A user intention prediction method and system based on commercial consumption data
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU ANJIE BIOTECHNOLOGY CO LTD
- Filing Date
- 2026-03-03
- Publication Date
- 2026-06-09
Smart Images
Figure CN122175619A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intent prediction technology, specifically relating to a method and system for predicting user intent based on commercial consumption data. Background Technology
[0002] With the deep integration of e-commerce and social media, commercial consumption data refers to the structured behavioral data of users on e-commerce platforms such as clicks, browsing, adding to cart, and purchasing, as well as the unstructured text data describing life status and consumption needs published on social platforms. Accurately predicting users' consumption intentions has significant economic value for optimizing the allocation of marketing resources. In the existing technology, predicting users' consumption intentions has the following shortcomings, such as (1) difficulty in multi-source data fusion: social text data and consumption behavior data are usually processed separately. The processing of social text data relies on keyword matching or shallow semantic classification to identify consumption intentions. The processing of consumption behavior data mostly adopts models such as logistic regression, factor decomposition machine or gradient boosting tree, and only estimates the conversion rate from the dimension of behavior frequency. It is impossible to make cross-modal semantic association between the two, resulting in a one-sided understanding of user intentions; (2) static intention prediction: traditional methods mostly make binary judgments of "whether to buy" based on the statistical characteristics of users' historical behavior, regard consumption intentions as static labels, and ignore the dynamic changes of users' consumption intentions with the evolution of life events, such as the shift of consumption focus from maternity clothes to infant formula in the process of "preparing for pregnancy" to "raising children".
[0003] Therefore, how to provide an effective technical solution to solve the difficulties in multimodal data fusion and the problem of static intention prediction in existing technologies has become an urgent problem to be solved in existing technologies. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for predicting user intent based on commercial consumption data, in order to solve the above-mentioned problems existing in the prior art.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting user intent based on commercial consumption data, comprising: The system acquires initial social text data and initial consumption behavior data of users across platforms. The initial social text data includes the original corpus of users' life events and consumption needs. The initial consumption behavior data includes the user's click, browsing, favorite, add to cart and purchase behavior sequence, time sequence information and corresponding product information. The initial social text data and initial consumption behavior data are preprocessed to obtain social text data and consumption behavior data. Pre-trained event extraction techniques are used to extract fragments from social text data to obtain event fragments describing life states. Relationships between event fragments are identified based on context. Based on the relationships between event fragments and the event fragments describing life states, an event knowledge graph is constructed. A bipartite graph embedding algorithm is used to extract the association relationships between preset consumption intention category nodes between consumption behavior data and social text data. A consumption graph is constructed based on the association relationships between preset consumption intention category nodes. The event knowledge graph and the consumption graph are fused to obtain a consumption event knowledge graph, which includes time nodes and consumption intention nodes. Historical behavior data of the target user is obtained, and features are extracted from the historical behavior data to obtain a high-dimensional feature set. The high-dimensional feature set includes long-term preference features and contextual features. The high-dimensional feature set is then filtered based on a feature elimination algorithm to obtain a key feature subset. The consumer knowledge graph and a subset of key features are input into a pre-trained intent prediction model to obtain prediction results, which include the predicted probabilities of target users’ various consumption intentions within a specified time period.
[0006] In one possible design, initial social text data and initial consumption behavior data of users across platforms are acquired, including: Web crawling technology is used to obtain users' initial public social text data and corresponding user identifiers from social media platforms; Obtain users' initial consumption behavior data and corresponding user identifiers from authorized datasets on e-commerce platforms; User profiles are obtained by associating initial consumer behavior data and initial social text data based on user identifiers.
[0007] In one possible design, the initial social text data and initial consumption behavior data are preprocessed to obtain social text data and consumption behavior data, including: Data cleaning is performed on the initial social text data to remove blank values and outliers, resulting in cleaned initial social text data. The initial social text data after cleaning is standardized to obtain standardized initial social text data. The initial social text data is then segmented to obtain social text data. Data cleaning is performed on the initial consumer behavior data to remove blank values, thus obtaining the final consumer behavior data.
