded when each derived marketing state is generated belong to different split user subgraphs by combining the user association edge dependency chain.

[0078] Based on user-related edge dependency chains and version mapping relationships, polluted derived marketing states generated from dependency target bridging edges or user subgraph behavior records after splitting are filtered and written to the rollback queue, including:

[0079] The user-related edge dependency chain includes derived marketing state identifier, derived marketing state type, user cluster identifier, generation time, user identifier record participating in the calculation, cross-channel user behavior record participating in the calculation, marketing event record participating in the calculation, and user-related edges participating in the calculation; the version mapping relationship includes the original user cluster version, target bridging edge, connection state change time, connection state change reason, user subgraph version after splitting, and the attribution relationship between user identifier record in the original user cluster version and user subgraph version after splitting; the user-related edge dependency chain is read one by one according to the derived marketing state identifier, so that each derived marketing state can correspond to the user identifier record, cross-channel user behavior record, marketing event record, and user-related edge actually participating in the calculation at the time of generation.

[0080] When reading user identifier records, cross-channel user behavior records, marketing event records, and user-related edges recorded when each derived marketing state is generated based on the user-related edge dependency chain, the read objects must be consistent with the objects saved when the user-related edge dependency chain is generated. When reading user identifier records involved in the calculation, the corresponding account identifier, device identifier, communication identifier, advertising identifier, network address, store member identifier, or channel temporary identifier is read. When reading cross-channel user behavior records involved in the calculation, the normalized time field, source marker, channel type, access object, access order, and behavior type of the cross-channel user behavior records are read. When reading marketing event records involved in the calculation, the marketing task identifier, marketing channel, marketing content type, reach time, reach object identifier, and reach result of the marketing event records are read. When reading user-related edges involved in the calculation, the two endpoints, source, validity time, confidence version, formation basis, and participation record of the user-related edge are read.

[0081] When determining whether the user-related edges recorded when each derived marketing state is generated contain the target bridging edge based on the version mapping relationship, each user-related edge recorded when each derived marketing state is generated is compared with the target bridging edge in the version mapping relationship: The comparison is made to see if the two endpoints of the user-related edge are consistent with the user identifier records at both ends of the target bridging edge, and to see if the source, validity time, and confidence version of the user-related edge are consistent with the source, validity time, and confidence version of the target bridging edge. If any user-related edge recorded when each derived marketing state is generated is consistent with the target bridging edge in terms of both endpoints, source, validity time, and confidence version, then the user-related edge recorded when each derived marketing state is generated contains the target bridging edge. If none of the user-related edges recorded when each derived marketing state is generated are consistent with the target bridging edge in terms of both endpoints, source, validity time, and confidence version, then the user-related edges recorded when each derived marketing state is generated do not contain the target bridging edge.

[0082] When determining whether user identifier records, cross-channel user behavior records, and marketing event records recorded when each derived marketing state is generated belong to different split user subgraphs based on version mapping relationships, the version of the split user subgraph corresponding to the user identifier record recorded when each derived marketing state is generated is determined according to the attribution relationship between the user identifier record within the original user cluster version and the split user subgraph version in the version mapping relationship. For cross-channel user behavior records recorded when each derived marketing state is generated, the record attribution relationship in the user behavior graph is read, and the version of the split user subgraph corresponding to the cross-channel user behavior record is determined according to the user identifier records connected by the record attribution relationship. For marketing event records recorded when each derived marketing state is generated, the event attribution relationship in the user behavior graph is read, and the version of the split user subgraph corresponding to the marketing event record is determined according to the user identifier records connected by the event attribution relationship.

[0083] When the user identifier record, cross-channel user behavior record, and marketing event record recorded when each derived marketing state is generated all correspond to the same split user subgraph version, it is determined that the user identifier record, cross-channel user behavior record, and marketing event record recorded when each derived marketing state is generated do not belong to different split user subgraphs; when the user identifier record recorded when each derived marketing state is generated corresponds to at least two split user subgraph versions, or the cross-channel user behavior record recorded when each derived marketing state is generated corresponds to at least two split user subgraph versions, or the marketing event record recorded when each derived marketing state is generated corresponds to at least two split user subgraph versions, it is determined that the user identifier record, cross-channel user behavior record, and marketing event record recorded when each derived marketing state is generated belong to different split user subgraphs; when a cross-channel user behavior record or marketing event record is connected to multiple user identifier records, the attribution is determined according to the multiple split user subgraph versions saved when the split user subgraph is generated.

