Marketing crowd feature accurate analysis method and system based on machine learning

Abstract

The invention relates to the technical field of machine learning, and particularly discloses a marketing crowd feature accurate analysis method and system based on machine learning, and the method comprises the following steps: constructing an initial feature vector of each user according to multi-source data; constructing a user attention network diagram based on the attention relationship data between the users, and calculating the weight of each edge in the user attention network diagram based on the initial feature vector; constructing a graph convolutional network model and adjusting parameters of the model until convergence; obtaining an updated feature vector of each user, and clustering the updated feature vectors to obtain a plurality of initial clusters and a center vector of each initial cluster; and obtaining an updated feature vector A of the to-be-analyzed user, and determining a marketing population feature to which the to-be-analyzed user belongs. According to the method, the reasonability and accuracy of marketing crowd feature recognition are improved.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, specifically to a method and system for precise analysis of marketing audience characteristics based on machine learning. Background Technology

[0002] In internet marketing scenarios, user behavior data is characterized by diverse sources, complex structures, and frequent changes. It includes interactions and content expressed on social media platforms, as well as browsing, clicking, and purchasing behaviors on e-commerce platforms. While these data exhibit significant heterogeneity and correlation, they are often used in isolation in practice, resulting in user profiles remaining at a single dimension and failing to form stable and comprehensive user personas.

[0003] Meanwhile, the network of relationships among users influences their interests and consumption tendencies. However, traditional audience segmentation methods struggle to effectively depict the impact of these relationships on feature distribution, resulting in inaccurate, unstable, and undynamically updated marketing audience segmentation results. Therefore, how to uniformly model user characteristics in a multi-source data environment and achieve refined identification and differentiation of marketing audience characteristics within complex relationship structures has become a pressing technical challenge in the field of marketing analytics. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for accurate analysis of marketing audience characteristics based on machine learning, thereby solving the aforementioned technical problems.

[0005] The objective of this invention can be achieved through the following technical solutions: A machine learning-based method for precise analysis of marketing audience characteristics includes the following steps: Acquire multi-source data of users, and construct an initial feature vector for each user based on the multi-source data, which includes text content published on social platforms, clickstream data and purchase records on e-commerce platforms; A user attention network graph is constructed based on the attention relationship data between users, and the weight of each edge in the user attention network graph is calculated based on the initial feature vector; A graph convolutional network model is constructed and trained using a self-supervised learning approach. During training, edges in the network graph that users follow are used as positive samples, and user pairs without follow relationships are used as negative samples. The model parameters are adjusted until convergence. The user attention network graph and the initial feature vector of each user are input into the trained graph convolutional network model to obtain the updated feature vector of each user. The updated feature vectors are then clustered to obtain multiple initial clusters and the center vector of each initial cluster. Obtain the initial feature vector of the user to be analyzed and update the user attention network graph based on the user's attention relationship data. Based on this, obtain the updated feature vector A of the user to be analyzed, calculate the distance between the updated feature vector A and each center vector, and take the initial cluster with the smallest distance as the marketing audience feature to which the user to be analyzed belongs.

[0006] Preferably, the process of constructing the initial feature vector is as follows: A word segmentation tool based on a hidden Markov model is used to segment the text content into word sequences. The word sequences are then matched with a preset sentiment dictionary, which includes a list of positive sentiment words and a list of negative sentiment words. The number of matched positive sentiment words and the number of matched negative sentiment words are counted, and the ratio of the number of positive sentiment words to the number of matched negative sentiment words is calculated. When the denominator is zero, the ratio is set to a preset fixed value to obtain the sentiment tendency value. The clickstream data is arranged in chronological order of click time. The product category codes corresponding to each click are combined into a click sequence. The frequency of each product category code in the click sequence is counted to generate a click frequency vector. Iterate through each order in the purchase history, extract the product category code from the order, count the number of times each product category code appears, and generate a purchase frequency vector; The sentiment index, click frequency vector, and purchase frequency vector are concatenated to form the initial feature vector.

[0007] Preferably, the process of calculating the weight of each edge in the user-focused network graph is as follows: A user attention network graph is constructed based on the attention relationship data between users. The nodes of the user attention network graph are users, and the edges represent the attention relationship between two users. For each edge, obtain the cosine similarity between the initial feature vectors of the first and second users connected by the edge, and use the cosine similarity as the weight of the edge; When the cosine similarity is less than a preset threshold, the edge is removed from the user-focused network graph.

[0008] Preferably, the graph convolutional network model includes a first graph convolutional layer and a second graph convolutional layer. The processing procedure of the first graph convolutional layer is as follows: for each node, obtain the initial feature vector of the node and the initial feature vectors of all first-order neighbor nodes, multiply the initial feature vectors of the first-order neighbor nodes by the weights of the corresponding edges and sum them, and then add them to the initial feature vector of the current node to obtain the first layer output vector. The processing procedure of the second convolutional layer is as follows: obtain the first layer output vector and the first layer output vectors of all first-order neighbor nodes, multiply the first layer output vectors of the first-order neighbor nodes with the weights of the corresponding edges and sum them, and then add them to the first layer output vector of the current node to obtain the updated feature vector.