[0008] In one possible design, pre-trained event extraction techniques are used to extract fragments from social text data to obtain event fragments describing life states. Relationships between these event fragments are identified based on context. Based on these relationships and the event fragments describing life states, an event graph is constructed. A bipartite graph embedding algorithm is used to extract the associations between predefined consumption intention category nodes and consumer behavior data and social text data. Finally, a consumer graph is constructed based on these predefined consumption intention category nodes and their associations, including: Social event blocks are extracted from social text data using a sequence labeling model. These social event blocks are used to represent events centered around verbs or verb-object phrases. The relationship classifier of the pre-trained model is used to determine the temporal relationship of social event blocks in order to determine whether there is a pre-set sequential temporal relationship between events in the same context. Event fragments with pre-set sequential temporal relationships are extracted to obtain event fragments describing life status. Based on the event fragments describing life status and the pre-set sequential temporal relationships, an event graph is constructed. Obtain specific event keywords, and count the frequency of these keywords in comment data corresponding to different consumption intentions. Construct an initial association weight matrix based on the frequency of occurrence, event, and consumption intention. An initial consumption graph is constructed based on an initial association weight matrix. The initial consumption graph includes event nodes and consumption intention nodes. The edge between the event node and the consumption intention node represents the probability of the event appearing in the comments of the product corresponding to the consumption intention. The binary embedding algorithm is used to calculate the embedding representation of each event node and consumption intention node in the initial consumption graph, and the event embedding vector and consumption intention embedding vector are obtained. The similarity between the event embedding vector and the consumption intention embedding vector is calculated to obtain the consumption intention ranking corresponding to a specific event. The initial consumption graph is then optimized based on the consumption intention ranking corresponding to the specific event to obtain the consumption graph.
[0009] In one possible design, features are extracted from the historical behavior data to obtain a high-dimensional feature set, including: The embedding technique is used to convert categorical features into dense vectors to obtain long-term preference features; Based on preset events, historical behavior data is extracted to obtain an event set. Based on the event set, the event nodes are located in the consumer behavior knowledge graph. Based on event nodes, a graph sampling algorithm is used to extract the local subgraph structure centered on the event nodes, thus obtaining the event subgraph. The type, data, and relationship path information of the neighboring nodes in the event subgraph are encoded into feature vectors to obtain the event context features.
[0010] In one possible design, the method further includes: acquiring a training dataset, which includes a historical consumption knowledge graph and key feature training data; and training the intent prediction model based on the training dataset using a binary cross-entropy loss function to obtain the trained intent prediction model.
[0011] In one possible design, the consumer behavior knowledge graph and a subset of key features are input into a pre-trained intent prediction model to obtain prediction results, including: The key feature subset and the consumption event knowledge graph are input into a pre-constructed homogeneous and heterogeneous relationship attention network model, which includes an event encoder and a consumption intention encoder. The target event node is obtained. The initial feature vector of the target event node is obtained using the event encoder. The first-order event neighbor node information of the target event node in the event graph is gathered using the homogeneous relation aggregator. The first-order consumption intention neighbor node information of the target event node in the consumption graph is gathered using the heterogeneous relation aggregator. The neighbor weights of the first-order event neighbor node information and the first-order consumption intention neighbor node information are calculated using the attention mechanism to obtain the event weight and consumption intention weight. The consumption intent encoder is used to aggregate the first-order event neighbor node information and the first-order consumption intent neighbor node information to obtain aggregated node information. The aggregated node information, event weights, consumption intention weights, and key feature subsets are concatenated to obtain concatenated data. This concatenated data is then input into a multilayer perceptron classifier to output prediction results, which include short-term and long-term prediction probabilities for multiple consumption intention categories.
[0012] Secondly, the present invention provides a user intent prediction system based on commercial consumption data, comprising: The preprocessing module is used to acquire initial social text data and initial consumption behavior data of users across platforms. The initial social text data includes the original corpus of users' life events and consumption needs. The initial consumption behavior data includes the user's click, browsing, favorite, add to cart and purchase behavior sequence, time sequence information and corresponding product information. The initial social text data and initial consumption behavior data are preprocessed to obtain social text data and consumption behavior data. The graph construction module is used to extract fragments from social text data using pre-trained event extraction technology to obtain event fragments describing life states, and to identify the relationships between event fragments based on context. Based on the relationships between event fragments and the event fragments describing life states, an event logic graph is constructed. The module also uses a bipartite graph embedding algorithm to extract the association relationships between preset consumption intention category nodes between consumption behavior data and social text data, and constructs a consumption graph based on the association relationships between preset consumption intention category nodes. Finally, the event logic graph and the consumption graph are fused to obtain a consumption event logic knowledge graph, which includes time nodes and consumption intention nodes. The feature extraction module is used to acquire the historical behavior data of the target user and extract three types of features from the historical behavior data to obtain a high-dimensional feature set. The high-dimensional feature set includes instantaneous behavior features, long-term preference features, and contextual features. The high-dimensional feature set is then filtered based on the feature elimination algorithm to obtain a key feature subset. The intent prediction module is used to input the consumer knowledge graph and a subset of key features into a pre-trained intent prediction model to obtain prediction results, which include the predicted probability of target users’ various consumption intentions within a specified time period.
