A Geographic Location Localization Method for Graph Convolutional Network Model Based on Attention Aggregation

Abstract

The present invention discloses a method for geographical location positioning of an attention aggregation graph convolutional network model. Based on the multi-classification prediction method of the traditional semi-supervised graph neural network model, the present invention optimizes the geographical location positioning problem of online social network users in real-life scenarios, and innovatively combines the attention mechanism with the graph convolutional neural network on this basis. In real application scenarios with a small number of samples with geographical location tags, model training for geographical location positioning of social network users is achieved. The method of the present invention does not require too much prior knowledge and does not need to obtain the geographical location tags of each user, meeting the requirements of privacy protection. Moreover, the present invention has good reusability in similar application scenarios and strong practical value of the invention.

Description

Technical Field

[0001] The present invention belongs to the field of information extraction in natural language processing, and particularly relates to a geographical location positioning method based on an attention aggregation graph convolutional network model. Background Art

[0002] The geographical location positioning of social media users, that is, identifying the geographical location of a user's "home", is the key to providing services for many downstream applications, such as public health monitoring, local event recommendation, or real-time emergency systems, etc., which all use social media as a hidden information resource of people. Social media services such as Twitter rely on IP addresses, WiFi footprints, and GPS data to locate users, but third-party service providers cannot easily obtain such information, so they have to rely on public sources of geolocation information, such as noisy and difficult-to-map profile location fields or geographically tagged tweets, but this information is only publicly available in 1% of tweets. The scarcity of publicly available location information has prompted people to predict the home geographical location of users from information such as tweet text and social interaction data.

[0003] Since the user data on social platforms is unstructured, an effective representation method has become the key to solving the user geographical location positioning. Currently, many methods have been proposed to construct and represent unstructured data to achieve better user geolocation. These mainly include mining location indicative information from the content posted by users, which relies on location indicative vocabulary to associate users with their locations. The representation of indicative vocabulary can utilize various natural language processing techniques, such as topic models and statistical models. A commonly used method for statistically measuring the distribution of location indicative vocabulary is to calculate the term frequency-inverse document frequency. In addition to the text content posted by users, users usually also actively establish connections with others and interact frequently to share experiences and feelings of life and work. Therefore, clues can be extracted from the social relationship network to infer the location of users, and this idea has inspired various network-based positioning methods. These methods are all based on user interactions and utilize various graph learning methods to obtain the feature representation of users, and then infer the location of users.

[0004] The emergence of graph neural network models represents a large amount of data with irregular spatial structures in most real - life scenarios, such as the interaction relationships of social network users, the relationships between online e - commerce users and products, the mutual citation network among documents, and so on. Graph neural network models can be used to fully mine the information in graph - structured data, learn the low - dimensional dense vector representations of nodes in the graph structure diagram, and use the learned node feature embedding vectors for many downstream tasks. Graph neural network models can generally be divided into three categories: spectral - domain - based methods, spatial - domain - based methods, and graph embedding techniques. Spectral - domain - based methods can be summarized as studying various properties of graphs by means of the Laplacian matrix of the graph, that is, eigenvectors and eigenvalues, with the graph convolutional network model as the representative. In spatial - domain - based methods, the representative model is the graph attention network model, in which each vertex updates its own attribute features by continuously integrating the information of its neighbors. The third type of graph embedding technology generally first learns an embedding vector for each node, and on this basis, trains a classifier. The biggest feature is that the training of the embedding vector and the classifier is independent, and the learning processes do not interfere with each other. The representative model is the Node2Vec model. Given the characteristics of parameter sharing and sparse connection in the graph convolutional network model in graph neural network models, and the characteristics of the attention mechanism having different weights for different nodes, the combination of the two has more advantages for the application scenarios of such sparse social network user graph structures. Therefore, this method proposes a geographical location positioning method of a graph convolutional network model based on attention aggregation. Summary of the Invention

[0005] The object of the present invention is to solve the problem of predicting the geographical location of users in an online social network, and propose a geographical location positioning method of a graph convolutional network model based on attention aggregation.

