A Marketing Link Shortcoming Location Method and System Based on Graph Neural Networks
By constructing a heterogeneous marketing link graph and introducing pseudo-activity, pendulum decision-making, and semantic mirroring factors, the problems of user behavior correlation and path propagation effect in the positioning of marketing link shortcomings are solved, and the precise positioning and root cause tracing of marketing nodes are achieved.
Patent Information
- Application Number
- CN202611000024.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-07
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-07-07
AI Technical Summary
Existing technologies fail to effectively consider the rhythmic changes in user back-and-forth jumping behavior and the spatial distribution characteristics of interaction trajectories when identifying shortcomings in the marketing chain. This results in a high misjudgment rate in high-frequency back-and-forth, highly discrete interaction scenarios, making it impossible to accurately identify the real shortcomings in touchpoints.
Construct a heterogeneous marketing link graph, and through pseudo-activity factors, pendulum decision factors, and semantic mirror factors, combined with user behavior paths and touchpoint interaction relationships, build a short-board propagation equation to achieve precise positioning of marketing nodes.
It significantly improves the accuracy of identifying shortcomings in the marketing chain and the ability to trace the root causes. It can accurately identify high-frequency, low-efficiency pseudo-active touchpoints, decision fatigue, and cross-touchpoint misleading traps, distinguish the inherent defects of touchpoints from their contributions to upstream and downstream dissemination, and is suitable for marketing chain scenarios with complex jump structures.
Smart Images

Figure CN122507551B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of marketing link weakness detection technology, and in particular to a marketing link weakness localization method and system based on graph neural networks. Background Technology
[0002] A digital marketing journey typically consists of multiple sequentially linked marketing touchpoints, with users navigating through each step to ultimately complete a conversion. Accurately identifying and optimizing the weakest touchpoints in this journey is crucial for improving conversion rates.
[0003] In existing technologies, the identification of weaknesses in the marketing chain mainly relies on two types of methods: one is a macro-assessment based on statistical indicators, which identifies weaknesses by analyzing indicators such as bounce rate, dwell time, and click-through rate at each touchpoint. This type of method is biased in identifying weaknesses under complex behavioral patterns because it does not consider the rhythmic changes in users' back-and-forth jumping behavior between touchpoints and the spatial distribution characteristics of interaction trajectories, especially in scenarios with high frequency of back-and-forth and high discreteness of interaction. The other type is a machine learning method based on conversion rate prediction models, which identifies touchpoints with large prediction errors as weaknesses by training the model. However, it ignores the propagation effect of weaknesses on the jumping path and the correlation of user behavior between different touchpoints. For example, experience problems at upstream touchpoints can lead to a decrease in conversion rates at downstream touchpoints, and existing models cannot distinguish between the inherent defects of the touchpoints and the cumulative effect propagated from upstream. Summary of the Invention
[0004] The purpose of this invention is to provide a marketing link bottleneck location method and system based on graph neural networks, which aims to solve the problem that traditional marketing link bottleneck location technology is difficult to effectively identify touchpoints with real bottlenecks because it does not consider the rhythm changes of users' back-and-forth jumping behavior between touchpoints and the spatial distribution characteristics of interaction trajectories.
[0005] In a first aspect, the present invention provides a marketing link weakness identification method based on graph neural networks, the method comprising: Using user nodes, marketing nodes, and content element nodes as vertices, and user jumping behavior between different marketing nodes as the first side, and the interaction behavior of different content elements under the same marketing node as the second side, a heterogeneous marketing link graph is constructed, in which each marketing node includes at least one content element. The behavior sequence matrix of each marketing node is extracted based on the heterogeneous marketing link graph. The behavior sequence matrix includes an interaction trajectory matrix composed of the two-dimensional coordinates of each user's interaction point on the marketing node, a time interval matrix composed of the time interval between two adjacent interactions, a pendulum state transition matrix composed of the time and direction of the user's round-trip jump between the current marketing node and its marketing node, and a semantic pair composed of the semantic embedding vectors of the content elements on the current marketing node and the previous marketing node. The pseudo-activity factor of the user is obtained based on the interaction trajectory matrix and the time interval matrix; the pendulum decision factor is obtained based on the pendulum state transition matrix; and the semantic mirror factor is obtained based on the semantic pair. The short-board propagation equation for each marketing node is constructed based on the pseudo-activity factor, pendulum decision factor, and semantic mirror factor. The short-board propagation equation is then solved based on the heterogeneous marketing link graph to obtain the short-board value for each marketing node. Finally, the short-board node is determined based on the short-board value.
[0006] Secondly, the present invention provides a marketing link bottleneck location system based on graph neural networks, the system comprising: The heterogeneous graph construction module is used to construct a marketing link heterogeneous graph with user nodes, marketing nodes, and content element nodes as vertices, user jumping behavior between different marketing nodes as the first side, and interaction behavior of different content elements under the same marketing node as the second side. Each marketing node includes at least one content element. The behavior matrix extraction module is used to extract the behavior sequence matrix of each marketing node based on the marketing link heterogeneous graph. The behavior sequence matrix includes an interaction trajectory matrix composed of the two-dimensional coordinates of each user's interaction point on the marketing node, a time interval matrix composed of the time interval between two adjacent interactions, a pendulum state transition matrix composed of the time and direction of the user's round-trip jump between the current marketing node and its marketing node, and a semantic pair composed of the semantic embedding vectors of the content elements on the current marketing node and the previous marketing node. The factor calculation module is used to obtain the user's pseudo-activity factor based on the interaction trajectory matrix and the time interval matrix, obtain the pendulum decision factor based on the pendulum state transition matrix, and obtain the semantic mirror factor based on the semantic pair. The propagation equation construction module is used to construct the short-board propagation equation for each marketing node based on the pseudo-activity factor, pendulum decision factor, and semantic mirror factor, and solve the short-board propagation equation based on the heterogeneous marketing link graph to obtain the short-board value for each marketing node, and determine the short-board node based on the short-board value.
