An advertisement effect evaluation method based on multi-modal behavior fusion

Abstract

The application discloses an advertisement effect evaluation method based on multi-modal behavior fusion, comprising the following steps: collecting multi-modal behavior data, and generating a user behavior sequence in chronological order; mapping the behavior sequence into a field set and inputting the improved AutoInt network, extracting the interaction relationship, constructing a structure diagram and generating a representation; constructing a cross-user projection field based on the behavior interaction representation, calculating the structural similarity and generating a set of potential behavior paths; performing a matching analysis on the potential paths, extracting common behavior migration chains and state nodes, and generating collapse results and intensities; performing item-by-item peeling verification on the conversion consensus body, identifying necessary migration chains and state nodes, and generating a supporting structure; comprehensively evaluating based on the path set, the collapse result and the supporting structure, and outputting the advertisement effect score and related information. Through multi-modal behavior fusion modeling and cross-user behavior projection analysis, the application realizes accurate evaluation of advertisement effect and identification of key conversion paths.

Description

Technical Field

[0001] This invention relates to the field of advertising data analysis technology, and in particular to an advertising effectiveness evaluation method based on multimodal behavior fusion. Background Technology

[0002] With the continuous expansion of internet advertising, advertising effectiveness evaluation has become a crucial aspect of advertising systems. Existing technologies typically evaluate advertising effectiveness based on metrics such as click-through rate (CTR) and conversion rate, and combine this with statistical analysis of user exposure, click, and conversion behaviors. Some methods incorporate machine learning or deep learning models to model historical user behavior to predict the probability of user response to ads, thereby assisting in ad optimization. These methods can achieve some success when the data scale is large and user behavior is complete, but their core remains based on single-user behavior statistics or simple feature interactions.

[0003] In practical applications, user behavior data often suffers from insufficient sample size, incomplete behavioral paths, and significant differences in behavioral patterns. Especially in scenarios involving cold-start users or low-frequency users, a single user's behavior records are insufficient to fully reflect their potential interests and conversion trends. Existing methods often employ simple feature concatenation or shallow interaction approaches to process multimodal behavior data, failing to effectively characterize the complex high-order relationships between exposure, browsing, clicks, interactions, and conversion behaviors, resulting in inaccurate model portrayals of user behavior evolution. Furthermore, existing technologies largely ignore the transferability of behavioral structures between different users, failing to leverage the behavioral paths of other users to complete the potential behavioral evolution trajectory of the target user.

[0004] Existing advertising effectiveness evaluation methods typically analyze based on a single path or independent samples, lacking the ability to characterize the consistency between multiple potential behavioral paths and failing to determine whether advertising has driven user behavior towards a stable conversion trend. For key behavioral stages influencing advertising conversion, existing methods rely heavily on feature importance or weight analysis, lacking structured path verification mechanisms. This makes it difficult to accurately identify the truly decisive behavioral migration chains and state nodes, resulting in insufficient interpretability of evaluation results, unclear optimization directions, and an inability to meet the needs of refined advertising placement and dynamic optimization.

[0005] Therefore, how to provide an advertising effectiveness evaluation method based on multimodal behavior fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an advertising effectiveness evaluation method based on multimodal behavior fusion. This invention introduces an improved AutoInt network to perform high-order interaction modeling of multimodal advertising behavior information and constructs a cross-user behavior projection-recycle field to generate potential behavioral evolution paths of target users. Through path consistency collapse analysis and conversion succession structure identification, it achieves a structured evaluation of advertising conversion trends and key behavioral chains, which has the advantages of high evaluation accuracy, strong behavioral path characterization ability, and strong interpretability.

[0007] An advertising effectiveness evaluation method based on multimodal behavior fusion according to an embodiment of the present invention includes:

[0008] Collect multimodal behavioral data generated during the advertising process, construct a behavioral data set, and form a user behavior sequence according to time order;

[0009] Multimodal advertising behavior information in user behavior sequences is mapped to a set of behavior fields, which are then input into an improved AutoInt network to perform cross-modal high-order interaction calculations. The interaction strength, interaction direction, and interaction level between behavior fields are extracted, a behavior interaction structure graph is constructed, and the corresponding behavior interaction representation is generated.

[0010] Based on behavioral interaction representation, a cross-user behavior projection-recovery field is constructed. The similarity of behavioral interaction structure between the target user and the candidate user is calculated. Candidate user behavior paths that meet the preset similarity conditions are selected. The candidate user behavior paths are projected onto the target user to generate a set of potential behavioral evolution paths for the target user.

[0011] Perform path alignment analysis on the set of potential behavioral evolution paths, extract common behavioral migration chains and common state nodes in different behavioral paths, combine common behavioral migration chains and common state nodes into a transformation consensus body, and generate path consistency collapse results and collapse strength.

[0012] Perform item-by-item stripping verification on the behavior migration chain and state node in the transformation consensus body, identify the behavior migration chain and state node necessary to maintain the path consistency collapse result, and combine the behavior migration chain and state node into the transformation acceptance structure.

[0013] Based on the set of potential behavioral evolution paths, path consistency collapse results, and conversion structure, the advertising effect is comprehensively evaluated, and the advertising effect score, potential behavioral paths, path convergence trends, and key behavioral impact information are output.

[0014] Optionally, the multimodal behavioral data includes ad exposure records, ad click records, ad browsing records, ad interaction records, ad conversion records, ad identification information, user identification information, terminal device information, access time information, access location information, and ad creative information.

[0015] Optionally, the construction of the behavioral data set, forming a user behavior sequence in chronological order, includes:

[0016] Read ad exposure records, ad click records, ad browsing records, ad interaction records, ad conversion records, and related scenario information; match and associate various records based on user identifiers and ad identifiers to generate corresponding behavioral record sets;

[0017] Extract the occurrence time, behavior type, behavior object, and scene attributes corresponding to each behavior record, and convert them according to a unified data format to generate standardized behavior records;

[0018] Standardized behavior records are sorted in ascending order according to the time of occurrence to determine the chronological relationship between each behavior record and generate a behavior timeline.

[0019] According to the order of the behavior time chain, adjacent behavior records are connected in sequence to generate the corresponding user's behavior trajectory;

[0020] The records of each behavior in the behavior trajectory are combined in chronological order to form a user behavior sequence.

[0021] Optionally, the step of constructing the behavior interaction structure diagram and generating the corresponding behavior interaction representation includes:

[0022] An improved AutoInt network is constructed, which consists of an embedding layer, a cross-modal gating interaction layer, a bidirectional causal masking interaction layer, and a graph-level attention pooling layer from bottom to top. The layers are connected through residual connections and layer normalization modules. The cross-modal gating interaction layer inserts a gating weight adjustment sublayer between the parallel multi-head self-attention sublayer and the residual output. The bidirectional causal masking interaction layer includes a forward masking branch and a backward masking branch. The graph-level attention pooling layer adopts a two-level attention mechanism to first aggregate nodes of the same modality and then aggregate nodes across modalities.

[0023] Independent embedding lookup tables are established for the exposure, click, browsing, interaction, conversion, advertising, terminal, and time fields in the user behavior sequence. Discrete fields are mapped to dense vectors, and continuous fields are mapped to vectors of the same dimension after min-max normalization, forming a field vector sequence.

[0024] Insert a behavior type marker vector at the beginning of the field vector sequence and a context scene marker vector at the end. Attach a relative time position code to each field vector to preserve the event time sequence information.

[0025] The field vector sequence containing the label vector and position encoding is input into the cross-modal gated interaction layer. Each cross-modal gated interaction layer consists of a parallel multi-head self-attention sub-layer, a cross-modal gated sub-layer and a residual normalization sub-layer cascaded together. The cross-modal gated sub-layer dynamically adjusts the weight coefficient of the self-attention sub-layer according to the modality to which the field belongs, so as to highlight the cross-modal interaction features.

[0026] The cross-modal interaction vector sequence is input into the bidirectional causal masking interaction layer. The forward masking branch captures the influence of the preceding behavior on the subsequent behavior, and the backward masking branch captures the retrospective influence of the subsequent behavior on the preceding behavior. The bidirectional output is fused through a depth-separable residual connection to generate a high-order behavior interaction representation containing interaction direction information.

[0027] The higher-order behavioral interaction representation is mapped to the attention level representation space. The interaction level is labeled according to the attention level index, and the interaction direction is labeled according to the bidirectional mask. The behavioral interaction structure graph is constructed using field vectors as nodes, interaction strength and direction between nodes, and interaction level as edge attributes.

[0028] A graph-level attention pooling layer is used to perform node aggregation on the behavior interaction structure graph. Nodes of the same modality are aggregated first, and then the aggregation results of different modalities are weighted and merged to output a behavior interaction representation with a unified dimension.