[0009] Preferably, the specific process of training the graph convolutional network model is as follows: A predetermined number of edges are randomly selected from the user attention network graph as the positive sample set, and the same number of user pairs without attention relationships are randomly selected as the negative sample set. The initial feature vectors of the two users corresponding to each edge in the positive sample set are input into the graph convolutional network model to obtain the updated feature vectors of the two users. The inner product of the two updated feature vectors is calculated as the positive sample similarity. The initial feature vectors of the two users contained in each user pair in the negative sample set are input into the graph convolutional network model to obtain the updated feature vectors of the two users. The inner product of the two updated feature vectors is calculated as the negative sample similarity. A loss function is constructed, which is the sum of the similarities of all negative samples minus the sum of the similarities of all positive samples. The parameters of the graph convolutional network model are updated using the gradient descent method. The above process is repeated until the value of the loss function no longer decreases.

[0010] Preferably, obtaining multiple initial clusters and center vectors includes: A cohesive hierarchical clustering algorithm is used to coarsely classify the updated feature vectors of all users. The Euclidean distance between every two updated feature vectors is calculated. Each user's updated feature vector is used as an initial cluster. The two clusters with the closest Euclidean distance are iteratively merged until the number of clusters is reduced to a preset first number, thus obtaining the coarsely classified clusters. For each coarse cluster, the K-means clustering algorithm is used to divide the updated feature vectors of all users within the cluster into multiple fine clusters. A set number of updated feature vectors are randomly selected as cluster centers. The Euclidean distance between each updated feature vector and each cluster center is calculated, and each updated feature vector is assigned to the cluster to which the nearest cluster center belongs. The average value of all updated feature vectors in each cluster is recalculated as the new cluster center. The above steps are repeated until the cluster centers no longer change or the preset number of iterations is reached, resulting in multiple fine clusters. Each fine cluster is used as the initial cluster, and the average value of all updated feature vectors in each fine cluster is used as the center vector of the initial cluster.

[0011] Preferably, the process of obtaining the updated feature vector A of the user to be analyzed is as follows: If the user to be analyzed has neighboring nodes in the user attention network graph, then the user to be analyzed is added as a new node to the user attention network graph, and the updated feature vector A of the user to be analyzed is obtained by inputting the updated user attention network graph and the initial feature vectors of all users into the trained graph convolutional network model. If the user to be analyzed has no neighboring nodes in the user attention network graph, then the initial feature vector of the user to be analyzed is directly input into the trained graph convolutional network model to obtain the updated feature vector A of the user to be analyzed.

[0012] A machine learning-based system for precise analysis of marketing audience characteristics includes: Data input module: Acquires multi-source data of users and constructs an initial feature vector for each user based on the multi-source data, which includes text content published on social platforms, clickstream data and purchase records on e-commerce platforms; Attention Fusion Module: Constructs a user attention network graph based on the attention relationship data between users, and calculates the weight of each edge in the user attention network graph based on the initial feature vector; Update module: Construct a graph convolutional network model and train it using a self-supervised learning approach. During training, edges in the user-focused network graph are used as positive samples, and user pairs without any focus relationship are used as negative samples. The model parameters are adjusted until convergence. The user attention network graph and the initial feature vector of each user are input into the trained graph convolutional network model to obtain the updated feature vector of each user. The updated feature vectors are then clustered to obtain multiple initial clusters and the center vector of each initial cluster. Analysis module: Obtain the initial feature vector of the user to be analyzed and update the user attention network graph according to the attention relationship data of the user to be analyzed. Based on this, obtain the updated feature vector A of the user to be analyzed, calculate the distance between the updated feature vector A and each center vector, and take the initial cluster with the smallest distance as the marketing audience feature to which the user to be analyzed belongs.

[0013] The beneficial effects of this invention compared to the prior art are as follows: This invention enables a more systematic and consistent characterization of marketing audience features in an environment where multi-source user behavior and relationship information coexist, thereby improving the accuracy and stability of audience feature identification. Through unified modeling and updating of user features, the invention comprehensively reflects users' interests and behavioral tendencies in different scenarios, avoiding the problem of one-sided features caused by a single data source. Furthermore, when relationships exist between users, this invention effectively transmits and supplements individual features within the group structure, thereby improving the ability to identify potential interests and consumption tendencies. The resulting audience segmentation has better discriminative power and consistency, facilitating the refined positioning of marketing targets and adapting to the feature update needs brought about by changes in user behavior, ensuring the continuity and scalability of audience analysis results. Moreover, this method reduces the impact of random factors on the results during audience feature classification and matching, making the identification of marketing audience features more stable and reliable, and providing a more valuable user feature foundation for subsequent marketing strategy formulation and implementation. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart illustrating the precise analysis method for marketing audience characteristics based on machine learning, as described in this invention.

[0016] Figure 2 This is a schematic diagram of the training graph convolutional network model of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figures 1-2 As shown, this invention is a method for precise analysis of marketing audience characteristics based on machine learning, including the following steps: Acquire multi-source data of users and construct an initial feature vector for each user based on the multi-source data, which includes text content published on social platforms, clickstream data and purchase records on e-commerce platforms.

