High-potential new customer group intelligent mining method and system based on full-scene behavior data analysis

Abstract

The application discloses a high-potential new customer group intelligent mining method and system based on full-scene behavior data analysis, first acquires cross-channel behavior track data sets, executes full-scene behavior mode deconstruction processing, generates a scene-related behavior chain network and extracts a cross-scene behavior migration path of a user, encodes the cross-scene behavior migration path into a behavior transition feature vector sequence, performs sequence mode matching on the behavior transition feature vector sequence and a potential activation mode template, allocates a potential tendency weight according to a migration mode matching degree, obtains a potential weighted feature vector sequence, performs multi-round propagation iteration based on a user social association relationship and the potential tendency weight, determines a high-potential new customer group according to a cumulative receiving weight after propagation convergence, and realizes collaborative analysis of cross-scene behavior characteristics and social propagation relationships.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and more specifically, to a method and system for intelligently identifying high-potential new customer groups based on full-scenario behavioral data analysis. Background Technology

[0002] With the continuous enrichment of business interaction channels such as online pages, offline devices, and self-service terminals, user behavior data exhibits characteristics of cross-channel, cross-scenario, and continuous migration. Existing methods for identifying high-potential new customers typically analyze behavior data from a single channel or static user characteristics, making it difficult to effectively depict the behavioral migration relationships between users across different business scenarios. Furthermore, existing methods underutilize social connections and potential influence propagation relationships between users, easily overlooking potential users whose behavioral characteristics are not yet fully manifested but who are continuously influenced by high-potential users, thus limiting the comprehensiveness and accuracy of high-potential new customer identification. Therefore, how to comprehensively utilize cross-scenario behavioral migration characteristics and user social propagation relationships for intelligent identification of high-potential new customers has become a technical problem that needs to be solved. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for intelligently identifying high-potential new customer groups based on full-scenario behavioral data analysis.

[0004] In a first aspect, embodiments of the present invention provide a method for intelligently identifying high-potential new customer groups based on full-scenario behavioral data analysis, the method comprising:

[0005] Obtain cross-channel behavior trajectory dataset;

[0006] Perform full-scenario behavior pattern deconstruction processing on the cross-channel behavior trajectory dataset to generate a scenario-related behavior chain network. Extract the cross-scenario behavior migration path of each user based on the scenario-related behavior chain network, and encode the cross-scenario behavior migration path into a sequence of behavior transition feature vectors.

[0007] A preset potential activation pattern template library is retrieved, and the behavioral transition feature vector sequence is matched with each potential activation pattern template in the potential activation pattern template library. The degree of migration pattern matching between the behavioral transition feature vector sequence and each potential activation pattern template is calculated, and potential tendency weights are assigned to each vector in the behavioral transition feature vector sequence according to the degree of migration pattern matching to obtain a potential weighted feature vector sequence.

[0008] The potential weighted feature vector sequence is input into a preset social propagation influence model. Based on the social relationship between users and the potential tendency weight, multiple rounds of propagation iterations are performed, and the cumulative reception weight of social nodes is updated until a preset propagation convergence state is reached. Then, users corresponding to social nodes whose cumulative reception weight exceeds a preset propagation threshold are extracted as high-potential new customer groups.

[0009] Secondly, embodiments of the present invention provide a high-potential new customer group intelligent mining system based on full-scenario behavioral data analysis, including at least one service node;

[0010] The service node includes a storage unit and a computing unit; the storage unit is used to store program code; the computing unit is used to run the program code to execute the intelligent mining method for high-potential new customer groups based on full-scenario behavioral data analysis as described in the first aspect.

[0011] Compared with existing technologies, the beneficial effects provided by this invention include: by deconstructing cross-channel behavioral trajectory data into full-scenario behavioral patterns, it can continuously depict the behavioral migration relationship of users in different business interaction scenarios, thereby improving the completeness of user potential characteristic representation; by matching potential activation pattern templates and allocating potential tendency weights, it can highlight key behavioral stages with high potential tendencies; further, by combining users' social relationships for multiple rounds of influence propagation iterations, it can discover potential users whose own behavioral characteristics have not yet fully manifested but who are continuously influenced by high-potential users, thereby improving the accuracy, comprehensiveness, and intelligence of high-potential new customer group mining. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating the steps of the intelligent mining method for high-potential new customer groups based on full-scenario behavioral data analysis provided in this embodiment of the invention.

[0014] Figure 2 This is a schematic diagram of the structure for constructing potential activation pattern templates and sequence pattern matching processing provided in an embodiment of the present invention;

[0015] Figure 3 This is a schematic diagram of the structure for high-value propagation path extraction processing provided in an embodiment of the present invention;

[0016] Figure 4A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0018] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the intelligent mining method for high-potential new customer groups based on full-scenario behavioral data analysis provided in this embodiment of the invention. The following is a detailed introduction to this intelligent mining method for high-potential new customer groups based on full-scenario behavioral data analysis.

[0020] Step S201: Obtain cross-channel behavior trajectory dataset;

[0021] Step S202: Perform full-scene behavior pattern deconstruction processing on the cross-channel behavior trajectory dataset to generate a scene-related behavior chain network. Extract the cross-scene behavior migration path of each user based on the scene-related behavior chain network and encode the cross-scene behavior migration path into a sequence of behavior transition feature vectors.

[0022] Step S203: Retrieve a preset potential activation pattern template library, perform sequence pattern matching processing on the behavior transition feature vector sequence and each potential activation pattern template in the potential activation pattern template library, calculate the migration pattern matching degree between the behavior transition feature vector sequence and each potential activation pattern template, and assign potential tendency weights to each vector in the behavior transition feature vector sequence according to the migration pattern matching degree to obtain a potential weighted feature vector sequence.

[0023] Step S204: Input the potential weighted feature vector sequence into the preset social propagation influence model, perform multiple rounds of propagation iterations based on the social relationship between users and the potential tendency weight, and update the cumulative reception weight of social nodes until the preset propagation convergence state is reached. Then, extract the users corresponding to the social nodes whose cumulative reception weight exceeds the preset propagation threshold as high-potential new customer groups.

[0024] In this embodiment, a new customer group intelligent mining server is used as the execution entity. Within the scope of the user authorization agreement, the server acquires user behavior data generated by online page interaction terminals, offline device interaction terminals, and self-service interaction terminals. It constructs a cross-channel behavior trajectory dataset according to user anonymization identifiers and behavior occurrence times, and builds a scene-related behavior chain network based on continuous behavior sequence fragments. It extracts the cross-scene behavior migration path and behavior transition feature vector sequence corresponding to each user. The server further calls a potential activation mode template library to perform sequence pattern matching on the behavior transition feature vector sequence, generates a potential weighted feature vector sequence based on the degree of migration pattern matching, and performs propagation iteration in conjunction with the social connections between users. When the social propagation influence model reaches a preset propagation convergence state, the server determines high-potential new customer groups based on the cumulative reception weight of each social node. The specific implementation methods of the aforementioned processing steps are described in the following embodiments.

[0025] In this embodiment of the invention, the cross-channel behavior trajectory dataset includes continuous behavior sequence fragments generated by users in multiple business interaction channels, and each continuous behavior sequence fragment is accompanied by a channel type identifier and a behavior trigger timestamp; the aforementioned step S201 can be implemented by the following example.

[0026] Within the scope of the user authorization agreement, original behavioral event streams are synchronously collected from online page interaction terminals, offline device interaction terminals, and self-service interaction terminals. Each original behavioral event in the original behavioral event stream records a user anonymization identifier, event trigger time, event type code, and interaction channel identifier.

[0027] The original behavioral event stream is aggregated according to the user anonymization identifier. All original behavioral events associated with the same user anonymization identifier are arranged in chronological order of event triggering time to form an event timeline sequence for each user.

[0028] Perform event semantic compression processing on the event timeline sequence, and merge the codes of the same type of event with an event trigger time interval of less than a preset interval threshold into behavior segment primitives. The behavior segment primitives encapsulate the start event type code, the end event type code, and the segment duration.

[0029] Based on the interaction channel identifier carried by the original behavior event within the behavior segment primitive, a channel type identifier is added to the behavior segment primitive to generate a continuous behavior sequence segment with the channel type identifier attached.

[0030] The continuous behavior sequence fragments with channel type identifiers are associated and stored with the user anonymization identifiers to construct a cross-channel behavior trajectory dataset with the user anonymization identifiers as the main index.

[0031] The starting event trigger time of each continuous behavior sequence segment in the cross-channel behavior trajectory dataset is used as the behavior trigger timestamp, and the behavior trigger timestamp is appended to the corresponding continuous behavior sequence segment.

[0032] In this embodiment, taking user anonymization identifier U1027 as an example, the server receives the original behavioral events generated by the user through online page interaction terminals, offline device interaction terminals, and self-service interaction terminals, and records the user anonymization identifier, event trigger time, event type code, and interaction channel identifier. The server collects the original behavioral events according to the user anonymization identifier and arranges them into an event timeline sequence according to the event trigger time.

[0033] For consecutive events in the event timeline sequence with time intervals less than a preset threshold and belonging to the same event type encoding, the server performs event semantic compression. For example, the content browsing events continuously generated by U1027 from 09:13 to 09:16 are merged into behavior segment primitives, and the starting event type encoding, ending event type encoding, and segment duration are determined based on the corresponding original behavior events. The server further adds a channel type identifier to the behavior segment primitives based on the interaction channel identifier to generate continuous behavior sequence segments, and stores the continuous behavior sequence segments in association with the user anonymization identifier. The server determines the starting event trigger time of each continuous behavior sequence segment as the behavior trigger timestamp, thereby forming a cross-channel behavior trajectory dataset with the user anonymization identifier as the primary index.

[0034] In this embodiment of the invention, the aforementioned step S202 can be implemented through the following examples.

[0035] Extract all continuous behavior sequence fragments associated with the same anonymous user from the cross-channel behavior trajectory dataset to form the user's full-channel behavior sequence set;

[0036] Each continuous behavior sequence segment in the omnichannel behavior sequence set is processed into a behavior sequence node, and each continuous behavior sequence segment is converted into a behavior sequence node. The behavior sequence node encapsulates the channel type identifier, behavior trigger timestamp and segment duration of the continuous behavior sequence segment.

[0037] Arrange all behavior sequence nodes corresponding to the same user anonymization identifier in chronological order of behavior trigger timestamps. Establish scene jump relationship edges between behavior sequence nodes with adjacent times after arrangement. The scene jump relationship edges record scene switching direction indicators from the preceding behavior sequence node to the following behavior sequence node.

[0038] For each scene jump relationship edge, calculate the time interval between the behavior trigger timestamp of the preceding behavior sequence node and the behavior trigger timestamp of the following behavior sequence node connected to it, and attach the time interval as the scene switching interval duration parameter to the scene jump relationship edge.

[0039] The user scenario behavior chain network is formed by combining all behavior sequence nodes and all scene jump relationship edges corresponding to the same user anonymization identifier. The user scenario behavior chain network represents the user's behavior migration relationship between various business interaction channels.

[0040] Perform cross-user merging processing on the user scenario behavior chain network corresponding to all user anonymization identifiers, aggregate behavior sequence nodes with the same channel identifier sequence pattern into shared behavior sequence nodes, and generate a scenario-related behavior chain network covering all users.

[0041] In the scene-related behavior chain network, for each user's anonymized identifier, path backtracking is performed along the user's behavior sequence nodes and scene jump relationship edges to extract the node sequence formed by the user's behavior sequence nodes, which serves as the user's cross-scene behavior migration path.

[0042] The channel type identifier, behavior trigger timestamp, and segment duration encapsulated in each behavior sequence node in the cross-scene behavior migration path are encoded into a multi-dimensional feature vector, and the multi-dimensional feature vector sequence corresponding to the cross-scene behavior migration path is used as the behavior jump feature vector sequence.

