An interactive behavior-based cross-domain user matching method
By generating dual anomaly signals and updating node classifications based on preliminary classification, attribution analysis, and anomaly detection based on user interaction behavior, the problem of interference from fake accounts in cross-domain user matching is solved, achieving highly accurate and robust cross-domain user association.
Patent Information
- Application Number
- CN202610707560.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-21
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-05-21
AI Technical Summary
Existing cross-domain user matching technologies struggle to distinguish between real and fake identities, lack effective differentiation between personal interests and the influence of others, resulting in low matching accuracy and a lack of adaptive feedback optimization mechanisms.
By acquiring users' historical interaction behavior data, a preliminary classification is performed to generate a candidate set of real nodes, fake nodes, and nodes to be detected. The driving factors of interaction behavior are distinguished as personal interests or the influence of others. The cross-domain interest migration strength is calculated, anomaly signals are generated, and the corrected confidence scores are calculated based on the anomaly signals to update the node classification and construct a closed-loop iterative mechanism.
It significantly suppresses the interference of fake accounts on matching results, improves the accuracy of the conversion of the node set to be detected into the real or fake node set, and achieves highly robust and accurate cross-domain user association.
Smart Images

Figure CN122451487B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of social network data analysis technology, and in particular to a cross-domain user matching method based on interactive behavior. Background Technology
[0002] In the context of increasingly diversified social networking platforms, the same individual often owns multiple accounts on different platforms. Cross-domain user matching technology aims to link these accounts to construct a complete user profile. However, existing methods mostly rely on single-dimensional user profile attributes or topological structures, making it difficult to address challenges such as fake accounts falsifying information, sparse behavioral data, and complex motivations for interaction behaviors. For example, fake accounts often interfere with the matching process by imitating the behavioral patterns of real users or forging social relationships, and existing methods lack an effective distinction between "self-interest driven" and "other-influenced" approaches, leading to attribution bias. At the same time, traditional methods are insufficient in matching quality assessment and adaptive parameter adjustment, making it difficult to form a closed-loop optimization. Therefore, there is an urgent need for a cross-domain user matching method that can integrate multi-source interactive behavioral data, identify fake accounts, and decouple behavioral driving factors to improve the accuracy and robustness of matching.
[0003] Chinese Patent Publication No. CN119025934A discloses an unsupervised cross-domain user matching method based on embedding correction, including: S1, data preprocessing using sorting, cleaning, transformation, and normalization; S2, obtaining multi-dimensional single-attribute embeddings of users based on pre-trained language models and graph neural network technology; S3, using an attention mechanism to fuse features of different single-attribute embeddings; S4, simultaneously integrating multiple types of information in both single-platform and cross-platform aspects to form multi-dimensional user discriminative features; S5, constructing similar user pairs using user discriminative features, training to generate the relevance information required in the correction process, and obtaining the final user embeddings; S6, calculating the cosine similarity of the embedding vectors between each pair of cross-platform users as the user matching score, and determining cross-network social network users based on the score. It is evident that existing cross-domain user matching technologies struggle to distinguish between real and fake identities, and to decouple personal interests from the influence of others, resulting in low accuracy in cross-domain user matching. Summary of the Invention
[0004] To address this issue, the present invention provides a cross-domain user matching method based on interactive behavior, which overcomes the problems of low matching accuracy caused by interference from fake accounts, complex behavioral motivations, and lack of adaptive feedback optimization mechanisms in the prior art.
[0005] To achieve the above objectives, the present invention provides a cross-domain user matching method based on interactive behavior, comprising:
[0006] Obtain historical interaction data for each user's account nodes on social networking platforms;
[0007] Based on historical interaction data, account nodes are initially classified to generate a candidate set of real nodes, a candidate set of fake nodes, and a candidate set of nodes to be detected.
[0008] Initial matching and attribution analysis are performed on account nodes in the real node candidate set to distinguish whether the driving factors of their interaction behavior are their own interests or the influence of others.
[0009] For behaviors determined to be driven by their own interests, the user's interest transfer patterns between the source and target domains are analyzed based on their own interest types, and the cross-domain interest migration intensity is calculated.
[0010] Among them, the types of self-interest include target domain interests, transferable source domain interests, and hybrid domain interests;
[0011] In response to the detection that the cross-domain interest migration intensity is greater than the migration intensity threshold, a first abnormal signal is generated;
[0012] For behaviors determined to be driven by the influence of others, trace the source account node of the influence;
[0013] In response to the detection that the affected source account node does not belong to the real node candidate set, a second abnormal signal is generated;
[0014] Calculate the corrected confidence score based on the first and second anomalous signals;
[0015] The account nodes are reclassified based on the corrected confidence scores, the real node candidate set, the fake node candidate set, and the node candidate set to be detected are updated, and the initial matching is performed again using the updated real node candidate set.
[0016] Furthermore, the process of initially classifying account nodes based on historical interaction behavior data includes:
[0017] Calculate the user profile integrity score;
[0018] The user profile integrity score is compared with the integrity score threshold to obtain the original confidence score based on the comparison result;
[0019] The comparison results include the first comparison result and the second comparison result;
[0020] Based on the original confidence scores, account nodes are divided into a real node candidate set, a fake node candidate set, and a node candidate set to be detected;
[0021] Compare the user profile integrity score with the integrity score threshold:
[0022] If the user profile integrity score is greater than or equal to the integrity score threshold, the first comparison result is obtained, three types of features are extracted, and the raw confidence score is obtained by inputting the logistic regression classifier.
[0023] If the user profile integrity score is less than the integrity score threshold, a second comparison result is obtained, and the original confidence score is calculated based on the associated account nodes.
[0024] Further, the process of extracting three types of features and inputting them into a logistic regression classifier to obtain the raw confidence scores includes:
[0025] The three types of features are behavioral regularity features, social authenticity features, and interaction depth features;
[0026] Calculate the regularity characteristic value of behavior based on the coefficient of variation of behavior interval, behavior concentration, and periodicity intensity;
[0027] The social authenticity feature value is calculated based on the fan / follower ratio, local clustering coefficient, and isolated node ratio.
[0028] The interaction depth feature value is calculated based on the interaction type entropy, average comment length, and content redundancy.
[0029] The behavioral regularity feature value, social authenticity feature value, and interaction depth feature value are input into the logistic regression model to obtain the original confidence score.
[0030] Furthermore, the process of calculating the original confidence score based on the associated account nodes includes:
[0031] Analyze the associated account nodes that are related to the account node to form associated account node pairs;
[0032] The original confidence score is calculated based on the historical interaction data of the associated account node pairs to perform preliminary classification of the account nodes.
[0033] Furthermore, the process of dividing account nodes into a candidate set of real nodes, a candidate set of fake nodes, and a candidate set of nodes to be detected based on the original confidence scores includes:
[0034] Compare the original confidence score with the first confidence score threshold and the second confidence score threshold:
[0035] If the original confidence score is less than or equal to the first confidence score threshold, the account node is classified into the candidate set of fake nodes;
[0036] If the original confidence score is between the first confidence score threshold and the second confidence score threshold, the account node is divided into a candidate set of nodes to be detected;
[0037] If the original confidence score is greater than or equal to the second confidence score threshold, the account node is classified into the real node candidate set.