[0019] In a preferred embodiment of the present invention, the process of constructing the initial feature vector is as follows: When acquiring multi-source user data, the user's unique identifier is used as the basis for association. Behavioral information from different platforms is uniformly aggregated. Multi-source data includes text content posted by users on social media platforms, clickstream data generated on e-commerce platforms, and purchase records. Clickstream data refers to continuous access behavior data generated by users while browsing products or pages, including click time, page identifier, product identifier, and corresponding product category code. Purchase records refer to order information generated by users after a transaction is completed, including order number, product category, quantity, and transaction time. During the data aggregation process, data from different sources are time-aligned and identity-mapped to ensure that the behavioral information of the same user can be described under a unified timeline. Based on this, the initial feature vector of each user is constructed.

[0020] For the text content, a word segmentation tool based on a Hidden Markov Model (HMM) is selected to process the user-posted text. A pre-trained state transition probability and emission probability model is used to segment the continuous character sequence, resulting in a word sequence with semantic boundaries. This word sequence is then matched one by one with a pre-built sentiment dictionary. The sentiment dictionary includes lists of positive and negative sentiment words, and common online expressions are standardized during dictionary construction to map different expressions to a unified term. During the matching process, the number of positive and negative sentiment words appearing in the word sequence is recorded, and their quantitative relationship is converted into a sentiment tendency value reflecting emotional inclination. When no negative sentiment word is matched, the corresponding value in the denominator is set to a preset fixed value to avoid interference from abnormal situations. This yields a feature value representing the user's emotional expression tendency. Subsequently, clickstream data is processed. The data is sorted according to the time sequence of clicks, and the product category codes corresponding to each click are sequentially arranged to form a click sequence. For example, a click sequence formed within a certain time period can be represented by codes such as 1, 3, 3, 5, 2, etc. Each product category in this sequence is then analyzed. The frequency of occurrence of codes is statistically analyzed to generate a click frequency vector reflecting the distribution of user browsing interests. Each dimension of this vector corresponds to a product category, and the value represents the number of times that category appears in the click sequence. Purchase records are then traversed, and the product category codes in each order are read and categorized. The number of times each product category appears in historical orders is recorded, forming a purchase frequency vector. This vector describes the distribution of user consumption tendencies at the actual transaction level. During the statistical process, multiple products in the same order are counted according to their corresponding categories, ensuring the vector accurately reflects the user's consumption structure. The sentiment tendency value, click frequency vector, and purchase frequency vector are concatenated in a unified order, allowing features from different sources to form a continuous expression in the same vector space. The concatenated vector dimension is composed of the sentiment feature dimension and the click and purchase statistics dimensions corresponding to each product category, thus forming the initial feature vector for each user. This vector reflects the user's emotional expression in the social environment, their interest preferences in browsing behavior, and the demand tendencies reflected in actual consumption behavior, enabling a unified description of user characteristics within the same structure. The sentiment tendency value can also be modified by combining sentiment intensity weights, text length normalization, or the output probability of the sentiment classification model to improve the stability of feature expression.

[0021] Users' expressions on social media platforms, browsing behavior in e-commerce environments, and actual transaction records reflect their emotional attitudes, interests, and consumption tendencies, respectively. These information essentially represent the behavior of the same individual in different scenarios. Unifying their representation allows disparate information to form a continuous data description within the same space, revealing the intrinsic relationships between individual behaviors. Emotional tendencies in text reflect changes in users' attitudes towards things; the frequency distribution of click behavior reflects the direction of attention and the scope of interests; and purchase records reflect real needs and decision-making outcomes. Transforming this information into structured features and uniformly encoding it can eliminate interference from differences in the original data format, making information from different sources comparable and computable. Simultaneously representing the characteristics of the expression layer, attention layer, and consumption layer in the same vector expression helps to form a more complete individual profile. When multidimensional features are represented in the same space, the similarity and differences between users can be directly characterized, making individuals with similar behavioral patterns tend to cluster in feature distribution, while individuals with obvious differences in behavioral features are naturally separated in feature space. This provides a stable basis for subsequent population segmentation, enabling marketing audience features to be extracted from discrete behaviors and formed into continuous expressions, providing a foundation for identifying people with similar interest structures and consumption tendencies.

[0022] A user attention network graph is constructed based on the attention relationship data between users, and the weight of each edge in the user attention network graph is calculated based on the initial feature vector.

[0023] In a preferred embodiment of the present invention, the process of calculating the weight of each edge in the user-focused network graph is as follows: After constructing the initial user feature vectors, a user attention network graph is built based on the attention relationship data formed between users. The attention relationship data comes from the attention behavior records between users on the social platform. This type of data is stored in the form of user identifier pairs. For example, if user 1001 follows user 2034, a record from 1001 to 2034 is formed. During the data collection process, the platform behavior log is read to obtain the attention establishment time, the identifiers of the following parties, and the relationship status. Data that has been unfollowed or has abnormal identifiers is removed. The data is then uniformly mapped using the user's unique identifier, so that users from different data sources can be mapped to the same node.