[0043] In this embodiment, taking user anonymization identifier U1027 as an example, the server extracts continuous behavioral sequence fragments corresponding to the user from the cross-channel behavioral trajectory dataset, such as online page content browsing, offline device business query, and self-service terminal business condition viewing, and converts them into behavioral sequence nodes. Each behavioral sequence node encapsulates the corresponding channel type identifier, behavior trigger timestamp, and fragment duration.

[0044] The server sorts each behavior sequence node according to the behavior trigger timestamp and establishes scene jump relationship edges between behavior sequence nodes with adjacent times. For example, a scene jump relationship edge is established between the first behavior sequence node and the second behavior sequence node, pointing from the online page scene to the offline device scene; a scene jump relationship edge is established between the second behavior sequence node and the third behavior sequence node, pointing from the offline device scene to the self-service scene. The server also calculates the corresponding scene switching interval parameter based on the adjacent behavior trigger timestamps, thus forming the user scene behavior chain network corresponding to U1027.

[0045] The server constructs a user scenario behavior chain network for other users in the same way, and aggregates behavior sequence nodes with the same channel identifier sequence pattern to form a scenario-related behavior chain network. Subsequently, the server backtracks the corresponding behavior sequence nodes and scenario jump relationship edges according to the user's anonymized identifier to obtain the cross-scenario behavior migration path for each user, and encodes the features of each behavior sequence node according to the node order in the path to form a behavior jump feature vector sequence.

[0046] For example, to establish a correspondence between the behavior transition feature vector and the template behavior type code in the potential activation mode template, the server further reads the start event type code and the end event type code contained in the behavior segment primitives constituting the continuous behavior sequence segments corresponding to the behavior sequence nodes, and determines the behavior type code of the corresponding behavior sequence node based on the start event type code and the end event type code. For a continuous behavior sequence segment composed of multiple behavior segment primitives, the server can form a behavior type code sequence according to the time arrangement of the behavior segment primitives, or determine the main behavior type code corresponding to the continuous behavior sequence segment based on the event type code that appears most frequently.

[0047] The server jointly encodes the channel type identifier, the behavior type code, the behavior trigger timestamp, and the segment duration to form a multi-dimensional feature vector for the corresponding behavior sequence node. For example, the multi-dimensional feature vector corresponding to the i-th behavior sequence node is represented as: , where, This represents the multidimensional feature vector corresponding to the i-th action sequence node; This represents the channel type identifier encoding corresponding to the i-th behavior sequence node; This represents the behavior type encoding corresponding to the i-th behavior sequence node; This represents the behavior trigger timestamp encoding corresponding to the i-th behavior sequence node; This represents the duration of the segment corresponding to the i-th behavior sequence node.

[0048] By incorporating behavior type encoding into multidimensional feature vectors, the semantic attributes of behavior in the current user's behavior transition feature vector sequence can establish a direct correspondence with the template behavior type encoding in the potential activation pattern template, thereby enabling subsequent stage matching to be performed simultaneously based on channel type, behavior type, and behavior persistence features.

[0049] Please refer to the following: Figure 2 , Figure 2This is a schematic diagram of the structure of potential activation pattern template construction and sequence pattern matching processing provided in an embodiment of the present invention. In this embodiment of the present invention, the potential activation pattern template records the cross-scenario behavior migration pattern of high-potential users from initial activation behavior to deep conversion behavior; the aforementioned step S203 can be implemented through the following example.

[0050] Retrieve all potential activation mode templates from the preset potential activation mode template library. Each potential activation mode template contains multiple template behavior stages arranged in chronological order. Each template behavior stage is marked with a template channel type identifier, template behavior type code, and template stage duration range.

[0051] For each multidimensional feature vector in the behavioral transition feature vector sequence, the channel type identifier and segment duration corresponding to the multidimensional feature vector are parsed. The parsing results are compared stage by stage with the template channel type identifier and template stage duration range of each template behavioral stage in each potential activation mode template to obtain the stage matching score between the multidimensional feature vector and each template behavioral stage.

[0052] The stage matching scores of each template behavior stage within the same potential activation mode template are accumulated sequentially according to the time order of the template behavior stages to generate the degree of matching between the behavior transition feature vector sequence and the migration mode of the potential activation mode template.

[0053] For each multidimensional feature vector in the behavior transition feature vector sequence, the template behavior stage aligned with the multidimensional feature vector is extracted from the potential activation mode template where the migration pattern matching degree exceeds a preset matching threshold. The stage matching score of the template behavior stage is used as the basis for assigning the potential tendency weight of the multidimensional feature vector.

[0054] The potential tendency weights of each multidimensional feature vector are assigned to the corresponding multidimensional feature vectors to generate a potential weighted feature vector sequence.

[0055] For example, the potential activation pattern template library can be pre-constructed based on historical user behavior data. The server obtains historical user behavior data and corresponding historical user behavior result tags within a preset historical analysis period, and determines a historical high-potential user sample set based on the historical user behavior result tags. Each historical high-potential user in the historical high-potential user sample set has a continuous historical behavior record that gradually develops from initial activation behavior to deep conversion behavior, and the initial activation behavior and the deep conversion behavior are identified by preset behavior type codes.

[0056] For each historically high-potential user, the server extracts the corresponding historical cross-scenario behavior migration path using the same processing method as for the current user, and forms a historical behavior transition feature vector sequence according to the order of behavior occurrence. Based on the channel type identifier and behavior type code in each historical behavior transition feature vector sequence, the server groups historical behavior transition feature vector sequences that have the same characteristics or meet preset sequence similarity conditions, forming multiple historical behavior migration pattern groups.

[0057] For each historical behavior migration pattern group, the server determines multiple template behavior stages in chronological order. For the j-th template behavior stage, the server counts the channel type identifiers corresponding to the j-th template behavior stage in the historical behavior migration pattern group, and determines the channel type identifier with the highest frequency as the template channel type identifier; it also counts the behavior type codes corresponding to the j-th template behavior stage, and determines the behavior type code with the highest frequency as the template behavior type code; simultaneously, it counts the duration of all segments corresponding to the j-th template behavior stage, and determines the duration range of the template stage based on the distribution of segment durations.

[0058] In one exemplary implementation, the server can determine the minimum and maximum duration of the segment corresponding to the j-th template behavior stage as the lower and upper limits of the template stage duration range, respectively, as follows: , ,in, This represents the lower limit of the duration range of the template phase corresponding to the j-th template behavior phase; This represents the upper limit of the duration range of the template phase corresponding to the j-th template behavior phase; The duration of the segment corresponding to the nth historical high-potential user in the jth template behavior stage is represented by N; N represents the number of historical high-potential users who participated in constructing the current potential activation pattern template.

[0059] The server will associate and store multiple template behavior stages arranged in chronological order, as well as the template channel type identifier, template behavior type code, and template stage duration range corresponding to each template behavior stage, to form a potential activation mode template. It will also store multiple potential activation mode templates formed by grouping different historical behavior migration modes together to construct the potential activation mode template library.

[0060] Therefore, the potential activation mode template is not arbitrarily set in advance, but is formed by aggregating the cross-scenario behavior migration sequence actually formed by historical high-potential users according to channel type, behavior type, stage sequence and stage continuity characteristics, thereby establishing a comparable relationship between historical high-potential behavior patterns and current user behavior leap feature vector sequences.

[0061] In one exemplary implementation, to clarify the method for determining the stage matching score, the server calculates the channel type matching value, behavior type matching value, and duration matching value respectively.

[0062] For the i-th multidimensional feature vector in the behavioral transition feature vector sequence and the j-th template behavioral stage in the k-th potential activation mode template, when the channel type identifier corresponding to the i-th multidimensional feature vector is consistent with the template channel type identifier corresponding to the j-th template behavioral stage, the channel type matching value is... Set to 1 otherwise; when the behavior type code corresponding to the i-th multidimensional feature vector matches the template behavior type code corresponding to the j-th template behavior stage, the behavior type matching value is set to 1. Set to 1 otherwise; when the duration of the segment corresponding to the i-th multidimensional feature vector is within the duration of the template stage corresponding to the j-th template behavior stage, the duration will be matched with the value. Set to 1, otherwise set to 0.

[0063] The server calculates a stage matching score based on the channel type matching value, the behavior type matching value, and the duration matching value, as follows: ,in, This represents the stage matching score between the i-th multidimensional feature vector and the j-th template behavior stage in the k-th potential activation pattern template; Indicates the channel type matching value; Indicates the behavior type matching value; Indicates the duration matching value; , and These represent the matching weights corresponding to channel type, behavior type, and duration, respectively, and satisfy the following conditions: .

[0064] The server performs monotonic sequence alignment while maintaining the temporal order of the behavior transition feature vector sequence and the template behavior stages. This alignment is performed for a behavior transition feature vector sequence consisting of n multidimensional feature vectors and a sequence consisting of... The server determines the target alignment relationship with the largest cumulative stage matching score from the candidate alignment relationships that satisfy temporal order consistency, which consists of the k-th potential activation pattern template composed of template behavior stages. The target alignment relationship is expressed as: ,in, This represents the candidate alignment relationship between the behavioral transition feature vector sequence and the k-th potential activation mode template, which satisfies temporal order consistency. This represents the target alignment relationship with the highest cumulative matching score at each stage.

[0065] The server calculates the migration pattern matching degree based on the stage matching score and effective alignment ratio in the target alignment relationship, expressed as: ,in, This indicates the degree of matching between the behavioral transition feature vector sequence and the k-th potential activation pattern template; The number of mutually aligned multidimensional feature vectors and template behavior stages in the target alignment relationship is represented by n; n represents the number of multidimensional feature vectors contained in the behavior transition feature vector sequence. This indicates the number of template behavior stages contained in the k-th potential activation pattern template.

[0066] The effective alignment ratio is used to correct the stage matching score so that users whose only behavioral stages are highly consistent with the potential activation pattern template, but whose overall migration paths are significantly different, do not receive an excessively high degree of migration pattern matching.

[0067] Furthermore, regarding the degree of migration pattern matching For potential activation pattern templates exceeding a preset matching threshold, the server determines the potential tendency weight based on the stage matching score corresponding to each multi-dimensional feature vector and the migration pattern matching degree of the corresponding potential activation pattern template. For example, the potential tendency weight corresponding to the i-th multi-dimensional feature vector is expressed as: ,in, This represents the potential tendency weight corresponding to the i-th multidimensional feature vector; This represents the set of potential activation pattern templates whose migration pattern matching degree exceeds a preset matching threshold and forms an effective alignment relationship with the i-th multidimensional feature vector; This represents the template behavior stage in the k-th potential activation pattern template that is aligned with the i-th multidimensional feature vector.

[0068] Therefore, when a multidimensional feature vector not only has a high stage matching score with the corresponding template behavior stage, but also has a high degree of migration pattern matching with the potential activation mode template in the entire behavior transition feature vector sequence, the multidimensional feature vector obtains a high potential tendency weight.

[0069] Taking U1027 as an example, the server, following the above calculation method, matches the multi-dimensional feature vectors corresponding to the user's online page content browsing, offline device service query, and self-service terminal service condition viewing with the corresponding template behavior stages in the potential activation mode template, and determines the target alignment relationship and migration mode matching degree while maintaining the consistency of the behavior occurrence order. For potential activation mode templates where the migration mode matching degree exceeds a preset matching threshold, the server determines the potential tendency weight based on the stage matching score corresponding to each multi-dimensional feature vector and the migration mode matching degree, and appends the potential tendency weight to the corresponding multi-dimensional feature vector to form the potential weighted feature vector sequence corresponding to U1027.

[0070] In this embodiment of the invention, the aforementioned step S204 can be implemented through the following examples.

[0071] Obtain a user social relationship dataset, wherein each social relationship record in the user social relationship dataset includes a first user anonymization identifier, a second user anonymization identifier, and a social interaction frequency parameter;

[0072] The user social relationship dataset is transformed into a user social association topology graph. Each node in the user social association topology graph corresponds to a user anonymization identifier, each edge corresponds to a social relationship record, and the weight value of the edge is the propagation coefficient mapped by the social interaction frequency parameter.