[0038] Furthermore, the initial matching process for account nodes in the candidate set of real nodes includes:
[0039] The account nodes in the real node candidate set are initially matched to form several matching account node pairs;
[0040] Calculate the actual similarity between each pair of matching account nodes, and mark the account nodes with actual similarity higher than the similarity threshold as the target matching nodes.
[0041] Furthermore, the process of attribution analysis on account nodes in the candidate set of real nodes includes:
[0042] A global user profile is generated based on the target matching nodes, and the correct matching probability of account nodes in the global user profile is calculated.
[0043] In response to a correct match probability less than a preset second threshold, attribution analysis is performed on account nodes in the real node candidate set.
[0044] Furthermore, in response to detecting that the cross-domain interest migration intensity is greater than a migration intensity threshold, a first anomalous signal is generated, specifically including:
[0045] When the cross-domain interest migration intensity is greater than the migration intensity threshold, it is determined that the interest migration pattern of the account node is abnormal, and the first abnormal signal is generated.
[0046] The first abnormal signal indicates that the account node does not belong to the real node candidate set, or that the interaction behavior of the account node should be judged as being driven by the influence of others rather than by its own interests.
[0047] Furthermore, the process of calculating the corrected confidence score based on the first and second anomalous signals includes:
[0048] The adjusted confidence score includes the first adjusted confidence score and the second adjusted confidence score;
[0049] When the second abnormal signal is received, the second corrected confidence score is calculated based on the second penalty value;
[0050] When the first abnormal signal is received, a first corrected confidence score is calculated based on the first penalty value.
[0051] Furthermore, the process of reclassifying account nodes based on the revised confidence score includes:
[0052] Account nodes whose corrected confidence scores are greater than the second confidence score threshold will be moved into the real node candidate set;
[0053] Account nodes whose corrected confidence scores are less than or equal to the first confidence score threshold are moved into the fake node candidate set;
[0054] The remaining account nodes will remain in the candidate set of nodes to be detected.
[0055] Compared with existing technologies, the beneficial effects of this invention are as follows: by constructing a closed-loop iterative mechanism of preliminary classification, attribution analysis, anomaly detection, and feedback correction, it can effectively distinguish between interactive behaviors driven by one's own interests and those driven by the influence of others. It generates dual anomaly signals based on the intensity of cross-domain interest migration and the authenticity of the influence source, performs punitive correction on the confidence score according to the anomaly signals, and updates the classification results of account nodes based on the corrected confidence score. This significantly suppresses the interference of fake accounts on the matching results and improves the accuracy of the conversion of the node set to be detected into the real or fake node set. At the same time, by iteratively optimizing the matching quality of the real node candidate set, it achieves high robustness and high accuracy of cross-domain association for the same natural person's accounts on multiple platforms. Attached Figure Description
[0056] Figure 1 This is a flowchart illustrating the cross-domain user matching method based on interactive behavior according to an embodiment of the present invention.
[0057] Figure 2 This is a schematic diagram illustrating the process of preliminary classification of account nodes based on the historical interaction behavior data in an embodiment of the present invention;
[0058] Figure 3 This is a schematic diagram of the process for obtaining the original confidence score in an embodiment of the present invention;
[0059] Figure 4 This is a schematic diagram illustrating the process of initial matching of account nodes in the real node candidate set according to an embodiment of the present invention. Detailed Implementation
[0060] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.
[0061] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.
[0062] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.
[0063] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0064] Please see Figure 1 The diagram shown is a flowchart illustrating a cross-domain user matching method based on interactive behavior according to an embodiment of the present invention. The present invention provides a cross-domain user matching method based on interactive behavior, comprising:
[0065] Step S1: Obtain historical interaction data corresponding to each user's account nodes on the social network platform;
[0066] Step S2: Based on the historical interaction behavior data, the account nodes are initially classified to generate a candidate set of real nodes, a candidate set of fake nodes, and a candidate set of nodes to be detected.
[0067] Step S3: Perform initial matching and attribution analysis on the account nodes in the real node candidate set to distinguish whether the driving factor of their interaction behavior is their own interest or the influence of others.
[0068] Step S4: For behaviors determined to be driven by their own interests, analyze the user's interest transfer patterns between the source and target domains based on their own interest types, and calculate the cross-domain interest migration intensity.
[0069] The self-interest types include target domain interests, transferable source domain interests, and hybrid domain interests;
[0070] In response to detecting that the cross-domain interest migration intensity is greater than the migration intensity threshold, a first abnormal signal is generated;
[0071] Step S5: For behaviors determined to be driven by the influence of others, trace the source account node of the influence;
[0072] In response to the detection that the affected source account node does not belong to the real node candidate set, a second abnormal signal is generated;
[0073] Step S6: Calculate the corrected confidence score based on the first abnormal signal and the second abnormal signal;
[0074] Step S7: Reclassify the account nodes according to the corrected confidence score to generate a candidate set of real nodes, a candidate set of fake nodes, and a candidate set of nodes to be detected;
[0075] Account nodes whose corrected confidence scores are greater than the score threshold will be moved into the real node candidate set;
[0076] Account nodes whose corrected confidence scores are less than or equal to the score threshold will be moved into the fake node candidate set.
[0077] The remaining account nodes will remain in the candidate set of nodes to be detected;
[0078] Then return to step S3 and perform the initial matching again using the updated set of real node candidates. The iteration stops when the preset maximum number of iterations, 10, is reached.
[0079] In this embodiment, the historical interaction data includes at least the type of each interaction, the timestamp of the occurrence, the content ID targeted, and the account ID of the content publisher. This data can be obtained through public interfaces or data logs provided by social network platforms. The types of interactions include likes, comments, reposts, and favorites. As an alternative or supplement, the iteration is terminated early when the change ratio of accounts in the real node candidate set and the fake node candidate set is less than 1% in two consecutive iterations.
[0080] By constructing a closed-loop iterative mechanism encompassing preliminary classification, attribution analysis, anomaly detection, and feedback correction, it can effectively distinguish between interactive behaviors driven by one's own interests and those influenced by others. It generates dual anomaly signals based on the intensity of cross-domain interest migration and the authenticity of the influence source. The confidence score is then penalized and corrected according to the anomaly signals, and the classification results of account nodes are updated based on the corrected confidence scores. This suppresses the interference of fake accounts on matching results and improves the accuracy of converting the node set to be detected into a real or fake node set. Simultaneously, by iteratively optimizing the matching quality of the real node candidate set, it achieves highly robust and accurate cross-domain association of the same individual across multiple platform accounts.
[0081] See Figure 2 As shown, it is a schematic diagram of the process of preliminary classification of account nodes based on the historical interaction behavior data in an embodiment of the present invention;
[0082] Specifically, the process of initially classifying account nodes based on the historical interaction behavior data includes:
[0083] Step S2001: Calculate the user profile integrity score;
[0084] Step S2002: Compare the user profile integrity score with the integrity score threshold to obtain the original confidence score based on the comparison result;
[0085] The comparison results include the first comparison result and the second comparison result;
[0086] Step S2003: Based on the original confidence score, the account nodes are divided into a real node candidate set, a fake node candidate set, and a node candidate set to be detected;
[0087] In this embodiment, the user profile integrity score is compared with the integrity score threshold:
[0088] If the user profile integrity score is greater than or equal to the integrity score threshold, the first comparison result is obtained;
[0089] Upon obtaining the first comparison result, three types of features are extracted and input into a logistic regression classifier to obtain the raw confidence scores.
