E-commerce platform advertisement effect tracking analysis method based on big data technology

Abstract

The invention discloses an e-commerce platform advertisement effect tracking analysis method based on a big data technology, and the method comprises the following steps: S1, collecting user data, and constructing a user behavior graph; s2, introducing an advertisement intervention path, comparing propagation states with or without advertisement intervention, calculating a disturbance consistency value and labeling an edge weight, and forming a structure chart; s3, inputting the structure diagram into a diagram causal reasoning network, and outputting advertisement individual effect estimation; s4, constructing a behavior disturbance sampler, generating a plurality of anti-fact behavior path training samples, and improving the causal fitting capability of the model; s5, executing path-level attribution on the actual user behavior graph by using the trained model, and outputting a causal weight matrix; and S6, generating an optimized advertisement putting strategy according to an attribution result. The advertisement causal attribution precision is improved, the putting strategy is optimized, and the conversion efficiency and the budget utilization rate are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of data mining and advertising technology, and in particular to a method for tracking and analyzing advertising effects on an e-commerce platform based on big data technology. Background Art

[0002] With the rapid development of e-commerce platforms, advertising has become a crucial tool for increasing conversion rates and boosting commercial revenue. The introduction of big data technologies has enabled platforms to record and analyze massive amounts of user behavior data, supporting advertising effectiveness evaluation. Existing advertising effectiveness evaluation methods mostly rely on superficial statistical indicators such as click-through rate and conversion rate, often overlooking the deeper impact of advertising on users' ultimate conversion decisions. Especially when user behavior chains are long and advertising intervention points are numerous, relying solely on static indicators or simple correlation analysis makes it difficult to accurately determine whether advertising has truly led to user purchases.

[0003] Traditional approaches often employ time series modeling or behavioral prediction methods based on recommendation systems. However, these approaches focus more on the order in which user behaviors occur rather than on how advertising insertions alter behavioral pathways. Furthermore, while causal reasoning is gradually being introduced into the field of advertising attribution, most current methods rely on strong modeling assumptions, lacking adaptability and struggling to function within complex and dynamic user behavior networks. This is especially true when faced with numerous perturbations and repeated exploratory browsing behaviors, making it difficult for models to determine which behaviors are directly triggered by advertising and which are simply a continuation of natural behavior.

[0004] On the other hand, the development of advertising strategies is generally plagued by a serious problem of generalization. Platforms often make display decisions based on historical conversion performance or user tags, but rarely consider the actual contribution of the ad in the user's behavioral path. In other words, even if an ad is shown multiple times, if it does not significantly drive conversions, existing systems find it difficult to distinguish such invalid impressions. Furthermore, the behavioral path is complex, with multiple hops, concurrency, and branches. Different ads may also compete for the same target, leading to priority conflicts and resource waste in delivery strategies.

[0005] Overall, current e-commerce advertising effectiveness analysis faces three major challenges: First, there is a lack of in-depth modeling of the causal impact of advertising, making it impossible to identify the true role of advertising in behavioral pathways; second, the coarse granularity of path attribution fails to reflect individual user differences and path perturbations; and third, the rigid logic of strategy optimization lacks a dynamic adjustment mechanism based on causal contributions. Therefore, it is necessary to propose a new method that can combine user behavior graphs with advertising intervention pathways. While ensuring the semantic integrity of behavior, it can measure the true effectiveness of advertising in different contexts and generate more reasonable and precise delivery strategies based on this.

[0006] Therefore, how to provide an e-commerce platform advertising effect tracking and analysis method based on big data technology is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0007] One purpose of the present invention is to propose a method for tracking and analyzing the advertising effects of e-commerce platforms based on big data technology. The present invention makes full use of big data technology, user behavior graph modeling, perturbation consistency calculation and multi-path causal estimation technology, and describes in detail how to construct a behavior graph structure with advertising intervention labels, and realizes causal attribution analysis of advertising nodes on user conversion behavior based on a graph neural inference network. It has the advantages of high attribution accuracy, strong interpretability and strong strategy adaptability.

[0008] The method for tracking and analyzing the advertising effectiveness of an e-commerce platform based on big data technology according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect user data and build a user behavior graph, where nodes represent behavior events and edges represent behavior sequence relationships;

[0010] S2. Introduce advertising intervention simulation paths into the user behavior graph. Compare the propagation states of the same behavior edge with and without advertising intervention. Calculate the perturbation consistency value based on the state difference. Annotate the perturbation consistency value as the dynamic weight of the edge in the user behavior graph to form a structural graph with perturbation consistency labels.

[0011] S3. Build a graph causal inference network, taking the structure graph as input. The graph causal inference network includes an edge weight adjustment layer, a path-aware convolutional layer, and a causal estimation layer. The edge weight adjustment layer adjusts the information propagation intensity based on the perturbation consistency value. The path-aware convolutional layer extracts the multi-hop causal path features from advertising to conversion. The causal estimation layer outputs the estimated value of the individual effect of the user-level advertisement.