[0013] Thirdly, the present invention provides a computer device comprising a memory, a processor, and a transceiver connected in sequence and communication, wherein the memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the user intent prediction method based on commercial consumption data as described in the first aspect above.
[0014] Fourthly, the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, perform the user intent prediction method based on commercial consumption data as described in the first aspect above.
[0015] Fifthly, the present invention provides a computer program product containing instructions that, when executed on a computer, cause the computer to perform the user intent prediction method based on commercial consumption data as described in the first aspect above.
[0016] The beneficial effects of this invention are as follows: This invention discloses a user intent prediction method and system based on commercial consumption data. It acquires initial social text data and initial consumption behavior data from cross-platform users. The initial social text data includes raw corpus of users' life events and consumption needs, while the initial consumption behavior data includes user clicks, browsing, favorites, adding to cart, and purchase behavior sequences, time-series information, and corresponding product information. The initial social text data and initial consumption behavior data are preprocessed to obtain social text data and consumption behavior data. Subsequently, a pre-trained event extraction technique is used to extract fragments from the social text data to obtain event fragments describing life states, and these event fragments are identified based on context. Based on the relationships between event fragments and event fragments describing life states, a knowledge graph of consumption events is constructed. A bipartite graph embedding algorithm is used to extract the association relationships between preset consumption intention category nodes between consumption behavior data and social text data. A consumption graph is then constructed based on the association relationships between preset consumption intention category nodes. The knowledge graph of consumption events and the consumption graph are fused to obtain a knowledge graph of consumption events. Historical behavior data of target users is obtained, and features are extracted from the historical behavior data to obtain a high-dimensional feature set. The high-dimensional feature set is filtered based on a feature elimination algorithm to obtain a key feature subset. The knowledge graph of consumption events and the key feature subset are input into a pre-trained intent prediction model to obtain the prediction result. This invention constructs a consumer event knowledge graph, which performs cross-modal semantic association between life events in users' social text data and consumption intentions in consumption behavior data. It also uses a bipartite graph embedding algorithm to solve the problem of separation between behavioral data and text data, which makes reasoning impossible. This significantly enhances the comprehensive perception of users' explicit and implicit consumption intentions. Furthermore, by constructing event and consumption graphs, it enables the inference of short-term intentions based on users' current time and the prediction of long-term intentions, significantly improving prediction accuracy and efficiency, and facilitating application and promotion. Attached Figure Description
[0017] Figure 1 A flowchart illustrating a user intent prediction method based on commercial consumption data provided in an embodiment of the present invention; Figure 2 This is a block diagram of a user intent prediction system based on commercial consumption data provided in an embodiment of the present invention. Detailed Implementation
[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.
[0019] It should be understood that although the terms first, second, etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit, without departing from the scope of the exemplary embodiments of the invention.
[0020] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.
[0021] Example: like Figure 1 As shown, the first aspect of this embodiment provides a user intent prediction method based on commercial consumption data, which can be executed, but is not limited to, by a computer device or virtual machine with certain computing resources, such as a personal computer or smartphone, or by a virtual machine; the user intent prediction method includes, but is not limited to, the following steps: S1. Obtain the user's initial social text data and initial consumption behavior data across platforms. The initial social text data includes the user's original corpus of life events and consumption needs. The initial consumption behavior data includes the user's click, browsing, favorite, add to cart and purchase behavior sequence, time sequence information and corresponding product information. Preprocess the initial social text data and initial consumption behavior data to obtain social text data and consumption behavior data. In step S1, the initial social text data and initial consumption behavior data of users across platforms are obtained, including: S11. Use web crawling technology to obtain users' publicly available initial social text data and corresponding user identifiers from social media platforms; S12. Obtain the user's initial consumption behavior data and the corresponding user identifier from the authorized dataset of the e-commerce platform; S13. Based on the user identifier, the initial consumption behavior data and the initial social text data are associated to obtain the user profile.