[0090] If the user profile integrity score is less than the integrity score threshold, a second comparison result is obtained.
[0091] In the second comparison, the associated account nodes that are related to the account node are analyzed to form associated account node pairs. Based on the historical interaction behavior data of the associated account node pairs, the original confidence score is calculated to perform preliminary classification of the account nodes.
[0092] In this embodiment, the user profile integrity score reflects the completeness of the static profile information actively filled in by the user or provided by the platform. It is a quantitative indicator used to measure whether the initial data of the account is scarce. If the user fills in relatively complete personal information, such as nickname, avatar, gender, geographical location, personal introduction, and bound mobile phone / email, it means that the initial data of the account is not scarce and richer feature analysis can be performed. If a large amount of user information is missing or is garbled or default value, it means that the initial data of the account is scarce and cannot be effectively classified directly through profile information. Therefore, it is necessary to use related accounts, such as common interactions, social relationships, devices / IP, number of interactions, and length of behavior sequences, to make indirect inferences.
[0093] In this embodiment, the integrity score threshold is set to 0.5, which can be set according to the actual application scenario and platform characteristics. In a preferred embodiment, statistical analysis is performed on historical account data from multiple social networking platforms to calculate the integrity score distribution of real and fake accounts, and the score value that best distinguishes the two types of accounts is selected as the threshold. Experiments show that when the threshold is set to 0.5, the proportion of real accounts with a score ≥0.5 exceeds 85%, while the proportion of fake accounts with a score ≥0.5 is less than 20%. Therefore, this embodiment preferably sets the integrity score threshold to 0.5. As an alternative or supplement, this threshold can also be set as an adjustable parameter, adaptively adjusted according to the accuracy feedback of subsequent matching results. For example, if an abnormal proportion of low integrity accounts is found in the real node candidate set during the iteration process, the threshold can be dynamically reduced; conversely, if the number of high integrity accounts increases in the fake node candidate set, the threshold can be dynamically increased. The adjustment range is usually 0.4 to 0.6, and those skilled in the art can determine a suitable threshold through simple experiments based on the characteristics of the platform data.
[0094] If the user profile completeness score is greater than or equal to the completeness score threshold, it indicates that most key attributes, such as username, avatar, and binding method, have been filled in, which can support the analysis of behavioral patterns and social authenticity, and thus enters the high completeness path. If the user profile completeness score is less than the completeness score threshold, it indicates that there are insufficient direct features, and inferences need to be made based on associated accounts, thus entering the low completeness path.
[0095] In this embodiment, the specific process for calculating the user profile integrity score is as follows: The account node's profile attribute set is defined as follows: Username, Nickname, Avatar, Gender, Location, Personal Profile, Linked Mobile Phone / Email, and Registration Time, totaling 8 attributes. Each attribute is assigned a weight: Username 0.15, Nickname 0.10, Avatar 0.10, Gender 0.05, Location 0.10, Personal Profile 0.15, Linked Mobile Phone / Email 0.25, and Registration Time 0.10. The user profile integrity score I = Valid attributes are defined as non-empty, non-garbled characters (length > 1 and does not contain consecutive number strings), and non-default avatar. For example, if an account only has a valid username and bound mobile phone number, and all other attributes are invalid, then the sum of the weights of the valid attributes = 0.15 + 0.25 = 0.40, and the sum of the weights of all attributes = 1.0, so the score = 0.40 < 0.5, thus entering the low integrity path.
[0096] See Figure 3 The diagram shown illustrates the process for obtaining the original confidence scores.
[0097] Specifically, in step S2002, the process of extracting three types of features and inputting them into the logistic regression classifier to obtain the original confidence scores includes:
[0098] The three types of features are behavioral regularity features, social authenticity features, and interaction depth features;
[0099] Step S2012: Calculate the behavioral regularity characteristic value based on the behavioral interval variation coefficient, behavioral concentration, and periodicity intensity;
[0100] Step S2022: Calculate the social authenticity feature value based on the fan / follower ratio, local clustering coefficient, and isolated node ratio;
[0101] Step S2032: Calculate the interaction depth feature value based on the interaction type entropy, average comment length, and content duplication.
[0102] Step S2042: Input the behavioral regularity feature value, social authenticity feature value, and interaction depth feature value into the logistic regression model to obtain the original confidence score.
[0103] In this embodiment, the behavioral regularity feature value Freg = (f1 + f2 + f3) / 3; the social authenticity feature value Fsoc = (f4 + f5 + f6) / 3; and the interaction depth feature value Fdepth = (f7 + f8 + f9) / 3. The mapping score f1 = max(0, min(1, (CV - 0.5) / 1.5)), where CV is the coefficient of variation of behavioral intervals. The formula for calculating the coefficient of variation of behavioral intervals is: CV = σΔt / μΔt, where the calculation is performed using the most recent 100 interactions. σΔt represents the standard deviation of the time interval (Δt) between two consecutive interactions, reflecting the dispersion of the behavioral intervals; μΔt represents the average value of the time intervals, reflecting the average frequency of the behavior. If the total number of historical interactions Ntotal is less than 100, then all available interaction data is used to calculate CV. If Ntotal < 3, then CV cannot be calculated, and the behavioral regularity feature value Freg is set to 0.5. Accounts are directly included in the candidate set of nodes to be detected and do not participate in the initial matching; the coefficient of variation of behavior interval is used to measure the regularity of user interaction behavior: the interaction interval of real users tends to fluctuate greatly (high CV), while the behavior interval of fake accounts is usually more uniform (low CV); generally, the CV of real users > 1.5, and the CV of fake bots < 0.5; f2 = 1 - behavior concentration, behavior concentration is the ratio of the maximum number of hours of interaction to the total number of interactions; f3 = 1 - periodicity intensity, periodicity intensity is the maximum peak value of autocorrelation; f4 = max(0, 1 - |lg[(r+1) - 0.5]| / 2), r is the fan / follow ratio; f5 = 1 - |Ca - 0.5| / 0.5, Ca is the local clustering coefficient, representing the average edge density between user friends; f6 = 1 - isolate, isolate is the proportion of isolated nodes, representing the proportion of friends without other friends; f7 = H / ln5, H is the interaction type entropy, H = - Normalized to 0-1, with a maximum of ln5, Pt represents the probability of the t-th interaction type occurring in all of the user's historical interaction behaviors, ∑Pt=1; f8 is the normalized average comment length, normalized as 1 for >50 characters and 0 for 0 characters; content duplication is the proportion of repeated sentences, obtained by calculating Jaccard similarity, Jaccard similarity >0.9 is considered duplication, f9=1-content duplication; logistic regression model:
[0104] =W0+WregFreg+WsocFsoc+WdepthFdepth;
[0105] s = σ(x) = 1 / (1 + ), W0=-0.5, Wreg=2, Wsoc=1.5, Wdepth=1;
[0106] For example, a user's profile integrity is 0.85, behavioral interval coefficient of variation (CV) is 2.1, behavioral concentration is 0.15, periodicity intensity is 0.3; follower / follower ratio (r) is 1.2, local clustering coefficient (Ca) is 0.4, isolated node ratio (isolate) is 0.1, and out of 100 interactions: 70 likes, 20 comments, and 10 shares, then: P 点赞 =0.7, P 评论 =0.2, P 转发 =0.1, then the interaction type entropy H = -(0.7 × ln0.7 + 0.2 × ln0.2 + 0.1 × ln0.1) ≈ 0.802, the average comment length is 80 characters, and the content duplication rate is 0.05.