[0073] The potential tendency weights of each multidimensional feature vector in the potential weighted feature vector sequence are summed and aggregated to obtain the potential tendency aggregate weight of each user anonymization identifier. The potential tendency aggregate weight is assigned to the node corresponding to the user anonymization identifier in the user social association topology graph as the propagation source strength of that node.

[0074] In the user social association topology graph, initialize the propagation state and set the initial value of the cumulative receiving weight of all nodes to their own propagation source strength.

[0075] Multiple rounds of propagation iterations are performed. In each round of propagation iteration, for each edge in the user social association topology graph, a portion of the propagation source strength is transmitted from the propagation source node to the target node. The amount of transmission is determined by the product of the current propagation source strength of the propagation source node and the propagation coefficient of the edge. The amount of transmission is accumulated to the cumulative reception weight of the target node.

[0076] After each round of propagation iteration, the propagation source strength of each node is updated to the product of the propagation source strength of the previous round and the propagation attenuation coefficient. When the change in the propagation source strength of all nodes is less than the preset attenuation threshold, the preset propagation convergence state is determined to be reached.

[0077] After reaching the preset propagation convergence state, extract the nodes in the user social association topology graph whose cumulative reception weight exceeds the preset propagation threshold, and mark the users pointed to by the user anonymization identifiers corresponding to these nodes as high-potential new customer groups.

[0078] In this embodiment, the server acquires a user social relationship dataset and establishes a user social association topology graph based on the social relationships between anonymized user identifiers, mapping the social interaction frequency parameter to the propagation coefficient of the corresponding social association edge. The server aggregates the potential tendency weights in the potential weighted feature vector sequence corresponding to each user and uses the resulting aggregated potential tendency weights as the initial propagation source strength of the corresponding user node.

[0079] The server initializes the cumulative reception weights of each user node and executes multiple rounds of propagation iterations. For example, the node corresponding to U1027 propagates its propagation weight to U1035 along the social connection edge between it and the node corresponding to U1035. The propagation weights received by U1035 are added to the corresponding cumulative reception weights and participate in subsequent propagation according to the propagation recursion method described later. The server continues to execute propagation iterations until a preset propagation convergence state is reached, and users corresponding to user nodes whose cumulative reception weights exceed a preset propagation threshold are marked as high-potential new customers.

[0080] For example, the server can normalize the propagation coefficient based on the frequency parameter of social interactions associated with the same propagation source node. For a social association edge pointing from user node i to user node j, its propagation coefficient is expressed as: ,in, This represents the propagation coefficient corresponding to the social connection edge between user node i and user node j; This parameter represents the frequency of social interactions between user node i and user node j. This represents the set of neighboring user nodes that have established social connections with user node i. This represents the frequency parameter of social interactions between user node i and the r-th user node in the set of neighboring user nodes.

[0081] Through the above normalization process, the sum of the propagation coefficients corresponding to different social association edges of the same user node is 1, so that the propagation coefficient can reflect the proportion of social association strength of different adjacent users relative to the current propagation source node.

[0082] In other embodiments, the propagation coefficient can be calculated after time decay processing based on the frequency of social interactions in different time periods within a preset historical time window. This embodiment does not limit this to a single method.

[0083] For example, the propagation source strength of user node i in the t-th propagation iteration is represented as The propagation amount of user node i along the social connection edge pointing to user node j is represented as: Then the transit quantity is expressed as: ,in, This represents the propagation coefficient corresponding to the social connection edge between user node i and user node j.

[0084] The cumulative reception weight of user node j after the t-th round of propagation iteration is expressed as: ,in, This represents the cumulative received weight of user node j before the start of the t-th round of propagation iteration; This represents the set of propagation source nodes that have social connection edges pointing to user node j.

[0085] To ensure that the potential influence received by a user node from other user nodes continues to propagate along social relationships to other user nodes in subsequent propagation iterations, the server determines the source strength for the next round of propagation based on the amount of transmission received in the current round and the propagation attenuation coefficient, expressed as: ,in, This represents the source strength of user node j in the (t+1)th propagation iteration; Let represent the propagation attenuation coefficient, and satisfy: .

[0086] Before the first round of propagation iterations begins, the potential tendency aggregate weight corresponding to each user node is used as the initial propagation source strength, i.e.: ,in, This represents the potential tendency aggregation weight corresponding to user node i.

[0087] Through the aforementioned propagation recursion process, the propagation influence received by a user node from an adjacent propagation source node in the current round, after propagation attenuation, can be used as the propagation source strength for the next round to continue propagating to subsequent adjacent user nodes. For example, after the node corresponding to U1027 transmits its potential influence to the node corresponding to U1035 in the first round, the node corresponding to U1035 forms the propagation source strength for the next round based on the amount of transmission received, and further propagates it along the social association edge between U1035 and U1084 to the node corresponding to U1084, thereby forming a multi-hop propagation relationship between U1027, U1035, and U1084.

[0088] After each round of propagation, the server compares the change in the propagation source strength of each user node between the current round and the previous round. When the change in the propagation source strength of all user nodes is less than a preset attenuation threshold, the server determines that the preset propagation convergence state has been reached, as shown below: ,in, This indicates the preset attenuation threshold.

[0089] In this embodiment of the invention, the method further includes:

[0090] The cross-scenario behavior migration path of users marked as high-potential new customer groups is decomposed into path segments, and the behavioral segment primitive sequence of high-potential new customer group users in various business interaction channels is extracted.

[0091] Frequent sequence pattern mining is performed on the behavioral fragment primitive sequences of all high-potential new customer groups. Subsequences of behavioral fragment primitives that occur more frequently than a preset frequency threshold are extracted as common patterns of high-potential behaviors.

[0092] The common high-potential behavior patterns are merged with existing potential activation pattern templates in the potential activation pattern template library, and high-potential behavior patterns not included in the existing potential activation pattern templates are added as potential activation pattern templates.

[0093] In this embodiment, the server extracts the cross-scenario behavior migration paths corresponding to all high-potential new customer groups, and reconstructs behavior fragment primitives based on the continuous behavior sequence fragments corresponding to each behavior sequence node, forming a behavior fragment primitive sequence according to the actual behavior occurrence order. The server performs frequent sequence pattern mining on the behavior fragment primitive sequences of all high-potential new customer groups, and identifies the behavior fragment primitive subsequences that occur more frequently than a preset frequency threshold as high-potential behavior common patterns. For example, when multiple high-potential new customer groups exhibit a continuous combination of behaviors such as browsing online page content, querying offline device services, and viewing self-service terminal service conditions, the server can extract the corresponding subsequences that meet the preset frequency condition as high-potential behavior common patterns.

[0094] The server compares the common high-potential behavior patterns with existing potential activation pattern templates. For common high-potential behavior patterns not yet included in the existing potential activation pattern templates, a new potential activation pattern template is generated according to the template conversion method described later and added to the potential activation pattern template library.

[0095] Before adding the common pattern of high-potential behavior as a potential activation pattern template, the server performs template-based conversion on the behavioral fragment primitives in the common pattern of high-potential behavior.

[0096] Specifically, the server divides one or more behavior fragment primitives with the same channel type identifier and continuous association of behavior types into the same candidate template behavior stage according to the time arrangement order of each behavior fragment primitive in the high-potential behavior common pattern, and forms a template behavior stage sequence according to the arrangement order of the candidate template behavior stages.

[0097] For each candidate template behavior stage, the server counts the channel type identifiers corresponding to the behavior fragment primitives constituting the candidate template behavior stage, and determines the channel type identifier with the highest frequency as the template channel type identifier; it counts the corresponding event type codes, and determines the event type code with the highest frequency as the template behavior type code; it counts the duration of the fragments corresponding to the candidate template behavior stage among all high-potential new customer groups, and determines the duration range of the template stage based on the duration of the fragments.

[0098] For example, the number of high-potential new customer group users who contain the aforementioned common behavioral patterns is denoted as... The total number of high-potential new customer group users participating in frequent sequence pattern mining is denoted as... The pattern support of the common pattern of high-potential behaviors is expressed as: .

[0099] When the pattern support exceeds the support threshold corresponding to the preset frequency threshold, and there is no template that satisfies the preset template repetition condition between the high-potential behavior common pattern and the existing potential activation pattern template, the server associates and stores the template behavior stage sequence obtained after template conversion with the template channel type identifier, template behavior type code and template stage duration range corresponding to each template behavior stage, and generates a new potential activation pattern template.

[0100] Therefore, the behavioral fragment primitive subsequences obtained from frequent sequence pattern mining are transformed into new potential activation pattern templates with the same data structure as existing potential activation pattern templates.

[0101] In this embodiment of the invention, the method further includes:

[0102] Obtain a list of users marked as high-potential new customer groups, and extract the channel type identifier sequence from the cross-scenario behavior migration path of each user in the user list;

[0103] Cluster the channel type identifier sequences of all high-potential new customer groups and group users whose channel type identifier sequences have a similarity exceeding a preset similarity threshold into the same channel migration preference subgroup.

[0104] For each channel migration preference subgroup, extract the high-frequency channel type identifier combinations that appear in the user channel type identifier sequence within the subgroup, and generate the channel preference descriptor set for that channel migration preference subgroup;

[0105] The set of channel preference descriptors is stored in the business interaction channel configuration module to adjust the order of channel content delivery for users corresponding to each business interaction channel.

[0106] In this embodiment of the invention, the method further includes:

[0107] Extract a subgraph composed of nodes corresponding to high-potential new customer groups from the user social association topology graph, and identify the boundary nodes in the subgraph that have direct edge connections with the nodes of high-potential new customer groups and whose own cumulative reception weight is lower than a preset propagation threshold.

[0108] For each boundary node, calculate the propagation coefficient of the edge between it and the adjacent high-potential new customer group user nodes and the weighted sum of the cumulative reception weights of the adjacent high-potential new customer group user nodes, as the potential diffusion penetration assessment value of the boundary node.

[0109] Extract boundary nodes whose potential diffusion penetration assessment values ​​exceed a preset penetration threshold, and supplement the anonymized user identifiers corresponding to the boundary nodes as high-potential new customer groups.

[0110] In this embodiment of the invention, for example, after completing multiple rounds of propagation iterations based on the social propagation influence model and obtaining the high-potential new customer group marking results, the server further utilizes the adjacency relationships of the high-potential new customer groups in the user social association topology graph to supplement the mining of users who have not yet reached the preset propagation threshold but have a strong direct association with high-potential users. Specifically, the server first locates the corresponding social nodes in the user social association topology graph based on the anonymized identifiers of users already marked as high-potential new customer groups, and extracts the social nodes and the existing social association relationship edges between the nodes to form a subgraph composed of high-potential new customer group user nodes. For example, after the aforementioned propagation iterations, the cumulative reception weights of the nodes corresponding to user anonymized identifiers U1027, U1035, and U1062 all exceed the preset propagation threshold. The server determines the above nodes as high-potential new customer group user nodes and further searches outwards along the social association relationship edges connected to the above nodes to find other user nodes directly connected to them.

[0111] The server reads the cumulative received weight of each of the retrieved neighboring user nodes and compares it with a preset propagation threshold. The node corresponding to user anonymization identifier U1084 has direct edge connections with the nodes corresponding to U1027 and U1035. However, the cumulative received weight of node U1084, even in the propagation convergence state, is still lower than the preset propagation threshold. Therefore, the server identifies node U1084 as a boundary node on the periphery of the high-potential new customer group user node subgraph. The server identifies all user nodes that have direct edge connections with high-potential new customer group user nodes and whose own cumulative received weight is lower than the preset propagation threshold in the same way, thus forming a set of boundary nodes to be further evaluated.

[0112] For each boundary node, the server further analyzes the degree of diffusion and penetration of the influence of adjacent high-potential new customer groups on that node. Taking the boundary node corresponding to U1084 as an example, the server reads the propagation coefficient of the social connection edge between U1084 and U1027, as well as the cumulative reception weight of the node corresponding to U1027. Simultaneously, it reads the propagation coefficient of the social connection edge between U1084 and U1035, as well as the cumulative reception weight of the node corresponding to U1035. The server weights the cumulative reception weight of each adjacent high-potential new customer node according to the propagation coefficient of the corresponding social connection edge, and sums the weighted results from multiple adjacent high-potential new customer nodes to obtain the potential diffusion and penetration assessment value of the boundary node corresponding to U1084. Therefore, the number of high-potential users directly influencing the boundary node simultaneously, the propagation capability of the corresponding social connection, and the degree of accumulated potential influence of adjacent high-potential users themselves are all reflected in the potential diffusion and penetration assessment value.