[0022] Furthermore, the initial social text data and initial consumption behavior data are preprocessed to obtain social text data and consumption behavior data, including: S14. Perform data cleaning on the initial social text data to remove blank values and outliers, and obtain the cleaned initial social text data; S15. Standardize the cleaned initial social text data to obtain standardized initial social text data. Then, perform word segmentation on the initial social text data to obtain social text data. S16. Perform data cleaning on the initial consumer behavior data to remove blank values and obtain the final consumer behavior data.
[0023] In practice, the process begins by using web crawling technology to obtain initial public social text data and corresponding user identifiers from social media platforms. It also involves obtaining initial consumer behavior data and corresponding user identifiers from authorized datasets on e-commerce platforms, containing sequences and time-series information of clicks, browsing, favorites, adding to cart, and purchase behaviors. These two types of data are then linked based on the user identifiers to form a unified user profile. Subsequently, the initial social text data undergoes data cleaning to remove blank values and outliers, standardization, and word segmentation to obtain structured social text data. Simultaneously, the initial consumer behavior data is cleaned to remove blank values, resulting in consumer behavior data that can be directly used for subsequent modeling.
[0024] S2. Using pre-trained event extraction technology, fragments are extracted from social text data to obtain event fragments describing life states. Relationships between event fragments are identified based on context. Based on the relationships between event fragments and the event fragments describing life states, an event knowledge graph is constructed. The association relationships between consumer behavior data and social text data based on the bipartite graph embedding algorithm are extracted. The association relationships between the preset consumer intention category nodes are constructed based on the preset consumer intention category nodes. The event knowledge graph and the consumer knowledge graph are fused to obtain a consumer event knowledge graph, which includes time nodes and consumer intention nodes. Specifically, in step S2, pre-trained event extraction techniques are used to extract fragments from social text data to obtain event fragments describing life states. Relationships between event fragments are identified based on context. Based on these relationships and the event fragments describing life states, an event graph is constructed. A bipartite graph embedding algorithm is used to extract the association relationships between preset consumption intention category nodes in consumer behavior data and social text data. Finally, a consumer graph is constructed based on these preset consumption intention category nodes and their association relationships, including: S21. Use a sequence labeling model to extract social event blocks from social text data, wherein the social event blocks are used to represent events centered on verbs or verb-object phrases; S22. Use a pre-trained model's relation classifier to determine the temporal relationship of social event blocks, in order to determine whether there is a pre-defined sequential temporal relationship between events in the same context, extract event fragments with a pre-defined sequential temporal relationship, obtain event fragments describing life status, and construct an event graph based on event fragments describing life status and pre-defined sequential temporal relationships; S23. Obtain the set specific event keywords, and count the frequency of the specific event keywords in the comment data corresponding to different consumption intentions, and construct an initial association weight matrix based on the frequency of occurrence, event and consumption intention; S24. Construct an initial consumption graph based on the initial association weight matrix. The initial consumption graph includes event nodes and consumption intention nodes. The edge between the event node and the consumption intention node represents the probability of the event appearing in the comments of the product corresponding to the consumption intention. S25. Use the binary embedding algorithm to calculate the embedding representation of each event node and consumption intention node in the initial consumption graph, obtain the event embedding vector and consumption intention embedding vector, calculate the similarity between the event embedding vector and the consumption intention embedding vector, obtain the consumption intention ranking corresponding to the specific event, and optimize the initial consumption graph based on the consumption intention ranking corresponding to the specific event to obtain the consumption graph.
[0025] It should be noted that the sequence labeling model is built on the BERT (Bidirectional Encoder Representations from Transformers) model. The principle of the BERT model is to generate dynamic, context-related word vector representations for natural language processing tasks through bidirectional context understanding and pre-training-fine-tuning paradigm. Sequential temporal relationship refers to the existence of a sequential relationship in a temporal sense. In this embodiment, it mainly focuses on the temporal relationship between events such as "after".