[0024] Based on this, each user is mapped to a node in the network graph, and an edge is established between two users who have a following relationship, thus forming a user following network graph containing a set of nodes and a set of edges. In this network graph, each edge only indicates that there is a following relationship between two users, without distinguishing the direction of following. The basic structure is formed by organizing the original following data.

[0025] For each edge in the network graph, the edge weight needs to be further calculated to reflect the similarity between two users at the feature level. The vector representations of the first and second users connected by the edge are extracted from the initial user feature vectors that have been constructed. The initial feature vectors include sentiment values, click frequency vectors, and purchase frequency vectors. Each dimension is arranged using a unified encoding method so that the vector dimensions of different users correspond one-to-one. During the calculation process, the directional consistency of the two user vectors is compared to obtain the cosine similarity value, which reflects the similarity between the two feature distributions. This value is used to represent the similarity between the two users in terms of interest structure and consumption tendency. This similarity is directly written into the network graph as the weight of the corresponding edge, so that each edge not only represents the existence of a relationship, but also contains information on the strength of the relationship.

[0026] After the edge weights are calculated, the network structure needs to be filtered. A threshold is pre-set to distinguish between highly relevant and low-relevance user relationships. The similarity of each edge is compared with this threshold. If the similarity is less than the threshold, it is considered that the two users have significantly different behavioral characteristics, and even if there is an interest relationship, it does not have a stable interest connection. This edge is then removed from the user interest network graph, and only edges with similarity higher than the threshold are retained, making the network graph structure more reflective of real interest connections. During the processing, all edges in the network graph are traversed sequentially. For example, in a network with 5000 nodes, edge information is read one by one and weights are calculated. The weight records are stored together with the edge identifiers. After the filtering is completed, a new user interest network graph is formed. The nodes in this network graph still correspond to all users, but the number of edges is reduced, and each edge is accompanied by weight information.

[0027] A graph convolutional network model is constructed and trained using a self-supervised learning approach. During training, edges in the user-focused network graph are used as positive samples, while user pairs without any focus relationship are used as negative samples. The model parameters are adjusted until convergence.

[0028] In a preferred embodiment of the present invention, in the first graph convolutional layer of the graph convolutional network, each node in the network graph is traversed, and the initial feature vector of the user corresponding to the node is read. The initial feature vector is formed by concatenating the sentiment value, click frequency vector, and purchase frequency vector, while maintaining a unified dimensional structure. At the same time, the set of first-order neighbor nodes of the node is retrieved according to the adjacency relationship of the network graph. First-order neighbor nodes refer to all nodes in the network graph that have a direct attention relationship with the current node. The initial feature vectors of each neighbor node are read one by one, and the weight values ​​of the corresponding edges are retrieved. The feature vectors of the neighbor nodes are weighted and accumulated with the corresponding weights, and then fused with the initial feature vector of the current node itself to obtain the output vector of the node in the first graph convolutional layer. The output vector contains a comprehensive representation of the node's own behavioral features and the behavioral features of the neighbor nodes formed under the action of the weights. After the first convolutional layer completes the processing of all nodes, the resulting output vector is used as the input to the second convolutional layer. In the second convolutional layer, the feature propagation process under the adjacency relationship is repeated. For each node, its first-layer output vector is read, and the first-layer output vectors of its first-order neighbor nodes are obtained again. The output vectors of the neighbor nodes are weighted according to the corresponding edge weights and accumulated. After being fused with the first-layer output vector of the current node, the output result of the second layer is formed. This result is the updated feature vector of each node. The updated feature vector is structurally consistent with the initial feature vector, but its numerical expression already includes the information propagation effect from the neighbor nodes. This makes each user's features not only reflect individual behavior, but also reflect their associated position and connection strength in the attention network, providing a unified data representation basis for subsequent feature space-based population segmentation and marketing analysis.

[0029] In another preferred embodiment of the present invention, the specific process of training the graph convolutional network model is as follows: After completing the construction of the user attention network graph and determining the initial feature vector of each user, the graph convolutional network model is trained. The user attention network graph uses users as nodes and the attention relationships between users as edges, with weight information based on the similarity of the users' initial feature vectors attached to the edges.

[0030] During training, a sample set needs to be constructed to guide the model in learning the structural association features between nodes. A set number of edges are randomly selected from the user attention network graph as the positive sample set. Each edge corresponds to a pair of users with an attention relationship. The set number can be determined according to the network size. For example, when there are 10,000 edges in the network, 2,000 of them can be selected as positive samples in a training batch. To ensure the effectiveness of contrastive learning, the same number of user pairs as the positive samples are randomly selected from all user pairs without attention relationships as the negative sample set. User pairs without attention relationships refer to two nodes that are not connected by an edge in the current user attention network graph. The selection of negative samples is randomized from all unconnected node pairs to avoid concentration in a certain local area and affecting the balance of model training.