[0029] Optionally, the set of potential behavioral evolution paths for generating the target user includes:

[0030] Read the behavioral interaction representations of the target user, the behavioral interaction representations of the candidate user, and the corresponding behavioral interaction structure diagram. Select the candidate users with consistent ad identifiers, consistent ad creative types, matching scene fields, and existing conversion field records as projection source users to generate a candidate user set.

[0031] Using the behavioral interaction representation of the target user as the projection receiver and the behavioral interaction representation of the candidate user as the projection source, the behavioral field nodes, interaction direction edges, interaction level markers, and the order of the occurrence of behaviors in the behavioral interaction structure diagram of the target user are written into the receiver, and the behavioral field nodes, interaction direction edges, interaction level markers, and the complete behavioral path in the behavioral interaction structure diagram of the candidate user are written into the source. Projection connection edges are established at positions where the field types are consistent, the advertising identifiers are consistent, and the scene fields are matched, thus constructing a cross-user behavior projection-recycle field.

[0032] In the cross-user behavior projection-recycle field, node correspondence, edge correspondence, and path segment correspondence are performed on the behavior interaction structure graphs of target users and candidate users. The similarity of behavior field types, node interaction intensity, interaction direction, interaction level, and behavior migration order are compared respectively. Behavior interaction structure similarity is generated according to the comparison results of field type, interaction intensity, interaction direction, interaction level, and migration order.

[0033] Candidate user behavior paths that meet the preset similarity conditions in terms of behavioral interaction structure and have path segments that match the order of behaviors already performed by the target user are marked as projectable behavior paths, and the ends of the matched path segments are determined as projection anchor points.

[0034] Starting from the projection anchor point, read the non-occurring path segments arranged in chronological order in the projectable behavior path, map the behavior migration chain and state nodes in the non-occurring path segments to the target user, generate candidate evolution paths, perform recycling verification on the candidate evolution paths, delete paths that are inconsistent with the order of the target user's already occurred behavior, ad identifiers, ad material types, scene fields or conversion fields, and combine the retained candidate evolution paths into a set of potential behavior evolution paths for the target user.

[0035] Optionally, the generation path consistency collapse result and collapse strength include:

[0036] Read the set of potential behavior evolution paths, write a path identifier, projection source user identifier, target user identifier, advertisement identifier and path sequence identifier for each potential behavior evolution path, and split each potential behavior evolution path into state nodes and behavior migration chains arranged in chronological order;

[0037] Using the projection anchor points corresponding to the target user's actions as the alignment starting point, the state nodes in different potential behavior evolution paths are aligned according to behavior type, advertising identifier, scene field, interaction direction, interaction level and path order to generate path alignment records.

[0038] Based on the path alignment record, the state nodes that appear repeatedly at the same alignment position are counted. State nodes with the same behavior type, the same advertising identifier, the same scene field, the same interaction direction, and the number of occurrences on the path reaches the preset number of nodes to be retained are marked as common state nodes.

[0039] Based on the path alignment record, the repeated behavior migration relationships between adjacent common state nodes are statistically analyzed. Behavior migration relationships with consistent start state nodes, consistent end state nodes, consistent migration direction, matching interaction level, and the number of occurrence paths reaching the preset migration retention number are marked as common behavior migration chains.

[0040] The common state nodes and common behavior migration chains are connected according to the path order to generate a transformation consensus body. The transformation consensus body includes a common state node identifier, a common behavior migration chain identifier, a corresponding advertisement identifier, a corresponding scene field, a corresponding projection source user identifier, and a corresponding path order identifier.

[0041] The path consistency collapse result is generated based on the number of potential behavioral evolution paths covered by the transformation consensus body, the number of common state nodes, the number of common behavioral migration chains, and the transformation field acceptance result. The collapse strength is generated based on the ratio between the number of paths covered by the transformation consensus body and the total number of paths in the potential behavioral evolution path set.

[0042] Optionally, the combination of the behavior transition chain and state nodes into a transformation receiving structure includes:

[0043] Read the common state nodes, common behavior migration chains, corresponding advertising identifiers, corresponding scene fields, corresponding path sequence identifiers, and path consistency collapse results in the conversion consensus body, and build a list of objects to be verified according to the path order;

[0044] Select a common state node or a common behavior migration chain from the list of objects to be verified as the current stripping object, temporarily remove the current stripping object from the transformation consensus body, and generate the stripped consensus body;

[0045] Perform a path connectivity check on the consensus body after stripping to determine whether there is still a continuous behavioral migration relationship between the remaining common state nodes in the consensus body after stripping, from exposure behavior, browsing behavior, click behavior, interaction behavior to conversion behavior, and generate path connectivity verification results;

[0046] Based on the path connectivity verification results, regenerate the path consistency collapse results after stripping, compare the path consistency collapse results after stripping with the path consistency collapse results before stripping, and if the path consistency collapse results after stripping disappear or the corresponding transformation field inheritance relationship is interrupted, mark the current stripped object as a necessary inheritance object.

[0047] Repeat the process of selecting, temporarily removing, checking path connectivity, and comparing consistency of the current stripped object until all common state nodes and common behavior migration chains in the list of objects to be verified have been stripped and verified, generating the necessary set of receiving objects.

[0048] The necessary common state nodes are used as conversion receiving nodes, and the necessary common behavior migration chains are used as conversion receiving edges. The conversion receiving nodes and conversion receiving edges are connected according to the path order to generate a conversion receiving chain. The advertising identifier, scene field, conversion field receiving mark and necessity verification mark are written into it to form a conversion receiving structure.

[0049] Optionally, the output of advertising effectiveness score, potential behavioral paths, path convergence trend, and key behavioral impact information includes:

[0050] Read the set of potential behavior evolution paths, path consistency collapse results, collapse strength and conversion acceptance structure, and extract the path identifier, projection source user identifier, advertisement identifier, scene field, path order identifier and conversion field acceptance mark corresponding to each potential behavior evolution path;

[0051] Based on the set of potential behavioral evolution paths, the number of projected paths, effective paths, and transformation paths corresponding to the target user are counted. Based on the path consistency collapse results, the number of paths covered by the transformation consensus body, the number of common state nodes, and the number of common behavioral migration chains are counted, and path convergence statistics are generated.

[0052] Based on the transformation and acceptance structure, read the transformation and acceptance nodes, transformation and acceptance edges, acceptance markers, and necessity verification markers. Count the number of times the transformation and acceptance nodes appear in the potential behavior evolution path set, the number of consecutive acceptances of the transformation and acceptance edges, and the number of acceptances of the transformation fields to generate key behavior impact records.

[0053] Based on the number of effective paths, the number of conversion paths, the collapse strength, the number of paths covered by the conversion consensus body, the number of times conversion receiving nodes appear, the number of consecutive receiving times of conversion receiving edges, and the number of receiving times of conversion fields, path coverage scores, path convergence scores, and conversion receiving scores are generated respectively. These scores are then weighted and synthesized according to preset scoring weights to generate an advertising performance score.

[0054] The system associates and writes advertising performance scores, potential behavior evolution path sets, path convergence statistics, conversion structure, and key behavior impact records to generate advertising performance evaluation results, outputting advertising performance scores, potential behavior paths, path convergence trends, and key behavior impact information.

[0055] The beneficial effects of this invention are:

[0056] This invention introduces an improved AutoInt network to perform high-order interaction modeling of multimodal advertising behavior information. It unifies previously scattered behaviors such as exposure, browsing, clicks, interactions, and conversions into a unified behavioral interaction structure graph, effectively characterizing the interaction intensity, direction, and hierarchical relationships between different behavioral fields. Compared to existing methods that rely solely on simple feature concatenation or low-order interactions, this invention more accurately reflects the complex relationships between user behaviors, elevating advertising effectiveness evaluation from single-indicator statistics to structured behavioral relationship analysis, thus improving the accuracy and stability of the evaluation results.

[0057] This invention constructs a cross-user behavior projection-recycle field to map the behavioral paths of candidate users to target users, generating multiple potential behavioral evolution paths. This addresses the issues of insufficient target user behavior samples and incomplete behavioral paths. By extracting a conversion consensus body through path alignment analysis and generating path consistency collapse results, advertising effectiveness evaluation no longer relies on a single user or path but is based on multi-path consistency. This effectively identifies whether advertising has driven user behavior towards a conversion trend, improving the robustness and generalization ability of the evaluation results.