[0026] In practice, a sequence labeling model is used to extract social event blocks centered on verbs or verb-object phrases from social text data, ensuring that the event fragments are semantically complete while maintaining concise expression, and supporting character-level non-continuous labeling decoding. Subsequently, a relation classifier is used to determine the sequential temporal relationship between events in the same context in order to extract event fragments with semantic associations, thereby forming a contextual graph with event fragments as nodes and sequential relationships as directed edges. In constructing the consumer graph, review data is used as a weakly supervised signal source: by statistically analyzing the co-occurrence frequency of product reviews corresponding to different consumer intent categories for specific event terms, an initial "event-consumption intent" association weight matrix is constructed. This initial association weight matrix is modeled as a bipartite graph composed of event nodes and consumer intent nodes. The Bipartite Network Embedding (BiNE) algorithm is used to iteratively optimize the vector representations of event nodes and consumer intents by simultaneously utilizing first-order explicit associations and second-order implicit adjacency relationships. The similarity between the event embedding vector and the consumer intent embedding vector is calculated to obtain the consumer intent ranking corresponding to a specific event. Subsequently, the initial association edges are optimized to obtain the consumer graph, which is then merged with the event logic graph to form a consumer event logic knowledge graph.
[0027] S3. Obtain the historical behavior data of the target user, and extract features from the historical behavior data to obtain a high-dimensional feature set. The high-dimensional feature set includes long-term preference features and contextual features. Based on the feature elimination algorithm, the high-dimensional feature set is filtered to obtain a key feature subset. Specifically, features are extracted from historical behavioral data to obtain a high-dimensional feature set, including: S31. Use Embedding technology to convert categorical features into dense vectors to obtain long-term preference features; S32. Extract events from historical behavior data based on preset events to obtain an event set, and locate the event nodes in the consumer behavior knowledge graph based on the event set; S33. Based on event nodes, a graph sampling algorithm is used to extract the local subgraph structure centered on the event nodes to obtain the event subgraph. The type, data and relationship path information of the neighboring nodes in the event subgraph are encoded into feature vectors to obtain the event context features.
[0028] In specific implementation, embedding technology is used to map the product category and user level, and other category feature data in the user's historical behavior data, into low-dimensional dense vectors to form a long-term preference feature representation of the user. Then, based on the user's recently published social text data, the specific life events in the user's current situation are identified to obtain the user's event set. Each event in the event set is used as a query anchor point to locate the node in the consumption event knowledge graph. Graph sampling algorithms (such as neighbor sampling) are used to extract the local subgraph structure centered on the event node. The neighbor node type (such as event or consumption intention), relationship path (such as sequential edge or mapping edge), hop count, and edge weight in the local subgraph are encoded into feature vectors to obtain the event context features. The event context features are used to characterize the user's event context features in the current situation.
[0029] S4. Input the consumer knowledge graph and key feature subset into a pre-trained intent prediction model to obtain prediction results, which include the predicted probabilities of target users’ various consumption intentions within a specified time period.
[0030] Specifically, the consumer knowledge graph and a subset of key features are input into a pre-trained intent prediction model to obtain prediction results, including: S41. Input the key feature subset and the consumption event knowledge graph into a pre-constructed homogeneous and heterogeneous relationship attention network model, wherein the homogeneous and heterogeneous relationship attention network model includes an event encoder and a consumption intention encoder; S42. Obtain the target event node, use the event encoder to obtain the initial feature vector of the target event node, use the homogeneous relation aggregator to gather the first-order event neighbor node information of the target event node in the event graph, use the heterogeneous relation aggregator to gather the first-order consumption intention neighbor node information of the target event node in the consumption graph, use the attention mechanism to calculate the neighbor weights of the first-order event neighbor node information and the first-order consumption intention neighbor node information respectively, and obtain the event weight and consumption intention weight. S43. Use a consumption intent encoder to aggregate the first-order event neighbor node information and the first-order consumption intent neighbor node information to obtain aggregated node information; S44. The aggregation node information, event weights, consumption intention weights, and key feature subsets are concatenated to obtain concatenated data. The concatenated data is then input into a multilayer perceptron classifier to output the short-term and long-term prediction probabilities for multiple consumption intention categories.