[0031] After obtaining the positive and negative sample sets, the initial feature vectors of the two users contained in each sample are input into a graph convolutional network model. The graph convolutional network model contains a first graph convolutional layer and a second graph convolutional layer. The first graph convolutional layer receives the initial feature vectors of all nodes and the adjacency information of the user-focused network graph. It aggregates the feature vectors of each node's first-order neighbors and combines them with the corresponding edge weights to form the first-layer output vector. The second graph convolutional layer uses the first-layer output vector as input and continues to aggregate features according to adjacency relationships, outputting an updated feature vector for each node. The model internally sets a trainable parameter matrix to process the aggregated vectors. A linear transformation is performed, and numerical stabilization is applied after each layer's output to ensure the training process can continue. During positive sample calculation, the two user nodes corresponding to each edge in the positive sample set are read, and the updated feature vectors after two layers of graph convolution are obtained. The inner product of the two updated feature vectors is then performed to obtain the positive sample similarity, which reflects the degree of feature proximity between two users with a relationship of interest under the current model parameters. Similarly, during negative sample calculation, the two initial feature vectors of each user pair in the negative sample set are input into the graph convolutional network model to obtain the corresponding two updated feature vectors, and the inner product is calculated as the negative sample similarity.

[0032] After calculating the similarity between all positive and negative samples, a loss function is constructed. This loss function is composed of the sum of the similarities of all negative samples minus the sum of the similarities of all positive samples. This loss function measures the degree of distinction between connected and unconnected users in the feature space under the current model parameters. Subsequently, gradient descent is used to update the parameter matrix in the graph convolutional network model. The gradient calculation is based on the backpropagation of the partial derivative of the loss function with respect to the parameters. After each round of parameter update, the sample selection, feature propagation, and similarity calculation steps are re-executed, forming a multi-round iterative training process. During training, a learning rate parameter can be set to control the magnitude of each parameter update, and a maximum number of iterations can be set as a stopping condition. When the loss function value no longer decreases in several consecutive iterations, the model is considered to have reached convergence. The graph convolutional network model obtained after training can map the initial user feature vector into an updated feature vector containing network structure information. This allows the updated feature vector to simultaneously reflect individual behavioral characteristics and their positional relationship in the attention network in a marketing scenario, thus providing a structured representation basis for subsequent audience segmentation and feature analysis.

[0033] In some implementations, the loss function may further include a normalization term, a regularization term, or a contrastive learning loss form to avoid numerical divergence or excessive scaling of the updated feature vectors during training. The loss function may be, but is not limited to, a contrastive loss function, a cross-entropy loss function, a BPR loss function, or a similarity optimization function with boundary constraints.

[0034] It should be noted that graph convolutional layers may also include normalization processing, linear mapping parameter matrices, and nonlinear activation functions to improve model expressiveness and training stability. In some implementations, graph convolution operations based on adjacency matrix normalization can be used.

[0035] It is understandable that the attention relationships formed between users essentially reflect the group structure characteristics of similar interests and behavioral associations. Individual behavioral characteristics do not exist in isolation within this structure, but rather exhibit mutual influence and similar clustering patterns within the relationships. By propagating and updating user characteristics within the relationship network, data originally derived solely from individual behavior can be further integrated with the associated information in its environment. This allows feature representation to not only reflect individual attributes but also its position and connection strength within the group. Strengthening feature proximity among users with attention relationships while simultaneously increasing feature distance between users without relationships helps form a distinctive feature distribution structure. This allows users with similar interests and consumption tendencies to gradually cluster in the feature space, while users with different preferences naturally separate. This relationship-driven feature adjustment process transforms data from discrete behavioral records into continuous representations with structural meaning, thereby more accurately depicting the potential association patterns between users in marketing scenarios. As features are continuously iterated and updated within the network structure, they can simultaneously reflect individual behavioral information and group relationship information, resulting in stronger consistency and identifiability of the formed population characteristics.

[0036] The user attention network graph and the initial feature vector of each user are input into the trained graph convolutional network model to obtain the updated feature vector of each user. The updated feature vectors are then clustered to obtain multiple initial clusters and the center vector of each initial cluster.

[0037] It's important to note that during the training phase, inputting two users as a pair is essentially just to construct positive and negative samples to calculate similarity and generate a loss signal. Its purpose is to continuously adjust the parameters of the graph convolutional network model through local samples, enabling the model to learn how to propagate and map node features within the attention relationship structure, rather than simply learning the mapping between the two users. The computational units of the graph convolutional network are designed for the entire graph structure. The parameters within the model are not dependent on a specific user pair, but rather correspond to the aggregation method and mapping rules of node features in adjacency relationships. During training, each time a pair of users is input into the model, the complete graph structure propagation mechanism is invoked, with node features passing information between their neighboring nodes. The model parameters are continuously updated based on the similarity feedback of the positive and negative sample pairs, gradually adapting the parameters to the structural patterns within the entire user attention network. Once training is complete, the model has learned the unified rules for feature propagation under this network structure. At this point, inputting the user-focused network graph and the initial feature vectors of all users into the model is equivalent to performing a complete graph convolution propagation process under fixed parameter conditions. All nodes will synchronously update their feature representations based on their neighbor relationships. The updated feature vector of each user node is formed by the combined effect of its own features and the features of its neighbors, and does not depend on whether a single user pair was used as input during training. Therefore, the updated feature vectors of all users can be obtained at once.