[0113] The server calculates the potential diffusion penetration assessment value of each boundary node using the same processing method and compares the obtained assessment value with the preset penetration threshold. For example, although the cumulative reception weight formed by U1084 in the aforementioned conventional multi-round propagation iterations has not exceeded the preset propagation threshold, it maintains direct social connections with two high-potential new customer group user nodes, U1027 and U1035, and the propagation coefficients corresponding to both edges are high. Furthermore, U1027 and U1035 themselves have high cumulative reception weights. Therefore, the potential diffusion penetration assessment value corresponding to U1084 exceeds the preset penetration threshold. Based on this, the server extracts the user anonymization identifier corresponding to U1084 and additionally marks this user as a high-potential new customer group.

[0114] Through the above processing, after completing the first round of high-potential new customer group identification based on the cumulative reception weight, the server can further expand the analysis around the social association boundary formed by the identified high-potential new customer groups. It can conduct potential diffusion and penetration assessment on boundary users who have not yet directly reached the preset propagation threshold but are simultaneously strongly and directly influenced by one or more high-potential users. This supplements the identification of high-potential users that may be missed by relying solely on conventional propagation thresholds, so that the final high-potential new customer group labeling results can further reflect the diffusion characteristics of high-potential behavior influence in the local boundary area of ​​the social association network.

[0115] In this embodiment of the invention, the method further includes:

[0116] The behavioral transition feature vector sequence corresponding to users marked as high-potential new customer groups is segmented in the time dimension, and the behavioral transition feature vector sub-sequence within the most recent preset time window is extracted.

[0117] The behavior transition feature vector subsequence is input into a preset behavior trend continuation prediction model, and the behavior trend continuation prediction model outputs a predicted behavior transition feature vector subsequence within a future preset time window.

[0118] The predicted behavior transition feature vector subsequence is matched with each potential activation pattern template in the potential activation pattern template library. The matching degree between the predicted behavior transition feature vector subsequence and each potential activation pattern template is calculated. High-potential new customer groups whose matching degree exceeds a preset maintenance threshold are included in the continuous potential tracking queue.

[0119] In this embodiment, after marking high-potential new customer groups, the server further analyzes the continuation trend of the potential behaviors of the marked users. The server first reads the behavioral transition feature vector sequence corresponding to each user from the list of high-potential new customer groups, and then segments it according to the time dimension based on the behavior trigger timestamps corresponding to each multi-dimensional feature vector. Taking user anonymization identifier U1027 as an example, this user's behavioral transition feature vector sequence records cross-scenario behaviors such as online page browsing, offline device queries, and self-service terminal business condition viewing formed at multiple time stages. The server determines the most recent preset time window based on the current analysis time, filters multi-dimensional feature vectors whose behavior trigger timestamps are located within the most recent preset time window from the complete behavioral transition feature vector sequence, and maintains the original time arrangement order of each multi-dimensional feature vector to form the behavioral transition feature vector sub-sequence corresponding to U1027. Thus, historical behaviors formed earlier and with weaker influence on the current behavioral development trend are excluded, retaining only the sequence data that reflects the user's recent cross-scenario behavioral changes.

[0120] Subsequently, the server inputs the behavioral transition feature vector sub-sequence into a preset behavioral trend continuation prediction model. The behavioral trend continuation prediction model processes the continuous change patterns of channel type, behavior occurrence sequence, and segment persistence features according to the temporal arrangement relationship between recent multi-dimensional feature vectors, and outputs a predicted behavioral transition feature vector sub-sequence for the future preset time window. For example, if U1027 continuously exhibits a migration from online page content browsing to offline device service queries within the recent preset time window, and further adds self-service terminal service condition viewing behavior, the server generates a predicted behavioral transition feature vector sub-sequence for the future preset time window based on the recent behavioral evolution relationship using the behavioral trend continuation prediction model. This predicted behavioral transition feature vector sub-sequence is arranged in prediction time order and is used to characterize the continuation state of the user's current cross-scenario behavioral development trend within the subsequent time range.

[0121] The server further retrieves the potential activation pattern template library and matches the predicted behavior transition feature vector subsequence corresponding to U1027 with each potential activation pattern template. The server compares the order of the predicted multi-dimensional feature vectors in the predicted behavior transition feature vector subsequence with the corresponding behavioral stages in the potential activation pattern templates, arranged from initial activation behavior to deep transformation behavior. It comprehensively judges the consistency between the channel migration order and behavioral continuity features corresponding to the predicted behavior and each potential activation pattern template, and calculates the matching degree between the predicted behavior transition feature vector subsequence and each potential activation pattern template. For example, if the predicted behavior transition feature vector subsequence of U1027 indicates that its subsequent behavior continues along the path of online attention, further offline inquiry, and in-depth understanding through self-service, and maintains a high degree of consistency with the aforementioned high-potential user cross-scenario behavior migration pattern, the server obtains a high template matching degree.

[0122] Finally, the server compares the predicted matching degree of each high-potential new customer group with a preset maintenance threshold. For users whose predicted behavioral transition feature vector subsequences, such as U1027, match at least one potential activation pattern template with a matching degree exceeding the preset maintenance threshold, the server determines that their current high-potential behavioral trend still has strong continuity within a preset future time window, and writes the corresponding user anonymization identifier into the continuous potential tracking queue. Users whose predicted matching degree does not reach the preset maintenance threshold are not included in the continuous potential tracking queue in this round. Thus, based on the existing high-potential new customer group identification results, the server can further combine the user's recent cross-scenario behavioral change trend to predict the continuity of their potential status, and select users with continuous potential behavior characteristics from the identified high-potential users, providing a corresponding user range for subsequent continuous tracking.

[0123] In this embodiment of the invention, the method further includes:

[0124] Obtain the cumulative reception weight distribution data of each node in the user social association topology graph when the social propagation influence model reaches the preset propagation convergence state;

[0125] Outlier node detection processing is performed on the cumulative received weight distribution data to identify outlier nodes whose cumulative received weight deviates from the cumulative received weight of neighboring social nodes by more than a preset deviation threshold.

[0126] The outlier nodes are removed from the high-potential new customer group marking results, and the original behavioral transition feature vector sequence corresponding to the outlier nodes is retrieved and matched with the potential activation mode template library for secondary matching. Outlier nodes whose migration mode matching degree of secondary matching is lower than the preset matching threshold are marked as propagation abnormal nodes and stored in the abnormal review queue.

[0127] In this embodiment, after the social propagation influence model completes multiple rounds of propagation iterations and reaches a preset propagation convergence state, the server performs anomaly verification on the high-potential new customer group labeling results obtained based on the propagation results. The server first reads the cumulative reception weights of all social nodes in the user's social association topology graph at the propagation convergence time, and establishes a correspondence between nodes and cumulative reception weights according to each user's anonymization identifier, forming cumulative reception weight distribution data. For example, the nodes corresponding to the aforementioned user anonymization identifiers U1027, U1035, and U1062 all form corresponding cumulative reception weights after multiple rounds of propagation. Simultaneously, the server obtains the connection relationships between each node and its adjacent social nodes, so that the cumulative reception weight distribution data not only reflects the final accumulated propagation influence of each node, but can also be used to compare the weight differences between different nodes within a local social association range.

[0128] The server then performs outlier detection processing on the accumulated received weight distribution data. Specifically, based on the social relationships in the user's social relationship topology graph, for each user node marked as a high-potential new customer group, the server extracts the neighboring social nodes directly connected to that node and their corresponding accumulated received weights, and compares the deviation between the current node's accumulated received weight and the accumulated received weights of its neighboring social nodes. For example, the node corresponding to user anonymization identifier U1096 forms a high accumulated received weight after propagation convergence and exceeds a preset propagation threshold. Therefore, it is marked as a high-potential new customer group in the aforementioned processing. However, further detection by the server reveals that the accumulated received weights of multiple neighboring social nodes directly connected to U1096 are all at a relatively concentrated low level, and the deviation between the accumulated received weight of the node corresponding to U1096 and the accumulated received weights of its neighboring social nodes exceeds a preset deviation threshold. Based on this, the server identifies the node corresponding to U1096 as an outlier node. The server uses the same method to traverse other nodes corresponding to high-potential new customer group users to complete the outlier node identification.

[0129] For example, for the user node i to be detected, the server calculates the local reference reception weight based on the neighboring user nodes that have a direct social relationship with the user node i to be detected, expressed as: , where, This represents the local reference receiving weight corresponding to user node i; This represents the set of adjacent user nodes corresponding to user node i. This represents the propagation coefficient corresponding to the social association edge between adjacent user node j and user node i; Let represent the cumulative reception weight of neighboring user node j in the propagation convergence state. The server further calculates the outlier bias value of user node i: , where This represents the outlier deviation value corresponding to user node i; This represents the cumulative received weight of user node i in the propagation convergence state. When When the deviation exceeds the preset threshold, the server will identify user node i as an outlier node.

[0130] For outlier nodes detected, the server first removes the corresponding user from the current high-potential new customer group labeling results, ensuring that users who accumulate high weights solely through anomalous propagation are not directly retained in the final high-potential new customer group results. Subsequently, the server uses the anonymized user identifier corresponding to the outlier node to trace back cross-channel behavioral trajectory data and scenario-related behavioral chain networks, retrieving the original behavioral transition feature vector sequence formed by the user before the potential tendency weight assignment. Taking U1096 as an example, the server re-extracts its original behavioral transition feature vector sequence formed by the encoding of behavioral sequence nodes from various business interaction channels, allowing the secondary verification to be directly based on the user's actual cross-scenario behavioral migration features, rather than using the accumulated reception weights formed from other user nodes during social propagation as the judgment basis.

[0131] The server performs a secondary matching process between the original behavioral transition feature vector sequence corresponding to U1096 and each potential activation pattern template in the potential activation pattern template library. Following the order of the multi-dimensional feature vectors in the original behavioral transition feature vector sequence, the server parses the corresponding channel type identifier and segment duration, compares them stage-by-stage with each template behavioral stage in the potential activation pattern templates, and calculates the migration pattern matching degree corresponding to the secondary matching based on the chronological order of the template behavioral stages. The server then determines whether the user's cross-scenario behavioral migration process corresponding to this outlier node actually possesses the potential to evolve from initial activation behavior to deep conversion behavior.

[0132] When the secondary matching result of the original behavioral transition feature vector sequence corresponding to U1096 and the potential activation mode template shows that its migration mode matching degree is lower than the preset matching threshold, the server determines that the user's own behavioral characteristics are insufficient to support its high-potential attributes, and the aforementioned high cumulative reception weight mainly manifests as an abnormal propagation result inconsistent with the local social network weight distribution. Based on this, the server marks the node corresponding to U1096 as a propagation anomaly node, and stores the corresponding user's anonymization identifier, cumulative reception weight, deviation results of the cumulative reception weight of adjacent social nodes, and the migration mode matching degree obtained from the secondary matching in an anomaly review queue. Through the above processing, after the social propagation influence model completes the high-potential new customer group mining, the server can use the local cumulative reception weight distribution to perform reverse verification of abnormally high-weight nodes, and combine this with the user's own original cross-scenario behavioral migration characteristics for secondary confirmation, thereby reducing the interference of abnormal propagation accumulation on the high-potential new customer group identification results and improving the reliability of the final high-potential new customer group labeling results.

[0133] In this embodiment of the invention, the method further includes:

[0134] The potential weighted feature vector sequence corresponding to users marked as high-potential new customer groups is subjected to sequence pattern reverse deconstruction. Following the order of the behavior trigger timestamps corresponding to each vector in the potential weighted feature vector sequence, the template behavior stage in the potential activation mode template aligned to each vector is parsed one by one. The template channel type identifier and template behavior type code corresponding to each template behavior stage are extracted to generate a high-potential behavior stage parsing sequence.