[0031] In practice, the key feature subset and the consumption event knowledge graph are input into a pre-constructed homogeneous-heterogeneous relationship attention network model. For a target event node, the homogeneous relationship aggregator in the event encoder gathers its first-order event neighbor node information in the event knowledge graph based on an attention mechanism. This first-order event neighbor node information reflects what events might occur after the event occurs. The heterogeneous relationship aggregator gathers its first-order consumption intention neighbor node information in the consumption graph. This first-order consumption intention neighbor node information reflects what needs to be purchased at the moment the event occurs. The attention weights of the attention mechanism are calculated by the interaction between the target node features and the neighbor node features, and are used to represent different... The contribution of neighbors to the current intention prediction is assessed. A consumption intention encoder is used to aggregate homogeneous and heterogeneous neighbors of the consumption intention node. Homogeneous neighbors are used to represent similar consumption intentions, while heterogeneous neighbors are used to represent the events that trigger the consumption intention. The output vectors of the event encoder and the consumption intention encoder are concatenated and deeply fused with long-term preference features and event context features to obtain a prediction feature tensor. The prediction feature tensor is input into a multilayer perceptron classifier and mapped by the Sigmoid function to output the short-term prediction probability and long-term prediction probability of the target user for multiple consumption intention categories. The short-term prediction probability is based on direct association with events, while the long-term prediction probability is based on multi-hop propagation of event paths.
[0032] In a preferred embodiment, the method further includes: acquiring a training dataset, the training dataset including a historical consumption knowledge graph and key feature training data; and training the intent prediction model based on the training dataset using a binary cross-entropy loss function as the loss function to obtain the trained intent prediction model.
[0033] like Figure 2 As shown, the second aspect of this embodiment provides a user intent prediction system based on commercial consumption data, including: The preprocessing module is used to acquire initial social text data and initial consumption behavior data of users across platforms. The initial social text data includes the original corpus of users' life events and consumption needs. The initial consumption behavior data includes the user's click, browsing, favorite, add to cart and purchase behavior sequence, time sequence information and corresponding product information. The initial social text data and initial consumption behavior data are preprocessed to obtain social text data and consumption behavior data. The graph construction module is used to extract fragments from social text data using pre-trained event extraction technology to obtain event fragments describing life states, and to identify the relationships between event fragments based on context. Based on the relationships between event fragments and the event fragments describing life states, an event logic graph is constructed. The module also uses a bipartite graph embedding algorithm to extract the association relationships between preset consumption intention category nodes between consumption behavior data and social text data, and constructs a consumption graph based on the association relationships between preset consumption intention category nodes. Finally, the event logic graph and the consumption graph are fused to obtain a consumption event logic knowledge graph, which includes time nodes and consumption intention nodes. The feature extraction module is used to acquire the historical behavior data of the target user and extract three types of features from the historical behavior data to obtain a high-dimensional feature set. The high-dimensional feature set includes instantaneous behavior features, long-term preference features, and contextual features. The high-dimensional feature set is then filtered based on the feature elimination algorithm to obtain a key feature subset. The intent prediction module is used to input the consumer knowledge graph and a subset of key features into a pre-trained intent prediction model to obtain prediction results, which include the predicted probability of target users’ various consumption intentions within a specified time period.
[0034] The working process, working details and technical effects of the user intent prediction system based on commercial consumption data provided in the second aspect of this embodiment can be found in the user intent prediction method based on commercial consumption data described in the first aspect, and will not be repeated here.
[0035] This embodiment provides a computer device including a memory, a processor, and a transceiver connected in sequence. The memory stores a computer program, the transceiver sends and receives messages, and the processor reads the computer program and executes the user intent prediction method based on commercial consumer data as described in the first aspect. Specifically, the memory may include, but is not limited to, random-access memory (RAM), read-only memory (ROM), flash memory, first-input first-output (FIFO), and / or first-input last-output (FILO), etc.; the processor may include, but is not limited to, a microprocessor of the STM32F105 series. Furthermore, the computer device may also include, but is not limited to, a power module, a display screen, and other necessary components.
[0036] The working process, working details and technical effects of the aforementioned computer device provided in the third aspect of this embodiment can be found in the user intent prediction method based on commercial consumption data described in the first aspect, and will not be repeated here.
[0037] The fourth aspect of this embodiment provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, perform the user intent prediction method based on commercial consumption data as described in the first aspect. The computer-readable storage medium refers to a data storage medium, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or Memory Sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.
[0038] The working process, working details and technical effects of the aforementioned computer-readable storage medium provided in the fourth aspect of this embodiment can be found in the user intent prediction method based on commercial consumption data as described in the first aspect, and will not be repeated here.