[0107] Then f1 = max(0, min(1, (2.1-0.5) / 1.5)) = 1.0; f2 = 1 - behavioral concentration = 1 - 0.15 = 0.85; f3 = 1 - periodicity intensity = 0.7, regularity feature value Freg = 0.85; f4 = max(0, 1 - |lg(r+0.5)| / 2) ≈ 0.9, f5 = 1 - |Ca-0.5| / 0.5 = 0.8, f6 = 0.9, social authenticity feature value Fsoc = 0.867; f7 = 0.802 / ln5 = 0.498, f8 = 1, f9 = 0.95, interaction depth feature value Fdepth = 0.816; x = -0.5 + 2 × 0.85 + 1.5 × 0.867 + 1 × 0.816 = 3.3165, s = ≈0.965.
[0108] The weight parameters W0, Wreg, Wsoc, and Wdepth in the above logistic regression model can be obtained through supervised training on a sample set labeled with real or fake accounts. The specific values given in this embodiment are preferred training results.
[0109] Logistic regression models can be obtained using conventional supervised training methods. Specifically, a sample set labeled with real and fake accounts is used to extract behavioral regularity features, social authenticity features, and interaction depth features as input. Binary cross-entropy is used as the loss function, and the model parameters are iteratively trained using gradient descent, such as the Adam optimizer. The training data can be labeled accounts collected from social network platforms, or it can be expanded by using high-confidence samples in a semi-supervised manner. Those skilled in the art can determine the training rounds, batch size, and other hyperparameters according to the actual data scale.
[0110] Optionally, the method described in this embodiment is applicable to account nodes that have at least a small amount of interactive behavior on each platform, such as ≥3 interactions; for newly registered accounts with no interactive behavior at all, they can be included in the candidate set of nodes to be detected, and automatically included in the matching process after sufficient data is accumulated.
[0111] Specifically, the process of calculating the raw confidence score based on the historical interaction data of associated account node pairs includes:
[0112] Based on the analysis of social data among users, it is determined whether the account nodes are associated with other account nodes;
[0113] When an account node has associated accounts and already has known tags, the original confidence score is calculated based on the weighted tag propagation model.
[0114] In this embodiment, the social data between users includes shared interactive items (at least one identical item ID), social neighbors (followers / followers / @ relationships), identical device fingerprints, or the same C-class IP address. Based on the social data between users, the analysis checks whether the account nodes have associated account nodes. Account nodes without associated account nodes are divided into a candidate set of nodes to be detected. For account nodes with associated account nodes, the tags of the social data between users of the associated account nodes are obtained. If they are not marked as real node candidate sets or fake node candidate sets, they are determined to have no associated tags. If they are marked as real node candidate sets or fake node candidate sets, it means that there are already known tags. Then, the original confidence score is calculated according to the weighted tag propagation model. Based on the original confidence score, the account nodes are divided into real node candidate sets, fake node candidate sets, and candidate nodes to be detected.
[0115] The weighted label propagation model uses the original confidence score = Where Wuv is the association strength between account u and associated account v; Yv is the label value, Yv=1 if v belongs to the set of real nodes, and Yv=0 if v belongs to the set of fake nodes; L is the set of associated accounts with known labels, and the association strength Wuv is calculated using the formula Wuv= +0.5×J+0.3×H, where D is the social distance; if it's a direct follow / fan / @ relationship, then the social distance D=1; if it's a second-degree friend, then the social distance D=2; otherwise, the social distance D=0; J is the shared item feature value; if there are at least one interaction record with the same item ID, then J=1; otherwise, 0; H is the feature value corresponding to the same device fingerprint or the same C-segment IP; if account u and associated account v have the same device fingerprint or the same C-segment IP, then H=1; otherwise, 0. For example, suppose account u has three associated accounts v1, v2, and v3, where associated account v1 has a social distance of 1, shares a common interaction item, and has the same device fingerprint or the same C-segment IP, then the corresponding association strength is Wuv1= +0.5×1+0.3×1=1.3; The social distance of associated account v2 is 2, there are no shared interactive items, no identical device fingerprints or the same C-class IP address, then the corresponding association strength is Wuv2= +0.5×0+0.3×0≈0.333; The social distance of the associated account v3 is 1, they have shared interactive items, and they do not have the same device fingerprint or the same C-class IP. Therefore, the corresponding association strength is Wuv3= +0.5×1+0.3×0=1.0; therefore, the original confidence score = If ≈0.874>0.7, then the account is classified into the real node candidate set.
[0116] The raw confidence score calculated by the weighted label propagation model represents the weighted average of the related account labels. Since a real account may be surrounded by a large number of fake accounts, possibly due to being surrounded by trolls, or a fake account may have fabricated associations with multiple real accounts, perhaps by purchasing followers or engaging in joint interactions, and since it is difficult for a fake account to simultaneously fabricate multiple strong associations, different weights are assigned to different association types to suppress this false propagation. If the related labels conflict, i.e., the ratio of real to fake is close, the calculated raw confidence score will fall between 0.3 and 0.7, and it will be included in the detection set without direct judgment. Furthermore, as the iteration progresses, the label quality of the real node candidate set and the fake node candidate set will continuously improve due to the correction of erroneous labels, and the accuracy of subsequent propagation will also improve.
[0117] Specifically, the process of dividing the account nodes into a real node candidate set, a fake node candidate set, and a node candidate set to be detected based on the original confidence scores includes:
[0118] Compare the original confidence score with the first confidence score threshold and the second confidence score threshold:
[0119] If the original confidence score is less than or equal to the first confidence score threshold, the account node is classified into a fake node candidate set;
[0120] If the original confidence score is between the first confidence score threshold and the second confidence score threshold, the account node is divided into a candidate set of nodes to be detected;
[0121] If the original confidence score is greater than or equal to the second confidence score threshold, the account node is divided into a real node candidate set.
[0122] In this embodiment, the first confidence score threshold is set to 0.3, and the second confidence score threshold is set to 0.7.
[0123] See Figure 4 As shown, it is a schematic diagram of the process of initial matching of account nodes in the real node candidate set according to an embodiment of the present invention.
[0124] Specifically, in step S3, the process of initially matching account nodes in the real node candidate set includes:
[0125] Step S3011: Perform initial matching on the account nodes in the real node candidate set to form several matching account node pairs;
[0126] Step S3012: Calculate the actual similarity between each pair of matching account nodes, and mark the account nodes with actual similarity higher than the similarity threshold as target matching nodes.
