Unsupervised cross-domain user matching method based on embedding correction
Through the unsupervised user identity association method based on embedding correction, user characteristics are extracted using multiple pre-trained language models and graph neural network technologies, the problem of tag data dependence and heterogeneity in user identity association across social networks is solved, and efficient Cross-domain user identification improves matching accuracy and efficiency.
Patent Information
- Application Number
- CN202411170592.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-11-26
AI Technical Summary
In the prior art, the cross-social network user identity association method mainly relies on supervised or semi-supervised strategies, requires label data, and few unsupervised methods, making it difficult to effectively deal with the impact of heterogeneous information and user behavior similarity, resulting in user matching. Inaccurate.
Using an unsupervised user identity association method based on embedding correction, user features are extracted through data preprocessing, multiple pre-trained language models and graph neural network technology, combined with attention mechanism and stochastic gradient descent, multi-dimensional discriminant features are generated, and through The repulsive force correction solution solves the user's offset problem and calculates the cosine similarity of the embedded vector to match.
In the unsupervised scenario, the accuracy of cross-domain user identification has been significantly improved. The accuracy of user identification has reached 74.3%, H@10 has reached 77.2%, and the MRR index has reached 75.4%, which is more than 30% higher than the idea of unused embedded correction. , realizes efficient cross-domain user matching.
Smart Images

Figure CN119025934A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to an unsupervised cross-domain user matching method based on embedding correction, and belongs to the technical field of social media data analysis. Background Art
[0002] The existence of multiple social network behaviors and overlapping users between networks has built a logical connection between the originally isolated platforms. The research content of social networks has gradually shifted and developed from the traditional single network to the combination of multiple networks. For various reasons, the information between multiple platforms usually lacks explicit correlation relationships. This sparse explicit relationship leads to information isolation between platforms.
[0003] According to the demand and use of labeled data, cross-social network user identity association methods can be divided into three categories: supervised, semi-supervised, and unsupervised methods. Most of the existing identity association methods adopt supervised or semi-supervised strategies. Such methods require labeled user pairs (i.e., real matching user identity pairs) as input during the learning and training process. However, in actual research and application, labeled data is usually not easy to obtain. Generally speaking, there are two methods to collect these labeled user pairs: one is manual judgment, that is, manually sorting a large amount of user data and finding matching users; the other is to use the user's self-declared information, such as other social platform accounts listed by the user in his personal profile, or user account association information in some personal network service aggregation websites (such as About.me, etc.). However, over-reliance on these data acquisition methods has certain disadvantages. For example, the first method is obviously time-consuming and labor-intensive, and both methods can only obtain a very small part of all users, which may bring potential training errors to the matching task.
[0004] Unsupervised user identity association methods do not rely on labeled user pairs, which can effectively avoid the above problems. However, there are relatively few unsupervised methods in existing work, so it is necessary and important to conduct efficient unsupervised user identity association research. Due to the characteristics of unsupervised methods that do not require labeled data, as well as the attributes and behavioral characteristics of cross-network social network users, while unsupervised user identity association reflects its advantages, it also faces three challenges: dealing with the heterogeneity of social network information, correctly and effectively using cross-network information, and dealing with the impact of user behavior similarity on the same platform. Therefore, it is necessary to design a new solution to meet actual needs. Summary of the invention
[0005] The purpose of the present invention is to develop an unsupervised user identity association scheme based on embedded correction and taking into account user discriminative features in an unsupervised situation to address the above shortcomings. It integrates multiple artificial intelligence technologies to obtain a systematic solution that solves the current difficulties in cross-domain user identification and matching tasks.
[0006] In order to achieve the above-mentioned purpose, the technical solution of the present invention is as follows: the main process of the present invention includes basic embedding generation, discriminative feature extraction and matching based on embedding correction.
[0007] First, data preprocessing technology is used to clean, select and organize the data; in the user basic embedding generation stage, a variety of pre-trained language models and graph neural network technologies are used to extract the features of user basic information and social relationships, and the attention mechanism is used to achieve the fusion of heterogeneous attributes; based on the characteristics of different user attributes, multiple types of attribute discriminant features are designed and generated to achieve the extraction of multi-dimensional deep features of user text, social, and data; the discriminant features are used to design correction operations for basic embeddings, and repulsion is introduced to design a correction scheme based on repulsion to solve the user offset phenomenon that occurs during the correction process; finally, the corrected embedding is obtained and the matching score is calculated for cross-network users to achieve cross-domain user matching. Combining the above multiple technologies and methods, the problems of incomplete attribute utilization and inaccurate user matching in the current unsupervised scenario can be solved to achieve cross-domain user identification.