[0135] The stage transition relationships between adjacent template behavior stages in the high-potential behavior stage parsing sequence are statistically analyzed. The frequency of each stage transition relationship in all high-potential new customer groups is counted. Stage transition relationships with a frequency exceeding a preset transition frequency threshold are marked as high-potential transition key edges. The high-potential transition key edges include the template channel type identifier of the preceding template behavior stage and the template channel type identifier of the following template behavior stage.

[0136] All high-potential transition critical edges are grouped according to the template channel type identifier of their preceding template behavior stage. Within each group, the high-potential transition critical edges are arranged from largest to smallest frequency of occurrence. The top few high-potential transition critical edges are extracted to form the set of subsequent transition guidance paths for that preceding template behavior stage.

[0137] The template channel type identifier and the corresponding set of subsequent transition guidance paths for each preceding template behavior stage are stored in the inter-channel behavior guidance rule library. The inter-channel behavior guidance rule library is used to push guidance content to the business interaction channel corresponding to the subsequent template behavior stage pointed to in the set of subsequent transition guidance paths after the current user completes the behavior corresponding to the preceding template behavior stage.

[0138] Extract the template channel type identifiers for the preceding and subsequent template behavior stages corresponding to each guidance rule from the inter-channel behavior guidance rule base, generate a cross-channel guidance trigger condition mapping table, and deploy the cross-channel guidance trigger condition mapping table to the real-time behavior monitoring module of each business interaction channel. When the system detects that a user has completed the behavior corresponding to the preceding template behavior stage, it triggers the display of the guidance entry of the corresponding business interaction channel for the subsequent template behavior stage.

[0139] In this embodiment, after marking high-potential new customer groups, the server further performs sequence pattern reverse deconstruction on the potential weighted feature vector sequence already formed by high-potential users to extract the key stage transition relationships that repeatedly occur during the migration of high-potential users from one business interaction channel to subsequent business interaction channels. Taking the user anonymization identifier U1027 as an example, the server reads the potential weighted feature vector sequence corresponding to U1027 and arranges it from earliest to latest according to the behavior trigger timestamps corresponding to each vector. The server reads the potential activation mode template and corresponding template behavior stage that each vector is aligned with in the aforementioned sequence pattern matching process. For example, the first vector of U1027 aligns with the initial attention stage of the online page, the second vector aligns with the further query stage of offline devices, and the third vector aligns with the in-depth understanding stage of the self-service terminal. The server extracts the template channel type identifier and template behavior type code corresponding to the above template behavior stages respectively, and arranges them in order of behavior trigger timestamps to form the high-potential behavior stage parsing sequence corresponding to U1027. The server completes the sequence reverse deconstruction of other high-potential new customer groups such as U1035 and U1062 in the same way.

[0140] Subsequently, the server performs stage transition relationship statistics on adjacent template behavior stages in the parsed sequence of all high-potential behavior stages. For example, among multiple high-potential new customer groups, the continuous relationship of migrating from the initial online page attention stage to the offline device further inquiry stage, and from the offline device further inquiry stage to the self-service terminal in-depth understanding stage, is repeatedly observed. The server counts the frequency of the above stage transition relationship among all high-potential new customer groups. When the frequency of a certain stage transition relationship exceeds a preset transition frequency threshold, the server marks the stage transition relationship as a high-potential transition key edge and records the template channel type identifier corresponding to the preceding template behavior stage and the template channel type identifier corresponding to the following template behavior stage, thereby obtaining key connection relationships that can reflect the common cross-channel migration directions of high-potential users.

[0141] The server further groups high-potential transition key edges according to the template channel type identifier of the preceding template behavior stage. For example, high-potential transition key edges using the online page channel type identifier as the preceding identifier are grouped into the same group, which includes key edges pointing from the online page channel to the offline device channel and from the online page channel to the self-service channel. The server ranks each high-potential transition key edge from highest to lowest frequency of occurrence in the high-potential new customer group and extracts the top-ranked high-potential transition key edges to form the set of subsequent transition guidance paths for the corresponding preceding template behavior stage. For the preceding template behavior stage corresponding to the offline device channel, the server uses the same method to obtain its corresponding set of subsequent transition guidance paths.

[0142] The server associates and stores the template channel type identifier for each preceding template behavior stage with the corresponding set of subsequent transition guidance paths, forming an inter-channel behavior guidance rule base. For example, if the set of subsequent transition guidance paths corresponding to the initial attention stage of the online page includes a high-frequency transition path pointing to the further query stage on offline devices, the server establishes a corresponding rule between the online page channel type identifier and this subsequent transition relationship. After the user completes the behavior corresponding to the initial attention stage of the online page, the server can determine the business interaction channel corresponding to the subsequent template behavior stage based on the inter-channel behavior guidance rule base, and configure guidance content to the corresponding business interaction channel to guide the user to continue migrating, so that the key channel migration relationships that repeatedly appear among historical high-potential users can be used for cross-channel behavior guidance of subsequent users.

[0143] Finally, the server extracts the template channel type identifiers for the preceding template behavior stage and the corresponding template channel type identifiers for the following template behavior stage from the inter-channel behavior guidance rule base, establishes the correspondence between the two, and generates a cross-channel guidance trigger condition mapping table. The server deploys the cross-channel guidance trigger condition mapping table to the real-time behavior monitoring modules corresponding to online page interaction terminals, offline device interaction terminals, and self-service interaction terminals. The real-time behavior monitoring module continuously identifies the template behavior stage corresponding to the user's current completed behavior. When it detects that a user has completed the behavior corresponding to the preceding template behavior stage, the server queries the corresponding following template behavior stage and its business interaction channel according to the cross-channel guidance trigger condition mapping table, and triggers the display of the guidance entry point of the corresponding business interaction channel. For example, after detecting that a user has completed the behavior corresponding to the initial attention stage of the online page, when the mapping result points to the further query stage of the offline device, the server triggers the display of the guidance entry point of the corresponding offline device business interaction channel. In this way, the server can transform the key cross-channel behavior transition patterns that repeatedly appear in the identified high-potential new customer groups into real-time executable inter-channel behavior guidance rules, realizing the connection from the high-potential behavior pattern analysis results to the cross-channel behavior guidance application.

[0144] Please refer to the following: Figure 3 , Figure 3 The diagram illustrates the structure of high-value propagation path extraction processing provided in this embodiment of the invention. In this embodiment, the method further includes:

[0145] The propagation iteration snapshot data generated during the propagation iteration process of the social propagation influence model is stored. The propagation iteration snapshot data includes the cumulative receiving weight and propagation source strength of each node in the user social association topology graph after the end of this round of propagation iteration.

[0146] Extract the cumulative reception weight growth curve of each high-potential new customer group user node in each round of propagation iteration from the propagation iteration snapshot data, perform growth pattern analysis on the cumulative reception weight growth curve, divide the cumulative reception weight growth curve into an initial slow growth stage, a rapid growth stage and a saturation convergence stage, and extract the starting iteration round and the ending iteration round of the rapid growth stage.

[0147] For each high-potential new customer group user node, trace back the propagation contribution records it received from adjacent social nodes within the time range of the rapid growth phase. The propagation contribution records include the identifier of the contribution source node, the amount of transmission in each propagation iteration, and the propagation source strength of the contribution source node. Arrange the contribution source nodes from largest to smallest according to the total amount of transmission, and extract the contribution source nodes whose total amount of transmission exceeds the preset contribution threshold as the key influence source nodes of the high-potential new customer group user.

[0148] The social connection edges between each high-potential new customer group user node and its key influence source nodes are marked as high-value propagation edges. All high-value propagation edges are extracted from the user social connection topology graph to form a high-value propagation subgraph. The high-value propagation subgraph reflects the key path structure of influence propagation among high-potential new customer groups.

[0149] The high-value propagation subgraph is subjected to propagation path compression processing. Multiple propagation paths with the same starting user node and the same ending user node are merged into a propagation backbone path. A path propagation flow parameter is added to each propagation backbone path. The path propagation flow parameter is the cumulative value of the transmission amount of each edge on the propagation backbone path in the propagation iteration. The propagation backbone path and the path propagation flow parameter are stored as a high-value propagation path set.

[0150] In this embodiment, during the multi-round propagation iteration of the social propagation influence model, the server continuously records the propagation state formed at the end of each propagation iteration. Specifically, after each round of propagation, the server reads the current cumulative receiving weight and updated propagation source strength of each social node in the user's social association topology graph, and establishes corresponding propagation iteration snapshot data according to the propagation iteration round. For example, when performing propagation iterations for users with anonymized identifiers such as U1027, U1035, and U1062, the server records the cumulative receiving weight and propagation source strength of each node in the first round, the second round, and until the preset propagation convergence state is reached, so that the weight changes of the same user node throughout the propagation process can be continuously restored according to the iteration rounds.

[0151] After identifying high-potential new customer groups, the server extracts the cumulative reception weights of each high-potential new customer group user node from the propagation iteration snapshot data at different propagation iteration rounds. For example, the server reads the cumulative reception weights of the node corresponding to U1035 from round 1 to the propagation convergence round and forms a cumulative reception weight growth curve according to the iteration round order. The server further analyzes the growth changes of the cumulative reception weights between adjacent iteration rounds, determining the interval with small weight increments in the early stage as the initial slow growth stage, the interval with consecutive large weight increments in the middle as the rapid growth stage, and the interval where the weight growth gradually slows down and approaches stability in the later stage as the saturation convergence stage. For example, U1035's cumulative reception weights increase slowly from round 1 to round 3, the growth rate increases significantly from round 4 to round 7, and gradually stabilizes after round 8. Based on this, the server extracts round 4 as the starting iteration round of the rapid growth stage and round 7 as the ending iteration round.

[0152] For example, for a high-potential new customer user node i, the server calculates the weight increase corresponding to round t based on the cumulative difference in received weights between adjacent propagation iterations, expressed as: , where, This represents the weight increase of user node i in the t-th round of propagation iteration; This represents the cumulative received weight after the t-th round of propagation iterations; This represents the cumulative reception weight after the (t-1)th round of propagation iterations.

[0153] The server is set with a preset rapid growth threshold. and preset saturation growth threshold ,in Greater than When the weight increase for a preset number of consecutive propagation iterations is greater than or equal to a preset rapid growth threshold. When the interval corresponding to the consecutive propagation iterations is defined as the rapid growth phase; when the weight increase corresponding to the subsequent preset number of consecutive propagation iterations is less than or equal to the preset saturation growth threshold. When the time interval is reached, the corresponding interval is defined as the saturated convergence phase; the propagation iteration interval before the rapid growth phase is defined as the initial slow growth phase.

[0154] The server determines the first propagation iteration round that continuously satisfies the rapid growth condition as the starting iteration round of the rapid growth phase, and determines the last propagation iteration round before entering the saturation convergence phase as the ending iteration round of the rapid growth phase.

[0155] Subsequently, for U1035 during the rapid growth phase corresponding to rounds 4 to 7, the server retrospectively traced the propagation contribution records received by the node from each adjacent social node. The server reads the identifier of the contributing source node, the corresponding transmission volume, and the propagation source strength of the contributing source node in that round, round by round. For example, U1027, U1062, and other adjacent nodes all transmit propagation weights to U1035 during this phase. The server sums the transmission volumes of each contributing source node in rounds 4 to 7 and sorts them from largest to smallest based on the total transmission volume. When U1027 transmits the highest cumulative weight to U1035 during this phase and exceeds a preset contribution threshold, the server identifies U1027 as a key influence source node for U1035; other contributing source nodes whose total transmission volume exceeds the preset contribution threshold are also identified as key influence source nodes.

[0156] For example, the rapid growth phase corresponding to user node v of a high-potential new customer group. The cumulative propagation contribution of the source node u to the high-potential new customer group user node v is expressed as: , where, This indicates the cumulative propagation contribution of the contributing node u to the high-potential new customer group node v during the rapid growth phase. This represents the amount of data transmitted from the contributing source node u to the high-potential new customer group user node v during the t-th round of propagation iteration.