[0058] This invention identifies the behavioral transition chains and state nodes necessary to maintain the path consistency collapse result by performing item-by-item stripping verification on the conversion consensus body, and constructs a conversion acceptance structure to achieve precise positioning of key conversion behavior chains. Compared with existing methods based on feature weights or simple contribution analysis, this invention can reveal the key paths of advertising conversion at the structural level, enhance the interpretability of evaluation results, provide clear behavioral adjustment basis for advertising optimization, and improve the targeting and optimization efficiency of advertising strategies. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of an advertising effectiveness evaluation method based on multimodal behavior fusion proposed in this invention;

[0061] Figure 2 This is a schematic diagram illustrating the process of constructing a behavioral interaction structure diagram based on an improved AutoInt network for an advertising effectiveness evaluation method based on multimodal behavior fusion proposed in this invention. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figure 1 and Figure 2 A method for evaluating advertising effectiveness based on multimodal behavior fusion, comprising:

[0064] Collect multimodal behavioral data generated during the advertising process, construct a behavioral data set, and form a user behavior sequence according to time order;

[0065] Multimodal advertising behavior information in user behavior sequences is mapped to a set of behavior fields, which are then input into an improved AutoInt network to perform cross-modal high-order interaction calculations. The interaction strength, interaction direction, and interaction level between behavior fields are extracted, a behavior interaction structure graph is constructed, and the corresponding behavior interaction representation is generated.

[0066] Based on behavioral interaction representation, a cross-user behavior projection-recovery field is constructed. The similarity of behavioral interaction structure between the target user and the candidate user is calculated. Candidate user behavior paths that meet the preset similarity conditions are selected. The candidate user behavior paths are projected onto the target user to generate a set of potential behavioral evolution paths for the target user.

[0067] Perform path alignment analysis on the set of potential behavioral evolution paths, extract common behavioral migration chains and common state nodes in different behavioral paths, combine common behavioral migration chains and common state nodes into a transformation consensus body, and generate path consistency collapse results and collapse strength.

[0068] Perform item-by-item stripping verification on the behavior migration chain and state node in the transformation consensus body, identify the behavior migration chain and state node necessary to maintain the path consistency collapse result, and combine the behavior migration chain and state node into the transformation acceptance structure.

[0069] Based on the set of potential behavioral evolution paths, path consistency collapse results, and conversion structure, the advertising effect is comprehensively evaluated, and the advertising effect score, potential behavioral paths, path convergence trends, and key behavioral impact information are output.

[0070] In this embodiment, the multimodal behavioral data includes ad exposure records, ad click records, ad browsing records, ad interaction records, ad conversion records, ad identification information, user identification information, terminal device information, access time information, access location information, and ad creative information.

[0071] In this embodiment, the construction of the behavioral data set, forming a user behavior sequence in chronological order, includes:

[0072] Read ad exposure records, ad click records, ad browsing records, ad interaction records, ad conversion records, and related scenario information; match and associate various records based on user identifiers and ad identifiers to generate corresponding behavioral record sets;

[0073] Extract the occurrence time, behavior type, behavior object, and scene attributes corresponding to each behavior record, and convert them according to a unified data format to generate standardized behavior records;

[0074] Standardized behavior records are sorted in ascending order according to the time of occurrence to determine the chronological relationship between each behavior record and generate a behavior timeline.

[0075] According to the order of the behavior time chain, adjacent behavior records are connected in sequence to generate the corresponding user's behavior trajectory;

[0076] The records of each behavior in the behavior trajectory are combined in chronological order to form a user behavior sequence.

[0077] In this embodiment, the step of constructing the behavior interaction structure diagram and generating the corresponding behavior interaction representation includes:

[0078] An improved AutoInt network is constructed, which consists of an embedding layer, a cross-modal gating interaction layer, a bidirectional causal masking interaction layer, and a graph-level attention pooling layer from bottom to top. The layers are connected through residual connections and layer normalization modules. The cross-modal gating interaction layer inserts a gating weight adjustment sublayer between the parallel multi-head self-attention sublayer and the residual output. The bidirectional causal masking interaction layer includes a forward masking branch and a backward masking branch. The graph-level attention pooling layer adopts a two-level attention mechanism to first aggregate nodes of the same modality and then aggregate nodes across modalities.

[0079] Independent embedding lookup tables are established for the exposure, click, browsing, interaction, conversion, advertising, terminal, and time fields in the user behavior sequence. Discrete fields are mapped to dense vectors, and continuous fields are mapped to vectors of the same dimension after min-max normalization, forming a field vector sequence.

[0080] Insert a behavior type marker vector at the beginning of the field vector sequence and a context scene marker vector at the end. Attach a relative time position code to each field vector to preserve the event time sequence information.

[0081] A sequence of field vectors containing label vectors and positional codes is input into a cross-modal gated interaction layer. Each cross-modal gated interaction layer consists of a cascaded parallel multi-head self-attention sublayer, a cross-modal gated sublayer, and a residual normalization sublayer. The cross-modal gated sublayer dynamically adjusts the weight coefficients of the self-attention sublayer according to the modality to which the field belongs, in order to highlight the cross-modal interaction features.

[0082] The parallel multi-head self-attention sublayer is used to receive a field vector sequence containing a label vector and a position encoding. The field vector sequence is input into multiple attention heads respectively. Within each attention head, the correlation strength between any behavior field and other behavior fields is calculated, and attention outputs between exposure field and click field, browsing field and conversion field, and interaction field and scene field are obtained respectively. The outputs of multiple attention heads are concatenated in the order of fields to generate a multi-head interaction vector sequence.

[0083] The cross-modal gating sublayer is used to read the multi-head interaction vector sequence and the modality label of each field. It assigns cross-modal gating coefficients to the interactions of fields belonging to different modalities and assigns same-modal gating coefficients to the interactions of fields belonging to the same modality. Based on the gating coefficients, it performs weighted adjustment on the field interaction results in the multi-head interaction vector sequence to generate a gated interaction vector sequence.

[0084] The residual normalization sublayer is used to add the residuals of the gated interaction vector sequence and the field vector sequence before the input cross-modal gated interaction layer, and perform layer normalization on the added vector to generate the cross-modal interaction vector sequence.

[0085] The cross-modal interaction vector sequence is input into the bidirectional causal masked interaction layer. The forward masking branch captures the influence of the preceding behavior on the subsequent behavior, and the backward masking branch captures the backtracking influence of the subsequent behavior on the preceding behavior. The bidirectional outputs are fused through depthwise separable residual connections to generate a high-order behavioral interaction representation containing interaction direction information. Specifically, the generation of this high-order behavioral interaction representation includes the following:

[0086] Read the cross-modal interaction vector sequence, write each behavior field vector into the forward sequence in chronological order, and write the same sequence into the backward sequence in reverse chronological order. In the forward sequence, for each time step, retain only the attention query range between the current behavior and the behaviors preceding it, calculate the forward attention score table, multiply the forward attention score table with the corresponding vectors in the forward sequence by time step and sum them to generate the preceding influence vector sequence. In the backward sequence, for each time step, retain only the attention query range between the current behavior and the behaviors following it, calculate the backward attention score table, and multiply the backward attention score table into the forward sequence vector sequence. The vectors corresponding to the numerical table and the backward sequence are multiplied and accumulated at each time step to generate a backtracking influence vector sequence. The preceding influence vector and the backtracking influence vector are then concatenated one-to-one according to their field positions. An input depth-separable linear mapping layer is used to group and map the behavioral feature channels and directional feature channels of the concatenated vectors to obtain a directional fusion vector. The directional fusion vector is then added to the cross-modal interaction vector sequence using residuals, and layer normalization is performed to output a vector sequence with directional labels. Finally, the output vector sequences are concatenated in chronological order to generate a high-order behavioral interaction representation containing interaction direction information.

[0087] The higher-order behavioral interaction representation is mapped to the attention level representation space. The interaction level is labeled according to the attention level index, and the interaction direction is labeled according to the bidirectional mask. The behavioral interaction structure graph is constructed using field vectors as nodes, interaction strength and direction between nodes, and interaction level as edge attributes. The construction of the behavioral interaction structure graph is as follows:

[0088] Read the sequential list of multiple high-order behavioral interaction vectors, register each vector in the list as a corresponding graph node, and record the field number, modality label, and time position index for each node; then process the interaction weight matrices of the attention outputs of layers 1, 2, and 3 in sequence, write the items with a weight value of not less than 12% in layer 1 into the first-level interaction candidate matrix, write the items with a weight value of not less than 8% in layer 2 into the second-level interaction candidate matrix, and write the items with a weight value of not less than 5% in layer 3 into the third-level interaction candidate matrix;

[0089] Traverse the three-level candidate matrix by row and column coordinates. For each candidate relationship, read the corresponding node pair, interaction strength value, interaction direction mark and interaction level number. If the node pair has not been recorded, create a new graph edge and write the interaction strength, interaction direction and interaction level. If the node pair already exists and the current interaction strength is higher than the recorded interaction strength, overwrite the original record with the current interaction strength, interaction direction and interaction level.

