Cross-channel consumer intention fusion analysis method and system

Abstract

The invention relates to the technical field of electric digital data processing, in particular to a cross-channel consumer intention fusion analysis method and system. The method comprises the following steps: constructing a consumer dynamic consumption journey map, wherein the dynamic consumption journey map at least comprises consumer nodes, contact nodes, intention nodes and distribution of target objects, and calculating intention uncertainty entropies of consumers corresponding to the consumer nodes; in response to the situation that the intention uncertainty entropy is smaller than a preset threshold value, aggregating feature vectors of all contact nodes connected with the same consumer node by adopting a graph attention network to generate an omni-channel feature vector; and determining the current consumption stage of the consumer based on the omni-channel feature vector, and pushing corresponding information to the consumer according to the consumption stage. The method and the device have the effects of reducing repeated pushing of marketing contents and saving marketing resources.

Description

Technical Field

[0001] This application relates to the field of digital data processing technology, and in particular to a method and system for cross-channel consumer intent fusion analysis. Background Technology

[0002] With the development of digital marketing and intelligent recommendation technologies, enterprises are using data-driven methods to analyze consumer behavior to support business decisions such as advertising, product recommendations, and customer relationship management. Among these technologies, consumer profiling and intent recognition typically rely on user behavior data collected by internet platforms, such as explicit behaviors like browsing, clicking, saving, and purchasing. Based on this data, user tagging systems or interest models are built, thereby enabling product or content recommendations through algorithms such as collaborative filtering.

[0003] Most existing consumer intent analysis and recommendation systems rely on data from a single platform or channel for analysis. For example, some advertising data management platforms (DMPs) primarily analyze user behavior data within single systems such as WeChat and Douyin, and complete ad targeting and recommendation decisions within that platform.

[0004] In real-world consumption scenarios, the process of purchasing goods often spans multiple online platforms, exhibiting distinct multi-stage and multi-touchpoint characteristics. For example, consumers may first obtain product information and develop initial interest on content community platforms, then experience or compare products in offline stores, and finally complete repeat purchases or long-term customer retention through instant messaging or e-commerce platforms. This "cognition-consideration-decision-loyalty" consumption process demonstrates cross-platform and cross-media continuity.

[0005] However, in related technologies, data collected from different channels is usually stored in separate, independent systems. This fragments consumer behavior across different platforms into independent events, making it impossible to identify the underlying consumer and the continuous evolution of their intentions. On the one hand, this data silo phenomenon makes it difficult for the system to accurately determine the consumer's actual consumption stage, easily misjudging consumers who have entered the decision-making or repurchase stage as only being in the initial interest stage. On the other hand, at the marketing execution level, the system may repeatedly push similar marketing content to the same consumer on different channels based on incomplete or fragmented intention judgments, resulting in inefficient use of marketing resources. Summary of the Invention

[0006] To address the issue of duplicate marketing content pushes due to misjudgment of consumer intent, this application provides a cross-channel consumer intent fusion analysis method and system.

[0007] Firstly, this application provides a cross-channel consumer intent fusion analysis method, employing the following technical solution: A cross-channel consumer intent fusion analysis method includes: Construct a dynamic consumer journey graph. The dynamic consumer journey graph includes at least consumer nodes, touchpoint nodes, intent nodes, and target objects. Touchpoints are the channels through which consumers corresponding to consumer nodes interact with target objects, and intent nodes are used to represent the psychological state of consumers corresponding to consumer nodes. For any consumer, determine all the touchpoint nodes corresponding to that consumer, and calculate the intent uncertainty entropy of the consumer corresponding to the consumer node based on the distribution of intent nodes connected to the touchpoint nodes. In response to the intent uncertainty entropy being less than a preset threshold, a graph attention network is used to aggregate the feature vectors of all touchpoint nodes connected to the same consumer node to generate an omnichannel feature vector; based on the omnichannel feature vector, the current consumption stage of the consumer is determined, and corresponding information is pushed to the consumer according to the consumption stage.

[0008] A dynamic consumer journey graph is constructed, comprising consumer nodes, touchpoint nodes, and intent nodes. This transforms consumer behavior at different touchpoint nodes from isolated events into structured graph data with temporal sequence and semantic relationships. By connecting touchpoint nodes with consumer nodes, specific consumer interactions across different platforms and media are explicitly modeled to differentiate the varying impacts of different touchpoints on consumer purchasing decisions. Furthermore, the introduction of intent nodes abstractly expresses the psychological states underlying these behaviors, achieving a transformation from behavioral data to intent expression.