[0007] Thirdly, the present invention provides a storage medium that stores one or more programs, which, when executed by a processor, implement the above-described marketing link bottleneck location method based on graph neural networks.
[0008] Fourthly, the present invention provides an electronic device, the electronic device comprising a memory and a processor, wherein: The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the above-mentioned marketing link bottleneck location method based on graph neural network.
[0009] Compared with the prior art, the present invention has the following advantages: This invention significantly improves the accuracy and root cause analysis capabilities of marketing link weakness localization by constructing a weakness localization system that integrates heterogeneous graphs and multi-factor fusion. It effectively addresses the problems of existing technologies that neglect the propagation effect of weaknesses along the jump path and the correlation of user behavior between different touchpoints. Specifically, a heterogeneous graph of the marketing link is first constructed. This model uses user nodes, marketing nodes, and content element nodes as vertices and jump behavior and interaction behavior as binary edges to achieve a structured expression of user behavior paths and touchpoint interaction relationships. Then, pseudo-activity factors, pendulum decision factors, and semantic mirroring factors are introduced. The pseudo-activity factor is constructed based on the joint deviation of the spatial distribution and temporal rhythm of user interactions at touchpoints. Its core principle is that non-converted users often click around on the page but never reach the effective area, and their interaction rhythm is chaotic, while converted users' interactions are concentrated in the effective area and have a stable rhythm. By quantifying the differences between non-converted and converted users in these two aspects, the pseudo-activity factor can effectively identify high-frequency, low-efficiency pseudo-active touchpoints. The pendulum decision factor is constructed based on the combined decay ratio of the round-trip time between touchpoints and the direction of departure. Its core principle is that when a user repeatedly swings between upstream and downstream, if the swing speed increases and eventually exits, then... The system identifies a decision fatigue issue at the touchpoint. If the user ultimately moves forward, it's considered normal exploratory behavior. By quantifying the acceleration ratio and conversion direction, the pendulum decision factor can accurately distinguish between accelerated exit and normal progress. The semantic mirror factor is constructed based on the semantic projection distance and functional contrast of cross-touchpoint content elements. Its core principle is that when the last interacted content element at the previous touchpoint is visually similar to but functionally opposite to the first seen content element at the current touchpoint, users are prone to misoperation due to inertia. By quantifying semantic similarity and functional oppositeness, the semantic mirror factor can accurately pinpoint cross-touchpoint misleading traps, thus constructing a more comprehensive and accurate weakness quantification system. Based on this, a two-way weakness propagation equation is constructed, enabling weakness values to distinguish between the touchpoint's own defects and upstream / downstream propagation contributions. Through progressive screening rules, normal nodes, nodes with upstream root causes, and nodes with propagation attenuation are eliminated level by level. The final output weakness nodes are all defect root cause nodes rather than passive manifestation nodes, thus effectively applicable to marketing links with complex jump structures and multiple paths, achieving accurate positioning of marketing nodes with genuine weaknesses. Attached Figure Description
[0010] Figure 1 This is a flowchart of a marketing link bottleneck location method based on graph neural networks proposed in an embodiment of the present invention; Figure 2This is a schematic diagram of a marketing link bottleneck location system based on graph neural networks, proposed in an embodiment of the present invention.
[0011] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation
[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.
[0013] like Figure 1 As shown, an embodiment of the present invention proposes a marketing link weakness localization method based on graph neural networks. The method includes steps S101 to S104, wherein: Step S101: Using user nodes, marketing nodes, and content element nodes as vertices, user jumping behavior between different marketing nodes as the first side, and interaction behavior of different content elements under the same marketing node as the second side, a heterogeneous marketing link graph is constructed, where each marketing node includes at least one content element. It should be noted that in this step, user nodes represent individuals participating in the interaction, marketing nodes represent different marketing scenarios or touchpoints, and content element nodes represent specific interactive components such as images, videos, and buttons within a marketing node. The edges in the diagram are divided into two categories: the first edge is used to depict the macroscopic path flow of users between different marketing nodes, and the second edge is used to capture the microscopic interactive behavior of users with different content elements within the same marketing node.
[0014] Furthermore, in some embodiments, for any two marketing nodes, the total number of times all users jump from one marketing node to the other and the dwell time after each jump are obtained, and the first-side path coefficient is defined according to the following formula: ; in, Let be the path coefficient of the first edge from marketing node a to marketing node b. Let this be the duration of the user's stay after the k-th jump from the a-th marketing node to the b-th marketing node. This represents the total number of times a user jumps from marketing node a to marketing node b. This represents the number of users who successfully converted from the a-th marketing node to the b-th marketing node. This represents the total number of users who ultimately complete the conversion starting from the a-th marketing node.
[0015] In summary, by integrating information from three dimensions—jump frequency, dwell time, and conversion contribution—into a unified edge weight, the importance of jump edges in the dissemination of information about shortcomings can be accurately quantified. Total dwell time reflects the user's attention level on this jump path, while the conversion percentage reflects the conversion efficiency of this path relative to other paths. Together, these two factors allow for the differentiation between jump edges with high dwell times but low conversion rates and those with low dwell times but high conversion rates.