[0008] The present invention is implemented by adopting the following technical solution: an unsupervised cross-domain user matching method based on embedding correction, the method is as follows:
[0009] S1. Preprocess data by sorting, cleaning, converting, and normalizing;
[0010] S2, by using the pre-trained RoBERTa, Canine model and graph neural network technology to obtain user multi-attribute embedding;
[0011] S3, use the attention mechanism and stochastic gradient descent method to fuse the features of different single attribute embeddings;
[0012] S4, integrating multiple types of information simultaneously on a single platform and across platforms to form multi-dimensional user discriminative features;
[0013] S5. Use user discriminative features to construct similar user pairs, train and generate the correlation information required in the correction process, and obtain the final user embedding;
[0014] S6. Calculate the cosine similarity of the embedding vectors between each pair of cross-platform users as the user matching score, and determine the cross-network social network users based on the score.
[0015] The specific cleaning process of step S1 includes the following steps:
[0016] S1-1. For the crawled cross-domain user pairs, the source platform and target platform users are re-sorted according to the unique correspondence of their user IDs to form sequential user pair data;
[0017] S1-2, identify different user attributes existing in different platforms, and analyze and screen the attributes;
[0018] S1-3. Use regular matching to clean data and clean user attributes of garbled characters, meaningless emoticons, meaningless test English alphanumeric strings, etc.
[0019] The specific data enhancement process of step S2 includes the following steps:
[0020] S2-1. Concatenate the nickname and username of each user, load the pre-trained RoBERTa model and use its built-in word segmenter to perform word segmentation, calculate and generate comprehensive embedding of username and nickname.
[0021] S2-2. For each user’s user description, load the pre-trained Canine model and use its built-in word segmenter to segment the text, calculate and generate the user description embedding, and truncate the text content that exceeds the length.
[0022] S2-3. Use user interaction relationships to build a topological network and add username embedding as a node attribute to the user node. Based on this topological network, use graph neural network technology and the attention mechanism to learn the interaction relationship between user nodes, calculate and generate user topological embedding.
[0023] Step S3: The specific coarse-grained SMS text classification technology includes the following steps:
[0024] S3-1, use the attention pooling mechanism to dynamically adjust the weights between different embeddings, and cross and concatenate the weighted single attribute embeddings;
[0025] S3-2, use a multi-layer neural network to perform nonlinear transformation and feature fusion on the fused features;
[0026] S3-3. Combining the idea of iterative optimization, use stochastic gradient descent (SGD) to fine-tune and optimize the fused feature representation to form the basic embedding of the user.
[0027] The specific cross-platform multi-dimensional user discriminative feature generation technology of step S4 includes the following steps:
[0028] S4-1, complete the generation of discriminative features based on user name data by preprocessing user name data, extracting n-gram features, constructing TF-IDF vectors, calculating Jaccard similarity coefficients, etc.;
[0029] S4-2, completing the generation of discriminative features based on user network topology data by building a network neighbor list and node degree dictionary, calculating the cosine similarity matrix, calculating the degree matrix, calculating the cross-platform similarity, and normalizing the similarity score;
[0030] S4-3, complete the generation of discriminative features based on user tweet data by loading vocabulary and filtering low-frequency words, building user-vocabulary mapping, extracting discriminative words, calculating the number of word overlaps and Jaccard similarity coefficients, and normalizing similarity scores;
[0031] S4-4, complete the generation of discriminative features based on user education and work experience data through steps such as text preprocessing, merging data and adding labels, calculating TF-IDF features, training random forest model, creating user feature importance matrix and mapping to users;
[0032] S4-5. The discriminative features based on the user quantity data are generated through the steps of text preprocessing, merging data and adding labels, calculating the skewness and kurtosis of features, discretization processing, creating a user feature discretization matrix and mapping it to users.