[0084] When marking a derived marketing state containing a target bridging edge or belonging to different split user subgraphs as a contaminated derived marketing state, it is determined whether each derived marketing state satisfies the condition that the user-related edge contains the target bridging edge, and then whether each derived marketing state satisfies the condition that the user identifier record, cross-channel user behavior record, and marketing event record belong to different split user subgraphs. A derived marketing state that satisfies any of these conditions is written into the contaminated derived marketing state marking result. The contaminated derived marketing state marking result includes the derived marketing state identifier, derived marketing state type, user cluster identifier, generation time, contamination source type, associated target bridging edge, associated split user subgraph version, user identifier record participating in the calculation, cross-channel user behavior record participating in the calculation, marketing event record participating in the calculation, and user-related edge participating in the calculation; the contamination source type includes containing a target bridging edge and belonging to different split user subgraphs.

[0085] When writing contaminated derived marketing states into the rollback queue, the order in the rollback queue is determined by the generation time and type of the contaminated derived marketing state. Contaminated derived marketing states with later generation times are arranged before those with earlier generation times. When the generation times are the same, they are arranged in the order of attribution status, marketing task status, reach frequency control, marketing segmentation, and profile features. Each queue record in the rollback queue includes the contaminated derived marketing state identifier, contaminated derived marketing state type, original user cluster version, split user subgraph version, target bridging edge, contamination source type, user identifier record participating in the calculation, cross-channel user behavior record participating in the calculation, marketing event record participating in the calculation, and user association edge participating in the calculation. The rollback queue is used to read contaminated derived marketing states according to the rollback queue.

[0086] The contaminated derived marketing state is revoked according to the rollback queue. The derived marketing state is then recalculated within the split user subgraph, and marketing service control instructions are generated, including:

[0087] When reading the contaminated derived marketing state according to the rollback queue, each queue record is read in the order of its arrangement. The queue record includes the contaminated derived marketing state identifier, the contaminated derived marketing state type, the original user cluster version, the split user subgraph version, the target bridging edge, the contamination source type, the user identifier record participating in the calculation, the cross-channel user behavior record participating in the calculation, the marketing event record participating in the calculation, and the user association edge participating in the calculation. After reading the contaminated derived marketing state, the corresponding profile features, marketing segments, reach frequency control, marketing task status, and attribution status are located based on the contaminated derived marketing state identifier. The profile features, marketing segments, reach frequency control, marketing task status, and attribution status corresponding to the contaminated derived marketing state are marked as revocation status. The revocation status includes the revocation time, revocation reason, associated target bridging edge, contamination source type, and corresponding rollback queue record. The revocation time is the processing time after the contaminated derived marketing state is written into the rollback queue, and the revocation reason is the contamination source type. The contamination source type includes those containing target bridging edges and those belonging to different split user subgraphs.

[0088] When the profile features, marketing segments, reach frequency control, marketing task status, and attribution status corresponding to the contaminated derived marketing state are marked as revoked, the user identifier records, cross-channel user behavior records, marketing event records, and user association edges recorded when the contaminated derived marketing state was generated are not deleted, and the user association edge dependency chain corresponding to the contaminated derived marketing state is retained; when a profile feature is marked as revoked, the profile feature no longer participates in the recalculation of marketing segments; when a marketing segment is marked as revoked, the marketing segment no longer serves as the input for the recalculation of marketing task status; when reach frequency control is marked as revoked, reach frequency control no longer serves as the input for the recalculation of marketing task status; when a marketing task status is marked as revoked, the marketing task status no longer serves as the input for the recalculation of attribution status; when an attribution status is marked as revoked, the attribution status no longer serves as the input for the generation of marketing service control instructions.

[0089] When reserving derived marketing states not written to the rollback queue to the corresponding split user subgraph based on the version mapping relationship, each derived marketing state identifier is read and compared with the contaminated derived marketing state identifier in the rollback queue. If the derived marketing state identifier does not appear in the rollback queue, the derived marketing state is determined to be a derived marketing state not written to the rollback queue. For derived marketing states not written to the rollback queue, the user association edge dependency chain corresponding to the derived marketing state not written to the rollback queue is read. Based on the user identifier record participating in the calculation in the user association edge dependency chain and the attribution relationship between the user identifier record in the original user cluster version and the split user subgraph version in the version mapping relationship, the split user subgraph version corresponding to the derived marketing state not written to the rollback queue is determined, and the state attribution relationship of the derived marketing state not written to the rollback queue is updated from the original user cluster version to the corresponding split user subgraph version.

[0090] When recalculating profile features, marketing segmentation, reach frequency control, marketing task status, and attribution status within the split user subgraph based on the user identifier record, cross-channel user behavior record, marketing event record, and user association edge, the user identifier record, user association edge, cross-channel user behavior record, and marketing event record corresponding to the split user subgraph are read from the split user subgraph version, excluding derived marketing statuses that have been marked as cancelled. When recalculating profile features, statistics are performed on the cross-channel user behavior record and marketing event record corresponding to the split user subgraph according to the behavior type, channel type, access object, and normalized time fields to obtain the number of views, clicks, favorites, add-to-cart, orders, payments, unsubscribes, complaints, inquiries, the time of the most recent behavior, the time of the most recent reach, and the number of responses after reach.