[0012] S4. Design a behavioral perturbation sampler. Set several intervention scenarios based on ad type, location, and frequency. Generate a complete set of counterfactual behavior paths in the behavior graph as training data to train the causal inference network to fit the causal relationship between different ad intervention paths and user behavior conversion.

[0013] S5. Use the trained causal inference network to perform path-level attribution reasoning on the actual user behavior graph, outputting a causal weight matrix of ad nodes and conversion results, and forming a nested attribution result that can be interpreted hierarchically.

[0014] S6. Generate an optimized advertising delivery strategy based on the nested attribution results.

[0015] Optionally, the user data collected in step S1 includes: collecting behavioral event sequence data of users on the e-commerce platform, wherein the behavioral events include identifiable actions such as page browsing, product clicking, ad exposure, adding to cart, favorites, ordering and payment, and constructing a user behavior graph according to user dimensions, wherein nodes represent behavioral events, edges represent the temporal relationship of the behaviors, and the number of times the advertising node is displayed and the event path of whether the display successfully leads to conversion behavior are recorded, which is used to calculate the actual conversion rate. The actual conversion rate is the ratio of the number of times the conversion node is successfully triggered after the advertising node is actually displayed to the total number of displays.

[0016] Optionally, step S2 specifically includes:

[0017] S21. Construct a user behavior graph G = (V, E) based on the user behavior data, where the node set V represents user behavior events, the edge set E represents the temporal sequence relationship between behavior events, and the user behavior graph is a directed graph structure;

[0018] S22, for each behavior edge e in the user behavior graph i,j ∈E, construct the following two types of path environments respectively:

[0019] S221, copy the included behavior edge e i,j The original user behavior path P orig , as the base path for building the path environment;

[0020] S222, in constructing advertising intervention path environment P adv When edge e in the original user behavior path m,n With e n,p Insert an advertising behavior node v between adv , so that the path topology changes to e m,n →e n,adv →e adv,p , where v adv ∈V represents the ad exposure event node. After the update, all downstream propagation edges in the path are renumbered and the timestamp order is rebuilt;

[0021] S223, in the construction of non-advertising intervention path environment P noadv When maintaining the original path P orig The structure and node order of remain unchanged, and the path does not contain any new advertising behavior nodes;

[0022] S224, recalculate the behavior edge e in the two types of paths i,j The propagation state vectors of adjacent nodes are recorded as and The propagation state is generated based on the combination of the node's contextual behavior characteristics and the node state, and the length of the state sequence is the number of time steps T;

[0023] S23, based on the propagation state vector and Calculate the perturbation consistency value PC(e i,j );

[0024] S24, the disturbance consistency value PC (e i,j ) as edge e i,j The dynamic edge weights of are annotated in the graph structure G, forming a graph G with perturbation consistency labels PC =(V,E PC ), where the edge set E PC Contains perturbation label information.

[0025] Optionally, step S3 specifically includes:

[0026] S31, the structure graph G with perturbation consistency label PC =(V,E PC ) is input into the graph causal reasoning network, where each edge e i,j ∈E PC The perturbation consistency value PC(e i,j ) is used as the initial propagation weight of the edge;

[0027] S32, in the edge weight adjustment layer, according to each edge e in the structure graph i,j ∈E PC The perturbation consistency value PC(e i,j ), construct the edge propagation weight set W PC , the propagation weight is used to adjust the propagation strength of the node state on the edge of the structure graph, where each edge e i,j The propagation weight is set to w i,j =PC(e i,j ), in the process of updating the node status in the graph structure, the propagation weight w i,j Participate in the side information aggregation operation as a multiplicative factor;

[0028] S33, in the path-aware convolutional layer, the initial node state representation of the node set V is used as input, combined with the edge propagation weight set W PC , perform state updates on multi-hop propagation paths, extract the causal path features from the advertising node to the conversion node in the user behavior graph layer by layer, and control the influence range of each hop state propagation based on the disturbance consistency value, dynamically integrating the node states and path structure characteristics of upstream and downstream in each hop propagation;

[0029] S34, in the causal estimation layer, the feature vector of the node state output by the path perception convolution layer is combined with the advertising node position identification matrix M advCombined, the advertising node position identification matrix M adv It is a one-dimensional Boolean vector. When the value of the node i is 1, it indicates that the corresponding node v i It belongs to the advertising node. When the value is 0, it means it is not an advertising node. adv Generate user-level estimates of individual advertising effects in, represents the user conversion improvement effect of node v∈V under advertising intervention;

[0030] S35. Combine the estimated values ​​of the individual effects of user-level advertisements of all nodes into a causal response vector in,[·] T Represents the transposed matrix.

[0031] Optionally, the behavior perturbation sampler designed in step S4 sets several advertising intervention scenarios in the user behavior graph, and constructs a set of counterfactual behavior paths with the same structure as the original path but different intervention conditions by inserting or removing advertising nodes before specific behavior nodes.