[0033] The specific pre-processing operation of step S4-1 includes the following steps:
[0034] S4-1-1, preprocessing user name data to standardize user names;
[0035] S4-1-2, extracting n-gram features of the user name;
[0036] S4-1-3. Calculate the TF-IDF value and generate the TF-IDF vector of the user name.
[0037] S4-1-4. Calculate the number of word overlaps and the Jaccard similarity coefficient and save them to a file.
[0038] The specific operation of step S4-2 includes the following steps:
[0039] S4-2-1, build a network neighbor list and node degree dictionary;
[0040] S4-2-2, establish an ID mapping table;
[0041] S4-2-3, calculate the cosine similarity matrix;
[0042] S4-2-4, construct the degree matrix according to the node degree dictionary;
[0043] S4-2-5. Calculate the similarity of the user's neighbor nodes, and calculate the similarity score by combining the degree and cosine similarity.
[0044] S4-2-6. Normalize the similarity score and save it to a file.
[0045] The specific operation of step S4-3 includes the following steps:
[0046] S4-3-1, load the vocabulary, and filter and generate a set of low-frequency words;
[0047] S4-3-2, based on the above steps, construct user-vocabulary mapping;
[0048] S4-3-3, low-frequency words are sorted by frequency and stored in the dictionary;
[0049] S4-3-4, calculate the number of word overlaps and the Jaccard similarity coefficient;
[0050] S4-3-5. Normalize the similarity score and save it to a file.
[0051] The specific operation of step S4-4 includes the following steps:
[0052] S4-4-1. Convert the JSON file of the data into Pandas DataFrame format and preprocess the data to remove special characters and spaces.
[0053] S4-4-2, merge data and add labels for classification tasks;
[0054] S4-4-3, convert text data into TF-IDF feature vector;
[0055] S4-4-4. Train the random forest model and extract feature importance;
[0056] S4-4-5. Create a user feature importance matrix and map it to users, and save it to a file.
[0057] The specific operation of step S4-5 includes the following steps:
[0058] S4-5-1. Convert the JSON file of the data into Pandas DataFrame format;
[0059] S4-5-2, merge data and add labels for classification tasks;
[0060] S4-5-3. Calculate the skewness and kurtosis of each user feature to determine the shape and characteristics of data distribution;
[0061] S4-5-4, divide the continuous attributes into several intervals or levels, and map the attribute value of each user to the corresponding discrete level;
[0062] S4-5-5. Create a user feature importance matrix and map it to users, and save it to a file.
[0063] Compared with the prior art, the advantages of the present invention are as follows: the invention can effectively improve the cross-domain user recognition capability. In a data set containing 9965 pairs of cross-domain users, the user recognition accuracy H@1 reaches 74.3%, i.e., accurate matching, and H@10 reaches 77.2%, i.e., the first ten users selected by the algorithm include the correct cross-domain users; the MRR index reaches 75.4%, and these indicators are all improved by more than 30% compared with those without the embedding correction idea. It can be seen that the algorithm can effectively and accurately judge cross-domain users in an unsupervised situation. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] Figure 1 It is a schematic diagram of the overall process of the present invention;
[0065] Figure 2 Schematic diagram of the movement of the user vector in space during the original correction process;
[0066] Figure 3 Schematic diagram of the movement of the user vector in space during the correction process of repulsive force enhancement. DETAILED DESCRIPTION
[0067] In order to deepen the recognition and understanding of the present invention, the scheme is described in detail below in conjunction with the accompanying drawings and implementation methods.
[0068] Example 1: Reference: Figure 1 , an unsupervised cross-domain user matching method based on embedding correction, the method is as follows:
[0069] This embodiment describes an unsupervised cross-domain user matching method based on embedding correction as follows:
[0070] 1) Data preprocessing: using methods such as sorting, cleaning, conversion, and normalization.
[0071] 1-1) Match the cross-domain user pairs according to the user ID. For the crawled cross-domain user pairs, the source platform and target platform users are re-sorted according to the unique correspondence of their user IDs to form sequential user pair data.
[0072] 1-2) Select attributes based on the intersection of platform attributes. Identify different user attributes existing in different platforms, analyze and filter the attributes, and select multiple common attributes including user name, description, and topological relationship.
[0073] 1-3) Use regular matching to clean data. Rely on Python and regular expressions to clean user attributes of garbled characters, meaningless emoticons, meaningless test English alphanumeric strings, etc.