[0127] In this embodiment, the initial matching of account nodes in the real node candidate set to obtain the target matching node is performed as follows: For each account node, three types of feature vectors for matching are extracted, including user profile attribute feature vector fpro, user topology feature vector ftopo, and behavior sequence time fingerprint feature vector ftime; the above three feature vectors are concatenated to obtain the comprehensive representation vector Vu∈ of the account node. Optionally, it can be compressed to 128 dimensions using a fully connected layer; for the two social network platforms A and B to be matched, there are real node candidate sets respectively. and Using block index matching, for For each account node, a hash bucket index is calculated using MinHash based on its profile characteristics (such as username prefix and geographical location), and only those nodes are indexed... Accounts within the same or adjacent hash buckets are considered as candidate matching objects, thus generating a candidate matching pair set C = {(UA, UB)}. For each candidate matching pair, profile similarity, topological similarity, and temporal fingerprint similarity are calculated separately. Then, a comprehensive similarity is calculated based on profile similarity, topological similarity, and temporal fingerprint similarity as the actual similarity. The similarity threshold is set to 0.75. Account nodes with actual similarity greater than or equal to the similarity threshold are marked as target matching nodes. The specific process of extracting the user profile attribute feature vector fpro is as follows: the text fields such as username and nickname are mapped to 128-dimensional semantic vectors through a pre-trained model (such as RoBERTa); the avatar image is mapped to 128-dimensional visual vectors through a CNN (such as ResNet-50); and the category attributes such as geographic location and gender are one-hot encoded and then concatenated. Then, a fully connected layer is used to compress it to 64 dimensions; finally, the profile feature vector is concatenated to obtain a dimension of 256; the specific process of extracting the user topological structure feature vector ftopo is as follows: a social graph is constructed based on the user's follower / fan relationship, and a graph convolutional network is used to generate a 64-dimensional embedding; structural indicators such as Jaccard similarity and Adamic-Adar index with neighbors are calculated, and after concatenation, a 128-dimensional feature is obtained; the final topological feature vector has a dimension of 128; the specific process of extracting the behavioral sequence time fingerprint feature vector ftime is as follows: the timestamp sequence of the user's historical interaction is extracted, the interval distribution (mean, standard deviation, quantile, autocorrelation coefficient) is calculated, and a 16-dimensional statistical feature is constructed; the behavioral sequence is encoded using BiLSTM to obtain a 64-dimensional sequence embedding; the final time fingerprint feature vector has a dimension of 80; profile similarity simprof = Topological similarity is normalized using common neighbors, and the topological similarity is simtopo = ×(1- ),in It is the set of first-order neighbors of a user on social networking platform A. Let be the set of first-order neighbors of a user on social networking platform B, dA be the degree of the user on social networking platform A, and dB be the degree of the user on social networking platform B. The degree represents the total number of first-order neighbor nodes of the corresponding account node on the platform, including people followed and followers. Calculate the Kolmogorov-Smirnov statistic for the behavioral interval distribution of the two accounts. The similarity is then converted to a time fingerprint, resulting in a time fingerprint similarity score of simtime=1- Actual similarity = simprof × wprof + simtopo × wtopo + simtime × wtime, where wprof is the weight corresponding to file similarity, wtopo is the weight corresponding to topological similarity, and wtime is the weight corresponding to time fingerprint similarity. wprof = 0.4, wtopo = 0.35, and wtime = 0.25.
[0128] Graph convolutional networks (GCNNs) are used to extract user topology feature vectors (ftopo). A two-layer LightGCN architecture can be employed, and unsupervised training can be performed using the Bayesian Personalized Ranking (BPR) loss function. Positive samples represent real user follow relationships, while negative samples are randomly sampled. If the platform graph structure is sparse, user profile features can be introduced as initial features, or virtual node technology can be used. All of the above models can be implemented using existing open-source frameworks such as PyTorch and DGL. Those skilled in the art can determine hyperparameters such as embedding dimension and learning rate based on the actual data scale. BiLSTM can be trained through a contrastive learning task using behavioral sequence encoding. Training data can use publicly available social network datasets, such as Twitter-15, Weibo-16, and Reddit, or can be annotated by those skilled in the art based on platform data.
[0129] Specifically, the process of performing attribution analysis on account nodes in the candidate set of real nodes includes:
[0130] A global user profile is generated based on the target matching node, and the correct matching probability of account nodes in the global user profile is calculated.
[0131] In response to the fact that the correct matching probability is less than a preset second threshold, attribution analysis is performed on the account nodes in the real node candidate set.
[0132] In this embodiment, the preset second threshold is used to determine whether the quality of the current matching result meets the requirements. According to experimental statistics, when the correct matching probability is lower than 0.7, the correction effect introduced by subsequent attribution analysis is improved; when the correct matching probability is higher than 0.9, the benefit brought by attribution analysis is small, and it can be skipped to save computing resources. Therefore, in this embodiment, the preset second threshold is preferably set to 0.7. This threshold can be adjusted according to the application scenario: for scenarios with extremely high matching accuracy requirements (such as financial anti-fraud), it can be set to 0.85; for scenarios that tolerate a certain error (such as content recommendation), it can be set to 0.6. The threshold adjustment range is recommended to be [0.5, 0.9]. Too low a threshold will lead to unnecessary attribution analysis overhead, while too high a threshold may miss potential fake accounts. The default value is 0.7 when there are no special requirements.
[0133] In this embodiment, the global user profile refers to the unified cross-platform user view formed by merging all target matching nodes (i.e., cross-platform account pairs) obtained through the initial matching, assuming they belong to the same natural person. This view includes graph construction, graph merging and clustering, consistency scoring, and estimation of the probability of correct matching. The specific construction method is as follows:
[0134] Graph construction involves using each target matching node pair (UA, UB) as an initial seed association, treating accounts of the same user on different platforms as the same entity, and constructing a cross-platform heterogeneous graph. Nodes represent accounts, and edges are divided into two categories: "matching edges," connecting the two matched accounts; and "interaction edges," preserving the original social interaction relationships within each platform. Graph merging and clustering employ connected component algorithms, such as disjoint-set data structures, to merge all account nodes directly or indirectly connected by matching edges into a "candidate natural person" cluster. Specifically, if account A1 has a matching edge with B1, and B1 has a matching edge with A2 (through B1's interaction edge or another matching edge), then A1, B1, and A2 belong to the same cluster, with each cluster representing a candidate cross-platform user identity. Consistency scoring: For each cluster, its internal consistency score is calculated. The formula is:
[0135]
[0136] in Let be the set of all matching edges within this cluster. The actual similarity calculated in step S3 is the global consistency score of the entire matching result set, which is the weighted average of the scores of all clusters.
[0137] Estimation of the correct match probability: When there is a known anchor node pair (an absolutely correct seed pair), the correct match probability is directly calculated using the validation set:
[0138]
[0139] If no known anchor node pair exists, the silhouette coefficient method is used: each cluster is treated as a category, and the average of all node silhouette coefficients is taken as the global silhouette coefficient S, with a value range of [-1, 1]. Let the probability of correct matching be... ;when If the matching quality is deemed insufficient, attribution analysis is triggered; this involves calculating the profile coefficient for each account node. When there is a matching edge or interaction edge between two nodes, the distance between the account nodes is defined as 1 minus the actual similarity calculated in step S3. For account nodes without direct matching edges, the distance between them is obtained through the shortest path in the graph. For example, A1—B1: matching edge, similarity=0.85, distance=0.15; B1—A2: interaction edge, similarity=0.78, distance=0.22; A2—B2: matching edge, similarity=0.90, distance=0.10. Then the shortest path from A1 to B2 is A1→B1→A2→B2, with a total distance of 0.15+0.22+0.10=0.47. If there is no edge directly from A1 to B2, i.e., no other shorter path, then the shortest path distance from A1 to B2 is 0.47. However, for two nodes that do not have any path between them, i.e., belong to different connected components, their distance is defined as 1.