[0032] Optionally, the step S5 includes: using the graph causal reasoning network trained by the counterfactual behavior path set generated by the behavior perturbation sampler as the reasoning model, receiving the structure graph G PC =(V,E PC ) as input, outputs the causal response vector composed of the estimated value of the individual advertising effect of each node, extracts the effective propagation path set P starting from the advertising node and ending at the conversion node according to the advertising node position identification matrix, and compares the individual estimated effect value of the node in the path with the disturbance consistency value PC(e i,j ) are jointly calculated to generate the causal weight C(i,j) of each path, and the causal weight matrix C from the advertising node to the conversion node is constructed in summary, and finally a nested attribution result that can be interpreted in a hierarchical manner is output.

[0033] Optionally, the advertising delivery strategy of step S6 specifically includes:

[0034] Advertisement node screening and frequency control strategy based on attribution strength: extract all ad nodes v that satisfy C(i,j)≥attribution strength threshold θ1 from the causal weight matrix C i As candidate optimization nodes, for these nodes, there is an edge perturbation consistency value PC(e m,n )≤disturbance consistency value threshold θ2, it is determined that the path disturbance stability is insufficient, and the display frequency is adjusted to not exceed the frequency threshold θ3;

[0035] Material type replacement strategy based on causal deviation: identifying the actual conversion rate R in candidate advertising nodesi Nodes that deviate from the attributed value C(i,j) satisfy R i <δ·C(i,j), where δ∈(0,1) is the attribution conversion deviation coefficient, then the ad creative type is replaced with a creative category that matches the same ad node in the path;

[0036] Advertisement position reordering strategy based on path competition: For the existence of several attribution weight values ​​C(i,j)≥θ1 and the terminal node v j For the same set of ad paths, the ad placements are rearranged according to the timestamp order of the user behavior events corresponding to the starting nodes of each path, giving priority to placements that were reached earlier in the user behavior graph.

[0037] Dynamic triggering strategy based on path stability: For all paths, the edge perturbation consistency value PC(e m,n )>θ2, extract the feature vector of the node state output by the path-aware convolutional layer in the behavior path, construct the feature representation of the path, and perform cosine similarity comparison with the feature mean vector of the similar advertising node path in the set of counterfactual behavior paths in the historical records. If the similarity is greater than the similarity threshold θ4, the display operation of the advertising node is dynamically triggered.

[0038] The beneficial effects of the present invention are:

[0039] (1) By constructing a user behavior graph with perturbation consistency labels, the present invention can distinguish the actual impact of advertising intervention on user behavior paths. This approach no longer relies on single click or conversion statistics, but analyzes the differences in advertising path propagation under intervention and non-intervention situations from the perspective of the user's complete behavior chain, which is closer to the actual value of advertising in real scenarios.

[0040] (2) The present invention introduces a graph causal reasoning network and combines it with an edge weight adjustment mechanism and a path-aware structure to solve the problems of unclear attribution and distorted information dissemination in complex paths in traditional algorithms. The present invention can capture the multi-hop causal path from advertising to conversion in a graph structure and output causal contribution results with hierarchical explanatory power, thereby improving the accuracy and credibility of the attribution process.

[0041] (3) In the advertising strategy generation process, the present invention not only focuses on historical data, but also introduces counterfactual path sampling and fitting training of individual behavior structures to form a combination of delivery strategies for different situations; the present invention can flexibly decide on display actions based on the causal weight of the advertising node and the current user status, thereby avoiding ineffective delivery, improving resource utilization efficiency, and achieving more responsive dynamic optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0043] Figure 1 This is the overall flow chart of the e-commerce platform advertising effect tracking and analysis method based on big data technology proposed by the present invention;

[0044] Figure 2 This is a flowchart of the dynamic optimization of advertising strategies for the e-commerce platform advertising effect tracking and analysis method based on big data technology proposed by the present invention. DETAILED DESCRIPTION

[0045] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0046] refer to Figure 1 and Figure 2 The e-commerce platform advertising effect tracking and analysis method based on big data technology includes the following steps:

[0047] S1. Collect user data and build a user behavior graph, where nodes represent behavior events and edges represent behavior sequence relationships;

[0048] S2. Introduce advertising intervention simulation paths into the user behavior graph. Compare the propagation states of the same behavior edge with and without advertising intervention. Calculate the perturbation consistency value based on the state difference. Annotate the perturbation consistency value as the dynamic weight of the edge in the user behavior graph to form a structural graph with perturbation consistency labels.

[0049] S3. Build a graph causal inference network, taking the structure graph as input. The graph causal inference network includes an edge weight adjustment layer, a path-aware convolutional layer, and a causal estimation layer. The edge weight adjustment layer adjusts the information propagation intensity based on the perturbation consistency value. The path-aware convolutional layer extracts the multi-hop causal path features from advertising to conversion. The causal estimation layer outputs the estimated value of the individual effect of the user-level advertisement.

[0050] S4. Design a behavioral perturbation sampler. Set several intervention scenarios based on ad type, location, and frequency. Generate a complete set of counterfactual behavior paths in the behavior graph as training data to train the causal inference network to fit the causal relationship between different ad intervention paths and user behavior conversion.

[0051] S5. Use the trained causal inference network to perform path-level attribution reasoning on the actual user behavior graph, outputting a causal weight matrix of ad nodes and conversion results, and forming a nested attribution result that can be interpreted hierarchically.