[0074] 2) User single attribute embedding generation. Use models such as RoBERTa, Canine, and GAT
[0075] 2-1) Generate embeddings based on username and nickname. Use the pre-trained BERT model.
[0076] After trying and comparing the word2vec model, BERT model, and RoBERTa model, this patent finally uses the RoBERTa pre-trained model for embedding generation, which can perform deep semantic understanding and encoding of text. The specific steps are as follows:
[0077] 2-1-1) Perform text splicing on the user name and user nickname;
[0078] 2-1-2) Use the RoBERTa model to load the open source pre-trained model;
[0079] 2-1-3) Load the dataset, use RoBERTa's built-in word segmenter to segment the text, and calculate and generate comprehensive embeddings of user names and user nicknames.
[0080] 2-2) Generate embeddings based on user descriptions. Use the pre-trained Canine model.
[0081] 2-2-1) Use the Canine model to load the open source pre-trained model;
[0082] 2-2-2) Load the dataset, use Canine's built-in word segmenter to segment the text, and calculate and generate user description embeddings.
[0083] 2-3) Generate embeddings based on user social relationships using graph neural network GAT.
[0084] 2-3-1) Use user interaction relationships to build a topological network and add user name embedding as a node attribute to the user node;
[0085] 2-3-2) Use graph neural network and attention mechanism to learn the interaction relationship between user nodes, calculate and claim user topology embedding.
[0086] 3) Fusion of user single attribute embeddings to form base embeddings. Use attention mechanism and stochastic gradient descent.
[0087] 3-1) Use the attention pooling mechanism to dynamically adjust the weights between different embeddings, and cross and concatenate the weighted single attribute embeddings;
[0088] 3-2) Use a multi-layer neural network to perform nonlinear transformation and feature fusion on the fused features;
[0089] 3-3) Combining the idea of iterative optimization, stochastic gradient descent (SGD) is used to fine-tune and optimize the fused feature representation to form the basic embedding of the user.
[0090] 4) Considering the impact of cross-platform information, that is, the possible inconsistency of user behavior on different platforms and the homogeneity of user behavior on a single platform, it is impossible to ensure that all information on a single platform contributes to the user identity association task. Therefore, the user embedding vector needs to be improved for the purpose of identity association tasks, so as to integrate multiple types of information on both a single platform and cross-platform to form user discriminative features;
[0091] 4-1) Generation of discriminative features based on user name data: This is done by preprocessing user name data, extracting n-gram features, constructing TF-IDF vectors, and calculating Jaccard similarity coefficients.
[0092] 4-1-1) Collect all users (U s ∪U t ) User name, preprocess the user name, use str.lower() to convert all characters in the string to lowercase, str.strip() to remove whitespace characters before and after the string, and the preprocess_fun() function to further remove periods, underscores, and spaces in the user name to standardize the user name;
[0093] 4-1-2) Generate n-gram features for each username, traverse the length range of 1 to 5, generate and merge all possible n-gram features;
[0094] 4-1-3) Use the python package sklearn.feature_extraction.text.TfidfVectorizer class to calculate the TF-IDF value in the document and generate the TF-IDF vector for each username.
[0095] 4-1-4) Measure the similarity between different users by calculating the number of word overlaps and the Jaccard similarity coefficient between users; and use the numpy.save() function to save the similarity scores to a file for subsequent use.
[0096] 4-2) Generate discriminative features based on user network topology data. This is done by building a network neighbor list and node degree dictionary, calculating the cosine similarity matrix, calculating the degree matrix, calculating cross-platform similarity, and normalizing the similarity score.
[0097] 4-2-1) Read the subnetwork files of the source platform and the target platform. Each line records the connection relationship between two nodes. By traversing the subnetwork files, build the neighbor list and degree dictionary of each node. The neighbor list is stored in a dictionary, and the value of each node is the list of neighbor nodes to which it is directly connected. The degree dictionary records the degree of each node.
[0098] 4-2-2) Read the user embedding file, which contains the vector representation of the user for similarity calculation. At the same time, map the internal ID to the original ID to establish an ID mapping table for subsequent processing.
[0099] 4-2-3) Use the python package sklearn.metrics.pairwise.cosine_similarity function to calculate the cosine similarity between the source platform and the target platform user embeddings. The result is a similarity matrix, where each element in the matrix represents the similarity between two user embedding vectors.