[0140] In this embodiment, the attribution analysis uses a joint judgment based on the "difference between the behavior occurrence time window and the source content publication time" and the "semantic similarity between the interactive content and the user's historical interests." The specific process is as follows: for each user interaction (such as liking, commenting, or forwarding), the publication time of the corresponding content (such as posts or videos) is obtained. And the account of the content publisher. If a user interacts with content within 30 minutes of its publication, and the semantic similarity between that interaction and the user's historical interactions over the past 7 days is below a threshold of 0.4, it is considered "driven by the influence of others"; otherwise, it is considered "driven by personal interest." Additionally, if the user interacts with the content publisher... The existence of a clear social relationship (such as following or mutual following) and the fact that the poster's historical behavior includes interest tags similar to the user's enhance the confidence level of the "influence-driven" judgment.
[0141] In a preferred embodiment, statistical analysis of a large amount of real user interaction data is performed to calculate the natural response time distribution of users to content of interest. Experiments show that the interval between the occurrence time of interaction driven by one's own interests and the content publication time is relatively dispersed, with the median usually being several hours or even several days. However, interaction driven by the influence of others, such as content published by influential figures or forwarded by friends, often occurs within a short period after the content is published, because users see the content instantly through the information stream or notification. By performing ROC curve analysis on the labeled dataset, the time point that balances the accuracy and recall of identification driven by the influence of others is selected. In this embodiment, the time window is preferably set to 30 minutes. This time window can also be adaptively adjusted according to the characteristics of the platform, such as the content exposure speed and user activity periods: for platforms with fast information dissemination and high user activity, such as Weibo and Twitter, the window can be appropriately shortened to 15 minutes; for platforms with slow information dissemination and large user interaction delays, such as forums and Q&A communities, the window can be appropriately extended to 60 minutes. The adjustment range is set to 10-120 minutes. Those skilled in the art can select the window value that optimizes the attribution analysis effect by analyzing the time interval distribution of user interaction behavior on the platform.
[0142] In this embodiment, the similarity threshold of 0.4 is an empirical value set by statistically analyzing the semantic similarity distribution between real users' own interest interactions and interactions influenced by others on a labeled dataset. In practical applications, this value can be adjusted according to platform characteristics. Specifically, the semantic similarity threshold can be determined through statistical analysis of the labeled dataset or empirical optimization. In a preferred embodiment, a certain number of user interaction samples are collected, and each interaction is manually labeled as either driven by the user's own interests or by the influence of others. For each sample, the semantic similarity between its interaction content and the user's historical interest content is calculated. The cosine similarity can be calculated using the Sentence-BERT model, and the similarity distribution of the two types of samples is statistically analyzed: interactions driven by the user's own interests are usually highly correlated with the user's historical interests, with similarities mostly distributed between 0.6 and 0.9; while interactions driven by the influence of others are often highly correlated with the user's historical interests. Historical interests deviate significantly, with similarity scores mostly ranging from 0.1 to 0.5. ROC curve analysis is used to select a point that balances classification accuracy and recall. In this embodiment, the semantic similarity threshold is preferably set to 0.4. As an alternative or supplement, this threshold can also be adaptively adjusted based on the actual effect of attribution analysis: if a high proportion of misclassifications is found in the results driven by others (e.g., a user's actual interest is judged as being influenced by others), the threshold can be appropriately lowered to 0.35 to reduce misclassifications; if a high number of missed classifications are found (e.g., an actual influence from others is not detected), the threshold can be appropriately increased to 0.45. The adjustment range is typically set between 0.3 and 0.6. Those skilled in the art can determine a suitable threshold based on the user content preference distribution of a specific platform through cross-validation or online learning, or make appropriate adjustments based on platform characteristics and actual application needs.
[0143] The above thresholds can be determined through simple experiments based on the platform data distribution. Different values can be used for different platforms, all of which can achieve the attribution analysis purpose of this invention.
[0144] In a preferred embodiment, semantic similarity is calculated by encoding the interactive content text and the user's historical interactive content text into sentence vectors using a pre-trained Sentence-BERT model, and then calculating the cosine similarity between the two sentence vectors.
[0145] Specifically, in response to detecting that the cross-domain interest migration intensity is greater than a migration intensity threshold, a first anomalous signal is generated, which specifically includes:
[0146] When the cross-domain interest migration intensity is greater than the migration intensity threshold, it is determined that the interest migration pattern of the account node is abnormal, and the first abnormal signal is generated.
[0147] The first abnormal signal is used to indicate that the account node does not belong to the real node candidate set, or that the interaction behavior of the account node should be determined to be driven by the influence of others rather than by its own interests.
[0148] Optionally, the aforementioned first abnormal signal may indicate one of the following two situations: the account node does not belong to the real node candidate set, or the interaction behavior of the account node should be judged as being driven by the influence of others rather than by its own interests; regardless of which situation is indicated, it essentially indicates that the original confidence score of the account node is overestimated and needs to be corrected downward; therefore, the present invention does not distinguish between these two situations and uniformly uses the same penalty logic to calculate the first corrected confidence score.
[0149] In this embodiment, the cross-domain interest migration strength is calculated as follows: assuming the user... The behavioral sequence in the source domain is represented as The sequence of behaviors in the target domain is represented as The mixed domain sequence is (Merged in chronological order), symbol This represents a sequence concatenation operation; specifically, it concatenates the source domain into a sequence. and target domain behavior sequence Merge them into a new mixed-domain sequence according to the order of timestamps. For example, if the user's actions in the source domain are in chronological order... The chronological order of actions in the target domain is as follows And the actual time sequence is The resulting hybrid domain sequence is: ; in generating mixed domain sequences First, timestamps need to be standardized, that is, timestamps obtained from all platforms are uniformly converted into standard Coordinated Universal Time (UTC) to ensure the accuracy of cross-platform behavior timing. This operation is used to generate a complete behavioral trajectory reflecting the evolution of users' cross-domain interests. Three types of interest vectors are extracted using a sequence encoder (such as SASRec).
[0150] Target domain interest vector: ;
[0151] Source domain interest vector: ;
[0152] Mixed domain interest vector: ;
[0153] The "transferable" portion of interest from the source domain is selected using an attention mechanism:
[0154] ;
[0155] Then, the cosine similarity between each pair of the three types of interests is calculated:
[0156]
[0157]
[0158] Cross-domain interest transfer intensity The combined contribution of transferable source domain interests and hybrid domain interests to target domain interests:
[0159]
[0160] The value ranges from [0, 1]. The closer it is to 1, the higher the consistency between the source domain and the mixed domain interests and the target domain interests, and the higher the migration intensity. However, an excessively high migration intensity (greater than the set threshold, such as 0.9) is considered abnormal. This indicates that the account's interest migration pattern is too "perfect" and lacks the natural fluctuations of real users, or that the account is a fake account that deliberately imitates the migration pattern.
[0161] For example, a user The user enjoys watching "suspense movies" on Douyin (source domain) and likes to save "suspense novels" notes on Xiaohongshu (target domain). The encoder produces an interest vector (simplified to a 2D diagram):
[0162] Target domain interest Transferable source domain interests (filtered by attention) Mixed-domain interests (merged sequences) Calculate cosine similarity:
[0163]
[0164] Pick ;
[0165] Then the intensity of cross-domain interest transfer:
[0166]
[0167] Therefore, the first anomalous signal is generated, indicating that the user's interest migration is too perfect and may be driven by a fake account or other people's influence.