[0052] S6. Generate an optimized advertising delivery strategy based on the nested attribution results.

[0053] This paper proposes an advertising effectiveness analysis method that combines user behavior paths with advertising intervention factors. By introducing simulated paths and intervention markers into the behavior graph, it not only shows whether an ad has been seen, but also further analyzes the actual impact of the ad on the user's final conversion. Unlike traditional methods that only look at clickthrough rates or conversion rates, this approach can more comprehensively reflect the actual value of the ad and provide a more detailed basis for strategy optimization.

[0054] In this embodiment, the user data collected in step S1 includes: collecting user behavioral event sequence data on the e-commerce platform, wherein the behavioral events include recognizable actions such as page browsing, product clicking, advertising exposure, adding to cart, favorites, ordering and payment, and constructing a user behavior graph according to the user dimension, in which nodes represent behavioral events, edges represent the temporal relationship of the occurrence of behaviors, and the number of times the advertising node is displayed and the event path of whether it successfully leads to conversion behavior after the display is recorded, which is used to calculate the actual conversion rate. The actual conversion rate is the ratio of the number of times the conversion node is successfully triggered after the advertising node is actually displayed to the total number of displays.

[0055] This paper establishes a behavioral graph model by structuredly collecting user behavioral events and introduces a method for recording ad node display and subsequent conversion behavior. The advantage of this approach is that the platform no longer relies on isolated click or order data, but instead extracts the actual conversion path from the behavioral chain, thereby calculating ad conversion rates that are more closely aligned with actual user behavior, improving the accuracy of the analysis.

[0056] In this embodiment, step S2 specifically includes:

[0057] S21. Construct a user behavior graph G = (V, E) based on the user behavior data, where the node set V represents user behavior events, the edge set E represents the temporal sequence relationship between behavior events, and the user behavior graph is a directed graph structure;

[0058] S22, for each behavior edge e in the user behavior graph i,j ∈E, construct the following two types of path environments respectively:

[0059] S221, copy the included behavior edge e i,j The original user behavior path P orig , as the base path for building the path environment;

[0060] S222, in constructing advertising intervention path environment P adv When edge e in the original user behavior path m,n With e n,pInsert an advertising behavior node v between adv , so that the path topology changes to e m,n →e n,adv →e adv,p , where v adv ∈V represents the ad exposure event node. After the update, all downstream propagation edges in the path are renumbered and the timestamp order is rebuilt;

[0061] S223, in the construction of non-advertising intervention path environment P noadv When maintaining the original path P orig The structure and node order of remain unchanged, and the path does not contain any new advertising behavior nodes;

[0062] S224, recalculate the behavior edge e in the two types of paths i,j The propagation state vectors of adjacent nodes are recorded as and The propagation state is generated based on the combination of the node's contextual behavior characteristics and the node state, and the length of the state sequence is the number of time steps T;

[0063] S23, based on the propagation state vector and Calculate the perturbation consistency value PC(e i,j ), calculated as:

[0064]

[0065] Among them, α t =exp(-λt) is the time attenuation factor, λ>0 is the preset attenuation coefficient, ω i,j is the behavior edge sensitivity weight, which reflects the sensitivity of the behavior type to the disturbance. is the normalization factor used to ensure that PC(e i,j )∈[0,1], ||·||2 represents the L2 norm;

[0066] This formula introduces a time decay factor and edge sensitivity weights to perform a weighted cumulative calculation of the differences in the state vectors of behavioral edges before and after an ad intervention, innovatively measuring the perturbation response strength of user behavior paths with and without ad intervention. Unlike traditional algorithms that rely on static state differences or single path similarity, this formula not only dynamically reflects the decreasing impact of an intervention over time but also adaptively adjusts the evaluation weights based on the importance of the behavioral event. This allows for more accurate and interpretable edge-level perturbation modeling while maintaining the stability of the normalized output, significantly improving causal reasoning's ability to identify path contributions.

[0067] S24, the disturbance consistency value PC (e i,j ) as edge ei,j The dynamic edge weights of are annotated in the graph structure G, forming a graph G with perturbation consistency labels PC =(V,E PC ), where the edge set E PC Contains perturbation label information.

[0068] This paper proposes a method for determining the impact of advertising interventions based on behavioral changes before and after the intervention, quantifying this impact using perturbation consistency values. Specifically, by incorporating a weighting mechanism based on time decay and behavioral sensitivity, the model not only determines whether an ad is effective but also identifies the behavioral nodes at which the effect is most pronounced. Compared to traditional graph modeling approaches, this approach makes subsequent causal analysis more stable and interpretable.