[0009] The algorithm introduces the calculation of intent uncertainty entropy and only performs graph attention aggregation operation when the uncertainty entropy is less than a preset threshold. This reduces the need for forced stage judgments when consumer intent is highly dispersed or ambiguous, which could lead to incorrect consumption stage judgments, push irrelevant content to consumers, and negatively impact the consumer experience.

[0010] Finally, a graph attention network is used to aggregate the feature vectors of all touchpoints connected to the same consumer node. The graph attention network can adaptively emphasize behavioral signals that are more discriminative of the current consumption stage based on the differences in touchpoint features, thereby generating a more representative omnichannel feature vector. This omnichannel feature vector is used to determine the consumer's current consumption stage. Differentiated marketing content is then pushed according to the consumer's consumption stage, ensuring that the marketing content matches the consumer's actual needs and reducing the problem of duplicate marketing content pushes caused by inaccurate identification of the consumer's consumption stage, a problem present in other technologies.

[0011] Optionally, a mask autoencoder can be used to associate heterogeneous identity identifiers on different consumer touchpoint nodes, and the connection edges between consumer nodes and touchpoints can be determined based on the association between different heterogeneous identities.

[0012] The mask autoencoder, through a self-supervised learning mechanism, mines potential correlations between different identity identifiers from platform behavior data across various channels, enabling cross-platform identity alignment even in the absence of unified login or strongly bound information. This method aggregates scattered behavioral data from different channels under a single consumer dimension by associating heterogeneous consumer identity identifiers, providing a stable data foundation for subsequent intent fusion and consumption stage judgment. This reduces biases in consumer consumption stage judgment caused by identity inconsistencies at the data source.

[0013] Optionally, the types of intent nodes in the dynamic consumer journey graph include at least: quality exploration intent nodes, price comparison intent nodes, and urgent purchase intent nodes.

[0014] By clearly distinguishing between intent types such as quality exploration, price comparison, and urgent purchase, the system can map different features to more semantically discriminative intent nodes when analyzing touchpoint behavior, thereby improving the interpretability of intent recognition. At the same time, this intent segmentation provides a clear basis for subsequent consumption stage judgment and marketing strategy formulation.

[0015] Optionally, a touchpoint node refers to the specific medium through which a consumer interacts with a target object. The feature vector of a touchpoint node includes: the individual dwell time of the consumer, the individual sentiment score, and a Boolean value indicating whether or not to share.

[0016] Multi-dimensional modeling of touchpoint feature vectors improves the effective information density of consumer behavior data for intent recognition, providing a more discriminative input basis for the generation of omnichannel feature vectors.

[0017] Optionally, the types of touchpoints include: platform interfaces where consumers interact with the target audience and QR code scanning points in physical stores.

[0018] Touchpoints include various online platforms and offline QR code scanning points, thereby aggregating online and offline information and further improving the accuracy of identifying the consumer's consumption node.

[0019] Optionally, calculating the consumer's intent uncertainty entropy includes: obtaining the implicit intent weights of each touchpoint node; for any intent node, summing the implicit intent weights of all its associated touchpoint nodes to obtain the cumulative intent intensity; using the ratio of the cumulative intent intensity corresponding to each intent to the sum of the cumulative intent intensities of all intents as the probability of the intent; and performing entropy calculation on the probabilities of all intents to obtain the consumer's intent uncertainty entropy.

[0020] For any given intention, its probability is the ratio of the weighted sum of the accumulated loudness of the intention to the total accumulated loudness of the intention. Here, different touchpoints are weighted to improve the accuracy of judging consumer intentions, thereby improving the accuracy of calculating the intention uncertainty entropy.

[0021] Optionally, the implicit intent weights of each touchpoint node are obtained, including: extracting multiple implicit feature values ​​corresponding to the touchpoint, and using the weighted sum of the multiple implicit feature values ​​as the implicit intent weights of the touchpoint node.

[0022] Multiple latent feature values ​​corresponding to touchpoints are extracted and their weighted sum is used as the latent intent weight of the touchpoint node, thereby realizing a comprehensive evaluation of the touchpoint value.