[0038] In a preferred embodiment of the present invention, obtaining multiple initial clusters and center vectors includes: A cohesive hierarchical clustering algorithm is used to coarsely classify the updated feature vectors of all users. The Euclidean distance between every two updated feature vectors is calculated, and each user's updated feature vector is used as an initial cluster. The two clusters with the closest Euclidean distance are iteratively merged until the number of clusters is reduced to a preset first number, thus obtaining the coarsely classified clusters.

[0039] For each coarse cluster, the K-means clustering algorithm is used to divide the updated feature vectors of all users within the cluster into multiple fine clusters. A set number of updated feature vectors are randomly selected as cluster centers. The Euclidean distance between each updated feature vector and each cluster center is calculated, and each updated feature vector is assigned to the cluster to which the nearest cluster center belongs. The average value of all updated feature vectors in each cluster is recalculated as the new cluster center. The above steps are repeated until the cluster centers no longer change or the preset number of iterations is reached, resulting in multiple fine clusters. Each fine cluster is used as the initial cluster, and the average value of all updated feature vectors in each fine cluster is used as the center vector of the initial cluster.

[0040] User updated feature vectors are continuous representations formed after integrating behavioral features and relational structures. Differences in interests and consumption tendencies among different users are represented by distance differences in this feature space. Users with similar feature distributions naturally cluster in the space, while users with significant differences are dispersed. By gradually merging vectors that are close in distance, individuals with similar behavioral patterns can form larger structured groups, transforming the overall data distribution from a disordered state to a hierarchical group form. Further subdivision within each group can identify more subtle differences while maintaining overall similarity. This allows the population structure to have both macro-level category distinctions and stable small-scale clusters in local areas. When vectors within the same group gradually converge around a certain center, that center can represent the typical characteristics of this type of user. Thus, discrete individual behavior is abstracted into representative vectors that can be used to describe group characteristics. Differences between different users can be measured by their proximity to the center, ultimately forming a hierarchical population feature structure. Users with similar interests and consumption tendencies are concentrated in the same structure, while users with significant behavioral differences are naturally separated.

[0041] Obtain the initial feature vector of the user to be analyzed and update the user attention network graph based on the user's attention relationship data. Based on this, obtain the updated feature vector A of the user to be analyzed, calculate the distance between the updated feature vector A and each center vector, and take the initial cluster with the smallest distance as the marketing audience feature to which the user to be analyzed belongs.

[0042] It's important to note that calculating the distance between the updated feature vector of the user to be analyzed and the center vectors of each initial cluster essentially determines which established demographic structure the user's position in the feature space is closest to. Marketing demographic features refer to the collective behavioral characteristics expressed by the updated feature vectors of all users within the same initial cluster. These features comprehensively reflect the overall performance of this user group in terms of interest distribution, consumption tendency structure, and positional characteristics in the relationship network, and are abstractly represented by the center vector. When the distance between a user to be analyzed and a certain center vector is minimized, it indicates that the user is closest to the typical characteristics of this group in terms of multidimensional feature expression, and therefore, the user is classified into the marketing demographic feature category corresponding to that initial cluster. The resulting marketing demographic features are not single labels, but rather a structured set of features represented by the center vector, used to describe the common behavioral patterns and interest structures of this demographic group in marketing scenarios.

[0043] In another preferred embodiment of the present invention, the process of obtaining the updated feature vector A of the user to be analyzed is as follows: If the user to be analyzed has neighboring nodes in the user attention network graph, then the user to be analyzed is added as a new node to the user attention network graph, and the updated feature vector A of the user to be analyzed is obtained by inputting the updated user attention network graph and the initial feature vectors of all users into the trained graph convolutional network model.

[0044] If the user to be analyzed has no neighboring nodes in the user attention network graph, then the initial feature vector of the user to be analyzed is directly input into the trained graph convolutional network model to obtain the updated feature vector A of the user to be analyzed.

[0045] In some implementations, the similarity calculation for updating feature vectors can employ normalized inner product, cosine similarity, or temperature-scaled dot product to reduce the impact of vector scale on similarity calculations. When the user to be analyzed has no neighboring nodes in the user's attention network graph, the graph convolutional layer degenerates into performing linear transformations or mappings based solely on the node's own feature vectors to obtain updated feature vectors.

[0046] A machine learning-based system for precise analysis of marketing audience characteristics includes: Data input module: Acquires multi-source data of users and constructs an initial feature vector for each user based on the multi-source data, including text content published on social platforms, clickstream data and purchase records on e-commerce platforms.

[0047] Attention Fusion Module: Constructs a user attention network graph based on the attention relationship data between users, and calculates the weight of each edge in the user attention network graph based on the initial feature vector.

[0048] Update module: Construct a graph convolutional network model and train it using a self-supervised learning approach. During training, edges in the user-focused network graph are used as positive samples, while user pairs without a focus relationship are used as negative samples. The model parameters are adjusted until convergence.