[0006] The data of the online social network users mainly includes the tweet messages they post and the content of forwarding other users' posts. The selected model verification dataset comes from the Twitter platform. Each tweet is a short text and some other content, such as photos and emojis. The additional information or symbols contained in the tweet text are usually related to content describing specific meanings. For example, "@" is used to mention Twitter users. Since the identification and judgment of the user's geographical location largely depend on the content they post, and their local social interactions, or interactions with friends also assist in the identification and judgment of the user's geographical location, therefore, the present invention proposes a geographical location positioning method of a graph convolutional network model based on attention aggregation. This method only requires very little prior knowledge and can obtain learning experience, train the model and make result predictions only by the user's tweet writing and network interactions with friends or netizens, and then obtain the geographical location results of social network users.

[0007] To achieve the above object, the technical solution adopted by the present invention is as follows: A geographical location positioning method for a graph convolutional network model based on attention aggregation, characterized by comprising the following steps:

[0008] (1) Collect the social text information posted by online social network users. For each user, collect their tweet content, including their tweet messages and the content of reposting other users' posts. Preprocess the data, filter out the text tags, photos, emojis, and punctuation marks of each user, perform word segmentation, remove stop words, etc. on the data, process the samples by combining TF-IDF and L2 normalization, and then extract the mention information between users from the preprocessed data, retain the user text content, and generate text embeddings.

[0009] (2) Convert the user text embeddings into a learnable matrix of single vectors, and use the multi-head attention mechanism to mine the information most relevant to geographical location contained in the user text, generate a user one-hot encoded vector, update the feature representation of the words related to geographical location according to the attention scores of the attention mechanism, and generate a user text view matrix. Then, according to the mention information extracted from the user text, construct a user mention matrix and process it to generate a social network view matrix.

[0010] (3) Build a graph convolutional network model with attention aggregation. The input of the model is the feature matrix and the adjacency matrix generated according to the tweet dataset. Introduce the attention mechanism into the graph convolutional network model, use the attention mechanism to capture the long-range interactions between neighbor nodes beyond multiple hops in each graph convolutional layer, so as to determine the importance of the aggregation results of each hop of neighbor nodes, improve the model's representation ability of nodes by increasing the receptive field in each layer of the network, capture the most valuable information from each single-hop neighbor node, and learn hierarchical local substructure features through aggregation.

[0011] (4) Input the result obtained by training with the graph convolutional network model into a geographical location predictor of a multi-layer perceptron with a fully connected layer with a softmax function of normalized exponential to predict the maximum probability of the geographical location cluster to which the user belongs.

[0012] Further, the generation representations of the user text view matrix and the social network view matrix in step (2) are specifically implemented as follows:

[0013] 2.1 Perform word segmentation on the user tweet statements, and convert the word sequence into a series of low-dimensional embedding vector sequences S.

[0014] 2.2 Divide the low-dimensional embedding vector sequence S into 8 parts and multiply it by the weight W h to obtain the input vector, and pass it through The query matrix, key parameter matrix, and value parameter matrix are calculated respectively, where Q, K, and V represent the initial representations of three vectors. represent the weight vectors of the three matrices;

[0015] 2.3 Use s h to represent the vector sequence in the h-th attention head, then the i-th word element in s h is represented as e i , and the relative importance score of the j-th word element to the i-th word element is calculated using softmax.

[0016] 2.4 Update the representation of the i-th word in s h by combining the features of all relevant words based on the importance scores;

[0017] 2.5 Calculate the new representation of the i-th word in s h by collecting the combined features learned by each attention head, and obtain the embedding representation of the sentence sequence;

[0018] 2.6 Generate the tweet content representation of the user through an additive linear transformation network, and finally obtain the user text view matrix.

[0019] 2.7 Based on the user mention information obtained from the extracted text, construct a social network matrix using networkx in Python and delete user nodes with too many edges.

[0020] Furthermore, the expression of the relative importance score of the j-th word element to the i-th word element in step 2.3 is as follows:

[0021]

[0022] where e i , e j respectively represent the i-th word element and the j-th word element in s h , Q h , K h respectively represent the query matrix and the key parameter matrix, and using the softmax operation and the division by the square root d k makes the relative importance scores between words have a more stable gradient.