[0023] Optional, implicit feature values ​​include: attention score calculated based on the duration of consumer dwell time at the touchpoint, and score representing the consumer's positive feelings toward the product expressed at that touchpoint.

[0024] Attention score represents the degree of consumer attention to a product at a touchpoint; favorability score represents the consumer's approval of the product at that touchpoint. If consumers show higher attention and higher approval at a touchpoint, it indicates that their intention at that touchpoint is stronger and more obvious, thus giving that touchpoint higher value. Simultaneously, by introducing a favorability score, the system can distinguish between positive interest and negative feedback. If a consumer gives a negative review at a touchpoint, they will still show a high dwell time at that touchpoint. Therefore, by using a favorability score ranging from -1 to 1, the system can reduce the occurrence of misjudging complaints or questioning behavior as purchases or other positive intentions.

[0025] Optionally, implicit feature values ​​may also include: a sharing score determined based on the consumer's sharing behavior at the touchpoint.

[0026] Secondly, this application provides a cross-channel consumer intent fusion analysis system, which adopts the following technical solution: A cross-channel consumer intent fusion analysis system includes a processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the cross-channel consumer intent fusion analysis method described above.

[0027] The aforementioned cross-channel consumer intent fusion analysis method generates a computer program, which is stored in memory for loading and execution by a processor. Thus, a system is created based on the memory and processor for easy use.

[0028] This application has the following technical effects: By leveraging a mask autoencoder, we can unify and integrate consumer identities across different platforms, constructing a dynamic consumer journey map of consumers, touchpoints, and intentions. By aggregating explicit and implicit characteristics such as dwell time, emotional inclination, and sharing behavior, we can accurately identify the consumer's current consumption stage, thereby enabling differentiated and precise marketing pushes across platforms and reducing the waste of marketing resources caused by duplicate marketing content. Attached Figure Description

[0029] Figure 1 This is a flowchart of a cross-channel consumer intent fusion analysis method according to Embodiment 1 of this application.

[0030] Figure 2 This is a flowchart illustrating a cross-channel consumer intent fusion analysis method according to Embodiment 2 of this application. Detailed Implementation

[0031] Example 1 This application discloses a cross-channel consumer intent fusion analysis method, which first uses a mask autoencoder (… This maps heterogeneous anonymous identifiers from various platforms (such as social media, e-commerce, and search) to a unified, unique consumer identifier, solving the problem of cross-channel entity alignment; then, through a graph attention network (... It aggregates global information and uses an attention mechanism to dynamically allocate the contribution of each platform's features. Finally, it obtains the consumer's consumption stage and pushes differentiated marketing content based on the different consumer stages, reducing repetitive content pushes to consumers, improving consumer experience, and saving marketing resources.

[0032] Reference Figure 1 A cross-channel consumer intent fusion analysis method includes steps S1-S3.

[0033] S1: Construct a dynamic consumer journey graph. The dynamic consumer journey graph includes at least consumer nodes, touchpoint nodes, intent nodes, and target objects. Touchpoints are the channels through which consumers corresponding to consumer nodes interact with target objects. Intent nodes are used to represent the psychological state of consumers corresponding to consumer nodes.

[0034] After cross-platform identity association is completed, this application constructs a cross-channel dynamic consumer journey graph based on the consumer's unique identifier. .in, This represents a dynamic consumer journey map that includes all consumers; The nodes in the graph include: consumers, touchpoints, intents, and target objects; The connecting edges in the graph include: browsing, interaction, transformation, and influence.

[0035] For consumers in a node, it refers to the unique identifier obtained after cross-platform identity association, which is the aggregation point of all behavioral signals.

[0036] Touchpoints refer to the specific mediums or platform interfaces through which consumers interact with a target audience, such as tweets on social media platforms, product detail pages on e-commerce platforms, and QR code scanning points in offline stores.

[0037] Intent is a psychological state node derived from consumer behavioral signals, including the following categories: Quality exploration intent node: the intent indicated by users' behavior of deeply reading ingredient lists, checking scientific endorsements or laboratory test reports; Price comparison intent node: the intent indicated by users' behavior of searching for coupons across platforms, comparing unit prices of different specifications, or paying attention to discount policies in live broadcast rooms; Urgent purchase intent node: the intent indicated by users' behavior of frequently checking logistics timeliness, confirming inventory status, or consulting customer service for same-day delivery.