[0016] Step S102: Extract the behavior sequence matrix of each marketing node based on the marketing link heterogeneous graph. The behavior sequence matrix includes an interaction trajectory matrix composed of the two-dimensional coordinates of each user's interaction point on the marketing node, a time interval matrix composed of the time interval between two adjacent interactions, a pendulum state transition matrix composed of the time and direction of the user's round trip jump between the current marketing node and its marketing nodes, and a semantic pair composed of the semantic embedding vectors of the content elements on the current marketing node and the previous marketing node. It's important to note that the interaction trajectory matrix captures the spatial movement pattern of user visual attention. These sequences of two-dimensional coordinates are not randomly distributed but rather map the rhythm of content layout in capturing and releasing user attention. The time interval matrix reveals the cognitive friction coefficient of user decision-making. Shorter intervals between interactions often indicate habitual responses under low cognitive load, while longer intervals suggest deeper processing or confusion under high cognitive load. The pendulum state transition matrix records the user's back-and-forth movement between the current node and the previous / next node. This back-and-forth behavior essentially reflects the user's hesitation at decision points, signaling that the user has not yet formed a clear decision path. Semantic pairs embed semantic connections at the content level into the behavioral sequence, ensuring that behavioral analysis is no longer divorced from the meaning of the content itself. Semantic similarities or abrupt changes directly relate to the continuity or discontinuity of the user's cognitive state.
[0017] Specifically, in some embodiments, the interaction trajectory matrix is constructed according to the following formula: ; in, Let be the interaction trajectory matrix of the u-th user at the n-th marketing node. , The first interaction and the second interaction, respectively. The two-dimensional coordinates of the interaction point on the marketing node during the second interaction. Let u be the total number of interactions by the u-th user at the n-th marketing node; Construct the time interval matrix according to the following formula: ; in, Let u be the time interval matrix for the u-th user at the n-th marketing node. , , , These refer to the 1st, 2nd, and 3rd marketing events occurring at the nth marketing node for the u-th user. sequence The moment of the next interaction; Construct the pendulum state transition matrix according to the following formula: ; in, Let be the pendulum state transition matrix for the u-th user at the n-th marketing node. Let u be the total number of round trips for the u-th user at the n-th marketing node. Let be the start time and return time of the first round trip for the u-th user at the n-th marketing node, respectively. These represent the first and second marketing events of the u-th user at the n-th marketing node. The departure direction after each round trip is incremented by +1 or -1. +1 indicates moving deeper into the marketing chain, while -1 indicates returning to the previous marketing node or exiting the marketing chain. , These are the 1st, 2nd, and 3rd marketing nodes for the u-th user. The start and return times of each round trip; Suppose the semantic embedding vector of the content element of the last interaction of the u-th user at the previous marketing node is... The semantic embedding vector of the content element of the first interaction of the u-th user at the current marketing node is: Then we get semantic pairs. .
[0018] In summary, by decoupling user behavior at marketing nodes along four mutually orthogonal dimensions—spatial dispersion, temporal rhythm stability, round-trip rhythm evolution trend, and cross-node semantic coherence—the field selection and granular design of each matrix precisely match the computational needs of the subsequent three short-board factors without retaining any redundant information. The interaction trajectory matrix and time interval matrix jointly amplify both spatial and temporal anomalous signals. The pendulum state transition matrix, by simultaneously retaining the start and end time sequences and directional information of the round trip, accurately distinguishes between two patterns that are indistinguishable in traditional statistical indicators: accelerated swing followed by exit and accelerated swing followed by forward movement. Semantic pairs, by taking the edge vectors of the tail of the previous node and the head of the current node, achieve pinpoint detection of cross-node semantic cliffs and form specific cross-validation with functional label vectors. These four matrices together constitute a lossless transformation layer from raw behavior logs to computable factors. Their orthogonality and extreme value sensitivity design enable subsequent short-board determination to perform multi-factor cross-validation without information loss, dimensional redundancy, or distinction between positive and negative behaviors, laying a data foundation for the accuracy of short-board location results and the interpretability of root causes.
[0019] Step S103: Obtain the user's pseudo-activity factor based on the interaction trajectory matrix and the time interval matrix, obtain the pendulum decision factor based on the pendulum state transition matrix, and obtain the semantic mirror factor based on the semantic pair; Specifically, in some embodiments, a set of users who access any marketing node is obtained, and the user set is divided into a subset of converted users and a subset of non-converted users based on the conversion results. A pseudo-activity factor is then calculated according to the following formula: ; in, For the pseudo-activity factor of the nth marketing node, Let be the two-dimensional coordinates of the interaction point on the marketing node during the i-th interaction. , Let the maximum and minimum values of the time interval matrix for the u-th user in the unconverted subset to access the n-th marketing node be given. Let the coordinates be the center coordinates of the trajectory of the u-th user in the unconverted user subset when visiting the n-th marketing node. Let the coordinates be the center coordinates of the trajectory of the u-th user in the converted user subset visiting the n-th marketing node. For the subset of unconverted users corresponding to the nth marketing node, For the subset of converted users corresponding to the nth marketing node, All are non-zero constants; In summary, the pseudo-activity factor formula uses the difference in the spatial distribution of interactions between converted and non-converted users on the page as a benchmark to calibrate pseudo-activity signals. Traditional methods use the average of all users as a reference, but this benchmark drifts with changes in the overall distribution, leading to unstable judgment standards. This formula uses the spatial distribution of converted users to define a dynamic anchor point for high-quality behavior. This anchor point adapts to the page layout of different marketing nodes, rather than a rigid global threshold. Its core logic is to compare the spatial focus of converted and non-converted users to determine whether non-converted users exhibit aimless wandering behavior. Then, it combines this spatial deviation with the degree of temporal rhythm disorder within the same group. Only when both aimless spatial wandering and inconsistent temporal rhythm occur simultaneously within the same user group will this factor exhibit a strong pseudo-activity signal. Its core effect is to accurately separate pseudo-activity behaviors with high-frequency interactions but low cognitive investment, and to adapt to dynamic benchmarks with different page layouts, avoiding the failure of fixed thresholds at different marketing nodes.