[0091] When recalculating marketing segments, the recalculated user profile features and the marketing event records corresponding to the split user subgraphs are used as input. The split user subgraphs are matched with pre-configured marketing segmentation conditions, and marketing segments are generated according to the pre-set segmentation priorities in the marketing segmentation conditions. When recalculating reach frequency control, the marketing event records corresponding to the split user subgraphs are used as input. The number of channel reach times, content reach times, the most recent reach time, and the time interval between adjacent reach times for the split user subgraphs within a specified time window are counted according to marketing channels, marketing content types, and reach times. The next allowed reach time is determined based on the most recent reach time and the interval duration configured for the corresponding marketing channel or corresponding marketing content type. When recalculating marketing task status, the marketing event records corresponding to the split user subgraphs and the recalculated reach frequency control are used as input. Using the recalculated marketing segments as input, marketing event records are aggregated according to marketing task identifiers. Based on the reach results, reach time, reach object identifiers, and recalculated reach frequency control of the marketing event records, the states of pending reach, reached reach, responded to, suppressed, and ended are determined. When recalculating the attribution state, the marketing event records corresponding to the split user subgraphs, the cross-channel user behavior records corresponding to the split user subgraphs, and the recalculated marketing task states are used as inputs. The reach time of the marketing event records is used as the attribution starting point. The cross-channel user behavior records corresponding to orders, payments, cancellations, complaints, and inquiries that occur after the reach time are used as the attribution behavior records. Attribution results are generated based on the matching relationship between the attribution behavior records and the marketing event records in terms of reach object identifiers, access objects, and marketing content types.

[0092] When the recalculated profile features, marketing segments, reach frequency control, marketing task status, and attribution status are used as the recalculated derived marketing status, a new derived marketing status identifier is generated for the recalculated derived marketing status. The recalculated derived marketing status type, the corresponding split user subgraph version, the generation time, the user identifier record participating in the recalculation, the cross-channel user behavior record participating in the recalculation, the marketing event record participating in the recalculation, and the user association edge participating in the recalculation are recorded. The recalculated derived marketing status establishes a replacement relationship with the contaminated derived marketing status of the revoked status. The replacement relationship includes the contaminated derived marketing status identifier of the revoked status, the recalculated derived marketing status identifier, the split user subgraph version, and the target bridging edge.

[0093] When generating marketing service control instructions, the recalculated derived marketing state is used as input, and the content of the marketing service control instructions is determined according to marketing segments, reach frequency control, marketing task status, and attribution status. Marketing service control instructions include target user identifier records, user subgraph versions after target splitting, marketing channels, marketing content types, allowed reach flags, suppressed reach flags, next allowed reach time, marketing task status update results, and attribution status update results. When reach frequency control shows the next allowed reach time is later than the planned reach time of the marketing task, the marketing service control instruction writes the suppressed reach flag; when reach frequency control shows the next allowed reach time is not later than the planned reach time of the marketing task and the marketing segment meets the marketing task configuration conditions, the marketing service control instruction writes the allowed reach flag; when the attribution status changes, the marketing service control instruction writes the attribution status update result. Marketing service control instructions and the recalculated derived marketing state are saved simultaneously and used to execute marketing service processing based on the split user subgraph versions.

[0094] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An AI-powered intelligent marketing service management system based on user behavior graphs, characterized in that, include: The graph construction module acquires cross-channel user behavior records, user identifier records, and marketing event records, generates user association edges according to user identifier relationships, and forms a user behavior graph from the user association edges; State Derivation Module: Generates user clusters based on the connectivity of the user behavior graph, calculates derived marketing states for user clusters, and records the user-related edge dependency chains corresponding to the derived marketing states. Bridge identification module: When the validity period of a user-related edge expires, the confidence version decreases, or the source status changes, it determines whether the user-related edge is a time-distorted bridge edge connecting different user subgraphs and obtains the target bridge edge. Subgraph splitting module: Deactivates the target bridging edge from the user behavior graph, performs connected component partitioning to generate the split user subgraph, and marks the version mapping relationship between the original user cluster version and the split user subgraph version; Pollution Filtering Module: Based on the user-related edge dependency chain and version mapping relationship, filter the polluted derived marketing states generated by the dependent target bridging edge or the user subgraph behavior records after the split, and write them to the rollback queue; Rollback and Recalculation Module: Undoes the contaminated derived marketing state according to the rollback queue, recalculates the derived marketing state in the split user subgraph, and generates marketing service control instructions.