[0100] 4-2-4) Construct a degree matrix based on the node degree dictionary. Each element in the matrix represents the degree of the corresponding node on the source platform and the target platform. It is used for similarity calculation of degree constraints.
[0101] 4-2-5) Traverse each user on the source platform and the target platform and calculate the similarity of their neighbor nodes. Select discriminative neighbor pairs through degree and cosine similarity constraints. Call a custom function to calculate the similarity score of each pair of users.
[0102] 4-2-6) Normalize the calculated similarity score. Use the numpy.max() function to adjust the score value to between 0 and 1. And use the numpy.save() function to save the normalized similarity score to a file for subsequent use.
[0103] 4-3) Generate discriminative features based on user tweet data. This is done by loading the vocabulary and filtering low-frequency words, building a user-vocabulary mapping, extracting discriminative words, calculating the number of word overlaps and the Jaccard similarity coefficient, and normalizing the similarity score.
[0104] 4-3-1) Use numpy.load() function to load the vocabulary files of the source platform and the target platform respectively. The vocabulary is stored in the form of a dictionary, with the key being the word and the value being the user and frequency of the word appearing in the tweet. Traverse the vocabulary, filter out the words whose frequency of appearance in the tweet is less than the threshold, and add these words to the low-frequency word set.
[0105] 4-3-2) For each user, build a low-frequency vocabulary. Traverse the vocabulary and record each low-frequency word and its frequency into the vocabulary of the corresponding user to build a user-vocabulary mapping.
[0106] 4-3-3) For each user, extract the low-frequency words in their tweets and sort them by frequency. The sorted low-frequency word list is stored in a dictionary with the user ID as the key and the sorted low-frequency word list as the value.
[0107] 4-3-4) For each pair of source and target platform users, calculate the intersection and union of their vocabulary sets. The size of the intersection is used to calculate the number of vocabulary overlaps, and the ratio of the union and intersection is used to calculate the Jaccard similarity coefficient.
[0108] 4-3-5) Normalize the calculated Jaccard similarity coefficient. Use the numpy.max() function to adjust the score value to between 0 and 1. The normalized similarity score is used for subsequent analysis. And use the numpy.save() function to save the normalized similarity score to a file for subsequent use.
[0109] 4-4) Generate discriminative features based on user education and work experience data. This is done through steps such as text preprocessing, merging data and adding labels, calculating TF-IDF features, training random forest models, creating user feature importance matrices and mapping them to users.
[0110] 4-4-1) Use the json.load() function to read the JSON file of education and work experience data and convert it to Pandas DataFrame format. Preprocess the description field as needed to remove special characters and spaces.
[0111] 4-4-2) Use the pandas.concat() function to merge the data of the source platform and the target platform into a DataFrame. And add labels to the data of each platform, with the source platform as 0 and the target platform as 1, for classification tasks.
[0112] 4-4-3) Use TfidfVectorizer to convert text data into TF-IDF feature vectors. Limit the maximum number of features to 1000 to reduce computational complexity.
[0113] 4-4-4) Use RandomForestClassifier to train a random forest model and extract feature importance. The model uses 100 decision trees and the random seed is 42 to ensure repeatable results.
[0114] 4-4-5) Use the numpy.zeros() function to create a matrix to store the feature importance of each user. Traverse the list of user IDs and map the feature importance to each user. According to the position of the user ID in the source platform and the target platform, assign the corresponding feature importance value to each user. And use the numpy.save() function to save the user feature importance matrix to a file for subsequent use.
[0115] 4-5) Generate discriminative features based on user quantity data. This is done through steps such as text preprocessing, merging data and adding labels, calculating the skewness and kurtosis of features, discretization, creating a user feature discretization matrix and mapping it to users.
[0116] 4-5-1) Use the json.load() function to read the JSON file containing the user quantity data and convert it into a Pandas DataFrame format.
[0117] 4-5-2) Use the pandas.concat() function to merge the data of the source platform and the target platform into a DataFrame. And add labels to the data of each platform, with the source platform as 0 and the target platform as 1, for classification tasks.
[0118] 4-5-3) Use scipy.stats.skew() and scipy.stats.kurtosis() functions to calculate the skewness and kurtosis of each feature in the user quantity data to determine the shape and characteristics of the data distribution.