[0038] The target object refers to a business entity that represents the object of consumer interaction. For example, the target object could be a product or a brand. The target object connects multiple touchpoints and intents, and is used to analyze the long-term impact of the target object and user loyalty.

[0039] For browsing in a connection edge, it represents the basic edge connecting the consumer and the touchpoint node. It represents the information contact between the consumer and the touchpoint, usually accompanied by implicit characteristics, such as dwell time, which can also be understood as reading time or browsing time.

[0040] Interaction refers to the enhanced edges that connect consumers to touchpoints. It represents deeper behaviors such as liking, commenting, sharing, or saving.

[0041] A conversion is the key outcome link connecting a "consumer" with a "target audience" or "intent." It represents the consumer completing a pre-defined target action, such as purchasing, registering, or placing an order. Conversion edges have the highest degree of intent certainty.

[0042] Impact is the attribution edge connecting "touchpoint" with "intent" or "touchpoint" with "target object." It reflects the facilitating effect of a touchpoint on the next stage of consumption.

[0043] The connection between consumers and touchpoints is primarily achieved by associating consumers' heterogeneous identities across different platforms. Specifically, this involves acquiring consumers' heterogeneous identities across multiple channels and aggregating these heterogeneous identities into a single unique consumer identifier.

[0044] In this embodiment, the system uses a preset interface ( This can be achieved by collecting raw behavioral signal streams generated by consumers on different platforms during a preset observation period using data tracking points. These raw behavioral signal streams primarily consist of signals generated when consumers interact with various platforms, such as clicks, keyword searches, and adding items to cart on e-commerce platforms, or likes, comments, dwell time on specific product detail pages, timestamps of visits, geographical locations, and device fingerprints used on social media platforms.

[0045] Because the data formats of different platforms vary greatly—for example, the data on social media platforms is mostly unstructured text comments and likes—their heterogeneous identifiers are usually encrypted. E-commerce platforms generate data including high-dimensional click heatmaps and search keywords; heterogeneous identifiers are typically browsers. Or the in-app account name. Payment and device-related data include location information and payment frequency; heterogeneous identity identifiers are... , .

[0046] Therefore, in order to logically aggregate data from multiple platforms to the same natural person, in one embodiment, a mask autoencoder can be introduced ( Cross-platform identity association is performed, which means linking the isolated identities of consumers on different platforms to confirm that they belong to the same natural person.

[0047] Specifically, this application will address heterogeneous identity identifiers from different platforms (such as...). , , Using features such as (e.g., access timestamp synchronization, browsing category similarity) as input, the system randomly masks the identifier information of some channels during the training phase, and reconstructs the masked identifiers using the behavioral sequences of the remaining channels (e.g., access timestamp synchronization, browsing category similarity). Through this self-supervised learning process, the system can learn the association relationships between different identifiers in the latent space.

[0048] As a preferred approach, this application maps the aligned data to a unified consumer unique identifier. In this embodiment, the mask ratio of the mask autoencoder is set to... When the mask ratio is lower than When the mask ratio is high, the model tends to learn simple correspondences between identifiers, resulting in weak generalization ability; when the mask ratio is higher than... At that time, insufficient remaining information caused the reconstruction to fail. (Settings...) The left and right sides can force the model to explore the deep causal logic between the behavioral characteristics of each platform and the user's identity, thereby improving the accuracy of cross-channel entity alignment.

[0049] In another embodiment, a unified login system can be integrated across different platforms. For example, when a consumer logs in using WeChat on a social media platform and also selects WeChat authorization login on an e-commerce platform, both platforms will obtain the same unique consumer identifier.

[0050] For example, if a Xiaohongshu account and a Douyin platform account belong to the same consumer, then all touchpoints generated by the consumer's subsequent interactions with the target on Xiaohongshu and Douyin will be connected to the same consumer.

[0051] In this embodiment, each specific consumer corresponds to a personalized sub-map. Among them, nodes Includes all platform touchpoint nodes associated with the consumer and their mapped intent nodes; connection edges The connection edges between each node are recorded.

[0052] S2: For any consumer, determine all touchpoint nodes corresponding to that consumer, and calculate the intent uncertainty entropy of the consumer corresponding to the consumer node based on the distribution of intent nodes connected to the touchpoint nodes.