[0049] The user attention network graph and the initial feature vector of each user are input into the trained graph convolutional network model to obtain the updated feature vector of each user. The updated feature vectors are then clustered to obtain multiple initial clusters and the center vector of each initial cluster.

[0050] Analysis module: Obtain the initial feature vector of the user to be analyzed and update the user attention network graph according to the attention relationship data of the user to be analyzed. Based on this, obtain the updated feature vector A of the user to be analyzed, calculate the distance between the updated feature vector A and each center vector, and take the initial cluster with the smallest distance as the marketing audience feature to which the user to be analyzed belongs.

[0051] It is important to note that traditional marketing audience segmentation methods typically analyze users' browsing history, purchase history, or tag information, treating each user as an independent individual and assuming that individual behavior is solely determined by their own preferences, while ignoring the behavioral influence and interest transmission effects brought about by social connections. In the real-world network environment, users maintain continuous attention relationships and interactions. Interests, consumption tendencies, and product choices often diffuse and converge among related users. If this structural connection is ignored during the modeling process, the resulting audience segmentation results are prone to deviating from the actual behavioral groups. This invention incorporates user relationship structures into the feature modeling process, allowing individual features to be integrated and adjusted within the relational network. This ensures that users' behavioral expressions not only reflect their own historical behavior but also their position and connection strength within the social structure, thereby constructing a more realistic group feature expression method that aligns with real-world consumption scenarios. This represents a shift from isolated modeling to structured modeling, improving the rationality and accuracy of marketing audience feature identification.

[0052] The overall idea of ​​this invention is to model user behavior information in different scenarios in a unified way with their positional relationship in social relationships. This makes user characteristics no longer solely derived from individual historical behavior, but rather formed into a more complete feature expression after propagation and modification in the relationship network. Users are then grouped based on this feature expression to obtain marketing audience characteristics that are consistent with actual behavior.

[0053] Traditionally, users are typically segmented based solely on browsing or purchase history. This approach only reflects past behavior and fails to capture their connections to others' interests. In real-world online environments, user relationships often indicate similar interests, frequent communication, and converging consumption choices within these relationships. Therefore, relying solely on individual data for segmentation can easily lead to bias. This new approach first transforms users' textual expressions, browsing behavior, and purchasing behavior into structured features, giving each user a unified initial feature representation. These features are then placed into a network structure based on user relationships. A graph convolutional model propagates each user's features among neighboring users, gradually incorporating the influence of their social connections. The resulting updated feature vector reflects both the user's own behavior and the behavioral patterns of their group, thus more closely approximating the true interest structure and consumption tendencies. Subsequently, clustering groups users with similar features are formed, each group represented by a central vector. When a new user joins, determining which group their features most closely resembles identifies their target marketing audience.

[0054] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation. The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A method for precise analysis of marketing audience characteristics based on machine learning, characterized in that, Includes the following steps: Acquire multi-source data of users, and construct an initial feature vector for each user based on the multi-source data, which includes text content published on social platforms, clickstream data and purchase records on e-commerce platforms; A user attention network graph is constructed based on the attention relationship data between users, and the weight of each edge in the user attention network graph is calculated based on the initial feature vector; A graph convolutional network model is constructed and trained using a self-supervised learning approach. During training, edges in the network graph that users follow are used as positive samples, and user pairs without follow relationships are used as negative samples. The model parameters are adjusted until convergence. The user attention network graph and the initial feature vector of each user are input into the trained graph convolutional network model to obtain the updated feature vector of each user. The updated feature vectors are then clustered to obtain multiple initial clusters and the center vector of each initial cluster. Obtain the initial feature vector of the user to be analyzed and update the user attention network graph based on the user's attention relationship data. Based on this, obtain the updated feature vector A of the user to be analyzed, calculate the distance between the updated feature vector A and each center vector, and take the initial cluster with the smallest distance as the marketing audience feature to which the user to be analyzed belongs.

2. The method for precise analysis of marketing audience characteristics based on machine learning according to claim 1, characterized in that, The process of constructing the initial feature vector is as follows: A word segmentation tool based on a hidden Markov model is used to segment the text content into word sequences. The word sequences are then matched with a preset sentiment dictionary, which includes a list of positive sentiment words and a list of negative sentiment words. The number of matched positive sentiment words and the number of matched negative sentiment words are counted, and the ratio of the number of positive sentiment words to the number of matched negative sentiment words is calculated. When the denominator is zero, the ratio is set to a preset fixed value to obtain the sentiment tendency value. The clickstream data is arranged in chronological order of click time. The product category codes corresponding to each click are combined into a click sequence. The frequency of each product category code in the click sequence is counted to generate a click frequency vector. Iterate through each order in the purchase history, extract the product category code from the order, count the number of times each product category code appears, and generate a purchase frequency vector; The sentiment index, click frequency vector, and purchase frequency vector are concatenated to form the initial feature vector.