[0023] Furthermore, the expression of the updated representation of the i-th word in step 2.4 is as follows:

[0024]

[0025] where t represents the length of a single sentence sequence, and V hRepresents the value parameter matrix, which is updated by combining the features of all relevant words based on the importance score to obtain s h The representation of the i-th word in

[0026] Furthermore, in step 2.5, according to the new representation e′ of the i-th word i The embedding representation s expression of the sentence sequence obtained is as follows:

[0027] e′ i = concat(s 1,i , s 2,i ,..., s 8,i )W O

[0028]

[0029] where concat means to calculate s by collecting the combined features learned by each head h The new representation of the i-th word in, that is, e′ i , W O is the output weight matrix.

[0030] Furthermore, the expression for generating the tweet content representation of the user in step 2.6 is as follows:

[0031] x = concat(s1, s2,..., s T )R S

[0032] where T is the total number of tweets of each user, and R S represents a learnable matrix that can convert multiple sentence embeddings into a single vector. The single-user tweet content representation is x, and the tweet content representations of all users are the user text view matrix X.

[0033] In step (3), introducing an attention mechanism in the graph convolutional network model to build a user geographical location prediction model can be expressed by the following formula:

[0034]

[0035] where is the adjacency matrix with self-connections for each node, is 's diagonal node degree matrix, represents the normalized graph structure, H is the feature of each layer, and H 0 = X, W represents the convolution parameter of the model to be trained, and σ represents the activation function. Here, vanilla attention is used to determine the importance of the aggregation result of each hop in the graph convolutional neural network, and α is the attention weight, αj Denote the attention weights of the j-hop neighbors, representing user u i in the local structure of the k-hop neighbors, and the final node representation contains hierarchical information.

[0036] Furthermore, the final node representation obtained in step (3) uses residual learning techniques to stack attention convolutional layers, and the input representation expression of each layer is as follows:

[0037]

[0038] where X represents the original input of the model, denotes the output node representation of the l-th layer. The input of each attention aggregation convolutional layer is the sum of the output of the previous layer and the original input X, so as to obtain the final node representation of the next layer, that is, the (l + 1)-th layer.

[0039] Furthermore, Adam is used as the stochastic gradient descent optimizer in the model training of step (3), and the output of the obtained final model and the combined representation expression of the output of each attention convolutional layer are as follows:

[0040]

[0041] where Dense() is a fully connected layer that combines the outputs of each attention graph convolutional layer. Y is the new representation of all user nodes obtained after model training.

[0042] The geographical location predictor in step (4) uses a multi-layer perceptron, and the expression of its predicted result value is as follows:

[0043] Y′ = softmax(MLP(Y))

[0044] where Y′ is the prediction for all users, and MLP is a multi-layer perceptron, that is, the geographical location predictor.

[0045] Furthermore, the geographical location predictor in step (4) uses cross-entropy as the loss function, and the expression is as follows:

[0046]

[0047] where y ij represents the probability that the i-th user belongs to the j-th cluster, n represents the number of users, and c represents the number of clusters.

[0048] Advantages of the present invention:

[0049] The present invention optimizes the problem of geographical location positioning of online social network users in real - life scenarios, and proposes a geographical location positioning method based on an attention - aggregation graph convolutional network model. In real - world application scenarios with a small number of samples with geographical location tags, based on the multi - classification prediction of traditional graph neural network models in semi - supervised learning, the model training for social network user geographical location positioning is achieved. Among them, the use of the attention mechanism in the text view matrix in step 2 of the present invention can not only capture the writing style of users, but also obtain the preferences of users for meaningful location - related words. The combination of the attention mechanism and the graph convolutional model in step 3 enables the model to utilize information from different - hop neighbor nodes, solves the weaknesses of traditional graph convolutional network models, and provides more feature representations of individual users. The method of the present invention does not require excessive prior knowledge and does not need to obtain the geographical location tags of each user, meeting the requirements of privacy protection. It has feasibility and superiority in the implementation scenarios of geographical location positioning based on user - generated content and social networks, and the present invention has good reusability in similar application scenarios, with strong practical value of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Schematic diagram of matrix generation provided by an embodiment of the present invention;

[0051] Figure 2 Schematic diagram of model construction provided by an embodiment of the present invention;

[0052] Figure 3 Schematic diagram of the attention convolutional layer provided by an embodiment of the present invention;

[0053] Figure 4 Schematic diagram of the model performance provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] In order to enable relevant personnel to more clearly understand the technical content of the present invention, the technical solution of the present invention will be described in detail below in conjunction with the drawings and specific embodiments.