[0053] First, for each consumer's touchpoints across various platforms, explicit intent features (such as clicks and order placements) and implicit intent features are extracted. Different touchpoints have varying importance to the consumer's final decision; therefore, this study analyzes the implicit features of each touchpoint connected to the consumer's corresponding consumer node and calculates the implicit intent weight for each touchpoint. Specifically, the implicit intent weight is obtained by weighted fusion of the implicit feature values.

[0054] In one embodiment, the latent features include at least the sentiment score and the duration of stay.

[0055] For any touchpoint, the formula for calculating its implicit intent weight can be expressed as: ; In the formula, Indicates contact point The weight of implicit intent; For consumers at touchpoints The duration of time spent on the page; This refers to the maximum historical dwell time of this channel within the preset statistical period; The favorability score generated by consumers at this touchpoint, as identified through a natural language processing algorithm; The first weighting coefficient is preset; This is the preset second weighting coefficient.

[0056] The steps for identifying consumer favorability scores at a specific touchpoint using natural language processing algorithms mainly include: obtaining the original text of the consumer's interaction with that touchpoint, including tweets, comments, and bullet screen messages generated by the consumer at that touchpoint; analyzing the original text and extracting features using a deep semantic model; and then classifying and mapping these features to... The favorability score is calculated based on the range of interactions. This score represents the degree to which a consumer subjectively identifies with the target object during a specific touchpoint interaction. For example, a positive score (such as...) ) reflects the intention to "plant grass" (i.e., negative scores are given). This reflects the intention to "question" or "avoid pitfalls." The favorability score can correct simple dwelling behavior, effectively distinguishing between positive attention and negative complaints, thereby significantly improving the accuracy of intention recognition.

[0057] In this embodiment, , Configure based on experience, for example... Set as , Set as .

[0058] In this formula, This represents the consumer's attention score, which is positively correlated with the time a consumer spends at a particular touchpoint. The longer a consumer stays at a touchpoint, the more attention they pay to the content within that touchpoint. Similarly, This reflects the consumer's level of identification with the target audience. Higher attention and favorability scores indicate a stronger underlying intent from the touchpoint. This allows the system to identify high-value intent signals from seemingly random browsing behavior.

[0059] In another embodiment, the implicit features of the touchpoint node may further include a sharing score, which characterizes whether the consumer has engaged in a sharing behavior at the touchpoint node. In this embodiment, in response to the consumer engaging in a sharing behavior at the touchpoint, the sharing score is set to 1; otherwise, the sharing score is set to 0.

[0060] The sharing score can be incorporated into the calculation of the implicit intent weight. Specifically, in this embodiment, the formula for calculating the implicit intent weight can be expressed as follows: In the formula, Indicates contact point The weight of implicit intent; For consumers at touchpoints The duration of time spent on the page; This refers to the maximum historical dwell time of this channel within the preset statistical period; The favorability score generated by consumers at this touchpoint, as identified through a natural language processing algorithm; To share scores, if consumers at the touchpoint When a sharing action occurs, the value is set to [value]. Otherwise ; The first weighting coefficient is preset; This is the preset second weighting coefficient; This is the preset third weighting coefficient.

[0061] In this embodiment, , Configure based on experience, for example... Set as , Set as , Set as .

[0062] Compared to dwell time and favorability score, sharing behavior further represents the degree to which consumers are willing to identify with the target, reflecting psychological identification and conversion certainty, thereby improving the accuracy of implicit intent weight calculation.

[0063] Subsequently, the intent uncertainty entropy of each consumer is determined based on the implicit intent weight corresponding to each touchpoint.

[0064] In one embodiment, the formula for calculating the intentional uncertainty entropy can be expressed as: ; In the formula, Consumers The intention of uncertainty entropy; Represents a given consumer subgraph Under the condition that the user belongs to the first The probability of a specific intention; This represents the total number of intent categories.

[0065] The user belongs to the The probability of a specific intention is a weighted probability. The steps for calculating the weighted probability of any given intention include: First, identify the touchpoints connected to the intent, and then accumulate the implicit intent weights corresponding to the touchpoints to obtain the cumulative intent strength of the intent.

[0066] Then, the sum of the cumulative intensity of each intention is calculated as the total intensity. For any intention, the ratio of the cumulative intensity of the intention to the total intensity is taken as the probability of that intention.