[0090] After traversal, the sum of the weights of incoming and outgoing edges for each node in the graph is calculated. Nodes with both incoming and outgoing edge weights less than 4% are marked as weakly related nodes and their related edges are deleted. The remaining graph edges are deduplicated to ensure that only the edge with the strongest interaction strength is retained for the same node pair. Finally, the node set, edge set, and the interaction strength, direction, and level of the edges, as well as the field number, modality label, and time position index of the nodes are written into the adjacency list and sparse adjacency matrix to obtain the complete behavioral interaction structure graph.

[0091] A graph-level attention pooling layer is used to perform node aggregation on the behavior interaction structure graph. Nodes of the same modality are aggregated first, and then the aggregation results of different modalities are weighted and merged to output a behavior interaction representation with a unified dimension. Specifically, the output of the unified behavior interaction representation is as follows:

[0092] The behavioral interaction structure graph is read, and nodes are divided into six groups according to modal labels: exposure modality, click modality, browsing modality, interaction modality, conversion modality, and scene modality. Within each group, a first-level attention aggregation is performed on the nodes, calculating the proportion of a node's interaction intensity to the total interaction intensity within the group as the node's attention weight. Nodes with a weight value of at least 10% participate in the aggregation, and the node vectors participating in the aggregation are summed using a weighted method to obtain the modality embedding vector corresponding to each modality. The six modality embedding vectors are arranged in modal order and input into the second-level cross-modal attention module. Four attention heads are set, and the inter-modal weight matrix is ​​calculated. The modality embedding vectors with a weight value of at least 15% are weighted and merged to obtain the cross-modal fusion vector. The cross-modal fusion vector is input into a two-layer linear mapping module. The first layer expands the vector dimension to 256, and the second layer compresses the dimension to 64. A layer normalization operation with residual connections is inserted between the two layers to output a 64-dimensional unified behavioral interaction representation.

[0093] In improving the training of the AutoInt network, historical ad delivery logs are used as training samples. Each training sample includes a user behavior sequence, ad identifier, ad creative type, terminal field, time field, scene field, click tag, and conversion tag. The click tag represents whether the user clicked or not, and the conversion tag represents whether the user converted or not. The multimodal ad behavior information in the user behavior sequence is input into the embedding layer, cross-modal gating interaction layer, bidirectional causal masking interaction layer, and graph-level attention pooling layer. The outputs are click prediction, conversion prediction, behavior interaction representation, and behavior interaction structure graph.

[0094] The loss function consists of click prediction loss, conversion prediction loss, interaction direction constraint loss, interaction level constraint loss, and parameter regularization loss. Click prediction loss constrains the network output's predicted click value to match the actual click label; conversion prediction loss constrains the network output's predicted conversion value to match the actual conversion label; interaction direction constraint loss constrains the interaction direction output of the bidirectional causal mask interaction layer to match the actual temporal order in the user behavior sequence; interaction level constraint loss constrains the association between shallow attention outputs and basic behavior fields, between mid-level attention outputs and cross-modal behavior, and between deep attention outputs and conversion continuation; and parameter regularization loss suppresses overfitting caused by excessively large network parameters.

[0095] For the i-th training sample, read the real click label, real conversion label, click prediction value, and conversion prediction value, and calculate the binary cross-entropy between the click label and the click prediction value, and between the conversion label and the conversion prediction value, respectively. Read the direction label corresponding to each interaction edge in the behavior interaction structure graph and compare it with the time sequence of behavior occurrence in the user behavior sequence. If the interaction direction is consistent with the time sequence, the direction loss is recorded as 0; if the interaction direction is inconsistent with the time sequence, the direction loss is recorded as 1. Read the attention level of each interaction edge, mark the interaction edges between the same modality field as the basic level, mark the interaction edges between different modality behavior fields as the cross-modality level, and mark the interaction edges pointing to the conversion field as the conversion receiving level. If the network output level is consistent with the above labels, the level loss is recorded as 0; if they are inconsistent, the level loss is recorded as 1. Add the click prediction loss, conversion prediction loss, interaction direction constraint loss, interaction level constraint loss, and parameter regularization loss according to the preset weights to obtain the total training loss of the improved AutoInt network.

[0096] During training, the embedding vector dimension was set to 32, the cross-modal gating interaction layer was set to 3 layers with 4 attention heads per layer, the bidirectional causal masking interaction layer was set to 2 layers, the graph-level attention pooling layer was set to a 2-level aggregation structure, the training batch size was set to 256, the initial learning rate was set to 0.001, the learning rate was multiplied by 0.9 every 5 training epochs, the parameter regularization coefficient was set to 0.0001, and the network parameters were updated using the Adam optimizer until the total training loss of the validation set did not decrease for 3 consecutive epochs or the number of training epochs reached 50 epochs, at which point training was stopped.

[0097] In this embodiment, generating the set of potential behavioral evolution paths for the target user includes:

[0098] Read the behavioral interaction representations of the target user, the behavioral interaction representations of the candidate user, and the corresponding behavioral interaction structure diagram. Select the candidate users with consistent ad identifiers, consistent ad creative types, matching scene fields, and existing conversion field records as projection source users to generate a candidate user set.

[0099] Using the target user's behavioral interaction representation as the projection receiver and the candidate user's behavioral interaction representation as the projection source, the behavioral field nodes, interaction direction edges, interaction level markers, and the order of occurring behaviors in the target user's behavioral interaction structure diagram are written to the receiver. The behavioral field nodes, interaction direction edges, interaction level markers, and complete behavioral paths in the candidate user's behavioral interaction structure diagram are written to the source. Projection connection edges are established at positions where the field type, advertising identifier, and scene field match, thus constructing a cross-user behavior projection-recycle field. Specifically, the construction of the cross-user behavior projection-recycle field is as follows:

[0100] Read the target user behavior interaction structure diagram, write all behavior field nodes into the receiving end node pool, and write the interaction direction edges between nodes into the receiving end edge pool. Add an interaction level marker and a sequence index of the target user's actions to each edge. Read the candidate user behavior interaction structure diagram, write all behavior field nodes into the source end node pool, and write the interaction direction edges between nodes into the source end edge pool. Add an interaction level marker and a sequence index of the candidate user's complete actions to each edge. Establish a node mapping list according to field type, ad identifier, and scene field, matching source end nodes with consistent field types, ad identifiers, and scene fields with receiving end nodes. Create a projection connection edge between each pair of corresponding nodes, writing the node mapping number and field class to the projection connection edge. The system includes type markers, ad identifiers, scene field markers, and projection weights, with the projection weight initially set to 1. It copies the complete behavior path segment containing the source node to the receiving end along the projection connection edge, preserving the source interaction direction edge and interaction level marker during copying. The system writes the projection source user ID and projection path ID to the copied edge. If the interaction direction of the copied edge conflicts with an existing interaction direction on the receiving end, they are merged according to the following priority: conversion direction takes precedence over interaction direction, interaction direction takes precedence over browsing direction, and browsing direction takes precedence over exposure direction. After completing all node mapping and path copying, the receiving end node pool, receiving end edge pool, projection connection edge, and copied edge are merged to obtain a cross-user behavior projection-recycle area containing receiving end nodes, source copy nodes, projection connection edges, and copied edges.

[0101] In the cross-user behavior projection-recycle field, node correspondence, edge correspondence, and path segment correspondence are performed on the behavior interaction structure graphs of the target user and candidate users. The similarity of behavior field types, node interaction intensity, interaction direction, interaction level, and behavior migration order are compared respectively. Behavior interaction structure similarity is generated based on the comparison results of field type, interaction intensity, interaction direction, interaction level, and migration order. Specifically, the generation of behavior interaction structure similarity involves:

[0102] The system sequentially scans the projected connection edges to locate the one-to-one correspondence between target user nodes and candidate user nodes, and counts the total number of node pairs. For each pair of nodes, if the field types are completely identical, it is recorded as a field-matching value of 1; otherwise, it is recorded as 0, and the field-matching value is accumulated. The system reads the interaction strength difference of the corresponding edges. If the difference does not exceed 0.05, it is recorded as an interaction strength close value of 1; otherwise, it is recorded as 0, and the interaction strength close value is accumulated. The system compares the interaction directions of the edges. If the direction labels are identical, it is recorded as an interaction direction matching value of 1; otherwise, it is recorded as 0, and the interaction direction matching value is accumulated. The system compares the interaction levels of the edges. If the level numbers are identical, it is recorded as an interaction level matching value of 1; otherwise, it is recorded as 0, and the interaction level matching value is accumulated. The system reads the time sequence index of the nodes in their respective behavior sequences. If the difference does not exceed 3, it is recorded as a migration order matching value of 1; otherwise, it is recorded as 0, and the migration order matching value is accumulated.