[0055] The present invention discloses a geographical location positioning method based on an attention - aggregation graph convolutional network model. Based on the multi - classification prediction method of traditional semi - supervised graph neural network models, the present invention optimizes the problem of geographical location positioning of online social network users in real - life scenarios, and innovatively combines the attention mechanism with the graph convolutional neural network on this basis. In real - world application scenarios with a small number of samples with geographical location tags, the model training for social network user geographical location positioning is achieved. The method of the present invention does not require excessive prior knowledge and does not need to obtain the geographical location tags of each user, meeting the requirements of privacy protection. And the present invention has good reusability in similar application scenarios, with strong practical value of the invention.

[0056] A geographical location positioning method for a graph convolutional network model based on attention aggregation. The implementation example includes the following steps:

[0057] Step 1: Collect the social text information posted by users of the online social network. For each user, collect their tweet content, including their tweet messages and the content of reposting other users' posts. Preprocess the data, filter out the text tags, photos, emojis, and punctuation marks of each user, perform word segmentation, remove stop words, etc. on the data, and process the samples by combining TF-IDF and L2 normalization. Then extract the mention information between users from the preprocessed data, retain the user text content, and generate text embeddings, as shown in the attached Figure 1 Schematic diagram of matrix generation;

[0058] Step 2: Use self-attention with multiple heads to process the tweet content of users. That is, perform word segmentation on the tweet statements of users, and then convert the word sequence into a series of low-dimensional embedding vector sequences S. Split S into 8 parts, and multiply it with the weight W h to form the input vector W h S. Calculate the query matrix, key parameter matrix, and value parameter matrix respectively through where Q, K, and V represent the initial representations of three vectors, represents the weight vectors of the three matrices;

[0059] Step 3: Use the softmax function to calculate the relative importance scores between word elements in the vector sequence in the attention head, as follows:

[0060]

[0061] where s h represents the vector sequence in the h-th attention head, e i and e j represent the i-th and j-th word elements in s h respectively, Q h and K h represent the query matrix and the key parameter matrix respectively. Using the softmax operation and division by the square root d k makes the relative importance scores between words have a more stable gradient.

[0062] Step 4: Based on the importance scores, combine the features of all relevant words to update the new representation of the i-th word in the vector sequence in the attention head, as follows:

[0063]

[0064] where t represents the length of a single sentence sequence, and V h represents the value parameter matrix.

[0065] Step 5: Calculate the new representation of the i-th word in the vector sequence in the attention head by collecting the combined features learned by each head, and thus obtain the final embedded representation of the tweet sentence sequence, as follows:

[0066] e′ i = concat(s 1,i , s 2,i ,..., s 8,i )W O

[0067]

[0068] where e′ i represents the new representation of the i-th word calculated from the combined features learned by multiple attention heads, and W O is the output weight matrix, and s represents the final embedded representation of the sentence sequence. Accordingly, the entire tweet content representation of the social network user can be obtained, as follows:

[0069] x = concat(s1, s2,..., s T )R S

[0070] where T is the total number of tweets of each user, which is a fixed value, and R S represents a learnable matrix that can transform multiple sentence embeddings into a single vector. The tweet content representation of a single user is x, and the tweet content representations of all users are represented by X, thereby obtaining the user text view matrix.

[0071] Step 6: Based on the user mention information obtained from the extracted text, use networkx in Python to construct a user interaction network to represent the social relationship between users. According to the mention network, if a user mentions another user, or they jointly mention another user, it means there is a connection between the two user nodes, that is, A uv = 1. Then, users with excessive edges are regarded as "celebrity" nodes and are deleted according to regulations to reduce the negative impact;

[0072] Step 7: Build an attention aggregation neural network training model based on the graph convolutional neural network and the attention mechanism. The input of the model is the feature matrix and the adjacency matrix in the previous step, as shown in the attached Figure 2 Schematic diagram of model construction. The most general form of graph convolution with depth k can be recursively represented by the convolution structure, and its propagation method between layers can be represented as follows:

[0073]