[0157] when When the contribution threshold is exceeded, the contribution source node u is identified as the key influence source node of the high-potential new customer group user node v, and the social connection edge between user node u and user node v is marked as a high-value dissemination edge.

[0158] For multiple propagation paths that share the same starting user node 'a' and the same ending user node 'b', the server merges the high-value propagation edges contained in each propagation path to form a set of high-value propagation edges for the corresponding main propagation path. The server retains the same high-value propagation edge that appears repeatedly in multiple propagation paths only once, and calculates the path propagation traffic parameters based on the cumulative transmission volume of each high-value propagation edge during the propagation iteration process, expressed as: ,in, This represents the path propagation traffic parameters for the main propagation path from the starting user node a to the ending user node b. This represents the set of high-value propagation edges after deduplication within the corresponding main propagation path; Let e ​​represent the amount of high-value propagation edge e transmitted in the tth round of propagation iteration; T represents the total number of propagation iterations when the preset propagation convergence state is reached.

[0159] By deduplicating duplicate high-value propagation edges before accumulating the propagation volume, the same propagation volume is avoided from being repeatedly accumulated when multiple propagation paths with the same starting user node and the same ending user node share the same high-value propagation edge.

[0160] The server further marks the social connection edges between key influence source nodes such as U1035 and U1027 as high-value propagation edges, and processes all high-potential new customer user nodes in the same way, extracting all high-value propagation edges and corresponding nodes from the user social connection topology graph to form a high-value propagation subgraph. For example, the high-value propagation subgraph forms multiple paths with high propagation contributions, such as U1027 propagating to U1035, U1035 propagating to U1084, and U1027 propagating to U1084 via U1062, thereby reflecting the main connection relationships and key propagation structures of potential influence propagation among different high-potential users.

[0161] Finally, the server performs propagation path compression processing on the high-value propagation subgraph. For multiple propagation paths with the same starting user node and the same ending user node, but passing through different nodes in between, the server merges them according to the common starting and ending nodes. For example, in the propagation path starting from U1027 and finally reaching U1084, there are paths where U1027 propagates directly to U1084, as well as paths where U1027 reaches U1084 via U1035 and U1027 reaches U1084 via U1062. The server compresses the above propagation paths with the same starting user node U1027 and ending user node U1084 into the corresponding propagation backbone path, and accumulates the transmission volume of each high-value propagation edge on the relevant path during the propagation iteration process to form the path propagation flow parameter corresponding to the propagation backbone path. The server compresses all high-value propagation paths in the same way, and stores the resulting main propagation path and corresponding path propagation traffic parameters as a set of high-value propagation paths, thereby forming structured propagation path data that can centrally reflect the main sources, key propagation directions and propagation intensity of the influence of high-potential new customer groups.

[0162] In this embodiment of the invention, the method further includes:

[0163] The behavioral transition feature vector sequence corresponding to users marked as high-potential new customer groups is sorted with the degree of matching between the transition patterns of each potential activation pattern template in the potential activation pattern template library. The potential activation pattern template with the highest matching degree for each high-potential new customer group user is extracted as the main matching template for that user, and the potential activation pattern template with the second highest matching degree is extracted as the secondary matching template for that user.

[0164] For each high-potential new customer group, analyze the differences between the main matching template and the secondary matching template in the template behavior stage sequence, locate the divergence stage position where the template behavior stage between the main matching template and the secondary matching template diverges, and extract the template behavior stage sequence of the main matching template after the divergence stage position and the template behavior stage sequence of the secondary matching template after the divergence stage position, respectively, into the main path subsequent stage sequence and the secondary path subsequent stage sequence.

[0165] Cluster all high-potential new customer group users according to the template channel type identifier corresponding to the divergence stage position. Perform frequent subsequence mining on the main path subsequent stage sequence and the secondary path subsequent stage sequence in each cluster. Extract the subsequent stage subsequence that appears more frequently than the preset frequency threshold in each cluster as the potential behavioral branch path corresponding to the divergence stage position.

[0166] For each divergence stage location, construct a behavior branch path decision point. The behavior branch path decision point is labeled with the template channel type identifier corresponding to the divergence stage location, the current template behavior stage, and the set of possible potential behavior branch paths.

[0167] All behavioral branch path decision points are combined into a behavioral branch path decision network. The behavioral branch path decision network is associated and stored with a potential activation pattern template library. During the potential activation pattern template matching process, when the user behavior transition feature vector sequence reaches the behavioral branch path decision point, the potential tendency weights corresponding to each path in the potential behavioral branch path set are predicted and evaluated, and the branch path recommendation ranking is output.

[0168] In this embodiment, after identifying high-potential new customer groups, the server further analyzes the different potential behavioral evolution directions that the same high-potential user may correspond to. The server first reads the behavioral transition feature vector sequence corresponding to each high-potential new customer group user, and the calculated migration pattern matching degree between the behavioral transition feature vector sequence and each potential activation pattern template in the potential activation pattern template library, and sorts them from high to low according to the migration pattern matching degree. Taking user anonymization identifier U1027 as an example, its behavioral transition feature vector sequence has different degrees of matching relationships with multiple potential activation pattern templates. The template with the highest matching degree records the cross-scenario migration pattern of "initial online page attention—offline device business inquiry—in-depth understanding of self-service terminal," and the server determines this template as the primary matching template for U1027. The template with the second highest matching degree records the migration pattern of "initial online page attention—offline device business inquiry—further online page interaction," and the server determines this template as the secondary matching template for U1027.

[0169] Subsequently, the server performs template behavior stage sequence difference analysis on the primary and secondary matching templates corresponding to U1027. The server compares the two templates stage by stage according to the order of the template behavior stages, identifying the common preceding stages and the first point of difference. Both templates have gone through the initial online page attention stage and the offline device business query stage. After that, the primary matching template enters the self-service terminal in-depth understanding stage, while the secondary matching template enters the online page further interaction stage. Therefore, the server determines the point where the path diverges after the offline device business query stage as the divergence stage position. The server extracts the template behavior stages of the primary matching template after this divergence stage position, such as the self-service terminal in-depth understanding stage, to form the main path subsequent stage sequence, and extracts the template behavior stages of the secondary matching template after this divergence stage position, such as the online page further interaction stage, to form the secondary path subsequent stage sequence.

[0170] The server processes other high-potential new customer groups such as U1035 and U1062 in the same way, and clusters them according to the template channel type identifier corresponding to the divergence stage position of each user. For example, if multiple users all diverge in subsequent behavior paths after the template behavior stage corresponding to offline device channels, the server will group these users into the same divergence stage cluster. In each cluster, the server collects the subsequent stage sequences of the main path and the subsequent stage sequences of the secondary path for each user, and performs frequent subsequence mining on the recurring template behavior stage combinations. For example, in the cluster with offline device channels as the divergence position, some high-potential users frequently continue to migrate along the direction of "offline device business inquiry - in-depth understanding of self-service terminals", while other high-potential users frequently continue to migrate along the direction of "offline device business inquiry - further interaction on online pages". The server counts the frequency of occurrence of the above subsequent stage subsequences, and extracts the subsequent stage subsequences whose frequency exceeds a preset frequency threshold as the potential behavior branch path corresponding to that divergence stage position.

[0171] After extracting potential behavioral branch paths, the server constructs behavioral branch path decision points for each divergence stage. For the aforementioned divergence stage position after offline device business query, the server marks the template channel type identifier corresponding to the offline device, the current offline device business query template behavior stage, and associates it with a set of potential behavioral branch paths composed of the self-service terminal's in-depth understanding direction and the online page's further interaction direction. The server constructs behavioral branch path decision points corresponding to other divergence stage positions in the same way, so that each decision point can represent multiple high-frequency behavioral migration directions that a high-potential user may continue to develop after reaching a specific template behavior stage.

[0172] Finally, the server combines all behavioral branch path decision points according to the migration relationships between corresponding template behavior stages to form a behavioral branch path decision network, and stores this network in association with the potential activation pattern template library. During subsequent potential activation pattern template matching, when the server detects that a user's behavioral transition feature vector sequence has matched a behavioral branch path decision point corresponding to the offline device business query stage, the server reads the set of potential behavioral branch paths associated with that decision point and predicts and evaluates the potential tendency weights corresponding to the user's continued migration along different potential behavioral branch paths. For example, the server evaluates the potential tendency weights corresponding to the user's migration towards a self-service terminal for deeper understanding and the migration towards returning to an online page for further interaction, and arranges each potential behavioral branch path from high to low according to the predicted potential tendency weights, outputting the corresponding branch path recommendation ranking. Thus, the server can further identify multiple subsequent evolution directions that high-potential users may generate in key behavioral stages based on traditional single-path template matching, and use the actual branching patterns formed by historical high-potential new customer groups to rank and evaluate the potential tendencies of different subsequent paths, thereby more completely representing the branching and multi-path evolution characteristics of high-potential users in the cross-scenario behavioral migration process.

[0173] In this embodiment of the invention, the method further includes:

[0174] The final cumulative reception weights of each node in the user social association topology graph in the social propagation influence model when reaching the preset propagation convergence state are globally sorted. Based on the global sorting results, the nodes in the user social association topology graph are divided into multiple propagation levels, and each propagation level corresponds to a range of cumulative reception weight values.

[0175] Extract the user anonymization identifiers corresponding to user nodes within each propagation level, obtain continuous behavior sequence fragments of user nodes within each propagation level in the cross-channel behavior trajectory dataset, calculate the average scene jump interval duration, average cross-channel behavior frequency, and average number of channel type identifiers of user nodes within each propagation level, and generate a hierarchical behavior profile description for each propagation level.

[0176] By performing a comparative analysis of the hierarchical behavior profiles of adjacent propagation levels, the direction and magnitude of change of higher levels relative to lower levels in terms of average scene jump interval duration, average cross-channel behavior frequency, and average number of channel type identifiers are extracted, and a gradient vector of behavior transition between levels is generated.

[0177] Behavioral indicators whose change magnitude exceeds a preset gradient threshold in the gradient vector of inter-level behavioral transitions are marked as key behavioral indicators of inter-level transitions. The key behavioral indicators of inter-level transitions and their corresponding change directions are combined into a set of key conditions for inter-level transitions. The set of key conditions for inter-level transitions is used to characterize the behavioral change conditions that users need to meet to migrate from a low propagation level to a high propagation level.

[0178] In this embodiment, after the social propagation influence model completes multiple rounds of propagation iterations and reaches a preset propagation convergence state, the server reads the final cumulative reception weights of all user nodes in the user's social association topology graph and performs a global sorting according to the final cumulative reception weights from high to low. Based on the global sorting results and the numerical ranges corresponding to different cumulative reception weights, the server divides the user nodes in the user's social association topology graph into multiple propagation levels. For example, user nodes such as U1027 and U1035, whose final cumulative reception weights are in a higher numerical range, are assigned to higher propagation levels; user nodes whose final cumulative reception weights are in a middle numerical range are assigned to middle propagation levels; and user nodes whose final cumulative reception weights are in a lower numerical range are assigned to lower propagation levels. This forms multiple propagation levels that can reflect the differences in the potential cumulative influence of different users in the social propagation influence model.

[0179] The server further extracts user anonymization identifiers corresponding to user nodes within each propagation level, and obtains continuous behavior sequence fragments of the corresponding users from the cross-channel behavior trajectory dataset based on these user anonymization identifiers. For each propagation level, the server counts the scene switching interval between adjacent behavior sequence nodes during the cross-scene behavior migration of each user, and calculates the average scene jump interval time corresponding to that propagation level; simultaneously, it counts the frequency of continuous behavior sequence fragments generated by each user in different business interaction channels, and calculates the average cross-channel behavior frequency; further, it counts the number of channel type identifier types actually involved by each user, and calculates the average number of channel type identifier types. For example, users in higher propagation levels typically form continuous migration behaviors across multiple channels such as online pages, offline devices, and self-service terminals. Based on this, the server obtains the average scene jump interval time, average cross-channel behavior frequency, and average number of channel type identifier types corresponding to that propagation level, and combines the above statistical results to form a hierarchical behavior profile description for that propagation level. The server generates hierarchical behavior profile descriptions for other propagation levels in the same way.