2. The AI-powered intelligent marketing service management system based on user behavior graphs according to claim 1, characterized in that, Acquire cross-channel user behavior records, user identifier records, and marketing event records; generate user association edges according to user identifier relationships; and form a user behavior graph from these user association edges, specifically as follows: Time normalization and source marking are performed on cross-channel user behavior records, user identification records, and marketing event records. User identification associations are determined based on account correspondence, device identification correspondence, communication identification correspondence, advertising identification correspondence, network co-location relationship, and behavior association. Configure the source, validity period, and confidence version for user identification associations to generate user association edges; A user behavior graph is formed by connecting the corresponding user identifier records with the user-related edges.

3. The AI-powered intelligent marketing service management system based on user behavior graphs according to claim 2, characterized in that, User clusters are generated based on the connectivity relationships in the user behavior graph. Derived marketing states are calculated for each user cluster, and the user-related edge dependency chains corresponding to these derived marketing states are recorded. Specifically: User clusters are divided based on user identifier records connected by user association edges in the user behavior graph. Cross-channel user behavior records and marketing event records within user clusters are aggregated, and derived marketing states, including profile features, marketing segmentation, reach frequency control, marketing task status, and attribution status, are calculated. When each derived marketing state is generated, the user identifiers, cross-channel user behavior records, marketing event records, and user association edges involved in the calculation are recorded to form a user association edge dependency chain corresponding to each derived marketing state.

4. The AI-powered intelligent marketing service management system based on user behavior graphs according to claim 3, characterized in that, When the validity period of a user-related edge expires, the confidence level decreases, or the source state changes, it is determined whether the user-related edge is a time-distortion bridging edge connecting different user subgraphs, and the target bridging edge is obtained, specifically: When the validity period of a user-related edge expires, the confidence level decreases, or the source status changes, the user-related edge that meets any of these conditions will be used as the user-related edge to be judged. Starting from the user identifier records at both ends of the user association edge to be determined, search for the set of reachable user identifier records along the user association edges other than the user association edge to be determined. When there is no intersection between the sets of reachable user identifier records, the user association edge to be judged is determined to be a time-distortion bridging edge connecting different user subgraphs, and the time-distortion bridging edge is taken as the target bridging edge.

5. The AI-powered intelligent marketing service management system based on user behavior graphs according to claim 4, characterized in that, The target bridging edge is deactivated from the user behavior graph, connected component partitioning is performed to generate a split user subgraph, and the version mapping relationship between the original user cluster version and the split user subgraph version is marked, specifically: The connection status of the target bridging edge is marked as inactive, and the source, validity period, and confidence version of the target bridging edge are retained in the user behavior graph. Based on the user identifier records that continue to connect to each other through user-associated edges after the target bridging edge is deactivated, perform connected component partitioning to generate the split user subgraph; Record the original user cluster version and the corresponding user subgraph version after splitting to form a version mapping relationship.

6. The AI-powered intelligent marketing service management system based on user behavior graphs according to claim 5, characterized in that, Based on the user-related edge dependency chain and version mapping relationship, filter the polluted derived marketing states generated from the dependent target bridging edges or user subgraph behavior records after the split, and write them into the rollback queue, specifically: Based on the user association edge dependency chain, read the user identifier record, cross-channel user behavior record, marketing event record and user association edge recorded when each derived marketing state is generated; Combine the version mapping relationship to determine whether the user association edge recorded when each derived marketing state is generated contains the target bridging edge, or to determine whether the user identification record, cross-channel user behavior record and marketing event record recorded when each derived marketing state is generated belong to different split user subgraphs; Mark derived marketing states that contain target bridging edges or belong to different split user subgraphs as contaminated derived marketing states and write contaminated derived marketing states to the rollback queue.

7. The AI-powered intelligent marketing service management system based on user behavior graphs according to claim 6, characterized in that, The contaminated derived marketing state is revoked according to the rollback queue. The derived marketing state is then recalculated within the split user subgraph, and marketing service control instructions are generated, specifically as follows: Read the contaminated derivative marketing status according to the withdrawal queue, and mark the profile features, marketing groups, reach frequency control, marketing task status and attribution status corresponding to the contaminated derivative marketing status as withdrawal status; Based on the version mapping relationship, the derived marketing states that are not written to the rollback queue are retained in the corresponding split user subgraph; Within the split user subgraph, recalculate profile features, marketing segmentation, reach frequency control, marketing task status, and attribution status based on user identification records, cross-channel user behavior records, marketing event records, and user association edges. The recalculated profile features, marketing segments, reach frequency control, marketing task status, and attribution status are used as the recalculated derived marketing status to generate marketing service control instructions.

Images