[0103] The field type consistency rate is obtained by dividing the cumulative value of consistent fields by the total number of node pairs; the interaction intensity closeness rate is obtained by dividing the cumulative value of similar interaction intensity by the total number of node pairs; the interaction direction consistency rate is obtained by dividing the cumulative value of consistent interaction direction by the total number of node pairs; the interaction level matching rate is obtained by dividing the cumulative value of consistent interaction level by the total number of node pairs; and the migration order consistency rate is obtained by dividing the cumulative value of consistent migration order by the total number of node pairs.

[0104] The five factors—field type consistency rate, interaction intensity similarity rate, interaction direction consistency rate, interaction level matching rate, and migration order consistency rate—are assigned weights of 0.20, 0.25, 0.20, 0.20, and 0.15, respectively. The products of these five weights are then multiplied and summed to generate a percentage value representing the similarity between the behavioral interaction structures of the target user and the candidate user.

[0105] Candidate user behavior paths that meet the preset similarity conditions and have path segments that match the order of actions already performed by the target user are marked as projectable behavior paths. The ends of the matched path segments are determined as projection anchor points. The preset similarity conditions are: field type consistency rate not less than 0.80, interaction intensity similarity rate not less than 0.75, interaction direction consistency rate not less than 0.75, interaction level matching rate not less than 0.70, and migration order consistency rate not less than 0.70. The overall similarity value of the behavior interaction structure obtained by weighted calculation is not less than 0.78.

[0106] Starting from the projection anchor point, read the non-occurring path segments arranged in chronological order in the projectable behavior path, map the behavior migration chain and state nodes in the non-occurring path segments to the target user, generate candidate evolution paths, perform recycling verification on the candidate evolution paths, delete paths that are inconsistent with the order of the target user's already occurred behavior, ad identifiers, ad material types, scene fields or conversion fields, and combine the retained candidate evolution paths into a set of potential behavior evolution paths for the target user.

[0107] In this embodiment, the generation path consistency collapse result and collapse strength include:

[0108] Read the set of potential behavior evolution paths, write a path identifier, projection source user identifier, target user identifier, advertisement identifier and path sequence identifier for each potential behavior evolution path, and split each potential behavior evolution path into state nodes and behavior migration chains arranged in chronological order;

[0109] Using the projection anchor points corresponding to the target user's actions as the alignment starting point, the state nodes in different potential behavior evolution paths are aligned according to behavior type, advertising identifier, scene field, interaction direction, interaction level and path order to generate path alignment records.

[0110] Based on the path alignment record, the state nodes that appear repeatedly at the same alignment position are counted. State nodes with the same behavior type, the same advertising identifier, the same scene field, the same interaction direction, and the number of occurrences on the path reaches the preset number of nodes to be retained are marked as common state nodes, where the preset number of nodes to be retained is 3.

[0111] Based on the path alignment record, the repeated behavior migration relationships between adjacent common state nodes are statistically analyzed. Behavior migration relationships with consistent start state nodes, consistent end state nodes, consistent migration direction, matching interaction level, and the number of occurrence paths reaching the preset migration retention number are marked as common behavior migration chains, where the preset migration retention number is 3.

[0112] Connecting common state nodes and common behavior migration chains according to the path sequence generates a transformation consensus body. This transformation consensus body includes a common state node identifier, a common behavior migration chain identifier, a corresponding advertisement identifier, a corresponding scene field, a corresponding projection source user identifier, and a corresponding path sequence identifier. Specifically, generating the transformation consensus body involves:

[0113] According to the path sequence number in the potential behavior evolution path set, the determined common state nodes are read sequentially from smallest to largest. Each common state node is written into the consensus node list, and the advertising identifier, scene field, projection source user identifier, and path sequence number of the node are recorded respectively. In the consensus node list, the determined common behavior migration chains are searched between adjacent nodes in chronological order. If there is a unique common behavior migration chain between adjacent nodes, the migration chain is written into the consensus chain list, and the starting node identifier, ending node identifier, migration direction mark, interaction level number, and corresponding advertising identifier and scene field are recorded for the migration chain. If there are multiple common behavior migration chains between adjacent nodes, the one with the highest interaction strength is selected and written into the consensus chain list, and the same information is recorded. If there is no common behavior migration chain between adjacent nodes, an empty chain mark is inserted into the consensus chain list.

[0114] After matching nodes with migration chains, the consensus node list and consensus chain list are reordered according to path sequence number. The position of the empty chain mark is checked. If the advertising identifier or scene field of the nodes before and after the empty chain mark is inconsistent, the nodes on both sides of the empty chain mark and related records are deleted. If they are consistent, the nodes on both sides of the empty chain mark are retained and the discontinuous chain identifier is recorded. The sorted consensus node list and consensus chain list are matched one by one. The node identifier is written into the common state node identifier field, the migration chain identifier is written into the common behavior migration chain identifier field, and the corresponding advertising identifier, corresponding scene field, corresponding projection source user identifier, and corresponding path sequence identifier are written into the consensus body structure table to generate the conversion consensus body.

[0115] Based on the number of potential behavioral evolution paths covered by the transformation consensus body, the number of common state nodes, the number of common behavioral migration chains, and the transformation field acceptance results, a path consistency collapse result is generated. The collapse strength is then generated based on the ratio between the number of paths covered by the transformation consensus body and the total number of paths in the potential behavioral evolution path set. Specifically, the generation of the collapse strength is as follows:

[0116] The number of potential behavioral evolution paths covered by the statistical transformation consensus body is divided by the total number of paths in the potential behavioral evolution path set to obtain the path coverage ratio. The number of common state nodes and common behavioral migration chains in the transformation consensus body are counted, and the sum of the two is divided by the maximum total of common state nodes and common behavioral migration chains to obtain the node chain density ratio. Then, the transformation field acceptance results are read, and the proportion of paths successfully guided to the transformation field by the transformation acceptance edge in the transformation consensus body to the number of paths covered by the transformation consensus body is calculated to obtain the transformation acceptance success ratio. The path coverage ratio, node chain density ratio, and transformation acceptance success ratio are assigned weights of 0.4, 0.3, and 0.3, respectively. The three ratios are weighted and summed to obtain a collapse strength percentage value between 0 and 1.

[0117] In this embodiment, combining the behavior transition chain and state nodes into a transformation receiving structure includes:

[0118] Read the common state nodes, common behavior migration chains, corresponding advertisement identifiers, corresponding scene fields, corresponding path sequence identifiers, and path consistency collapse results from the conversion consensus body, and build a list of objects to be verified according to the path order. Specifically, building the list of objects to be verified according to the path order is as follows:

[0119] The path sequence identifiers recorded in the consensus body are read sequentially and sorted from smallest to largest. Within each sorted path sequence identifier, common state nodes on the path are written first in chronological order, and then common behavior migration chains between adjacent common state nodes are written in chronological order, forming an entry sequence of alternating nodes and migration chains. The entry sequence of each path is appended to a unified list, and an object type marker, sequence number marker, corresponding advertisement identifier, corresponding scene field, and path sequence identifier are written to the list entries to complete the establishment of the list of objects to be verified.

[0120] Select a common state node or a common behavior migration chain from the list of objects to be verified as the current stripping object, temporarily remove the current stripping object from the transformation consensus body, and generate a stripped consensus body. The specific steps for generating the stripped consensus body are as follows:

[0121] Read the current stripped object according to the record position in the list of objects to be verified. If the current stripped object is a common state node, delete the node entry and one common behavior migration chain entry before and after it in the transformation consensus body. If the current stripped object is a common behavior migration chain, delete the migration chain entry in the transformation consensus body. After deleting the object, reorder the remaining entries according to the original path order. Insert path discontinuity placeholders between two consecutive common behavior migration chains or two consecutive common state nodes to maintain the integrity of the path order. Update the number of common state nodes, the number of common behavior migration chains, and the path coverage statistics in the transformation consensus body to generate the stripped consensus body.