[0119] 4-5-4) Use the numpy.linspace() function to divide the continuous attribute into several intervals or levels. Define the number of intervals as 5. And use the numpy.digitize() function to map each user's attribute value to the corresponding discrete level. The discrete level starts from 0.
[0120] 4-5-5) Use the numpy.zeros() function to create a matrix to store the feature importance of each user. Traverse the list of user IDs and map the feature importance to each user. According to the position of the user ID in the source platform and the target platform, assign the corresponding feature importance value to each user. And use the numpy.save() function to save the user feature importance matrix to a file for subsequent use.
[0121] 5) Use user discriminative features to construct similar user pairs, thereby generating the correlation information required in the correction process. In the Tensorflow framework, we use progressive iterative and hierarchical correction learning to directly integrate various attributes of users across platforms, adjust the position of users in the embedding space, and push similar user pairs across platforms closer to each other, while ensuring that the embedding of other users does not change as much as possible. Figure 2 and attached Figure 3 As shown in the figure, the attribute heterogeneity challenge of user identity association problem in unsupervised scenarios is effectively solved. In addition, the stochastic gradient descent method is used to optimize the objective function. Through training, the final user embedding can be obtained.
[0122] 6) Calculate the cosine similarity of the embedding vectors between each pair of cross-platform users as the user matching score. The larger the user matching score, the more likely it is that the two belong to the same offline entity user.
[0123] It should be noted that the above embodiments are not intended to limit the protection scope of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.
Claims
1. An unsupervised cross-domain user matching method based on embedding correction, characterized in that: The steps include: S1. Data preprocessing by sorting, cleaning, conversion and normalization; S2, by using the pre-trained RoBERTa, Canine model and graph neural network technology to obtain user multi-attribute embedding; S3, use the attention mechanism and stochastic gradient descent method to fuse the features of different single attribute embeddings; S4, integrating multiple types of information simultaneously on a single platform and across platforms to form multi-dimensional user discriminative features; S5. Use user discriminative features to construct similar user pairs, train and generate the correlation information required in the correction process, and obtain the final user embedding; S6. Calculate the cosine similarity of the embedding vectors between each pair of cross-platform users as the user matching score, and determine the cross-network social network users based on the score.
2. The unsupervised cross-domain user matching method based on embedding correction according to claim 1 is characterized in that: The specific cleaning process of step S1 includes the following steps: S1-1. For the crawled cross-domain user pairs, the source platform and target platform users are re-sorted according to the unique correspondence of their user IDs to form sequential user pair data; S1-2, identify different user attributes existing in different platforms, and analyze and screen the attributes; S1-3. Use regular matching to clean data and clean user attributes of garbled characters, meaningless emoticons, meaningless test English alphanumeric strings, etc.
3. The unsupervised cross-domain user matching method based on embedding correction according to claim 1 is characterized in that: The specific operation of step S2 includes the following steps: S2-1. For each user, concatenate the user nickname and user name, load the pre-trained RoBERTa model and use its built-in word segmenter to perform word segmentation, calculate and generate the comprehensive embedding of the user name and user nickname, S2-2. For each user’s user description, load the pre-trained Canine model and use its built-in word segmenter to segment the text, calculate and generate the user description embedding, and truncate the text content that exceeds the length. S2-3. Use user interaction relationships to build a topological network, add username embeddings as node attributes to user nodes, use graph neural network technology based on this topological network, use the attention mechanism to learn the interaction relationships between user nodes, and calculate and generate user topological embeddings.
4. The unsupervised cross-domain user matching method based on embedding correction according to claim 1 is characterized in that: The specific feature fusion process based on the attention mechanism and stochastic gradient descent in step S3 includes the following steps: S3-1, use the attention pooling mechanism to dynamically adjust the weights between different embeddings, and cross and concatenate the weighted single attribute embeddings; S3-2, use a multi-layer neural network to perform nonlinear transformation and feature fusion on the fused features; S3-3. Combining the idea of iterative optimization, use stochastic gradient descent (SGD) to fine-tune and optimize the fused feature representation to form the basic embedding of the user.