[0069] In this embodiment, step S3 specifically includes:

[0070] S31, the structure graph G with perturbation consistency label PC =(V,E PC ) is input into the graph causal reasoning network, where each edge e i,j ∈E PC The perturbation consistency value PC(e i,j ) is used as the initial propagation weight of the edge;

[0071] S32, in the edge weight adjustment layer, according to each edge e in the structure graph i,j ∈E PC The perturbation consistency value PC(e i,j ), construct the edge propagation weight set W PC , the propagation weight is used to adjust the propagation strength of the node state on the edge of the structure graph, where each edge e i,j The propagation weight is set to w i,j =PC(e i,j ), in the process of updating the node status in the graph structure, the propagation weight w i,j It participates in the edge information aggregation operation as a multiplicative factor, so that the edge with a higher perturbation consistency value has a greater impact on the state propagation, and the edge with a lower perturbation consistency value has a weaker contribution in the propagation;

[0072] S33, in the path-aware convolutional layer, the initial node state representation of the node set V is used as input, combined with the edge propagation weight set W PC , perform state updates on multi-hop propagation paths, extract the causal path features from the advertising node to the conversion node in the user behavior graph layer by layer, and control the influence range of each hop state propagation based on the disturbance consistency value, dynamically integrating the node states and path structure characteristics of upstream and downstream in each hop propagation;

[0073] S34, in the causal estimation layer, the feature vector of the node state output by the path-aware convolutional layer is combined with the advertising node position identification matrix M adv The advertising node position identification matrix M adv is a one-dimensional Boolean vector, which indicates which nodes in the graph structure are advertising intervention nodes, and when the value of the i-th position of the vector is 1, it indicates that the corresponding node v i is an advertising node, and a value of 0 indicates a non-advertising node; based on M adv generate user-level advertising individual effect estimation values wherein, represents the user conversion promotion effect of node v∈V in the advertising intervention scenario;

[0074] S35, combine the user-level advertising individual effect estimation values of all nodes into a causal response vector wherein, [·] T represents the transposed matrix.

[0075] The present application builds a graph neural network architecture specially used for advertising attribution analysis, which can automatically identify the influence path of advertising nodes on the final conversion in user behavior by setting edge weight adjustment and path awareness module. Compared with conventional machine learning models, this structure is more suitable for processing user behavior data with time and structure dependence, and is particularly suitable for multi-hop causal identification in advertising tasks.

[0076] In the present embodiment, the behavior disturbance sampler designed in step S4 sets several advertising intervention scenarios in the user behavior graph, and constructs a counterfactual behavior path set with the same original path structure but different intervention conditions by inserting or removing advertising nodes before specific behavior nodes.

[0077] A disturbance sampling mechanism is designed in the present application, which can construct counterfactual behavior paths containing advertising intervention and non-intervention scenarios without affecting the original behavior graph structure. This mechanism allows the model to "see" the influence of various advertising combinations on user behavior, thereby obtaining more robust and widely applicable judgment ability during the training process, which greatly helps to improve the accuracy of attribution analysis.

[0078] In the present embodiment, step S5 includes: using the graph causal reasoning network trained by the counterfactual behavior path set generated by the behavior disturbance sampler as a reasoning model, receiving the structure graph G PC =(V,E PC) as input, outputs the causal response vector composed of the estimated value of the individual advertising effect of each node, extracts the effective propagation path set P starting from the advertising node and ending at the conversion node according to the advertising node position identification matrix, and compares the individual estimated effect value of the node in the path with the disturbance consistency value PC(e i,j ) are jointly calculated to generate the causal weight C(i,j) of each path, and the causal weight matrix C from the advertising node to the conversion node is constructed in summary, and finally a nested attribution result that can be interpreted in a hierarchical manner is output.

[0079] Through the trained causal model, the present invention establishes a causal connection between the advertising nodes and conversion results in each user behavior path and outputs a clear attribution matrix. This result not only tells us which advertisement played a role, but also explains in layers at which step it began to influence user decisions. This approach is more instructive than simply outputting a conversion probability and makes subsequent strategy adjustments more targeted.

[0080] In this embodiment, the advertising delivery strategy of step S6 specifically includes:

[0081] Advertisement node screening and frequency control strategy based on attribution strength: extract all ad nodes v that satisfy C(i,j)≥attribution strength threshold θ1 from the causal weight matrix C i As candidate optimization nodes, for these nodes, there is an edge perturbation consistency value PC(e m,n )≤disturbance consistency value threshold θ2, it is determined that the path disturbance stability is insufficient, and the display frequency is adjusted to not exceed the frequency threshold θ3;

[0082] Material type replacement strategy based on causal deviation: identifying the actual conversion rate R in candidate advertising nodes i Nodes that deviate from the attributed value C(i,j) satisfy R i <δ·C(i,j), where δ∈(0,1) is the attribution conversion deviation coefficient, then the ad creative type is replaced with a creative category that matches the same ad node in the path;

[0083] Advertisement position reordering strategy based on path competition: For the existence of several attribution weight values ​​C(i,j)≥θ1 and the terminal node v j For the same set of ad paths, the ad placements are rearranged according to the timestamp order of the user behavior events corresponding to the starting nodes of each path, giving priority to placements that were reached earlier in the user behavior graph.