[0020] In addition, the pendulum decision factor needs to be calculated according to the following formula: ; in, For the pendulum decision factor at the nth marketing node, For the set of users who visit the nth marketing node, Let $\mathbf{u}$ be the start time and the return time, respectively, for the $u$-th user in the user set during the $r$-th round trip at the $n$-th marketing node. Let u be the departure direction of the u-th user in the user set after the r-th round trip at the n-th marketing node. For a sign function, when The value is 1 when it is positive, -1 when it is negative, and 0 when it is zero. It is a non-zero constant; It's important to note that the round-trip time for users between marketing nodes is influenced by numerous factors, including content length, page loading speed, and individual reading speed. Any external standard will be biased by these factors, while using the user's own first round-trip time as a baseline naturally eliminates all the aforementioned individual and environmental differences. The pendulum decision factor formula mentioned above considers both the evolution direction of the round-trip rhythm and the final destination. Only when the rhythm continuously accelerates and ultimately exits the chain will this factor manifest as a "decision fatigue signal." If the rhythm accelerates but ultimately continues deeper into the chain, the signal nature is completely different. The former represents a hasty exit due to exhaustion of patience, while the latter represents a decisive advancement of accelerated decision-making. Although they appear the same in the raw data, their business implications are diametrically opposed. This allows for a precise distinction between normal in-depth exploration and abnormal decision fatigue in high-frequency round-trips, fundamentally avoiding the problem of traditional methods misjudging prudent decision-making behavior as a weakness.
[0021] The semantic mirror factor is obtained using the following formula: ; in, For the semantic mirror factor of the nth marketing node, They are respectively with , The corresponding functional label vector, , m is the functional label dimension, Let u be the number of times the u-th user returns to the previous marketing node from the n-th marketing node. This represents the total number of redirects from the u-th user to other marketing nodes starting from the n-th marketing touchpoint.
[0022] It's important to note that the semantic mirror factor formula targets a very specific user experience disruption scenario, where the semantic connection between the preceding and following content is problematic when a user jumps from one marketing node to the next. However, it's crucial to distinguish between two entirely different semantic disruption situations.
[0023] The first scenario is when the semantics are completely unrelated. For example, the previous node talks about promotional discounts, and the next node talks about logistics and delivery. Users can tell at a glance that they have entered a completely different topic area, and their brains will automatically switch cognitive frameworks. Although this break is abrupt, users are psychologically prepared and will not make serious operational mistakes.
[0024] The second scenario involves highly related or even similar semantics, but with completely opposite functions. For example, the previous node displays a button to confirm an order, and the next node displays a similarly styled confirmation button, but its actual function is to cancel the subscription. Users browsing at a fast pace may develop habitual expectations, seeing similar visual elements and performing the actions based on their previous habits, resulting in completely opposite business operations. This misoperation significantly degrades the user experience.
[0025] The semantic mirroring factor is designed specifically for the second scenario. This factor creates a mutually constraining relationship between semantic similarity and functional consistency. When two pieces of content are semantically close, users develop a habitual expectation, which is the psychological premise for misoperation. At this point, checking if the functional tags of the two pieces of content are consistent reveals a normal cross-node connection without any misleading issues. However, if the semantic distance is very close but the functional tags are completely opposite, this constitutes a very subtle misleading trap.
[0026] This factor is actually insensitive to semantically unrelated situations because users are already sufficiently alert when the semantic distance is large, preventing the formation of habitual expectations and thus reducing the risk of misoperation. What truly needs to be flagged are specific combinations of semantically similar but functionally opposite elements. Such combinations are not identified as anomalies in traditional content analysis or behavioral analysis because the content alone appears coherent, and the behavior alone shows that the user is navigating normally. Only by cross-validating both can the misleading design flaws be discovered.
[0027] Furthermore, in some embodiments, the functional label dimension is a predefined multi-dimensional business semantic space used to standardize the labeling of functional attributes for each content element. This label space typically corresponds to predefined core functional categories in the business, such as incentive guidance, information notification, behavior triggering, risk warning, process advancement, and several other types. Each dimension takes two values: positive one or negative one. A positive one indicates that the content element has the functional attribute of that dimension, while a negative one indicates that it does not have that attribute. In this way, each content element is mapped to a multi-dimensional vector composed of positive and negative values, which precisely describes the position of the content in the business functional space.
[0028] Step S104: Construct the short-board propagation equation for each marketing node based on the pseudo-activity factor, pendulum decision factor, and semantic mirror factor, and solve the short-board propagation equation based on the heterogeneous marketing link graph to obtain the short-board value for each marketing node, and determine the short-board node based on the short-board value.