[0168] Specifically, the attention mechanism is used to filter out the "transferable" portion from the source domain interests. The specific process is as follows: Let the length of the user's behavioral sequence in the source domain be... (i.e., includes) (1 project), each location The corresponding source domain interest vector is The formula for calculating attention weights is:
[0169] ;
[0170] in: Represents the first in the source domain sequence The position (i.e., the first position) (Time step of each project or behavior), take the range ; It is a summation index variable, which also iterates through all positions in the source domain sequence. arrive , used for normalizing the summation of the denominator; Calculate the target domain interest vector and the source domain interest vector. The dot product (or cosine similarity, depending on the implementation) of the interest vectors is used to measure the degree of matching between the source domain interest and the target domain interest; ultimately, the transferable source domain interest vectors are obtained. The weighted sum of all source domain interest vectors:
[0171]
[0172] For example, if the source domain sequence length is 3 ( The dot products of the target domain interest vector and the three source domain interest vectors are respectively Then, after exponentiation, we get The sum of the denominators is Attention weight Interests can be transferred In this way, the source domain interests (the first one) that are highly matched with the target domain interests receive higher weights, thereby filtering out the transferable parts.
[0173] In this embodiment, the migration strength threshold can be set to When the migration intensity exceeds a threshold, the system determines that the interest migration pattern is abnormal and generates a first abnormal signal. This signal triggers subsequent confidence score adjustments to more accurately identify such fake accounts in the next iteration. It should be noted that the migration intensity threshold can be set according to the actual needs of cross-domain user matching and the platform's data distribution. In a preferred embodiment, statistical analysis is performed on cross-domain behavior datasets labeled with real users and fake accounts to calculate the cross-domain interest migration intensity distribution for real users and fake accounts respectively. Experiments show that the migration intensity of real users is usually distributed between 0.5 and 0.8, with natural fluctuations between different interest domains; while fake accounts, due to deliberate simulation or mechanical migration, often have a migration intensity higher than 0. Since the threshold value is 0.9 or lower than 0.3, this embodiment preferably sets the migration intensity threshold to 0.9 to effectively capture overly perfect abnormal migration patterns. As an alternative or supplement, this threshold can also be set as an adjustable parameter, which can be adaptively adjusted according to the triggering frequency of the first abnormal signal during the iteration process: if the triggering ratio of the first abnormal signal is too high, for example, more than 30%, the threshold can be appropriately increased, and can be adjusted to 0.92 to avoid excessive punishment; if the triggering ratio is too low, for example, less than 5%, the threshold can be appropriately decreased, and can be adjusted to 0.88 to enhance the sensitivity of anomaly detection; the adjustment range is usually 0.85 to 0.95. Those skilled in the art can determine the appropriate threshold through simple experiments or cross-validation based on the characteristics of fake account behavior on a specific platform.
[0174] In this embodiment, SASRec is used as the sequence encoder for sequences of length 10 ... input behavior sequence The encoder outputs the corresponding interest vector sequence. .in, Represents the front of the sequence The aggregate representation of the items reflects the user's experience up to the [number]th [item]. Dynamic interests at each time step. Specifically:
[0175] Source domain interest vector Take the encoder at position Output This indicates that the user is the [number]th [unit] in the source domain sequence. Interest status at each time step;
[0176] Target domain interest vector Take the last position of the target domain sequence encoder The output serves as a global representation of the user's interest in the target domain;
[0177] Mixed domain interest vector Take the last position of the mixed-domain sequence encoder The output serves as a representation of the user's cross-domain global interest.
[0178] All of the encoders mentioned above use the same SASRec architecture and do not share parameters.
[0179] In this embodiment, the self-interest type is divided into three interest vector categories based on the user behavior sequence encoding results, and the specific definitions are as follows:
[0180] Target domain interest: Long-term interest representation of users on the target platform. It is obtained by inputting the user's historical behavior sequence in the target domain into a sequence encoder (such as SASRec) and taking the output vector of the last time step as the target domain interest vector. This vector reflects the user's core preferences in the target domain.
[0181] Transferable source domain interests: The portion of a user's interests in the source domain that is highly correlated with the interests in the target domain and can be transferred to the target domain is selected from the interest vectors at each time step of the source domain behavior sequence through an attention mechanism. Attention weights The vector is calculated based on the dot product of the source domain interest and the target domain interest. This vector represents the interest components in the source domain that can be inherited by the target domain.
[0182] Hybrid domain interest: The user's behavioral sequences in the source and target domains are concatenated in chronological order to form a hybrid domain sequence. This sequence is then input into the same encoder, and the output of the last time step is taken to obtain the hybrid domain interest vector. This vector reflects the evolution of users' overall cross-domain interests.
[0183] The above three interest types are based entirely on the output of the behavior sequence encoder, without the need for manual annotation or external knowledge. Among them, the "target domain interest" and "hybrid domain interest" are directly taken from the last frame of the encoder, and the "transferable source domain interest" is obtained through attention-weighted aggregation.
[0184] The SASRec sequence encoder can be trained using a self-supervised learning approach. It takes a sequence of user behavior as input and predicts the next behavior item as its objective. The model parameters are optimized using a negative log-likelihood loss function. The encoders in the source domain, target domain, and mixture domain are trained independently without sharing parameters. After training, the output of the encoder at the last time step is extracted as the interest vector. If there is insufficient training data, publicly available pre-trained behavior sequence models can be used for fine-tuning.
[0185] Specifically, the process of calculating the corrected confidence score based on the first and second anomalous signals includes:
[0186] The adjusted confidence score includes the first adjusted confidence score and the second adjusted confidence score;
[0187] When the second abnormal signal is received, the second corrected confidence score is calculated based on the second penalty value;
[0188] When the first abnormal signal is received, a first corrected confidence score is calculated based on the first penalty value.
[0189] In this embodiment, the first corrected confidence score corresponds to the penalty result of the first anomalous signal, and the second corrected confidence score corresponds to the penalty result of the second anomalous signal. The process of calculating the corrected confidence score based on the first and second anomalous signals is as follows:
[0190] When the first anomalous signal (abnormal cross-domain interest migration strength) is detected, the first penalty value is calculated:
[0191]
[0192] in , , The intensity of cross-domain interest migration;
[0193] When a second abnormal signal (abnormal source of influence) is detected, a second penalty value is calculated:
[0194]
[0195] in , Indicates account node The social authenticity feature value Fsoc;
[0196] If two abnormal signals are triggered simultaneously, the total penalty... If only one is triggered, the other penalty value is 0.
[0197] Corrected confidence score:
[0198]
[0199] Subsequently, the ownership of the account node is re-determined using the same thresholds (0.7 and 0.3) as in step S2.
[0200] In this embodiment, to avoid excessive penalty, the penalty intensity coefficient can vary with the number of iterations. Decrease (e.g., multiply by 0.9 per round):
[0201]
[0202] initial value , When no abnormal signal of a certain type is detected for two consecutive rounds, the corresponding coefficient is restored to its initial value.
[0203] For example, an account The original confidence score in the first iteration was (If the value is higher than 0.7, it should originally be included in the real node candidate set.) After matching and attribution analysis, the system detected the following: the cross-domain interest migration strength of this account. (Threshold 0.9) triggers the first abnormal signal; simultaneously, the source account node affected does not belong to the real node candidate set, triggering the second abnormal signal, and the social authenticity feature value of the account... (Too low), take the initial penalty coefficient: , Migration intensity exceeds threshold Calculate the penalty value:
[0204]
[0205]
[0206] Total Penalty .