[0084] Dynamic triggering strategy based on path stability: For all paths, the edge perturbation consistency value PC(e m,n)>θ2, extract the feature vector of the node state output by the path-aware convolutional layer in the behavior path, construct the feature representation of the path, and perform cosine similarity comparison with the feature mean vector of the same type of advertising node path in the set of counterfactual behavior paths in the historical records. If the similarity is greater than the similarity threshold θ4, the display operation of the advertising node is dynamically triggered. The display operation includes: submitting a display request to the advertising scheduling system, loading the display material corresponding to the advertising node, and executing the advertising content presentation in the user interface. At the same time, the display behavior is recorded as the user behavior graph update node.

[0085] Based on the attribution results, the present invention proposes a variety of specific delivery optimization methods, including adjusting display frequency, replacing advertising materials, reordering advertising positions, and setting dynamic trigger conditions. Each strategy is executed based on the advertising effect results actually inferred by the model. This advertising strategy optimization method driven by causal reasoning is more in line with business logic and can significantly improve the efficiency of advertising resource utilization.

[0086] Example 1:

[0087] In order to verify the feasibility of the present invention in implementation, the present invention is applied to the homepage recommendation advertising delivery system of an e-commerce platform. Through modeling and advertising attribution analysis of real user behavior data, the actual performance of traditional CTR (CTR is a traditional strategy for evaluating and optimizing advertising effects based on click-through rate) click-through rate logic and the causal reasoning method of the present invention are compared in terms of attribution accuracy, strategy optimization effect, and delivery resource utilization.

[0088] The platform's advertising system originally employed a strategy optimization approach based on statistical indicators, primarily evaluating ad effectiveness through click-through rate and conversion rate. However, in practice, ads with high click-through rates but very low conversion rates often resulted in ads being clicked but not actually driving users to complete orders or payments. This was particularly true for behavior paths involving multiple views and cross-category switching, making it difficult for the platform to determine whether an ad node had a causal effect on the final purchase, leading to wasted budgets and ineffective strategy adjustments.

[0089] This paper processes the behavioral data of approximately 2.85 million users on the platform over the past month, extracting the complete user behavior path, including events such as browsing, adding to cart, ad exposure, clicks, and ordering, resulting in approximately 41 million behavioral edges. This data is then used to construct a user behavior graph, where advertising node insertion points include multiple locations, such as the recommended position on the homepage, the top of the search page, and the middle of the product details page. To measure the actual impact of advertising under different intervention conditions, we introduce simulated advertising intervention paths into the platform's behavior graph. By calculating the perturbation consistency value and dynamically annotating the edge weights of the behavioral edges, we construct a structurally complete input graph structure with semantic labels.

[0090] Next, we input the training data into the proposed graph causal inference network and train it using multiple ad scenario paths generated by counterfactual sampling. The trained model outputs estimates of the individual causal effects of each ad node on the conversion event, ultimately generating an attribution matrix. Using these attribution results, we reorder the ads currently displayed in the system, replace creatives for ads with high causal weight but low conversion rates, and reposition or restrict trigger conditions for ads with high causal paths but high path perturbations, ultimately optimizing the delivery strategy.

[0091] In order to compare the effects, we tested the present invention in parallel with the platform's original CTR delivery logic. The control results show that in a week-long A / B experiment, the click-through rate of ads under the system of the present invention decreased slightly, but the order conversion rate increased significantly, with an average increase of 3.8%, an average increase of 1.4 items per person, and an increase of 12.6% in the efficiency of advertising budget utilization. More importantly, in highly competitive ad positions, the strategy of the present invention effectively avoids repeated display of similar ads, optimizes ad selection through the attribution weight matrix, and improves the overall display diversity and effectiveness of advertising paths. The average ad path depth of the system was reduced from 3.2 hops to 2.4 hops, proving that path compression optimization is effective. At the same time, some "accidental touch ads" that have no substantial impact on conversions were filtered out through perturbation consistency weight calculation, and the user's stay time in the controlled area increased by nearly 5.2 seconds, further illustrating that the delivery content is more in line with behavioral interests.

[0092] The following table shows the comparative data of the method of the present invention and the traditional method under key indicators in the above experiments:

[0093] Table 1: Comparison of advertising effect attribution and strategy optimization based on graph causal reasoning

[0094] Indicator Category Original CTR strategy system Method of the present invention Index change range Average click-through rate (%) 7.24 7.19 -0.7％ Order conversion rate (%) 2.15 2.23 +3.8％ Average ad impressions 5.6 4.9 -12.5％ Average number of orders per person 2.2 3.6 +63.6％ Average path hop count 3.2 2.4 -25％ Efficiency of advertising budget utilization 81.7％ 92.0％ +12.6％ User stay time (seconds) 62.1 67.3 +5.2 seconds High causal advertising hit rate none 91.3％ /

[0095] Comparing our approach with traditional CTR strategies, while the average click-through rate decreased slightly from 7.24% to 7.19%, the overall ad quality and conversion efficiency showed more positive changes. The order conversion rate increased from 2.15% to 2.23%, a 3.8% increase, demonstrating a stronger conversion rate. This improvement no longer relies on superficial clickthroughs, but rather, guided by causal analysis, precisely targeting ads at the pathways that directly impact the final action, thereby more effectively driving purchases.