[0029] It should be noted that, in order to construct the short-board propagation equation, the short-board source of the nth marketing node is first defined according to the following formula: ; in, The bottleneck source for the nth marketing node; The positive propagation index from the current marketing node to its downstream marketing nodes is defined using the following formula: ; in, Let be the positive propagation index from the nth marketing node to its downstream kth marketing node. Let the first set of users jump from the nth marketing node to its downstream kth marketing node be defined. Let the second set of users be the ones moving towards conversion from the nth marketing node. Let be the path coefficient of the first edge from the nth marketing node to the kth marketing node; The backpropagation index from the current marketing node to its upstream marketing node is defined using the following formula: ; in, Let be the backpropagation index from the nth marketing node to its upstream mth marketing node. For the nth marketing node, directly jump to the third set of users of its upstream mth marketing node. For the nth marketing node, return the fourth set of users to its upstream mth marketing node. Let be the path coefficient of the first edge from the nth marketing node to the mth marketing node; Construct the short-board propagation equation based on the following formula: ; Simplifying, we get: ; in, Let n be the bottleneck value of the nth marketing node. Let m be the shortest link value of the upstream marketing node. Let be the set of upstream nodes whose first edge points to the nth marketing node. Let be the set of downstream nodes whose first edge points to the nth marketing node; When n=1, .
[0030] In summary, the Shortcoming Source integrates three factors into a single comprehensive score. The integration logic lies in the different types of defects indicated by these factors: a high pseudo-activity factor suggests insufficient content appeal or excessively high cognitive barriers, leading to superficial user engagement; a low pendulum decision factor indicates the node failed to help users establish a sufficiently strong comparative advantage, resulting in decision-making stagnation; and a high semantic mirroring factor suggests misleading design in cross-node connections. These three factors complement each other rather than overlap in their business implications, hence the integration rather than selection approach. The Positive Propagation Index specifically quantifies the negative impact of a current node on its downstream nodes. Its construction logic is that a higher proportion of users who jump from the current node to the downstream node and ultimately complete the conversion indicates higher quality traffic delivered downstream from the current node. Conversely, a lower proportion indicates a large amount of low-quality traffic being output downstream from the current node. These low-quality users drag down the performance of the downstream nodes, which is the mechanism of the Shortcoming Propagation downstream. The backpropagation index specifically quantifies the degree to which a current node is negatively affected by its upstream nodes. Its construction logic lies in the fact that if there are numerous round-trip behaviors among users returning upstream from the current node, it indicates that users have already experienced decision-making confusion upstream. This confusion is carried over to the current node, making the current node's poor performance not due to its own defects, but rather a result of the transmission effect of upstream problems. The short-board propagation equation integrates the short-board source, forward propagation index, and backpropagation index into a solvable system. The equation's structure ensures that each node's short-board value is determined by three sources: its own short-board source, the negative impact transmitted from upstream, and the negative impact diffused downstream, all of which in turn affect the node itself. This two-way propagation mechanism allows the short-board value to distinguish whether a poorly performing node has inherent defects or is merely experiencing the propagation effect from upstream. The short-board value obtained after solving this equation accurately separates a node's own defects from the upstream and downstream transmission contributions, ensuring that the final output short-board nodes are all root cause nodes of defects rather than passively transmitted nodes. This avoids the systematic bias of traditional methods that misjudge downstream poor performance caused by upstream problems as the downstream's own shortcomings.
[0031] Furthermore, in some embodiments, the bottleneck value of any marketing node is traversed to determine whether the bottleneck value of the marketing node is less than a first threshold. If the bottleneck value of a marketing node is greater than or equal to the first threshold, then determine whether the bottleneck values of all upstream marketing nodes of that marketing node are less than the second threshold. If there is at least one upstream marketing node whose short board value is greater than or equal to the second threshold, then obtain the ratio of the short board value of that marketing node to the short board value of all marketing nodes whose short board value is greater than or equal to the second threshold, and filter out the ratio of the largest short board value from all the ratios of short board values, and determine whether the ratio of the largest short board value is less than the third threshold. If the ratio of the maximum shortest value is greater than or equal to the third threshold, the marketing node is determined to be a shortest node, and the maximum value is obtained from the three factors of the shortest node to determine the shortest type of the shortest node based on the maximum value.
[0032] For example, the first threshold can be equal to the sum of the mean and standard deviation of the short board values of all nodes, so that only nodes whose short board values deviate significantly from the mean are included in the subsequent judgment; the second threshold can be equal to the first threshold multiplied by a coefficient less than 1, used to determine whether the low performance of the current node may be due to upstream transmission; the third threshold is a preset lower limit threshold.
[0033] In summary, a three-layer progressive screening rule was constructed to accurately isolate propagation interference and pinpoint the true root cause of the short board from the short board value, a comprehensive value that combines the node's own defects with upstream and downstream transmission contributions. The first layer directly excludes nodes with normal performance based on their short board value being below a first threshold. The second layer checks whether the short board values of all upstream nodes of a node are below the second threshold. If all are below, it indicates that the node is not affected by upstream transmission and can be directly identified as the root cause of the short board. If there are upstream nodes with similarly high short board values, the process proceeds to the third layer. This layer calculates the ratio of the short board value of the node to that of all upstream nodes with high short board values and selects the largest ratio. If this ratio is below the third threshold, it indicates that the node's poor performance is mainly due to upstream transmission contributions and should be excluded. If it is above or equal to the threshold, it indicates that the node's own defects have exceeded the upstream influence and should still be identified as an independent root cause of the short board. This progressive and step-by-step elimination screening logic ensures that the final output of weak nodes are all root cause nodes rather than passively transmitted nodes. It completely solves the systematic bias in traditional methods that misjudges the poor performance of downstream caused by upstream problems as the weak nodes of the downstream itself. Furthermore, for nodes identified as root cause nodes, the maximum value is taken from three factors. By comparing the relative magnitudes of the pseudo-activity factor, pendulum decision factor, and semantic mirror factor, the type of defect of the weak node is determined, whether it is insufficient content attractiveness, decision locking failure, or cross-node connection misleading. This achieves the one-time completion of weak node root cause location and defect type attribution.