3. The method for precise analysis of marketing audience characteristics based on machine learning according to claim 2, characterized in that, The process of calculating the weight of each edge in the user attention network graph is as follows: A user attention network graph is constructed based on the attention relationship data between users. The nodes of the user attention network graph are users, and the edges represent the attention relationship between two users. For each edge, obtain the cosine similarity between the initial feature vectors of the first and second users connected by the edge, and use the cosine similarity as the weight of the edge; When the cosine similarity is less than a preset threshold, the edge is removed from the user-focused network graph.

4. The method for precise analysis of marketing audience characteristics based on machine learning according to claim 3, characterized in that, The graph convolutional network model includes a first graph convolutional layer and a second graph convolutional layer. The processing procedure of the first graph convolutional layer is as follows: for each node, obtain the initial feature vector of the node and the initial feature vectors of all first-order neighbor nodes, multiply the initial feature vectors of the first-order neighbor nodes with the weights of the corresponding edges and sum them, and then add them to the initial feature vector of the current node to obtain the first layer output vector. The processing procedure of the second convolutional layer is as follows: obtain the first layer output vector and the first layer output vectors of all first-order neighbor nodes, multiply the first layer output vectors of the first-order neighbor nodes with the weights of the corresponding edges and sum them, and then add them to the first layer output vector of the current node to obtain the updated feature vector.

5. The method for precise analysis of marketing audience characteristics based on machine learning according to claim 4, characterized in that, The specific process of training a graphical convolutional network model is as follows: A predetermined number of edges are randomly selected from the user attention network graph as the positive sample set, and the same number of user pairs without attention relationships are randomly selected as the negative sample set. The initial feature vectors of the two users corresponding to each edge in the positive sample set are input into the graph convolutional network model to obtain the updated feature vectors of the two users. The inner product of the two updated feature vectors is calculated as the positive sample similarity. The initial feature vectors of the two users contained in each user pair in the negative sample set are input into the graph convolutional network model to obtain the updated feature vectors of the two users. The inner product of the two updated feature vectors is calculated as the negative sample similarity. A loss function is constructed, which is the sum of the similarities of all negative samples minus the sum of the similarities of all positive samples. The parameters of the graph convolutional network model are updated using the gradient descent method. The above process is repeated until the value of the loss function no longer decreases.

6. The method for precise analysis of marketing audience characteristics based on machine learning according to claim 5, characterized in that, Obtaining multiple initial clusters and their center vectors includes: A cohesive hierarchical clustering algorithm is used to coarsely classify the updated feature vectors of all users. The Euclidean distance between every two updated feature vectors is calculated. Each user's updated feature vector is used as an initial cluster. The two clusters with the closest Euclidean distance are iteratively merged until the number of clusters is reduced to a preset first number, thus obtaining the coarsely classified clusters. For each coarse cluster, the K-means clustering algorithm is used to divide the updated feature vectors of all users within the cluster into multiple fine clusters. A set number of updated feature vectors are randomly selected as cluster centers. The Euclidean distance between each updated feature vector and each cluster center is calculated, and each updated feature vector is assigned to the cluster to which the nearest cluster center belongs. The average value of all updated feature vectors in each cluster is recalculated as the new cluster center. The above steps are repeated until the cluster centers no longer change or the preset number of iterations is reached, resulting in multiple fine clusters. Each fine cluster is used as the initial cluster, and the average value of all updated feature vectors in each fine cluster is used as the center vector of the initial cluster.

7. The method for precise analysis of marketing audience characteristics based on machine learning according to claim 6, characterized in that, The process of obtaining the updated feature vector A of the user to be analyzed is as follows: If the user to be analyzed has neighboring nodes in the user attention network graph, then the user to be analyzed is added as a new node to the user attention network graph, and the updated feature vector A of the user to be analyzed is obtained by inputting the updated user attention network graph and the initial feature vectors of all users into the trained graph convolutional network model. If the user to be analyzed has no neighboring nodes in the user attention network graph, then the initial feature vector of the user to be analyzed is directly input into the trained graph convolutional network model to obtain the updated feature vector A of the user to be analyzed.

8. A precise analysis system for marketing audience characteristics based on machine learning, characterized in that: include: Data input module: Acquires multi-source data of users and constructs an initial feature vector for each user based on the multi-source data, which includes text content published on social platforms, clickstream data and purchase records on e-commerce platforms; Attention Fusion Module: Constructs a user attention network graph based on the attention relationship data between users, and calculates the weight of each edge in the user attention network graph based on the initial feature vector; Update module: Construct a graph convolutional network model and train it using a self-supervised learning approach. During training, edges in the user-focused network graph are used as positive samples, and user pairs without any focus relationship are used as negative samples. The model parameters are adjusted until convergence. The user attention network graph and the initial feature vector of each user are input into the trained graph convolutional network model to obtain the updated feature vector of each user. The updated feature vectors are then clustered to obtain multiple initial clusters and the center vector of each initial cluster. Analysis module: Obtain the initial feature vector of the user to be analyzed and update the user attention network graph according to the attention relationship data of the user to be analyzed. Based on this, obtain the updated feature vector A of the user to be analyzed, calculate the distance between the updated feature vector A and each center vector, and take the initial cluster with the smallest distance as the marketing audience feature to which the user to be analyzed belongs.

Images