[0067] For example, suppose a consumer subgraph includes three touchpoints, each pointing to one of two latent intentions: quality exploration and price comparison. Through the calculation in step S2, the latent intention weight corresponding to the first touchpoint is 0.8; the latent intention weight corresponding to the second touchpoint is 0.6; and the latent intention weight corresponding to the third touchpoint is 0.2. Adding the first and second touchpoints associated with the quality exploration intention node yields a cumulative intention strength of 1.4. Therefore, the probability of this intention is... .

[0068] Intent uncertainty entropy reflects the dispersion of consumer intentions. When consumer signals across platforms are extremely chaotic, the probabilities of each intention tend to be equal, leading to an increase in uncertainty entropy. This reflects that the consumer is currently in a period of decision-making confusion. When the entropy value is low, it means that the probability is concentrated on a specific intention, and the system can determine that the intention identified at this time has a very high degree of confidence.

[0069] In response to an intent uncertainty entropy less than a preset threshold, a graph attention network is used to aggregate the feature vectors of all touchpoint nodes connected to the same consumer node, generating an omnichannel feature vector. Based on the omnichannel feature vector, the current consumption stage of the consumer is determined, and differentiated marketing content is pushed according to the consumption stage. Otherwise, behavioral data corresponding to the continuous consumer continues to be collected in real time.

[0070] S3: In response to the intent uncertainty entropy being less than a preset threshold, a graph attention network is used to aggregate the feature vectors of all touchpoint nodes connected to the same consumer node to generate an omnichannel feature vector; based on the omnichannel feature vector, the current consumption stage of the consumer is determined, and corresponding information is pushed to the consumer according to the consumption stage.

[0071] After obtaining the features of each touchpoint node, in order to transform the discrete touchpoint information in the sub-graph corresponding to each consumer into a unified intent vector, this application uses a Graph Attention Network (GAT) to aggregate and obtain the full-channel feature vector.

[0072] For any consumer, the formula for calculating its corresponding omnichannel feature vector is as follows: In the formula, Consumers The omnichannel feature vector; To represent a nonlinear activation function, it is usually expressed as... or ; This indicates the consumer node in the dynamic consumer journey graph. A collection of directly connected contact nodes; Represents the learnable weight transformation matrix; Represents contact node eigenvectors; Represents contact node For consumer nodes Attention weight coefficients.

[0073] Attention weight coefficient Represents the contact node For consumer nodes The contribution of the overall consumption stage analysis is calculated using the following logic: ; In the formula, Represents contact node For consumers Attention weight coefficient; For consumer nodes eigenvectors; For contact nodes eigenvectors; This is a vector concatenation symbol; It is a non-linear activation function; This is the transpose of the attention mechanism's weight vector; This indicates the consumer node in the dynamic consumer journey graph. A collection of directly associated contact nodes; Represents contact node eigenvectors; This represents the learnable weight transformation matrix.

[0074] In one embodiment, the contact node The feature vector includes information such as the consumer's dwell time, individual sentiment score, and whether they shared the information.

[0075] Subsequently, the probability of each consumption stage to which the consumer currently belongs is generated based on the consumer's omnichannel feature vector.

[0076] In the formula, Consumers In the first The probability of each consumption stage; The output layer weight matrix is ​​a learnable matrix. This is a bias term. Consumers The omnichannel feature vector, where the vector dimension is consistent with the number of types of consumption stages. In this embodiment, the consumption stages are artificially divided into four stages: "cognition", "consideration", "decision", and "advocacy". This represents the probability normalization function.

[0077] After identifying the consumption stage, the system pushes different marketing messages to consumers based on the probability of them being in that stage.

[0078] Specifically, based on the four stages of consumer purchasing: awareness, consideration, decision-making, and loyalty, when the probability of the "decision-making" stage is greater than 0.8, it is considered that the consumer has already passed the "awareness" and "consideration" stages, meaning that the consumer does not currently need product-related content. At this point, all awareness-based content, such as product encyclopedias and brand stories, is prohibited from being pushed to the entire platform. Instead, information that can promote the consumer's decision-making is pushed, such as limited-time coupons and exclusive discount codes. This method can, on the one hand, reduce interference with basic information when consumers already have a strong desire to buy, thus improving the consumer experience; and on the other hand, reduce wasted marketing budgets.