[0034] Finally, all the bottleneck nodes are sorted from largest to smallest value, and the sorting results, along with the bottleneck type and value of each bottleneck node under the sorting results, are summarized to obtain a repair priority list. The generation of the repair priority list upgrades marketing optimization work from qualitative decision-making based on experience to precise decision-making based on quantitative ranking. Operations personnel can directly implement targeted repairs on each bottleneck node in the order of the list, and the bottleneck type corresponding to each node provides clear directional guidance for the design of specific repair solutions. For example, nodes with high pseudo-activity factors should optimize page layout to lower the cognitive threshold, nodes with low pendulum decision factors should strengthen value propositions to help users establish a comparative advantage, and nodes with high semantic mirror factors should adjust the visual and functional consistency of cross-node content to avoid misleading. Thus, the three links of bottleneck identification, cause attribution, and repair guidance are fully closed in the final output.
[0035] In summary, based on the aforementioned graph neural network-based marketing link bottleneck localization method, this invention significantly improves the accuracy and root cause analysis capabilities of marketing link bottleneck localization by constructing a bottleneck localization system that integrates heterogeneous graphs and multi-factor fusion. This effectively solves the problems of existing technologies neglecting the propagation effect of bottlenecks on the jump path and the correlation of user behavior between different touchpoints. Specifically, a heterogeneous marketing link graph is first constructed. This model uses user nodes, marketing nodes, and content element nodes as vertices and jump behavior and interaction behavior as binary edges to achieve a structured expression of user behavior paths and touchpoint interaction relationships. Then, pseudo-activity factors, pendulum decision factors, and semantic mirroring factors are introduced. The pseudo-activity factor is constructed based on the joint deviation of the spatial distribution and temporal rhythm of user interactions at touchpoints. Its core principle is that non-converted users often click around on the page but never reach the effective area, and their interaction rhythm is chaotic, while converted users' interactions are concentrated in the effective area and have a stable rhythm. By quantifying the differences between non-converted and converted users in these two aspects, the pseudo-activity factor can effectively identify high-frequency, low-efficiency pseudo-active touchpoints. The pendulum decision factor is constructed based on the combined decay ratio of the round-trip time between touchpoints and the direction of departure. Its core principle is that when a user repeatedly swings between upstream and downstream, if the swing speed increases and eventually exits, then... The system identifies a decision fatigue issue at the touchpoint. If the user ultimately moves forward, it's considered normal exploratory behavior. By quantifying the acceleration ratio and conversion direction, the pendulum decision factor can accurately distinguish between accelerated exit and normal progress. The semantic mirror factor is constructed based on the semantic projection distance and functional contrast of cross-touchpoint content elements. Its core principle is that when the last interacted content element at the previous touchpoint is visually similar to but functionally opposite to the first seen content element at the current touchpoint, users are prone to misoperation due to inertia. By quantifying semantic similarity and functional oppositeness, the semantic mirror factor can accurately pinpoint cross-touchpoint misleading traps, thus constructing a more comprehensive and accurate weakness quantification system. Based on this, a two-way weakness propagation equation is constructed, enabling weakness values to distinguish between the touchpoint's own defects and upstream / downstream propagation contributions. Through progressive screening rules, normal nodes, nodes with upstream root causes, and nodes with propagation attenuation are eliminated level by level. The final output weakness nodes are all defect root cause nodes rather than passive manifestation nodes, thus effectively applicable to marketing links with complex jump structures and multiple paths, achieving accurate positioning of marketing nodes with genuine weaknesses.
[0036] like Figure 2 As shown, one embodiment of the present invention proposes a marketing link bottleneck localization system based on graph neural networks, the system comprising: The heterogeneous graph construction module 10 is used to construct a marketing link heterogeneous graph with user nodes, marketing nodes, and content element nodes as vertices, user jumping behavior between different marketing nodes as the first side, and interaction behavior of different content elements under the same marketing node as the second side. Each marketing node includes at least one content element. The behavior matrix extraction module 20 is used to extract the behavior sequence matrix of each marketing node according to the marketing link heterogeneous graph. The behavior sequence matrix includes an interaction trajectory matrix composed of the two-dimensional coordinates of each user's interaction point on the marketing node, a time interval matrix composed of the time interval between two adjacent interactions, a pendulum state transition matrix composed of the time and direction of the user's round-trip jump between the current marketing node and its marketing node, and a semantic pair composed of the semantic embedding vectors of the content elements on the current marketing node and the previous marketing node. Factor calculation module 30 is used to obtain the user's pseudo-activity factor based on the interaction trajectory matrix and the time interval matrix, obtain the pendulum decision factor based on the pendulum state transition matrix, and obtain the semantic mirror factor based on the semantic pair; The propagation equation construction module 40 is used to construct the short-board propagation equation for each marketing node based on the pseudo-activity factor, pendulum decision factor, and semantic mirror factor, and solve the short-board propagation equation based on the marketing link heterogeneity graph to obtain the short-board value of each marketing node, and determine the short-board node based on the short-board value.