[0122] Perform a path connectivity check on the stripped consensus body to determine whether a continuous behavioral transition relationship still exists between the remaining common state nodes in the stripped consensus body, from exposure behavior, browsing behavior, click behavior, interaction behavior to conversion behavior, and generate a path connectivity verification result. Specifically, the generation of the path connectivity verification result is as follows:

[0123] The remaining common state nodes in the consensus body after stripping are grouped according to the path sequence, and the node type sequence is read group by group. Within each group of node type sequences, the node types are retrieved in the order of five actions: exposure, browsing, clicking, interaction, and conversion. If the next expected behavior type is missing during the retrieval, the behavior chain is marked as broken on the path. For the same group of node type sequences, if the same behavior type appears twice consecutively before the conversion node is retrieved, the behavior is marked as duplicate on the path. When the five-step behavior sequence is complete and there are no behavior chain breaks or behavior duplicates, the path is marked as a connected path. After completing the retrieval of all paths, the number of connected paths, the number of paths marked with behavior chain breaks, and the number of paths marked with behavior duplicates are counted. The three statistical results are compared with the corresponding statistical results of the consensus body before stripping. If the number of connected paths remains unchanged and the number of paths marked with behavior chain breaks and behavior duplicates does not increase, a path connectivity verification result is generated to maintain connectivity; otherwise, a path connectivity verification result is generated to indicate that connectivity is interrupted.

[0124] Based on the path connectivity verification results, regenerate the path consistency collapse results after stripping, compare the path consistency collapse results after stripping with the path consistency collapse results before stripping, and if the path consistency collapse results after stripping disappear or the corresponding transformation field inheritance relationship is interrupted, mark the current stripped object as a necessary inheritance object.

[0125] Repeat the process of selecting, temporarily removing, checking path connectivity, and comparing consistency of the current stripped object until all common state nodes and common behavior migration chains in the list of objects to be verified have been stripped and verified, generating the necessary set of receiving objects.

[0126] Necessary common state nodes are used as conversion receiving nodes, and necessary common behavior migration chains are used as conversion receiving edges. Conversion receiving nodes and edges are connected according to the path order to generate a conversion receiving chain. This chain is then written with ad identifiers, scene fields, conversion field receiving tags, and necessity verification tags to form a conversion receiving structure. Specifically, the conversion receiving structure is formed as follows:

[0127] Read all necessary common state nodes and necessary common behavior migration chains. Number the necessary common state nodes according to the path sequence and write them into the receiving node table. Number the necessary common behavior migration chains according to the path sequence and write them into the receiving edge table. Select the node with the smallest sequence number from the receiving node table as the chain start point. Select the first receiving edge after the start point from the receiving edge table as the chain head edge. Connect nodes and edges in sequence until all receiving nodes and edges are connected, generating a branchless conversion receiving chain. Write the corresponding advertisement identifier, scene field, and path sequence number into each receiving node and receiving edge entry in the conversion receiving chain. Write the conversion field receiving mark into the receiving edge entry. Write the necessity verification mark as necessary into the receiving node and receiving edge entry. Finally, merge the receiving node table, receiving edge table, and conversion receiving chain sequence table to output the conversion receiving structure.

[0128] In this embodiment, the output of advertising effectiveness score, potential behavioral path, path convergence trend, and key behavioral impact information includes:

[0129] Read the set of potential behavior evolution paths, path consistency collapse results, collapse strength and conversion acceptance structure, and extract the path identifier, projection source user identifier, advertisement identifier, scene field, path order identifier and conversion field acceptance mark corresponding to each potential behavior evolution path;

[0130] Based on the set of potential behavioral evolution paths, the number of projected paths, effective paths, and transformation paths corresponding to the target user are counted. Based on the path consistency collapse results, the number of paths covered by the transformation consensus body, the number of common state nodes, and the number of common behavioral migration chains are counted, and path convergence statistics are generated.

[0131] Based on the transformation and acceptance structure, read the transformation and acceptance nodes, transformation and acceptance edges, acceptance markers, and necessity verification markers. Count the number of times the transformation and acceptance nodes appear in the potential behavior evolution path set, the number of consecutive acceptances of the transformation and acceptance edges, and the number of acceptances of the transformation fields to generate key behavior impact records.

[0132] Based on the number of effective paths, the number of conversion paths, the collapse strength, the number of paths covered by the conversion consensus body, the number of times conversion receiving nodes appear, the number of consecutive receiving steps by conversion receiving edges, and the number of receiving steps by conversion fields, path coverage scores, path convergence scores, and conversion receiving scores are generated respectively. These scores are then weighted and synthesized according to preset scoring weights to generate an advertising performance score, specifically including path coverage scores, path convergence scores, and conversion receiving scores.

[0133] The path coverage ratio is obtained by dividing the number of effective paths by the total number of paths in the potential behavior evolution path set; the transformation path ratio is obtained by dividing the number of transformed paths by the number of effective paths; the path coverage ratio and transformation path ratio are added together and then divided by 2 to generate the path coverage score; subsequently, the collapse intensity is directly used as the path convergence score; the consensus body coverage ratio is obtained by dividing the number of paths covered by the transformation consensus body by the total number of paths in the potential behavior evolution path set; the acceptance density ratio is obtained by adding the number of occurrences of transformation receiving nodes, the number of consecutive acceptances of transformation receiving edges, and the number of acceptances of transformation fields and then dividing by the total number of paths; the acceptance density ratio is obtained by adding the consensus body coverage ratio and the acceptance density ratio and then dividing by 2 to generate the transformation acceptance score.

[0134] The process of generating an advertising performance score is as follows: read the path coverage score, path convergence score, and conversion acceptance score, and weight the three scores according to the preset scoring weights. Assign a weight of 0.40 to the path coverage score, 0.35 to the path convergence score, and 0.25 to the conversion acceptance score. Calculate the path coverage score by multiplying it by 0.40, the path convergence score by multiplying it by 0.35, and the conversion acceptance score by multiplying it by 0.25. Sum the three products and keep two decimal places to obtain the advertising performance score out of 100.

[0135] The system associates and writes advertising performance scores, potential behavior evolution path sets, path convergence statistics, conversion structure, and key behavior impact records to generate advertising performance evaluation results, outputting advertising performance scores, potential behavior paths, path convergence trends, and key behavior impact information.

[0136] Example 1: During an evaluation period for a news feed ad campaign, the ad platform received ad behavior logs from mobile terminals, involving one target ad campaign and three ad creative formats: short video creatives, static image creatives, and carousel creatives. The system received a total of 486,320 raw behavior logs, including 302,800 exposure records, 94,160 browsing records, 51,420 click records, 27,680 interaction records, and 10,260 conversion records. This batch of data exhibits a significant problem of insufficient target user behavior samples. 12,800 target users only generated exposure, browsing, or minimal interaction behavior, averaging only 2.8 behavior records per target user, with a complete click-to-conversion path rate of only 6.9%. In traditional click-through rate and conversion rate evaluations, these users are typically directly classified as low-value users due to no clicks or payments, making it impossible to determine whether the ad has led them into a potential conversion path.

[0137] The system cleaned the raw logs, deleting 6,820 records lacking user identifiers, ad identifiers, or behavior timestamps, and deleting 4,310 duplicate reports, retaining 475,190 valid logs. These logs were then grouped by user identifier and ad identifier, and arranged chronologically to form user behavior sequences. For example, a target user's initial behavior was exposure, browsing, and swiping, with a dwell time of 27 seconds and 4 swipes, without any clicks or conversions. Traditional methods would read this as 0 clicks and 0 conversions, giving an ad performance score of 26.5, but this result cannot indicate whether the user had a tendency to click or add to cart later.

[0138] The system maps valid logs to exposure, click, browsing, interaction, conversion, advertising, terminal, and time fields. Discrete fields are entered into independent embedding lookup tables, while continuous fields undergo min-max normalization. The original range for dwell time is 0 to 120 seconds; for this target user, 27 seconds is mapped to 0.225. The original range for swipe count is 0 to 12; for this target user, 4 swipes are mapped to 0.333. Each field is mapped to a 32-dimensional vector, resulting in 8 field vectors for this target user, with a total size of 8×32. The system adds a behavior type marker vector at the beginning of the field vector sequence and a scene marker vector at the end, and writes relative time position encodings for 10 vectors.

[0139] The improved AutoInt network comprises three cross-modal gating interaction layers, two bidirectional causal masking interaction layers, and one graph-level attention pooling layer. In the first layer's attention output, the interaction weight between the browsing and click fields is 0.137, reaching the 12% first-level interaction retention threshold. In the second layer's attention output, the interaction weight between the swipe and conversion fields is 0.091, reaching the 8% second-level interaction retention threshold. In the third layer's attention output, the interaction weight between the ad creative and conversion fields is 0.063, reaching the 5% third-level interaction retention threshold. The system removes two weakly related nodes with both inbound and outbound edge weights less than 4%, retaining eight behavioral field nodes and 13 interaction edges, forming the behavioral interaction structure graph for the target user. After processing by the bidirectional causal masking interaction layer, the average forward influence response of the browsing and click fields is 0.61, the average backward response of the click and browsing fields is 0.38, and the average response of the conversion-related dimensions after fusion increases from 0.24 to 0.47.