[0074] Among them is the adjacency matrix where each node has a self-connection, is the diagonal node degree matrix of, represents the normalized graph structure, H is the feature of each layer, and H 0 = X, W represents the parameters of the convolution of the model to be trained, and σ represents the activation function. The attention mechanism is introduced into the graph convolution model to capture the long-range interactions between neighbor nodes beyond multiple hops in each graph convolution layer, so as to determine the importance of the aggregation results of neighbor nodes in each hop. By increasing the receptive field in each layer of the network, the model's representation ability for nodes is improved, and the most valuable information is captured from each single-hop neighbor, and hierarchical local sub-structure features are learned through aggregation. Therefore, the model convolution result will be a hierarchical representation containing the most valuable information from different-hop convolution processes, and the formed hierarchical node representation is as follows:

[0075]

[0076] Here, vanilla attention is used to determine the importance of the aggregation results of each hop, where α is the attention weight, and α j represents the attention weight of the j-hop neighbor, represents the tweet user u i in the local structure of the k-hop neighbor. The final node representation contains hierarchical information, as shown in the attached Figure 3 attention convolution layer schematic diagram.

[0077] Step 8: Use the residual learning technique to stack an attention-aggregated graph convolution layer to maximize the advantages of deep learning and learn deeper potential features, and construct an attention-aggregated graph convolution module to obtain a better final node representation. The input of each module is the sum of the output of the previous layer and the original input X, which is represented as follows:

[0078]

[0079] Among them, X represents the original input of the model, represents the output node representation of the l-th layer. The input of each attention-aggregated convolution layer is the sum of the output of the previous layer and the original input X, so as to obtain the final node representation of the next layer, that is, the (l + 1)-th layer.

[0080] Step 9: Use a fully connected layer at the end of the model to process the output combination of each convolution layer, which is represented as follows:

[0081]

[0082] Among them, Dense() is a fully connected layer that combines the outputs of the graph convolutional layers for each attention aggregation. The resulting Y is the new representation of all user nodes now, which is used as the input to the geographical location predictor;

[0083] Step 10: Construct a multi-layer perceptron as the geographical location predictor. Its goal is to predict the maximum probability of the location cluster to which the user belongs. Based on the learned user feature representation Y, the result is predicted as follows:

[0084] Y′ = softmax(MLP(Y))

[0085] where Y′ ∈ R n×c is the prediction for all users. n represents the number of all users, and c represents the number of predicted clusters. Here, cross-entropy is used as the loss function, which is expressed as follows:

[0086]

[0087] where y ij represents the probability that the i-th user belongs to the j-th cluster. During the training process, Adam is used as the stochastic gradient descent optimizer, and the final prediction result is obtained.

[0088] After completing the above steps, this method analyzes the experimental performance on three classic Twitter datasets containing a large number of users, and at the same time compares it with the existing online social network user location method models. The performance is shown in the attached Figure 4 Schematic diagram of model performance. The results show that this method is superior to other methods in all three performance metric values.

[0089] It should be known that the parts not elaborated in detail in this specification all belong to the prior art. Relevant technical personnel should understand that the above embodiments are only for helping readers understand the principle and implementation method of the present invention. The scope protected by the present invention is not limited to such embodiments. Any equivalent replacement made on the basis of the present invention is within the scope of protection of the rights of the present invention.

Claims

1. A method for geolocation of a graph convolutional network model based on attention aggregation, characterized in that It includes the following steps: (1) Collect the social text information posted by online social network users, and preprocess the data, including removing text tags, photos, symbols, and punctuation marks; then extract the mention information between users from the preprocessed data, retain the user text content, and generate text embeddings; (2) Convert the text embeddings of users into a learnable matrix of single vectors, use the multi-head attention mechanism to mine the information most relevant to geolocation, generate user one-hot encoded vectors, thereby updating the feature representation of words related to geolocation, and generating a user text view matrix; use the extracted mention information to construct a user mention matrix and generate a social network view matrix; (3) Introduce the attention mechanism into the graph convolutional network model, use the attention mechanism to capture the long-range interactions between neighbor nodes beyond multiple hops, improve the model's representation ability of nodes by increasing the receptive field in each layer of the network, and then capture the most valuable information from each single-hop neighbor node; (4) Input the result obtained from training the graph convolutional network model into a geographical location predictor of a multi-layer perceptron, so as to predict the maximum probability of the geographical location cluster to which the user belongs; The generation representations of the user text view matrix and the social network view matrix in step (2) are as follows. The specific process is as follows: 2.1 Perform word segmentation on the user tweet statements, and convert the word sequence into a series of low-dimensional embedding vector sequences S; 2.2 Split the low-dimensional embedding vector sequence S into 8 parts, multiply it with the weight W h to obtain the input vector, and calculate the query matrix, key parameter matrix, and value parameter matrix respectively, where Q, K, and V are the initial representations of the three vectors, representing the weight vectors of the three matrices respectively; 2.3 Using s h to represent the vector sequence in the h-th attention head, obtaining s h The i-th word element in i i is denoted as e, and the relative importance score of the j-th word element to the i-th word element is calculated using softmax; 2.4 Update s by combining the features of all relevant words based on the importance score h The representation of the i-th word in 2.5 Calculate s by collecting the combined features learned by each attention head h to obtain the new representation of the i-th word in h , and get the embedding representation of the sentence sequence; 2.6 Generate the representation of the user's tweet content through an additive linear transformation network, and finally obtain the user text view matrix; 2.7 According to the mention information between users obtained from the extracted text, use networkx in Python to construct a social network matrix, and delete the user nodes with too many edges.