[0096] From the perspective of advertising intensity, the average number of ad exposures decreased from 5.6 to 4.9, a 12.5% ​​decrease. This means that better conversion results were achieved with fewer exposure resources, demonstrating the advantages of this invention in terms of resource efficiency. At the same time, the average number of items ordered per person increased from 2.2 to 3.6, a 63.6% increase, further confirming the increased repurchase potential brought about by the causal enhancement of user behavior paths.

[0097] In terms of path structure optimization, this invention significantly reduces the depth of the advertising impact path from an average of 3.2 hops to 2.4 hops, a path compression of 25%. This reflects that the path for ads to reach target users is shorter and more direct, effectively avoiding the "ineffective jump" phenomenon and improving the efficiency of advertising information transmission.

[0098] Ad budget efficiency increased from 81.7% in the original system to 92.0%, a 12.6% increase. This demonstrates that the proposed causal attribution strategy more accurately guides budget allocation and reduces waste. Furthermore, the average user duration on the platform increased from 62.1 seconds to 67.3 seconds, a 5.2-second increase. This suggests that recommended content is more relevant to user interests and that ads do not cause users to quickly leave or bounce.

[0099] Finally, the newly added metric of high causal ad hit rate reached 91.3%, which is not reflected in traditional CTR strategies. This shows that this invention effectively identifies and prioritizes the ads with the greatest practical impact on conversions, thereby comprehensively improving the overall value of platform advertising.

[0100] In summary, the present invention not only improves delivery efficiency and user experience, but also realizes the transition of advertising strategy from "correlation logic" to "causality logic", forming a clear distinction in the existing technology.

[0101] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for tracking and analyzing the advertising effect of an e-commerce platform based on big data technology, characterized in that: The steps include: S1. Collect user data and build a user behavior graph, where nodes represent behavior events and edges represent behavior sequence relationships; S2. Introduce advertising intervention simulation paths into the user behavior graph. Compare the propagation states of the same behavior edge with and without advertising intervention. Calculate the perturbation consistency value based on the state difference. Annotate the perturbation consistency value as the dynamic weight of the edge in the user behavior graph to form a structural graph with perturbation consistency labels. S3. Build a graph causal inference network, taking the structure graph as input. The graph causal inference network includes an edge weight adjustment layer, a path-aware convolutional layer, and a causal estimation layer. The edge weight adjustment layer adjusts the information propagation intensity based on the perturbation consistency value. The path-aware convolutional layer extracts the multi-hop causal path features from advertising to conversion. The causal estimation layer outputs the estimated value of the individual effect of the user-level advertisement. S4. Design a behavioral perturbation sampler. Set several intervention scenarios based on ad type, location, and frequency. Generate a complete set of counterfactual behavior paths in the behavior graph as training data to train the causal inference network to fit the causal relationship between different ad intervention paths and user behavior conversion. S5. Use the trained causal inference network to perform path-level attribution reasoning on the actual user behavior graph, outputting a causal weight matrix of ad nodes and conversion results, and forming a nested attribution result that can be interpreted hierarchically. S6. Generate an optimized advertising delivery strategy based on the nested attribution results.

2. The method for tracking and analyzing the advertising effect of an e-commerce platform based on big data technology according to claim 1, characterized in that: The user data collected in step S1 includes: collecting user behavioral event sequence data on the e-commerce platform, wherein the behavioral events include identifiable actions such as page browsing, product clicking, ad exposure, adding to cart, favorites, ordering and payment, constructing a user behavior graph according to user dimensions, wherein nodes represent behavioral events, edges represent the temporal relationship of the behaviors, and the number of times an advertising node is displayed and the event path of whether the display successfully leads to conversion behavior is recorded, which is used to calculate the actual conversion rate. The actual conversion rate is the ratio of the number of times a conversion node is successfully triggered after the advertising node is actually displayed to the total number of displays.

3. The method for tracking and analyzing the advertising effect of an e-commerce platform based on big data technology according to claim 2 is characterized in that: The step S2 specifically includes: S21. Construct a user behavior graph G = (V, E) based on the user behavior data, where the node set V represents user behavior events, the edge set E represents the temporal sequence relationship between behavior events, and the user behavior graph is a directed graph structure; S22, for each behavior edge e in the user behavior graph i,j ∈E, construct the following two types of path environments respectively: S221, copy the included behavior edge e i,j The original user behavior path P orig , as the base path for building the path environment; S222, in constructing advertising intervention path environment P adv When edge e in the original user behavior path m,n With e n,p Insert an advertising behavior node v between adv , so that the path topology changes to e m,n →e n,adv →e adv,p , where v adv ∈V represents the ad exposure event node. After the update, all downstream propagation edges in the path are renumbered and the timestamp order is rebuilt; S223, in building a non-advertising intervention path environment P noadv When maintaining the original path P orig The structure and node order of remain unchanged, and the path does not contain any new advertising behavior nodes; S224, recalculate the behavior edge e in the two types of paths i,j The propagation state vectors of adjacent nodes are recorded as and The propagation state vector is generated based on the node's contextual behavior characteristics and the node state. The length of the state sequence is the number of time steps T. S23, based on the propagation state vector and Calculate the perturbation consistency value PC(e i,j ); S24, the disturbance consistency value PC (e i,j ) as edge e i,j The dynamic edge weights of are annotated in the graph structure G, forming a graph G with perturbation consistency labels PC =(V,E PC ), where the edge set E PC Contains perturbation label information.