[0079] When the probability of the "endorsement" stage is greater than 0.8, it means that the consumer has transcended the role of a mere buyer and transformed into a loyal supporter and active promoter of the target audience. At this point, it is no longer necessary to push marketing content that guides purchases or promotes prices; instead, content that enhances social capital can be pushed, such as inviting them to participate in user beta testing programs or pushing advanced styling guides based on their purchasing preferences.

[0080] Example 3 differs from Example 1 in that it employs a graph attention network to aggregate the feature vectors of all touchpoint nodes connected to the same consumer node, generating an omnichannel feature vector. During this process, the feature vectors of touchpoint nodes also include the implicit intent weights calculated in step S2. In the aggregation process, the implicit value weights are input into the model as part of the touchpoint node's feature vector. Through learning, the model automatically assigns higher attention coefficients to touchpoints with high implicit value (e.g., those that remained on the page for a long time and shared positive reviews even without making a purchase). This allows the generated omnichannel feature vectors to more realistically represent consumer psychology after cross-channel integration.

[0081] This application also discloses a cross-channel consumer intent fusion analysis system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a cross-channel consumer intent fusion analysis method according to this application.

[0082] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0083] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A cross-channel consumer intent fusion analysis method, characterized in that, Construct a dynamic consumer journey graph, which includes at least consumer nodes, touchpoint nodes, intent nodes, and target objects. Touchpoints are the channels through which consumers corresponding to consumer nodes interact with target objects; intent nodes are used to represent the psychological state of consumers corresponding to consumer nodes. For any consumer, determine all the touchpoint nodes corresponding to that consumer, and calculate the intent uncertainty entropy of the consumer corresponding to the consumer node based on the distribution of intent nodes connected to the touchpoint nodes. In response to the intent uncertainty entropy being less than a preset threshold, a graph attention network is used to aggregate the feature vectors of all touchpoint nodes connected to the same consumer node to generate an omnichannel feature vector; based on the omnichannel feature vector, the current consumption stage of the consumer is determined, and corresponding information is pushed to the consumer according to the consumption stage.

2. The cross-channel consumer intent fusion analysis method according to claim 1, characterized in that, A masked autoencoder is used to associate heterogeneous identity identifiers on different consumer touchpoint nodes. Based on the association between different heterogeneous identities, the connection edges between consumer nodes and touchpoints are determined.

3. The cross-channel consumer intent fusion analysis method according to claim 1, characterized in that, The types of intent nodes in a dynamic consumer journey map include at least: quality exploration intent nodes, price comparison intent nodes, and urgent purchase intent nodes.

4. The cross-channel consumer intent fusion analysis method according to claim 1, characterized in that, Touchpoints refer to the specific medium through which consumers interact with the target object. The feature vectors of touchpoints include: individual dwell time of consumers, individual sentiment score, and Boolean value indicating whether or not to share.

5. The cross-channel consumer intent fusion analysis method according to claim 1, characterized in that, The types of touchpoints include: platform interfaces where consumers interact with target objects and QR code scanning points in offline stores.

6. The cross-channel consumer intent fusion analysis method according to claim 1, characterized in that, Calculating the consumer's intent uncertainty entropy includes: obtaining the implicit intent weights of each touchpoint node; for any intent node, summing the implicit intent weights of all its associated touchpoint nodes to obtain the cumulative intent strength; using the ratio of the cumulative intent strength of each intent to the sum of the cumulative intent strengths of all intents as the probability of the intent; and performing entropy calculation on the probabilities of all intents to obtain the consumer's intent uncertainty entropy.

7. The cross-channel consumer intent fusion analysis method according to claim 6, characterized in that, Obtain the implicit intent weight of each touchpoint node, including: extracting multiple implicit feature values ​​corresponding to the touchpoint, and using the weighted sum of the multiple implicit feature values ​​as the implicit intent weight of the touchpoint node.

8. The cross-channel consumer intent fusion analysis method according to claim 7, characterized in that, Implicit features include: attention score calculated based on the duration of consumer dwell time at touchpoints, and score representing the consumer's positive feelings toward the product at that touchpoint.

9. The cross-channel consumer intent fusion analysis method according to claim 8, characterized in that, Implicit features also include: sharing scores determined based on consumers' sharing behavior at touchpoint nodes.

10. A cross-channel consumer intent fusion analysis system, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement a cross-channel consumer intent fusion analysis method according to any one of claims 1-9.

Images