[0140] The system constructs a cross-user behavior projection-recycle bin. The candidate user pool contains 38,400 users with relatively complete behavior paths. After initial screening based on consistent ad identifiers, consistent creative types, consistent terminal types, and matching scene fields, 6,420 users remain. The system performs node mapping, edge mapping, and path segment mapping on the behavior interaction structure graph of the target user and candidate users. A candidate user's path is exposure, browsing, swiping, clicking, adding to cart, and payment. The field type consistency rate is 0.875, the interaction intensity similarity rate is 0.812, the interaction direction consistency rate is 0.786, the interaction level matching rate is 0.741, the migration order consistency rate is 0.833, and the overall similarity is 0.807, meeting the projection condition of 0.78. Therefore, it is marked as a projectable path.

[0141] For this target user, the system identified 9 projectable paths: 6 of which are exposure, browsing, swiping, clicking, adding to cart, payment, or adjacent variations; 2 are exposure, browsing, swiping, clicking, favorites, and placing an order; and 1 is exposure, browsing, and exiting. During the verification process, the system deleted one path with inconsistent material types, retaining 8 potential behavior evolution paths. Thus, the target user, who originally had only 3 active behavior nodes, now has 8 potential paths, 34 state nodes, and 29 behavior migration chains after projection.

[0142] The system performs path alignment analysis on 8 potential behavioral evolution paths. Statistical results show that the click node appears 7 times in the 8 paths, the add-to-cart node appears 5 times, the payment node appears 4 times, the favorite node appears 2 times, and the exit node appears once. Since the preset node retention quantity is 3, click, add-to-cart, and payment are marked as common state nodes, while favorite and exit are excluded. Between adjacent nodes, browsing to click occurs 7 times, click to add-to-cart occurs 5 times, and add-to-cart to payment occurs 4 times, all reaching the preset migration retention quantity of 3, and therefore are marked as common behavioral migration chains. The system connects the 3 common state nodes and 3 common behavioral migration chains in path order to generate a transformation consensus body. This consensus body covers 6 of the 8 potential paths, with a path coverage ratio of 0.75; the density ratio of common nodes to common migration chains is 0.86; 4 paths successfully lead to the transformation field, with a transformation success rate of 0.67. After weighting by 0.4, 0.3, and 0.3, a collapse strength of 0.75 is obtained.

[0143] The system continues to perform item-by-item stripping verification on the conversion consensus body. After stripping the click node, the number of connected paths decreases from 6 to 1, and the path consistency collapse result disappears; therefore, the click node is marked as a necessary receiving object. After stripping the add-to-cart node, the number of payment receiving paths decreases from 4 to 0, and the conversion field receiving relationship is interrupted; therefore, the add-to-cart node is marked as a necessary receiving object. After stripping the browse to click migration chain, only 1 of the 8 potential paths remains that can connect to the click node; this migration chain is also marked as a necessary receiving object. The system uses necessary common state nodes as conversion receiving nodes and necessary common behavior migration chains as conversion receiving edges, connecting them in the path order to generate the conversion receiving structure.

[0144] During the scoring phase, the target user had 8 valid paths and 4 conversion paths, resulting in a path coverage score of 0.625 and a path convergence score of 0.75. The conversion connection node appeared 11 times, the conversion connection edge was consecutively connected 9 times, and the conversion field was connected 4 times, generating a conversion connection score of 0.71. The system weighted the scores (path coverage 0.40, path convergence 0.35, and conversion connection 0.25) to obtain an advertising effectiveness score of 69.00. The user actually clicked and added items to their cart in the verification data but did not make a payment within the same period, indicating that the high-potential conversion judgment provided by this invention is consistent with the actual behavioral evolution.

[0145] In the overall comparative experiment, the system sampled 8,000 low-frequency target users, retaining only the first three behavioral records for each user, with the remaining behaviors used as verification results. The traditional click-through rate (CTR) conversion rate method identified 2,140 potential converting users, with a hit rate of 1,320 (61.7%); the traditional similar user statistics method identified 3,580 potential users, with a hit rate of 2,180 (60.9%); and the method of this invention identified 5,870 potential users, with a hit rate of 4,604 (78.4%). Regarding key behavioral chain identification, the traditional CTR conversion rate method could not output behavioral chains, the traditional similar user statistics method output 1,260 similar paths, and this invention output 6,540 conversion succession structures, of which 5,012 were consistent with actual subsequent behaviors, a consistency rate of 76.6%. After adjusting the advertising entry point based on the conversion structure output by this invention, the click-through rate of low-frequency target users increased from 8.63% to 10.91%, the add-to-cart rate increased from 3.28% to 4.76%, the payment conversion rate increased from 1.42% to 2.18%, and the exit rate after repeated exposure decreased from 38.20% to 29.75%.

[0146] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for evaluating advertising effectiveness based on multimodal behavior fusion, characterized in that, include: Collect multimodal behavioral data generated during the advertising process, construct a behavioral data set, and form a user behavior sequence according to time order; Multimodal advertising behavior information in user behavior sequences is mapped to a set of behavior fields, which are then input into an improved AutoInt network to perform cross-modal high-order interaction calculations. The interaction strength, interaction direction, and interaction level between behavior fields are extracted, a behavior interaction structure graph is constructed, and the corresponding behavior interaction representation is generated. Based on behavioral interaction representation, a cross-user behavior projection-recovery field is constructed. The similarity of behavioral interaction structure between the target user and the candidate user is calculated. Candidate user behavior paths that meet the preset similarity conditions are selected. The candidate user behavior paths are projected onto the target user to generate a set of potential behavioral evolution paths for the target user. Perform path alignment analysis on the set of potential behavioral evolution paths, extract common behavioral migration chains and common state nodes in different behavioral paths, combine common behavioral migration chains and common state nodes into a transformation consensus body, and generate path consistency collapse results and collapse strength. Perform item-by-item stripping verification on the behavior migration chain and state node in the transformation consensus body, identify the behavior migration chain and state node necessary to maintain the path consistency collapse result, and combine the behavior migration chain and state node into the transformation acceptance structure. Based on the set of potential behavioral evolution paths, path consistency collapse results, and conversion structure, the advertising effect is comprehensively evaluated, and the advertising effect score, potential behavioral paths, path convergence trends, and key behavioral impact information are output.

2. The advertising effectiveness evaluation method based on multimodal behavior fusion according to claim 1, characterized in that, The multimodal behavioral data includes ad exposure records, ad click records, ad browsing records, ad interaction records, ad conversion records, ad identification information, user identification information, terminal device information, access time information, access location information, and ad creative information.

3. The advertising effectiveness evaluation method based on multimodal behavior fusion according to claim 1, characterized in that, The constructed behavioral data set, arranged in chronological order to form a user behavior sequence, includes: Read ad exposure records, ad click records, ad browsing records, ad interaction records, ad conversion records, and related scenario information; match and associate various records based on user identifiers and ad identifiers to generate corresponding behavioral record sets; Extract the occurrence time, behavior type, behavior object, and scene attributes corresponding to each behavior record, and convert them according to a unified data format to generate standardized behavior records; Standardized behavior records are sorted in ascending order according to the time of occurrence to determine the chronological relationship between each behavior record and generate a behavior timeline. According to the order of the behavior time chain, adjacent behavior records are connected in sequence to generate the corresponding user's behavior trajectory; The records of each behavior in the behavior trajectory are combined in chronological order to form a user behavior sequence.