[0037] In another aspect, the present invention also proposes a storage medium on which one or more programs are stored, which, when executed by a processor, implement the above-described marketing link bottleneck location method based on graph neural networks.
[0038] In another aspect, the present invention also proposes an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so as to realize the above-mentioned marketing link bottleneck location method based on graph neural network.
[0039] Those skilled in the art will understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain stored, communicated, propagated, or transmitted programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.
[0040] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.
[0041] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0042] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.
Claims
1. A method for locating weaknesses in the marketing chain based on graph neural networks, characterized in that, The method includes: Using user nodes, marketing nodes, and content element nodes as vertices, and user jumping behavior between different marketing nodes as the first side, and the interaction behavior of different content elements under the same marketing node as the second side, a heterogeneous marketing link graph is constructed, in which each marketing node includes at least one content element. The behavior sequence matrix of each marketing node is extracted based on the heterogeneous marketing link graph. The behavior sequence matrix includes an interaction trajectory matrix composed of the two-dimensional coordinates of each user's interaction point on the marketing node, a time interval matrix composed of the time interval between two adjacent interactions, a pendulum state transition matrix composed of the time and direction of the user's round-trip jump between the current marketing node and other marketing nodes, and a semantic pair composed of the semantic embedding vectors of the content elements of the current marketing node and the previous marketing node. The user's pseudo-activity factor is obtained based on the interaction trajectory matrix and the time interval matrix; the pendulum decision factor is obtained based on the pendulum state transition matrix; and the semantic mirror factor is obtained based on the semantic pair. Specifically, this includes: Obtain the set of users who access any marketing node, and divide the user set into a subset of converted users and a subset of non-converted users based on the conversion results. Calculate the pseudo-activity factor according to the following formula: ; in, Let be the total number of interactions of the u-th user at the n-th marketing node. For the pseudo-activity factor of the nth marketing node, Let be the two-dimensional coordinates of the interaction point on the marketing node during the i-th interaction. , Let the maximum and minimum values of the time interval matrix for the u-th user in the unconverted subset to access the n-th marketing node be given. Let the coordinates be the center coordinates of the trajectory of the u-th user in the unconverted user subset when visiting the n-th marketing node. Let the coordinates be the center coordinates of the trajectory of the u-th user in the converted user subset visiting the n-th marketing node. For the subset of unconverted users corresponding to the nth marketing node, For the subset of converted users corresponding to the nth marketing node, All are non-zero constants. Let u be the time interval matrix for the u-th user at the n-th marketing node; The pendulum decision factor is calculated using the following formula: ; in, For the pendulum decision factor at the nth marketing node, For the set of users who visit the nth marketing node, Let $\mathbf{u}$ be the start time and the return time, respectively, for the $u$-th user in the user set during the $r$-th round trip at the $n$-th marketing node. This represents the departure direction of the u-th user in the user set after the r-th round trip at the n-th marketing node. It is incremented by +1 or -1, where +1 indicates moving deeper into the marketing chain, and -1 indicates returning to the previous marketing node or exiting the marketing chain. For a sign function, when The value is 1 when it is positive, -1 when it is negative, and 0 when it is zero. It is a non-zero constant. Let u be the total number of round trips for the u-th user at the n-th marketing node. These are the start and return times of the first round trip for the u-th user at the n-th marketing node, respectively. The semantic mirror factor is obtained using the following formula: ; in, For the semantic mirror factor of the nth marketing node, They are respectively with , The corresponding functional label vector, , m is the functional label dimension, Let u be the number of times the u-th user returns to the previous marketing node from the n-th marketing node. Let u be the total number of redirects from the nth marketing touchpoint to other marketing nodes. Let be the semantic embedding vector of the content element of the last interaction of the u-th user at the previous marketing node. Let be the semantic embedding vector of the content element of the first interaction of the u-th user at the current marketing node; Based on the pseudo-activity factor, pendulum decision factor, and semantic mirror factor, a short-board propagation equation is constructed for each marketing node. This short-board propagation equation is then solved using the heterogeneous marketing link graph to obtain the short-board value for each marketing node. Finally, the short-board node is determined based on this short-board value. Specifically, this includes: The bottleneck source of the nth marketing node is defined according to the following formula: ; in, The bottleneck source for the nth marketing node; The positive propagation index from the current marketing node to its downstream marketing nodes is defined using the following formula: ; in, Let be the positive propagation index from the nth marketing node to its downstream kth marketing node. Let the first set of users jump from the nth marketing node to its downstream kth marketing node be defined. Let the second set of users be the ones moving towards conversion from the nth marketing node. Let be the path coefficient of the first edge from the nth marketing node to the kth marketing node; The backpropagation index from the current marketing node to its upstream marketing node is defined using the following formula: ; in, Let be the backpropagation index from the nth marketing node to its upstream mth marketing node. For the nth marketing node, directly jump to the third set of users of its upstream mth marketing node. For the nth marketing node, return the fourth set of users to its upstream marketing nodes. Let be the path coefficient of the first edge from the nth marketing node to the mth marketing node; Construct the short-board propagation equation based on the following formula: ; Simplifying, we get: ; in, Let n be the bottleneck value of the nth marketing node. Let m be the shortest link value of the upstream marketing node. Let be the set of upstream nodes whose first edge points to the nth marketing node. Let be the set of downstream nodes whose first edge points to the nth marketing node; When n=1, .