2. The geographical location positioning method of a graph convolutional network model based on attention aggregation according to claim 1, characterized in that: The expression for the relative importance score of the j-th word element to the i-th word element in step 2.3 is as follows: where e i and e j represent the i-th word element and the j-th word element in s h respectively, and Q h and K h represent the query matrix and the key parameter matrix respectively. Using the softmax operation and the division by the square root of d k makes the relative importance scores between words have a more stable gradient.

3. The geographical location positioning method of a graph convolutional network model based on attention aggregation according to claim 2, characterized in that: The updated representation of the i-th word in step 2.4 is as follows: Among them, t represents the length of a single sentence sequence, and V h represents the value parameter matrix, which is updated by combining the features of all relevant words based on the importance score to obtain s h the representation of the i-th word in 4. A method for geolocation of a graph convolutional network model based on attention aggregation according to claim 3, characterized in that: In step 2.5, according to the new representation e′ of the i-th word i The expression for obtaining the embedded representation s of the sentence sequence is as follows: e′ i = concat(s 1,i , s 2,i ,..., s 8,i )W O (3) Among them, concat represents calculating s by collecting the combined features learned from each head h The new representation of the i-th word in h , that is, e′ i , W O is the output weight matrix 5. A geographical location positioning method for a graph convolutional network model based on attention aggregation according to claim 2, characterized in that: The expression for generating the representation of the user's tweet content in step 2.6 is as follows: x = concat(s1, s2,..., s T )R S (5) where T is the total number of tweets for each user, and R S represents a learnable matrix that can transform multiple sentence embeddings into a single vector. The tweet content of a single user is represented as x, and the tweet content of all users is represented as the user text view matrix X.

6. A method for geolocation of a graph convolutional network model based on attention aggregation according to claim 5, characterized in that: In step (3), introduce the attention mechanism into the graph convolutional network model and build a user geographical location prediction model. The expression is as follows: Among them, vanilla attention is used to determine the importance of the aggregation results of each hop in the graph convolutional neural network. α is the attention weight, and α j represents the attention weight of the j-hop neighbor, represents user u i in the local structure of the k-hop neighbor, which can be recursively represented by the convolutional structure of the classical graph convolutional network to obtain the propagation method between layers. The final node representation contains hierarchical information.

7. A method for geolocation of a graph convolutional network model based on attention aggregation according to claim 6, characterized in that The obtained final node representation uses the residual learning technique to stack attention convolutional layers, so the input representation expression of each layer is as follows: Among them, X represents the original input of the model, represents the output node representation of the l-th layer. The input of each attention aggregation convolutional layer is the sum of the output of the previous layer and the original input X, so as to obtain the final node representation of the next layer, that is, the (l + 1)-th layer.

8. A method for geolocation of a graph convolutional network model based on attention aggregation according to claim 7, characterized in that: In step (3), Adam is used as the stochastic gradient descent optimizer for model training. The expression for the combined representation of the output of the final model and the output of each attention convolutional layer is as follows: Where Dense() is a fully connected layer that combines the outputs of each attention graph convolutional layer; Y is the new representation form of all user nodes obtained after model training.

Images