5. The unsupervised cross-domain user matching method based on embedding correction according to claim 1 is characterized in that: The specific cross-platform multi-dimensional user discriminative feature generation technology of step S4 includes the following steps: S4-1, complete the generation of discriminative features based on user name data by preprocessing user name data, extracting n-gram features, constructing TF-IDF vectors, calculating Jaccard similarity coefficients, etc.; S4-2, completing the generation of discriminative features based on user network topology data by building a network neighbor list and node degree dictionary, calculating the cosine similarity matrix, calculating the degree matrix, calculating the cross-platform similarity, and normalizing the similarity score; S4-3, complete the generation of discriminative features based on user tweet data by loading vocabulary and filtering low-frequency words, building user-vocabulary mapping, extracting discriminative words, calculating the number of word overlaps and Jaccard similarity coefficients, and normalizing similarity scores; S4-4, complete the generation of discriminative features based on user education and work experience data through steps such as text preprocessing, merging data and adding labels, calculating TF-IDF features, training random forest model, creating user feature importance matrix and mapping to users; S4-5. The discriminative features based on the user quantity data are generated through the steps of text preprocessing, merging data and adding labels, calculating the skewness and kurtosis of features, discretization processing, creating a user feature discretization matrix and mapping it to users.
6. The unsupervised cross-domain user matching method based on embedding correction according to claim 5 is characterized in that: The specific processing process of step S4-1 includes the following steps: S4-1-1, preprocessing user name data to standardize user names; S4-1-2, extracting n-gram features of the user name; S4-1-3. Calculate the TF-IDF value and generate the TF-IDF vector of the user name. S4-1-4. Calculate the number of word overlaps and the Jaccard similarity coefficient and save them to a file.
7. The unsupervised cross-domain user matching method based on embedding correction according to claim 5 is characterized in that: The specific processing process of step S4-2 includes the following steps: S4-2-1, build a network neighbor list and node degree dictionary; S4-2-2, establish an ID mapping table; S4-2-3, calculate the cosine similarity matrix; S4-2-4, construct the degree matrix according to the node degree dictionary; S4-2-5, calculate the similarity of the user's neighbor nodes, and calculate the similarity score by combining the degree and cosine similarity. S4-2-6. Normalize the similarity score and save it to a file.
8. The unsupervised cross-domain user matching method based on embedding correction according to claim 5 is characterized in that: The specific processing process of step S4-3 includes the following steps: S4-3-1, load the vocabulary, and filter and generate a set of low-frequency words; S4-3-2, based on the above steps, construct user-vocabulary mapping; S4-3-3, low-frequency words are sorted by frequency and stored in the dictionary; S4-3-4, calculate the number of word overlaps and the Jaccard similarity coefficient; S4-3-5. Normalize the similarity score and save it to a file.
9. The unsupervised cross-domain user matching method based on embedding correction according to claim 5 is characterized in that: The specific processing process of step S4-4 includes the following steps: S4-4-1. Convert the JSON file of the data into Pandas DataFrame format and preprocess the data to remove special characters and spaces. S4-4-2, merge data and add labels for classification tasks; S4-4-3, convert text data into TF-IDF feature vector; S4-4-4. Train the random forest model and extract feature importance; S4-4-5. Create a user feature importance matrix and map it to users, and save it to a file.
10. The unsupervised cross-domain user matching method based on embedding correction according to claim 5, characterized in that: The specific knowledge supplementation process of step S4-5 includes the following steps: S4-5-1. Convert the JSON file of the data into Pandas DataFrame format; S4-5-2, merge data and add labels for classification tasks; S4-5-3. Calculate the skewness and kurtosis of each user feature to determine the shape and characteristics of data distribution; S4-5-4, divide the continuous attributes into several intervals or levels, and map the attribute value of each user to the corresponding discrete level; S4-5-5. Create a user feature importance matrix and map it to users, and save it to a file.
Citation Information
Patent Citations
Cross-social network user identity matching method based on entropy weight method, medium and device
CN115048563A
Power business system data alignment method based on graph convolutional neural network
CN115935941A
Binary function similarity detection method using hypergraph twin neural network
CN117475180A
Network embedding method based on attention mechanism
CN118094252A
Multi-granularity fusion model-based intelligent semantic chinese sentence matching method, and device
WO2021164199A1
Cited By
Cross-domain user matching method based on interactive behaviors
CN122451487A
An interactive behavior-based cross-domain user matching method
CN122451487B