[0039] The fifth aspect of this embodiment provides a computer program product, including a computer program or instructions, which, when executed by a computer, are used to implement the user intent prediction method based on commercial consumption data as described in the first aspect.
[0040] The working process, working details, and technical effects of the aforementioned computer program product provided in this embodiment can be found in the user intent prediction method based on commercial consumption data as described in the first aspect, and will not be repeated here.
[0041] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for predicting user intent based on commercial consumption data, characterized in that, include: The system acquires initial social text data and initial consumption behavior data of users across platforms. The initial social text data includes the original corpus of users' life events and consumption needs. The initial consumption behavior data includes the user's click, browsing, favorite, add to cart and purchase behavior sequence, time sequence information and corresponding product information. The initial social text data and initial consumption behavior data are preprocessed to obtain social text data and consumption behavior data. Pre-trained event extraction techniques are used to extract fragments from social text data to obtain event fragments describing life states. Relationships between event fragments are identified based on context. Based on the relationships between event fragments and the event fragments describing life states, an event knowledge graph is constructed. A bipartite graph embedding algorithm is used to extract the association relationships between preset consumption intention category nodes between consumption behavior data and social text data. A consumption graph is constructed based on the association relationships between preset consumption intention category nodes. The event knowledge graph and the consumption graph are fused to obtain a consumption event knowledge graph, which includes time nodes and consumption intention nodes. Historical behavior data of the target user is obtained, and features are extracted from the historical behavior data to obtain a high-dimensional feature set. The high-dimensional feature set includes long-term preference features and contextual features. The high-dimensional feature set is then filtered based on a feature elimination algorithm to obtain a key feature subset. The consumer knowledge graph and a subset of key features are input into a pre-trained intent prediction model to obtain prediction results, which include the predicted probabilities of target users’ various consumption intentions within a specified time period.
2. The user intent prediction method based on commercial consumption data according to claim 1, characterized in that, Acquire users' initial social text data and initial consumption behavior data across platforms, including: Web crawling technology is used to obtain users' initial public social text data and corresponding user identifiers from social media platforms; Obtain users' initial consumption behavior data and corresponding user identifiers from authorized datasets on e-commerce platforms; User profiles are obtained by associating initial consumer behavior data and initial social text data based on user identifiers.
3. The user intent prediction method based on commercial consumption data according to claim 1, characterized in that, Preprocessing the initial social text data and initial consumption behavior data yields the following: Data cleaning is performed on the initial social text data to remove blank values and outliers, resulting in cleaned initial social text data. The initial social text data after cleaning is standardized to obtain standardized initial social text data. The initial social text data is then segmented to obtain social text data. Data cleaning is performed on the initial consumer behavior data to remove blank values, thus obtaining the final consumer behavior data.
4. The user intent prediction method based on commercial consumption data according to claim 1, characterized in that, Pre-trained event extraction techniques are used to extract fragments from social text data, resulting in event fragments describing life states. Relationships between these event fragments are identified based on context. Based on these relationships and the event fragments describing life states, an event graph is constructed. A bipartite graph embedding algorithm is used to extract the associations between consumer behavior data and social text data for preset consumer intention category nodes. Finally, a consumer graph is constructed based on these preset consumer intention category nodes and their associations, including: Social event blocks are extracted from social text data using a sequence labeling model. These social event blocks are used to represent events centered around verbs or verb-object phrases. The relationship classifier of the pre-trained model is used to determine the temporal relationship of social event blocks in order to determine whether there is a pre-set sequential temporal relationship between events in the same context. Event fragments with pre-set sequential temporal relationships are extracted to obtain event fragments describing life status. Based on the event fragments describing life status and the pre-set sequential temporal relationships, an event graph is constructed. Obtain specific event keywords, and count the frequency of these keywords in comment data corresponding to different consumption intentions. Construct an initial association weight matrix based on the frequency of occurrence, event, and consumption intention. An initial consumption graph is constructed based on an initial association weight matrix. The initial consumption graph includes event nodes and consumption intention nodes. The edge between the event node and the consumption intention node represents the probability of the event appearing in the comments of the product corresponding to the consumption intention. The binary embedding algorithm is used to calculate the embedding representation of each event node and consumption intention node in the initial consumption graph, and the event embedding vector and consumption intention embedding vector are obtained. The similarity between the event embedding vector and the consumption intention embedding vector is calculated to obtain the consumption intention ranking corresponding to a specific event. The initial consumption graph is then optimized based on the consumption intention ranking corresponding to the specific event to obtain the consumption graph.