2. The marketing link bottleneck location method based on graph neural networks according to claim 1, characterized in that, The steps to construct a heterogeneous marketing link graph, using user nodes, marketing nodes, and content element nodes as vertices, user navigation between different marketing nodes as the first side, and interaction between different content elements within the same marketing node as the second side, include: For any two marketing nodes, obtain the total number of times all users jump from one marketing node to the other and the dwell time after each jump, and define the first-side path coefficient according to the following formula: ; in, Let be the path coefficient of the first edge from marketing node a to marketing node b. This represents the duration of a user's stay after the k-th jump from the a-th marketing node to the b-th marketing node. This represents the total number of times a user jumps from marketing node a to marketing node b. This represents the number of users who successfully converted from the a-th marketing node to the b-th marketing node. This represents the total number of users who ultimately complete the conversion starting from the a-th marketing node.
3. The marketing link bottleneck location method based on graph neural networks according to claim 2, characterized in that, The step of extracting the behavior sequence matrix of each marketing node based on the heterogeneous marketing link graph includes: Construct the interaction trajectory matrix according to the following formula: ; in, Let be the interaction trajectory matrix of the u-th user at the n-th marketing node. , The first interaction and the second interaction, respectively. The two-dimensional coordinates of the interaction point on the marketing node during the second interaction; Construct the time interval matrix according to the following formula: ; in, , , , These refer to the 1st, 2nd, and 3rd marketing events occurring at the nth marketing node for the u-th user. sequence The moment of the next interaction; Construct the pendulum state transition matrix according to the following formula: ; in, Let be the pendulum state transition matrix for the u-th user at the n-th marketing node. These represent the first and second marketing events of the u-th user at the n-th marketing node. The direction of departure after the round trip. , These are the 1st, 2nd, and 3rd marketing nodes for the u-th user. The start and return times of each round trip.
4. The marketing link bottleneck location method based on graph neural networks according to claim 3, characterized in that, The step of solving the bottleneck propagation equation based on the heterogeneous marketing link diagram to obtain the bottleneck value of each marketing node, and determining the bottleneck node based on the bottleneck value, includes: Iterate through the bottleneck values of any marketing node to determine whether the bottleneck value of the marketing node is less than the first threshold. If the bottleneck value of a marketing node is greater than or equal to the first threshold, then determine whether the bottleneck values of all upstream marketing nodes of that marketing node are less than the second threshold. If there is at least one upstream marketing node whose short board value is greater than or equal to the second threshold, then obtain the ratio of the short board value of that marketing node to the short board value of all marketing nodes whose short board value is greater than or equal to the second threshold, and filter out the ratio of the largest short board value from all the ratios of short board values, and determine whether the ratio of the largest short board value is less than the third threshold. If the ratio of the maximum shortest value is greater than or equal to the third threshold, the marketing node is determined to be a shortest node, and the maximum value is obtained from the three factors of the shortest node to determine the shortest type of the shortest node based on the maximum value.
5. The marketing link bottleneck location method based on graph neural networks according to claim 4, characterized in that, The method further includes: Sort all the short board nodes by their short board values from largest to smallest, and summarize the sorting results, along with the short board type and short board value of each short board node under the sorting results, to obtain a repair priority list.
6. A marketing link bottleneck localization system based on graph neural networks, used to implement the marketing link bottleneck localization method based on graph neural networks as described in any one of claims 1-5, characterized in that, The system includes: The heterogeneous graph construction module is used to construct a marketing link heterogeneous graph with user nodes, marketing nodes, and content element nodes as vertices, user jumping behavior between different marketing nodes as the first side, and interaction behavior of different content elements under the same marketing node as the second side. Each marketing node includes at least one content element. The behavior matrix extraction module is used to extract the behavior sequence matrix of each marketing node based on the marketing link heterogeneous graph. The behavior sequence matrix includes an interaction trajectory matrix composed of the two-dimensional coordinates of each user's interaction point on the marketing node, a time interval matrix composed of the time interval between two adjacent interactions, a pendulum state transition matrix composed of the time and direction of the user's round-trip jump between the current marketing node and other marketing nodes, and a semantic pair composed of the semantic embedding vectors of the content elements on the current marketing node and the previous marketing node. The factor calculation module is used to obtain the user's pseudo-activity factor based on the interaction trajectory matrix and the time interval matrix, obtain the pendulum decision factor based on the pendulum state transition matrix, and obtain the semantic mirror factor based on the semantic pair. The propagation equation construction module is used to construct the short-board propagation equation for each marketing node based on the pseudo-activity factor, pendulum decision factor, and semantic mirror factor, and solve the short-board propagation equation based on the heterogeneous marketing link graph to obtain the short-board value for each marketing node, and determine the short-board node based on the short-board value.
7. A storage medium, characterized in that, The storage medium stores one or more programs that, when executed by a processor, implement the marketing link bottleneck location method based on any one of claims 1-5.
8. An electronic device, characterized in that, The electronic device includes a memory and a processor, wherein: The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the marketing link bottleneck location method based on graph neural networks as described in any one of claims 1-5.
Citation Information
Patent Citations
Private domain user behavior big data layering and fine operation method based on machine learning
CN120707187A
Content operation supervision system of intensive website platform
CN121660672A