[0207] Corrected confidence score:
[0208]
[0209] because Within the interval (0.3, 0.7), the account was moved to the candidate set of nodes to be detected, instead of the original candidate set of real nodes. This indicates that the anomalous signal suggests it may be a fake account and requires further verification. In the second iteration, the account was still detected with the same anomalous signal, but the penalty coefficient had been reduced to:
[0210]
[0211] If other parameters remain unchanged, then:
[0212]
[0213] Corrected score = They are still in the set to be tested. As the iteration continues, the penalty gradually decreases. If the authenticity of the account is confirmed by other means, it may eventually re-enter the real set.
[0214] In this embodiment, the source account node represents the original content publisher or disseminator that directly triggered the interaction in the interaction behavior determined to be driven by others. The specific tracing rules are as follows:
[0215] If the interaction type is "comment" or "like" on a piece of content, the source of influence is the original author's account. If the interaction type is "share," the source of influence is the original content publisher to whom the share is directed, not an intermediate sharer. If the interaction type is "participate in a topic / challenge," the top 3 content creators with the highest exposure under that topic are extracted, and the account with the most recent interaction time with the user and a social connection is taken as the source of influence (if none, the topic initiator is taken). The tracing process can be achieved based on the interaction metadata (such as in_reply_to_user_id, retweeted_user_id) provided by the social network platform.
[0216] The model parameters in this embodiment, such as the weights of logistic regression and the embedding dimension of SASRec, are only preferred examples. In actual applications, they can be appropriately adjusted according to the data scale and computing resources.
[0217] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A cross-domain user matching method based on interactive behavior, characterized in that, include: Obtain historical interaction data for each user's account nodes on social networking platforms; Based on the historical interaction data, the account nodes are initially classified to generate a candidate set of real nodes, a candidate set of fake nodes, and a candidate set of nodes to be detected. The account nodes in the real node candidate set are initially matched and attribution analyzed to distinguish whether their interaction behavior is driven by their own interests or the influence of others. For behaviors determined to be driven by their own interests, the user's interest transfer patterns between the source and target domains are analyzed based on their own interest types, and the cross-domain interest migration intensity is calculated. The self-interest types include target domain interests, transferable source domain interests, and hybrid domain interests; In response to detecting that the cross-domain interest migration intensity is greater than the migration intensity threshold, a first abnormal signal is generated; For behaviors determined to be driven by the influence of others, trace the source account node of the influence; In response to the detection that the affected source account node does not belong to the real node candidate set, a second abnormal signal is generated; Calculate the corrected confidence score based on the first and second anomalous signals; The account nodes are reclassified based on the corrected confidence score, the real node candidate set, the fake node candidate set and the node candidate set to be detected are updated, and the initial matching is performed again using the updated real node candidate set. The process of initially classifying account nodes based on the historical interaction behavior data includes: Calculate the user profile integrity score; The user profile integrity score is compared with the integrity score threshold to obtain the original confidence score based on the comparison result; The comparison results include a first comparison result and a second comparison result; Based on the original confidence scores, the account nodes are divided into a real node candidate set, a fake node candidate set, and a node candidate set to be detected; Compare the user profile integrity score with the integrity score threshold: If the user profile integrity score is greater than or equal to the integrity score threshold, the first comparison result is obtained, three types of features are extracted, and the raw confidence score is obtained by inputting the logistic regression classifier. The three types of features are behavioral regularity features, social authenticity features, and interaction depth features; If the user profile integrity score is less than the integrity score threshold, a second comparison result is obtained, and the original confidence score is calculated based on the associated account nodes.
2. The cross-domain user matching method based on interactive behavior according to claim 1, characterized in that, The process of extracting three types of features and inputting them into a logistic regression classifier to obtain the raw confidence scores includes: Calculate the regularity characteristic value of behavior based on the coefficient of variation of behavior interval, behavior concentration, and periodicity intensity; The social authenticity feature value is calculated based on the fan / follower ratio, local clustering coefficient, and isolated node ratio. The interaction depth feature value is calculated based on the interaction type entropy, average comment length, and content redundancy. The behavioral regularity feature value, social authenticity feature value, and interaction depth feature value are input into the logistic regression model to obtain the original confidence score.
3. The cross-domain user matching method based on interactive behavior according to claim 1, characterized in that, The process of calculating the original confidence score based on associated account nodes includes: Analyze the associated account nodes that are related to the account node to form associated account node pairs; The original confidence score is calculated based on the historical interaction data of the associated account node pairs to perform preliminary classification of the account nodes.
4. The cross-domain user matching method based on interactive behavior according to claim 3, characterized in that, The process of dividing the account nodes into a real node candidate set, a fake node candidate set, and a node candidate set to be detected based on the original confidence scores includes: Compare the original confidence score with the first confidence score threshold and the second confidence score threshold: If the original confidence score is less than or equal to the first confidence score threshold, the account node is classified into a fake node candidate set; If the original confidence score is between the first confidence score threshold and the second confidence score threshold, the account node is divided into a candidate set of nodes to be detected; If the original confidence score is greater than or equal to the second confidence score threshold, the account node is divided into a real node candidate set.
5. The cross-domain user matching method based on interactive behavior according to claim 1, characterized in that, The process of initial matching of account nodes in the real node candidate set includes: The account nodes in the real node candidate set are initially matched to form several matching account node pairs; Calculate the actual similarity between each pair of matching account nodes, and mark the account nodes with actual similarity higher than the similarity threshold as the target matching nodes.
6. The cross-domain user matching method based on interactive behavior according to claim 5, characterized in that, The process of performing attribution analysis on account nodes in the candidate set of real nodes includes: A global user profile is generated based on the target matching node, and the correct matching probability of account nodes in the global user profile is calculated. In response to the fact that the correct matching probability is less than a preset second threshold, attribution analysis is performed on the account nodes in the real node candidate set.
7. The cross-domain user matching method based on interactive behavior according to claim 1, characterized in that, In response to detecting that the cross-domain interest migration intensity is greater than a migration intensity threshold, a first abnormal signal is generated, specifically including: When the cross-domain interest migration intensity is greater than the migration intensity threshold, it is determined that the interest migration pattern of the account node is abnormal, and the first abnormal signal is generated. The first abnormal signal is used to indicate that the account node does not belong to the real node candidate set, or that the interaction behavior of the account node should be determined to be driven by the influence of others rather than by its own interests.
8. The cross-domain user matching method based on interactive behavior according to claim 1, characterized in that, The process of calculating the corrected confidence score based on the first and second anomalous signals includes: The adjusted confidence score includes the first adjusted confidence score and the second adjusted confidence score; When the second abnormal signal is received, the second corrected confidence score is calculated based on the second penalty value; When the first abnormal signal is received, a first corrected confidence score is calculated based on the first penalty value.
9. The cross-domain user matching method based on interactive behavior according to claim 1, characterized in that, The process of reclassifying account nodes based on the revised confidence score includes: Account nodes whose corrected confidence scores are greater than the second confidence score threshold will be moved into the real node candidate set; Account nodes whose corrected confidence scores are less than or equal to the first confidence score threshold are moved into the fake node candidate set; The remaining account nodes will remain in the candidate set of nodes to be detected.
Citation Information
Patent Citations
Unsupervised cross-domain user matching method based on embedding correction
CN119025934A
Cross-social network account matching method based on knowledge graph label propagation
CN121301950A
Cross-domain sequence recommendation method based on causal inference and preference evolution
CN121350366A