[0180] Subsequently, the server compares the hierarchical behavioral profiles of adjacent propagation levels according to their ascending propagation hierarchy. The server calculates changes in average scenario jump interval duration, average cross-channel behavior frequency, and average number of channel type identifiers for higher propagation levels relative to their adjacent lower propagation levels. For example, when comparing from lower to higher propagation levels, the server detects a change in average scenario jump interval duration from longer to shorter, indicating tighter behavioral connections between different business interaction channels for users at higher propagation levels; a significant increase in average cross-channel behavior frequency, indicating more frequent cross-channel interactions for users at higher propagation levels; and a simultaneous increase in the average number of channel type identifiers, indicating a richer variety of business interaction channels for users at higher propagation levels. The server records the direction and magnitude of these behavioral indicators and combines the results to generate an inter-level behavioral transition gradient vector between adjacent propagation levels.

[0181] Finally, the server judges the magnitude of changes in behavioral indicators in the gradient vector of behavioral transitions between each level, and marks behavioral indicators whose changes exceed preset gradient thresholds as key behavioral indicators for level transitions. For example, if the server determines that when migrating from a lower propagation level to an adjacent higher propagation level, the increase in average cross-channel behavior frequency and the increase in average number of channel type identifiers both exceed preset gradient thresholds, and the decrease in average scene jump interval time also exceeds a preset gradient threshold, then these three behavioral indicators are identified as key behavioral indicators for level transitions between the corresponding levels, and the directions of change for the increase in average cross-channel behavior frequency, the increase in average number of channel type identifiers, and the decrease in average scene jump interval time are recorded respectively. The server combines these key behavioral indicators for level transitions and their corresponding directions of change to form a set of key conditions for level transitions. Thus, the server can extract key behavioral conditions that show significant changes during the migration of users from lower to higher propagation levels based on the actual behavioral differences already formed by users at different propagation levels, thereby representing the cross-scene behavioral transition patterns between different potential propagation levels in a structured manner.

[0182] In this embodiment of the invention, the method further includes:

[0183] Path length analysis is performed on the cross-scenario behavior migration path of users marked as high-potential new customer groups. The number of behavior sequence nodes and the number of scene jump relationship edges contained in the cross-scenario behavior migration path of each high-potential new customer group user are counted. The number of behavior sequence nodes and the number of scene jump relationship edges are combined to describe the migration path complexity of the user.

[0184] Extract the scene switching interval duration parameter of the scene jump relationship edge from the cross-scene behavior migration path of all high-potential new customer groups. Cross-group according to the channel type identifier corresponding to the preceding behavior sequence node and the channel type identifier corresponding to the following behavior sequence node of the scene jump relationship edge. Calculate the distribution characteristics of the scene switching interval duration parameter in each cross group. Extract the distribution concentration interval of the scene switching interval duration parameter as the standard switching duration interval corresponding to the cross group.

[0185] Obtain the cross-scenario behavior migration path corresponding to the user who is not currently marked as a high-potential new customer group, calculate the migration path complexity description of the user, compare the deviation of the user's migration path complexity description with the median value of the migration path complexity description of the high-potential new customer group users, and extract the unmarked users whose deviation is lower than the preset deviation threshold and whose proportion of the scene switching interval duration parameter of each scene jump relationship edge in the user's cross-scenario behavior migration path falling into the corresponding cross-grouping standard switching duration interval exceeds the preset proportion threshold as potential missed candidate users;

[0186] The behavioral transition feature vector sequence of the potential omission candidate users is recalculated to match the migration pattern of each potential activation pattern template in the potential activation pattern template library. Potential omission candidate users whose migration pattern matching degree exceeds the preset matching threshold are additionally marked as high-potential new customers.

[0187] For example, the server counts the number of behavioral sequence nodes and the number of scene jump edges corresponding to all high-potential new customer groups, and calculates the median value of the number of behavioral sequence nodes and the median value of the number of scene jump edges. The two together constitute the baseline migration path complexity description corresponding to the high-potential new customer groups.

[0188] The median value of the number of nodes in the behavior sequence is represented as The median value of the number of scene jump relationship edges is represented as For the currently unlabeled user u, the number of nodes in its behavior sequence is represented as... The number of scene jump relationship edges is represented as The server then calculates the complexity deviation between the unlabeled user u and the baseline migration path complexity description: ,in, This represents the complexity deviation value corresponding to the unlabeled user u; The deviation weight corresponding to the number of nodes in the behavior sequence; The deviation weight represents the number of edges representing jump relationships in the scene, and satisfies: .when When the error is below a preset deviation threshold, the migration path complexity of the unmarked user u is determined to be similar to the baseline migration path complexity corresponding to the high-potential new customer group.

[0189] By calculating the median values ​​of the number of nodes in the behavior sequence and the number of edges in the scene jump relationship, and forming a complexity deviation value based on the relative deviation of the two dimensions, we avoid directly comparing a single undefined median value for a migration path complexity description that includes multiple dimensions.

[0190] In this embodiment, after marking high-potential new customer groups, the server performs path length analysis on the cross-scenario behavior migration paths of the marked users. The server reads the behavior sequence nodes and scenario jump edges corresponding to each high-potential new customer group user, and counts the number of behavior sequence nodes and scenario jump edges contained in each cross-scenario behavior migration path. Taking user anonymization identifier U1027 as an example, its cross-scenario behavior migration path sequentially includes behavior sequence nodes corresponding to online page content browsing, offline device business query, and self-service terminal business condition viewing. It also includes two scenario jump edges corresponding to online page to offline device and offline device to self-service terminal. Based on this, the server forms a migration path complexity description for U1027. The server completes the migration path complexity statistics for all high-potential new customer group users in the same way.

[0191] Subsequently, the server extracts the scene switching interval duration parameters carried by each scene jump relationship edge from the cross-scenario behavioral migration paths of all high-potential new customer groups, and cross-groups them according to the channel type identifiers corresponding to the preceding and following behavior sequence nodes. For example, all scene jump relationship edges migrating from online page channels to offline device channels are grouped into one cross group, and scene jump relationship edges migrating from offline device channels to self-service channels are grouped into another cross group. The server statistically analyzes the distribution of scene switching interval duration parameters within each cross group, identifies the duration range in which most high-potential users are concentrated, and determines this concentrated distribution interval as the standard switching duration interval for the corresponding channel migration combination. Thus, the server obtains relatively typical time rhythm characteristics of high-potential users migrating between different channels.

[0192] For example, for a preceding channel type identifier of The subsequent channel type identifier is The server arranges all scene switching interval duration parameters in the cross-group in ascending order, and then sorts them according to a preset lower quantile ratio. and preset upper quantile ratio Determine the lower quantile duration respectively and upper percentile duration ,in: The server determines the interval between the lower quantile duration and the upper quantile duration as the standard handover duration interval corresponding to the cross-group, expressed as: ,in, Indicates the preceding channel type Migration to subsequent channel types The corresponding standard switching time range.

[0193] For the currently unlabeled user u, the server counts the number of all scene jump relationship edges in its cross-scene behavior migration path. And count the number of scene jump relationship edges that fall within the corresponding channel cross-grouping standard switching time interval for the corresponding scene switching interval parameter. The standard switching time is calculated in proportion to the following: When the complexity deviation value The deviation is below the preset threshold, and the standard switching time meets the ratio. When the ratio exceeds a preset threshold, the unmarked user u is identified as a potential missing candidate user.

[0194] The server further obtains the cross-scenario behavior migration paths of users who are not yet marked as high-potential new customer groups, and calculates the migration path complexity description for each unmarked user in the same way as for high-potential new customer groups. Taking user anonymization identifier U1118 as an example, the server counts the number of behavior sequence nodes and the number of scene jump relationship edges, and compares the deviation with the median value of the migration path complexity descriptions of all high-potential new customer group users. When the deviation between the migration path complexity description corresponding to U1118 and the median value is lower than a preset deviation threshold, the server further checks the scene jump relationship edges in its cross-scenario behavior migration path one by one. For example, U1118 has a behavior process of migrating from online page channels to offline device channels and from offline device channels to self-service channels. The server determines whether the corresponding scene switching interval duration parameter falls within the standard switching duration range of the aforementioned corresponding cross-grouping, and counts the proportion of scene jump relationship edges that meet the standard switching duration range to the total number of scene jump relationship edges of the user.

[0195] When the migration path complexity deviation of U1118 is lower than a preset deviation threshold, and the proportion of its scene switching interval parameter falling within the corresponding standard switching duration range exceeds a preset proportion threshold, the server extracts U1118 as a potential missed candidate user. At this point, the server further retrieves the behavioral transition feature vector sequence corresponding to U1118 and performs sequence pattern matching again with each potential activation pattern template in the potential activation pattern template library, recalculating the corresponding migration pattern matching degree. The server compares U1118's channel migration order, behavioral occurrence timing, and segment persistence features according to the behavioral stages of each template. When its migration pattern matching degree with at least one potential activation pattern template exceeds a preset matching threshold, the server additionally marks U1118 as a high-potential new customer. Through the above processing, the server can utilize the typical path complexity features and cross-channel switching time patterns formed by the identified high-potential new customer groups to re-screen unmarked users who did not meet the marking conditions in the initial identification process but whose behavioral structures are highly similar to high-potential users, and perform secondary confirmation in conjunction with potential activation pattern templates, thereby reducing the probability of omissions in the high-potential new customer group identification process.

[0196] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned intelligent mining method for high-potential new customer groups based on full-scenario behavioral data analysis. Figure 4 As shown, Figure 4 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0197] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussion is not intended to be exhaustive or to limit the invention to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the invention and its practical application, thereby enabling those skilled in the art to best utilize the invention and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A method for intelligently identifying high-potential new customer groups based on full-scenario behavioral data analysis, characterized in that: The method includes: Obtain cross-channel behavior trajectory dataset; Perform full-scenario behavior pattern deconstruction processing on the cross-channel behavior trajectory dataset to generate a scenario-related behavior chain network. Extract the cross-scenario behavior migration path of each user based on the scenario-related behavior chain network, and encode the cross-scenario behavior migration path into a sequence of behavior transition feature vectors. A preset potential activation pattern template library is retrieved, and the behavioral transition feature vector sequence is matched with each potential activation pattern template in the potential activation pattern template library. The degree of migration pattern matching between the behavioral transition feature vector sequence and each potential activation pattern template is calculated, and potential tendency weights are assigned to each vector in the behavioral transition feature vector sequence according to the degree of migration pattern matching to obtain a potential weighted feature vector sequence. The potential weighted feature vector sequence is input into a preset social propagation influence model. Based on the social relationship between users and the potential tendency weight, multiple rounds of propagation iterations are performed, and the cumulative reception weight of social nodes is updated until a preset propagation convergence state is reached. Then, users corresponding to social nodes whose cumulative reception weight exceeds a preset propagation threshold are extracted as high-potential new customer groups.

2. The method according to claim 1, characterized in that, The cross-channel behavior trajectory dataset includes continuous behavior sequence fragments generated by users in multiple business interaction channels, and each continuous behavior sequence fragment is accompanied by a channel type identifier and a behavior trigger timestamp; The acquisition of the cross-channel behavior trajectory dataset includes: Within the scope of the user authorization agreement, original behavioral event streams are synchronously collected from online page interaction terminals, offline device interaction terminals, and self-service interaction terminals. Each original behavioral event in the original behavioral event stream records a user anonymization identifier, event trigger time, event type code, and interaction channel identifier. The original behavioral event stream is aggregated according to the user anonymization identifier. All original behavioral events associated with the same user anonymization identifier are arranged in chronological order of event triggering time to form an event timeline sequence for each user. Perform event semantic compression processing on the event timeline sequence, and merge the codes of the same type of event with an event trigger time interval of less than a preset interval threshold into behavior segment primitives. The behavior segment primitives encapsulate the start event type code, the end event type code, and the segment duration. Based on the interaction channel identifier carried by the original behavior event within the behavior segment primitive, a channel type identifier is added to the behavior segment primitive to generate a continuous behavior sequence segment with the channel type identifier attached. The continuous behavior sequence fragments with channel type identifiers are associated and stored with the user anonymization identifiers to construct a cross-channel behavior trajectory dataset with the user anonymization identifiers as the main index. The starting event trigger time of each continuous behavior sequence segment in the cross-channel behavior trajectory dataset is used as the behavior trigger timestamp, and the behavior trigger timestamp is appended to the corresponding continuous behavior sequence segment.