4. The method for tracking and analyzing the advertising effect of an e-commerce platform based on big data technology according to claim 3 is characterized in that: The step S3 specifically includes: S31, the structure graph G with perturbation consistency label PC =(V,E PC ) is input into the graph causal reasoning network, where each edge e i,j ∈E PC The perturbation consistency value PC(e i,j ) is used as the initial propagation weight of the edge; S32, in the edge weight adjustment layer, according to each edge e in the structure graph i,j ∈E PC The perturbation consistency value PC(e i,j ), construct the edge propagation weight set W PC , the propagation weight is used to adjust the propagation strength of the node state on the edge of the structure graph, where each edge e i,j The propagation weight is set to w i,j =PC(e i,j ), in the process of updating the node status in the graph structure, the propagation weight w i,j Participate in the side information aggregation operation as a multiplicative factor; S33, in the path-aware convolutional layer, the initial node state representation of the node set V is used as input, combined with the edge propagation weight set W PC , perform state updates on multi-hop propagation paths, extract the causal path features from the advertising node to the conversion node in the user behavior graph layer by layer, and control the influence range of each hop state propagation based on the disturbance consistency value, dynamically integrating the node states and path structure characteristics of upstream and downstream in each hop propagation; S34, in the causal estimation layer, the feature vector of the node state output by the path perception convolution layer is combined with the advertising node position identification matrix M adv Combined, the advertising node position identification matrix M adv It is a one-dimensional Boolean vector. When the value of the node i is 1, it indicates that the corresponding node v i It belongs to the advertising node. When the value is 0, it means it is not an advertising node. adv Generate user-level estimates of individual advertising effects in, represents the user conversion improvement effect of node v∈V under advertising intervention; S35. Combine the estimated values ​​of the individual effects of user-level advertisements of all nodes into a causal response vector in,[·] T Represents the transposed matrix.

5. The method for tracking and analyzing the advertising effect of an e-commerce platform based on big data technology according to claim 4 is characterized in that: The behavior perturbation sampler designed in step S4 sets several advertising intervention scenarios in the user behavior graph, and constructs a set of counterfactual behavior paths with the same structure as the original path but different intervention conditions by inserting or removing advertising nodes before specific behavior nodes.

6. The method for tracking and analyzing the advertising effect of an e-commerce platform based on big data technology according to claim 5 is characterized in that: The step S5 includes: using the graph causal reasoning network trained by the counterfactual behavior path set generated by the behavior perturbation sampler as the reasoning model, receiving the structure graph G PC =(V,E PC ) as input, outputs the causal response vector composed of the estimated value of the individual advertising effect of each node, extracts the effective propagation path set P starting from the advertising node and ending at the conversion node according to the advertising node position identification matrix, and compares the individual estimated effect value of the node in the path with the disturbance consistency value PC(e i,j ) are jointly calculated to generate the causal weight C(i,j) of each path, and the causal weight matrix C from the advertising node to the conversion node is constructed in summary, and finally a nested attribution result that can be interpreted in a hierarchical manner is output.

7. The method for tracking and analyzing the advertising effect of an e-commerce platform based on big data technology according to claim 6 is characterized in that: The advertising delivery strategy in step S6 specifically includes: Advertisement node screening and frequency control strategy based on attribution strength: extract all ad nodes v that satisfy C(i,j)≥attribution strength threshold θ1 from the causal weight matrix C i As candidate optimization nodes, for these nodes, there is an edge perturbation consistency value PC(e m,n )≤disturbance consistency value threshold θ2, it is determined that the path disturbance stability is insufficient, and the display frequency is adjusted to not exceed the frequency threshold θ3; Material type replacement strategy based on causal deviation: identifying the actual conversion rate R in candidate advertising nodes i Nodes that deviate from the attributed value C(i,j) satisfy R i <δ·C(i,j), where δ∈(0,1) is the attribution conversion deviation coefficient, then the ad creative type is replaced with a creative category that matches the same ad node in the path; Advertisement position reordering strategy based on path competition: For the existence of several attribution weight values ​​C(i,j)≥θ1 and the terminal node v j For the same set of ad paths, the ad placements are rearranged according to the timestamp order of the user behavior events corresponding to the starting nodes of each path, giving priority to placements that were reached earlier in the user behavior graph. Dynamic triggering strategy based on path stability: For all paths, the edge perturbation consistency value PC(e m,n )>θ2, extract the feature vector of the node state output by the path-aware convolutional layer in the behavior path, construct the feature representation of the path, and perform cosine similarity comparison with the feature mean vector of the similar advertising node path in the set of counterfactual behavior paths in the historical records. If the similarity is greater than the similarity threshold θ4, the display operation of the advertising node is dynamically triggered.

Images