5. The user intent prediction method based on commercial consumption data according to claim 1, characterized in that, Feature extraction was performed on the historical behavior data to obtain a high-dimensional feature set, including: The embedding technique is used to convert categorical features into dense vectors to obtain long-term preference features; Based on preset events, historical behavior data is extracted to obtain an event set. Based on the event set, the event nodes are located in the consumer behavior knowledge graph. Based on event nodes, a graph sampling algorithm is used to extract the local subgraph structure centered on the event nodes, thus obtaining the event subgraph. The type, data, and relationship path information of the neighboring nodes in the event subgraph are encoded into feature vectors to obtain the event context features.
6. The user intent prediction method based on commercial consumption data according to claim 1, characterized in that, Also includes: A training dataset is obtained, which includes a historical consumption knowledge graph and key feature training data. Based on the training dataset, the intention prediction model is trained using a binary cross-entropy loss function to obtain the trained intention prediction model.
7. The user intent prediction method based on commercial consumption data according to claim 1, characterized in that, By inputting the consumer behavior knowledge graph and a subset of key features into a pre-trained intent prediction model, prediction results are obtained, including: The key feature subset and the consumption event knowledge graph are input into a pre-constructed homogeneous and heterogeneous relationship attention network model, which includes an event encoder and a consumption intention encoder. The target event node is obtained. The initial feature vector of the target event node is obtained using the event encoder. The first-order event neighbor node information of the target event node in the event graph is gathered using the homogeneous relation aggregator. The first-order consumption intention neighbor node information of the target event node in the consumption graph is gathered using the heterogeneous relation aggregator. The neighbor weights of the first-order event neighbor node information and the first-order consumption intention neighbor node information are calculated using the attention mechanism to obtain the event weight and consumption intention weight. The consumption intent encoder is used to aggregate the first-order event neighbor node information and the first-order consumption intent neighbor node information to obtain aggregated node information. The aggregated node information, event weights, consumption intention weights, and key feature subsets are concatenated to obtain concatenated data. This concatenated data is then input into a multilayer perceptron classifier to output prediction results, which include short-term and long-term prediction probabilities for multiple consumption intention categories.
8. A user intent prediction system based on commercial consumption data, used to implement the method according to any one of claims 1 to 7, characterized in that, include: The preprocessing module is used to acquire initial social text data and initial consumption behavior data of users across platforms. The initial social text data includes the original corpus of users' life events and consumption needs. The initial consumption behavior data includes the user's click, browsing, favorite, add to cart and purchase behavior sequence, time sequence information and corresponding product information. The initial social text data and initial consumption behavior data are preprocessed to obtain social text data and consumption behavior data. The graph construction module is used to extract fragments from social text data using pre-trained event extraction technology to obtain event fragments describing life states, and to identify the relationships between event fragments based on context. Based on the relationships between event fragments and the event fragments describing life states, an event logic graph is constructed. The module also uses a bipartite graph embedding algorithm to extract the association relationships between preset consumption intention category nodes between consumption behavior data and social text data, and constructs a consumption graph based on the association relationships between preset consumption intention category nodes. Finally, the event logic graph and the consumption graph are fused to obtain a consumption event logic knowledge graph, which includes time nodes and consumption intention nodes. The feature extraction module is used to acquire the historical behavior data of the target user and extract three types of features from the historical behavior data to obtain a high-dimensional feature set. The high-dimensional feature set includes instantaneous behavior features, long-term preference features, and contextual features. The high-dimensional feature set is then filtered based on the feature elimination algorithm to obtain a key feature subset. The intent prediction module is used to input the consumer knowledge graph and a subset of key features into a pre-trained intent prediction model to obtain prediction results, which include the predicted probability of target users’ various consumption intentions within a specified time period.
9. A computer device, characterized in that, The device includes a memory, a processor, and a transceiver that are sequentially and communicatively connected. The memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the user intent prediction method based on commercial consumption data as described in any one of claims 1 to 7.
10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or the instructions are executed by the computer, they implement the user intent prediction method based on commercial consumption data as described in any one of claims 1 to 7.