3. The method according to claim 1, characterized in that, The process involves performing full-scenario behavior pattern deconstruction on the cross-channel behavior trajectory dataset to generate a scenario-related behavior chain network. Based on this network, the cross-scenario behavior migration path for each user is extracted, and the cross-scenario behavior migration path is encoded into a sequence of behavior transition feature vectors, including: Extract all continuous behavior sequence fragments associated with the same anonymous user from the cross-channel behavior trajectory dataset to form the user's full-channel behavior sequence set; Each continuous behavior sequence segment in the omnichannel behavior sequence set is processed into a behavior sequence node, and each continuous behavior sequence segment is converted into a behavior sequence node. The behavior sequence node encapsulates the channel type identifier, behavior trigger timestamp and segment duration of the continuous behavior sequence segment. Arrange all behavior sequence nodes corresponding to the same user anonymization identifier in chronological order of behavior trigger timestamps. Establish scene jump relationship edges between behavior sequence nodes with adjacent times after arrangement. The scene jump relationship edges record scene switching direction indicators from the preceding behavior sequence node to the following behavior sequence node. For each scene jump relationship edge, calculate the time interval between the behavior trigger timestamp of the preceding behavior sequence node and the behavior trigger timestamp of the following behavior sequence node connected to it, and attach the time interval as the scene switching interval duration parameter to the scene jump relationship edge. The user scenario behavior chain network is formed by combining all behavior sequence nodes and all scenario jump relationship edges corresponding to the same user anonymization identifier. The user scenario behavior chain network represents the user's behavior migration relationship between various business interaction channels. The user scenario behavior chain network corresponding to all user anonymization identifiers is merged across users, and behavior sequence nodes with the same channel identifier sequence pattern are aggregated into shared behavior sequence nodes to generate a scenario-related behavior chain network covering all users. In the scene-related behavior chain network, for each user's anonymized identifier, path backtracking is performed along the user's behavior sequence nodes and scene jump relationship edges to extract the node sequence formed by the user's behavior sequence nodes, which serves as the user's cross-scene behavior migration path. The channel type identifier, behavior trigger timestamp, and segment duration encapsulated in each behavior sequence node in the cross-scene behavior migration path are encoded into a multi-dimensional feature vector, and the multi-dimensional feature vector sequence corresponding to the cross-scene behavior migration path is used as the behavior jump feature vector sequence.

4. The method according to claim 1, characterized in that, The potential activation pattern template records the cross-scenario behavior migration patterns of high-potential users from initial activation behavior to deep conversion behavior. The process involves retrieving a preset potential activation pattern template library, performing sequence pattern matching between the behavioral transition feature vector sequence and each potential activation pattern template in the library, calculating the degree of migration pattern matching between the behavioral transition feature vector sequence and each potential activation pattern template, and assigning potential tendency weights to each vector in the behavioral transition feature vector sequence based on the degree of migration pattern matching to obtain a potential-weighted feature vector sequence, including: Retrieve all potential activation mode templates from the preset potential activation mode template library. Each potential activation mode template contains multiple template behavior stages arranged in chronological order. Each template behavior stage is marked with a template channel type identifier, template behavior type code, and template stage duration range. For each multidimensional feature vector in the behavioral transition feature vector sequence, the channel type identifier and segment duration corresponding to the multidimensional feature vector are parsed. The parsing results are compared stage by stage with the template channel type identifier and template stage duration range of each template behavioral stage in each potential activation mode template to obtain the stage matching score between the multidimensional feature vector and each template behavioral stage. The stage matching scores of each template behavior stage within the same potential activation mode template are accumulated sequentially according to the time order of the template behavior stages to generate the degree of matching between the behavior transition feature vector sequence and the migration mode of the potential activation mode template. For each multidimensional feature vector in the behavior transition feature vector sequence, the template behavior stage aligned with the multidimensional feature vector is extracted from the potential activation mode template where the migration pattern matching degree exceeds a preset matching threshold. The stage matching score of the template behavior stage is used as the basis for assigning the potential tendency weight of the multidimensional feature vector. The potential tendency weights of each multidimensional feature vector are assigned to the corresponding multidimensional feature vectors to generate a potential weighted feature vector sequence.

5. The method according to claim 1, characterized in that, The process involves inputting the potential weighted feature vector sequence into a preset social propagation influence model, performing multiple rounds of propagation iterations based on the social connections between users and the potential tendency weights, and updating the cumulative reception weights of social nodes until a preset propagation convergence state is reached. Then, users corresponding to social nodes whose cumulative reception weights exceed a preset propagation threshold are extracted as high-potential new customer groups, including: Obtain a user social relationship dataset, wherein each social relationship record in the user social relationship dataset includes a first user anonymization identifier, a second user anonymization identifier, and a social interaction frequency parameter; The user social relationship dataset is transformed into a user social association topology graph. Each node in the user social association topology graph corresponds to a user anonymization identifier, each edge corresponds to a social relationship record, and the weight value of the edge is the propagation coefficient mapped by the social interaction frequency parameter. The potential tendency weights of each multidimensional feature vector in the potential weighted feature vector sequence are summed and aggregated to obtain the potential tendency aggregate weight of each user anonymization identifier. The potential tendency aggregate weight is assigned to the node corresponding to the user anonymization identifier in the user social association topology graph as the propagation source strength of that node. In the user social association topology graph, initialize the propagation state and set the initial value of the cumulative receiving weight of all nodes to their own propagation source strength. Multiple rounds of propagation iterations are performed. In each round of propagation iteration, for each edge in the user social association topology graph, a portion of the propagation source strength is transmitted from the propagation source node to the target node. The amount of transmission is determined by the product of the current propagation source strength of the propagation source node and the propagation coefficient of the edge. The amount of transmission is accumulated to the cumulative reception weight of the target node. After each round of propagation iteration, the propagation source strength of each node is updated to the product of the propagation source strength of the previous round and the propagation attenuation coefficient. When the change in the propagation source strength of all nodes is less than the preset attenuation threshold, the preset propagation convergence state is determined to be reached. After reaching the preset propagation convergence state, extract the nodes in the user social association topology graph whose cumulative reception weight exceeds the preset propagation threshold, and mark the users pointed to by the user anonymization identifiers corresponding to these nodes as high-potential new customer groups.

6. The method according to claim 1, characterized in that, The method further includes: The cross-scenario behavior migration path of users marked as high-potential new customer groups is decomposed into path segments, and the behavioral segment primitive sequence of high-potential new customer group users in various business interaction channels is extracted. Frequent sequence pattern mining is performed on the behavioral fragment primitive sequences of all high-potential new customer groups. Subsequences of behavioral fragment primitives that occur more frequently than a preset frequency threshold are extracted as common patterns of high-potential behaviors. The common high-potential behavior patterns are merged with existing potential activation pattern templates in the potential activation pattern template library, and high-potential behavior patterns not included in the existing potential activation pattern templates are added as potential activation pattern templates.

7. The method according to claim 1, characterized in that, The method further includes: Obtain the cumulative reception weight distribution data of each node in the user social association topology graph when the social propagation influence model reaches the preset propagation convergence state; Outlier node detection processing is performed on the cumulative received weight distribution data to identify outlier nodes whose cumulative received weight deviates from the cumulative received weight of neighboring social nodes by more than a preset deviation threshold. The outlier nodes are removed from the high-potential new customer group marking results, and the original behavioral transition feature vector sequence corresponding to the outlier nodes is retrieved and matched with the potential activation mode template library for secondary matching. Outlier nodes whose migration mode matching degree of secondary matching is lower than the preset matching threshold are marked as propagation abnormal nodes and stored in the abnormal review queue.

8. The method according to claim 1, characterized in that, The method further includes: The propagation iteration snapshot data generated during the propagation iteration process of the social propagation influence model is stored. The propagation iteration snapshot data includes the cumulative receiving weight and propagation source strength of each node in the user social association topology graph after the end of this round of propagation iteration. Extract the cumulative reception weight growth curve of each high-potential new customer group user node in each round of propagation iteration from the propagation iteration snapshot data, perform growth pattern analysis on the cumulative reception weight growth curve, divide the cumulative reception weight growth curve into an initial slow growth stage, a rapid growth stage and a saturation convergence stage, and extract the starting iteration round and the ending iteration round of the rapid growth stage. For each high-potential new customer group user node, trace back the propagation contribution records it received from adjacent social nodes within the time range of the rapid growth phase. The propagation contribution records include the identifier of the contribution source node, the amount of transmission in each propagation iteration, and the propagation source strength of the contribution source node. Arrange the contribution source nodes from largest to smallest according to the total amount of transmission, and extract the contribution source nodes whose total amount of transmission exceeds the preset contribution threshold as the key influence source nodes of the high-potential new customer group user. The social connection edges between each high-potential new customer group user node and its key influence source nodes are marked as high-value propagation edges. All high-value propagation edges are extracted from the user social connection topology graph to form a high-value propagation subgraph. The high-value propagation subgraph reflects the key path structure of influence propagation among high-potential new customer groups. The high-value propagation subgraph is subjected to propagation path compression processing. Multiple propagation paths with the same starting user node and the same ending user node are merged into a propagation backbone path. A path propagation flow parameter is added to each propagation backbone path. The path propagation flow parameter is the cumulative value of the transmission amount of each edge on the propagation backbone path in the propagation iteration. The propagation backbone path and the path propagation flow parameter are stored as a high-value propagation path set.

9. The method according to claim 1, characterized in that, The method further includes: Path length analysis is performed on the cross-scenario behavior migration path of users marked as high-potential new customer groups. The number of behavior sequence nodes and the number of scene jump relationship edges contained in the cross-scenario behavior migration path of each high-potential new customer group user are counted. The number of behavior sequence nodes and the number of scene jump relationship edges are combined to describe the migration path complexity of the user. Extract the scene switching interval duration parameter of the scene jump relationship edge from the cross-scene behavior migration path of all high-potential new customer groups. Cross-group according to the channel type identifier corresponding to the preceding behavior sequence node and the channel type identifier corresponding to the following behavior sequence node of the scene jump relationship edge. Calculate the distribution characteristics of the scene switching interval duration parameter in each cross group. Extract the distribution concentration interval of the scene switching interval duration parameter as the standard switching duration interval corresponding to the cross group. Obtain the cross-scenario behavior migration path corresponding to the user who is not currently marked as a high-potential new customer group, calculate the migration path complexity description of the user, compare the deviation of the user's migration path complexity description with the median value of the migration path complexity description of the high-potential new customer group users, and extract the unmarked users whose deviation is lower than the preset deviation threshold and whose proportion of the scene switching interval duration parameter of each scene jump relationship edge in the user's cross-scenario behavior migration path falling into the corresponding cross-grouping standard switching duration interval exceeds the preset proportion threshold as potential missed candidate users; The behavioral transition feature vector sequence of the potential omission candidate users is recalculated to match the migration pattern of each potential activation pattern template in the potential activation pattern template library. Potential omission candidate users whose migration pattern matching degree exceeds the preset matching threshold are additionally marked as high-potential new customers.

10. A high-potential new customer group intelligent mining system based on full-scenario behavioral data analysis, characterized in that, Includes at least one service node; The service node includes a storage unit and a computing unit; the storage unit is used to store program code. The computing unit is used to run the program code to execute the intelligent mining method for high-potential new customer groups based on full-scenario behavioral data analysis, as described in any one of claims 1 to 9.

Images