4. The advertising effectiveness evaluation method based on multimodal behavior fusion according to claim 1, characterized in that, The construction of the behavior interaction structure diagram and the generation of the corresponding behavior interaction representation include: An improved AutoInt network is constructed, which consists of an embedding layer, a cross-modal gating interaction layer, a bidirectional causal masking interaction layer, and a graph-level attention pooling layer from bottom to top. The layers are connected through residual connections and layer normalization modules. The cross-modal gating interaction layer inserts a gating weight adjustment sublayer between the parallel multi-head self-attention sublayer and the residual output. The bidirectional causal masking interaction layer includes a forward masking branch and a backward masking branch. The graph-level attention pooling layer adopts a two-level attention mechanism to first aggregate nodes of the same modality and then aggregate nodes across modalities. Independent embedding lookup tables are established for the exposure, click, browsing, interaction, conversion, advertising, terminal, and time fields in the user behavior sequence. Discrete fields are mapped to dense vectors, and continuous fields are mapped to vectors of the same dimension after min-max normalization, forming a field vector sequence. Insert a behavior type marker vector at the beginning of the field vector sequence and a context scene marker vector at the end. Attach a relative time position code to each field vector to preserve the event time sequence information. The field vector sequence containing the label vector and position encoding is input into the cross-modal gated interaction layer. Each cross-modal gated interaction layer consists of a parallel multi-head self-attention sub-layer, a cross-modal gated sub-layer and a residual normalization sub-layer cascaded together. The cross-modal gated sub-layer dynamically adjusts the weight coefficient of the self-attention sub-layer according to the modality to which the field belongs, so as to highlight the cross-modal interaction features. The cross-modal interaction vector sequence is input into the bidirectional causal masking interaction layer. The forward masking branch captures the influence of the preceding behavior on the subsequent behavior, and the backward masking branch captures the retrospective influence of the subsequent behavior on the preceding behavior. The bidirectional output is fused through a depth-separable residual connection to generate a high-order behavior interaction representation containing interaction direction information. The higher-order behavioral interaction representation is mapped to the attention level representation space. The interaction level is labeled according to the attention level index, and the interaction direction is labeled according to the bidirectional mask. The behavioral interaction structure graph is constructed using field vectors as nodes, interaction strength and direction between nodes, and interaction level as edge attributes. A graph-level attention pooling layer is used to perform node aggregation on the behavior interaction structure graph. Nodes of the same modality are aggregated first, and then the aggregation results of different modalities are weighted and merged to output a behavior interaction representation with a unified dimension.

5. The advertising effectiveness evaluation method based on multimodal behavior fusion according to claim 1, characterized in that, The set of potential behavioral evolution paths for generating the target user includes: Read the behavioral interaction representations of the target user, the behavioral interaction representations of the candidate user, and the corresponding behavioral interaction structure diagram. Select the candidate users with consistent ad identifiers, consistent ad creative types, matching scene fields, and existing conversion field records as projection source users to generate a candidate user set. Using the behavioral interaction representation of the target user as the projection receiver and the behavioral interaction representation of the candidate user as the projection source, the behavioral field nodes, interaction direction edges, interaction level markers, and the order of the occurrence of behaviors in the behavioral interaction structure diagram of the target user are written into the receiver, and the behavioral field nodes, interaction direction edges, interaction level markers, and the complete behavioral path in the behavioral interaction structure diagram of the candidate user are written into the source. Projection connection edges are established at positions where the field types are consistent, the advertising identifiers are consistent, and the scene fields are matched, thus constructing a cross-user behavior projection-recycle field. In the cross-user behavior projection-recycle field, node correspondence, edge correspondence, and path segment correspondence are performed on the behavior interaction structure graphs of target users and candidate users. The similarity of behavior field types, node interaction intensity, interaction direction, interaction level, and behavior migration order are compared respectively. Behavior interaction structure similarity is generated according to the comparison results of field type, interaction intensity, interaction direction, interaction level, and migration order. Candidate user behavior paths that meet the preset similarity conditions in terms of behavioral interaction structure and have path segments that match the order of behaviors already performed by the target user are marked as projectable behavior paths, and the ends of the matched path segments are determined as projection anchor points. Starting from the projection anchor point, read the non-occurring path segments arranged in chronological order in the projectable behavior path, map the behavior migration chain and state nodes in the non-occurring path segments to the target user, generate candidate evolution paths, perform recycling verification on the candidate evolution paths, delete paths that are inconsistent with the order of the target user's already occurred behavior, ad identifiers, ad material types, scene fields or conversion fields, and combine the retained candidate evolution paths into a set of potential behavior evolution paths for the target user.

6. The advertising effectiveness evaluation method based on multimodal behavior fusion according to claim 1, characterized in that, The generated path consistency collapse result and collapse strength include: Read the set of potential behavior evolution paths, write a path identifier, projection source user identifier, target user identifier, advertisement identifier and path sequence identifier for each potential behavior evolution path, and split each potential behavior evolution path into state nodes and behavior migration chains arranged in chronological order; Using the projection anchor points corresponding to the target user's actions as the alignment starting point, the state nodes in different potential behavior evolution paths are aligned according to behavior type, advertising identifier, scene field, interaction direction, interaction level and path order to generate path alignment records. Based on the path alignment record, the state nodes that appear repeatedly at the same alignment position are counted. State nodes with the same behavior type, the same advertising identifier, the same scene field, the same interaction direction, and the number of occurrences on the path reaches the preset number of nodes to be retained are marked as common state nodes. Based on the path alignment record, the repeated behavior migration relationships between adjacent common state nodes are statistically analyzed. Behavior migration relationships with consistent start state nodes, consistent end state nodes, consistent migration direction, matching interaction level, and the number of occurrence paths reaching the preset migration retention number are marked as common behavior migration chains. The common state nodes and common behavior migration chains are connected according to the path order to generate a transformation consensus body. The transformation consensus body includes a common state node identifier, a common behavior migration chain identifier, a corresponding advertisement identifier, a corresponding scene field, a corresponding projection source user identifier, and a corresponding path order identifier. The path consistency collapse result is generated based on the number of potential behavioral evolution paths covered by the transformation consensus body, the number of common state nodes, the number of common behavioral migration chains, and the transformation field acceptance result. The collapse strength is generated based on the ratio between the number of paths covered by the transformation consensus body and the total number of paths in the potential behavioral evolution path set.

7. The advertising effectiveness evaluation method based on multimodal behavior fusion according to claim 1, characterized in that, The combination of behavior transition chains and state nodes into a transformation receiving structure includes: Read the common state nodes, common behavior migration chains, corresponding advertising identifiers, corresponding scene fields, corresponding path sequence identifiers, and path consistency collapse results in the conversion consensus body, and build a list of objects to be verified according to the path order; Select a common state node or a common behavior migration chain from the list of objects to be verified as the current stripping object, temporarily remove the current stripping object from the transformation consensus body, and generate the stripped consensus body; Perform a path connectivity check on the consensus body after stripping to determine whether there is still a continuous behavioral migration relationship between the remaining common state nodes in the consensus body after stripping, from exposure behavior, browsing behavior, click behavior, interaction behavior to conversion behavior, and generate path connectivity verification results; Based on the path connectivity verification results, regenerate the path consistency collapse results after stripping, compare the path consistency collapse results after stripping with the path consistency collapse results before stripping, and if the path consistency collapse results after stripping disappear or the corresponding transformation field inheritance relationship is interrupted, mark the current stripped object as a necessary inheritance object. Repeat the process of selecting, temporarily removing, checking path connectivity, and comparing consistency of the current stripped object until all common state nodes and common behavior migration chains in the list of objects to be verified have been stripped and verified, generating the necessary set of receiving objects. The necessary common state nodes are used as conversion receiving nodes, and the necessary common behavior migration chains are used as conversion receiving edges. The conversion receiving nodes and conversion receiving edges are connected according to the path order to generate a conversion receiving chain. The advertising identifier, scene field, conversion field receiving mark and necessity verification mark are written into it to form a conversion receiving structure.

8. The advertising effectiveness evaluation method based on multimodal behavior fusion according to claim 1, characterized in that, The output advertising performance score, potential behavioral paths, path convergence trends, and key behavioral impact information include: Read the set of potential behavior evolution paths, path consistency collapse results, collapse strength and conversion acceptance structure, and extract the path identifier, projection source user identifier, advertisement identifier, scene field, path order identifier and conversion field acceptance mark corresponding to each potential behavior evolution path; Based on the set of potential behavioral evolution paths, the number of projected paths, effective paths, and transformation paths corresponding to the target user are counted. Based on the path consistency collapse results, the number of paths covered by the transformation consensus body, the number of common state nodes, and the number of common behavioral migration chains are counted, and path convergence statistics are generated. Based on the transformation and acceptance structure, read the transformation and acceptance nodes, transformation and acceptance edges, acceptance markers, and necessity verification markers. Count the number of times the transformation and acceptance nodes appear in the potential behavior evolution path set, the number of consecutive acceptances of the transformation and acceptance edges, and the number of acceptances of the transformation fields to generate key behavior impact records. Based on the number of effective paths, the number of conversion paths, the collapse strength, the number of paths covered by the conversion consensus body, the number of times conversion receiving nodes appear, the number of consecutive receiving times of conversion receiving edges, and the number of receiving times of conversion fields, path coverage scores, path convergence scores, and conversion receiving scores are generated respectively. These scores are then weighted and synthesized according to preset scoring weights to generate an advertising performance score. The system associates and writes advertising performance scores, potential behavior evolution path sets, path convergence statistics, conversion structure, and key behavior impact records to generate advertising performance evaluation results, outputting advertising performance scores, potential behavior paths, path convergence trends, and key behavior impact information.

Images