A method for correcting the trend of the number of followers of an account

Abstract

The present invention proposes a method for correcting the trend of the number of account fans. First, multi-dimensional data related to the number of account fans is collected and the collected multi-dimensional data is pre-processed to ensure the accuracy and consistency of the data; the dependency relationship between account fans and interactive behaviors is constructed using a time graph network, and the trend of future fan count changes is predicted by combining time dependency and interactive behaviors, so as to more accurately capture the time series changes of the number of fans and the impact of account interactive behaviors on the number of fans; the attention mechanism is used to weight key events to further improve the accuracy of trend prediction and ensure that the model can fully consider the event time points that have a greater impact on the fluctuation of the number of fans. The present invention can significantly improve the accuracy of account fan count prediction by introducing technologies such as the collection and processing of multi-dimensional data, the construction of a time graph network, the weighting of the attention mechanism, and causal inference, while correcting the impact of abnormal fluctuations, and providing users with a more accurate trend of fan count changes.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis and processing, and more specifically, to a method for correcting a trend of the number of followers of an account. Background Art

[0002] In current short video platforms, the number of fans of KOL (Key Opinion Leader) is one of the important indicators for evaluating their influence. However, the number of fans is often disturbed by abnormal factors, such as the adjustment of platform algorithms, packet loss during data capture, network delays, or sudden data fluctuations caused by incorrect operations. These abnormal data will have a serious impact on the number of fans of KOL, and thus affect the actual commercial value evaluation of KOL.

[0003] In the prior art, these abnormal data are generally corrected by removing data points that are obviously beyond the expected range through a simple mean method or a 3-times standard deviation method. However, these traditional methods have obvious limitations, and they are prone to large errors when dealing with nonlinear fluctuations in the number of KOL fans. Especially when the number of fans fluctuates drastically in a short period of time, the fitting effect of these static models is poor and cannot accurately reflect the true trend of the number of fans. In addition, these methods usually do not adequately handle the time dependency between data, and cannot accurately predict and fill in missing values ​​during abnormal data, thus affecting the accuracy of the overall data. Summary of the invention

[0004] The present invention provides a method for correcting the trend of the number of fans of an account, which solves the technical problem that the static model in the prior art has a poor fitting effect and cannot accurately reflect the real trend of the number of fans.

[0005] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0006] The present invention provides a method for correcting the trend of the number of followers of an account, comprising the following steps:

[0007] Collect multi-dimensional data of the account on the social platform and pre-process the multi-dimensional data;

[0008] Constructing a time graph network according to the preprocessed multi-dimensional data, wherein each node in the time graph network represents a data dimension in the multi-dimensional data, and the edges in the time graph network represent the degree of influence between data of different dimensions;

[0009] Calculating the attention weight corresponding to each data dimension in the multi-dimensional data;

[0010] According to the time graph network and the attention weights corresponding to each data dimension, a preliminary result of the trend of the number of followers of the account is obtained;

[0011] The preliminary account follower number trend result is corrected using causal inference analysis and a federated learning model to obtain a corrected account follower number trend result.

[0012] Furthermore, the multi-dimensional data includes historical data on the number of fans, account work interaction data, account popular work data, account negative information, account hot search list data and external factor data.

[0013] Furthermore, the multi-dimensional data is preprocessed, including:

[0014] Cleaning the multi-dimensional data;

[0015] Align the cleaned multi-dimensional data in the time dimension;

[0016] The historical data of the number of fans in the aligned multi-dimensional data is monitored for abnormal data points, and the abnormal data points are corrected.

[0017] Furthermore, monitoring abnormal data points of the historical data of the number of fans in the aligned multi-dimensional data and correcting the abnormal data points include:

[0018] Calculate the rate of change of the number of fans:

[0019]

[0020] In the formula, R(t) represents the rate of change of the number of fans, F(t) represents the number of fans at time point t, and F(t-1) represents the number of fans at time point t-1;

[0021] Use unsupervised learning models to monitor the time series data of the change rate of the number of fans and obtain abnormal data points;

[0022] Correction of the anomalous data point:

[0023]

[0024] In the formula, F′(t) is the corrected number of fans at time point t, and F(t+1) is the number of fans at time point t+1.

[0025] Furthermore, preprocessing the multi-dimensional data further includes:

[0026] Perform entity recognition processing on the text in the account hot search list data to obtain the account entity and additional information, wherein the additional information includes the time when the account was listed, the topic and the ranking;

[0027] Use the BERT model to perform sentiment recognition and classification on the text in the interactive data of the account works to obtain sentiment classification and sentiment intensity.

[0028] Furthermore, the time graph network includes:

[0029] G=(V,E)

[0030] V={F′(t),I′(t),H′(t),N′(t),T i ,H i ,R i}

[0031] E={(i,j,w ij )}

[0032] In the formula, G represents the time graph network, V is the node set, which represents the preprocessed multi-dimensional data, and E is the edge set, which represents the dependency relationship between nodes; I′(t) represents the preprocessed account work interaction data, H′(t) represents the preprocessed account popular work data, N′(t) represents the preprocessed account negative information, and T i represents the time information obtained by entity recognition processing, H i represents the additional information obtained by entity recognition processing, R i Represents sentiment classification S i and emotional intensity W i ;(i,j,w ij ) represents the weight w of the edge between node i and node j ij .

[0033] Further, calculating the attention weight corresponding to each data dimension in the multi-dimensional data includes:

[0034]

[0035] e i =w T ·T i +w H ·H i +w S ·S i +w W ·W i

[0036] In the formula, a i represents the attention weight of the i-th dimension data in the multi-dimensional data, e i is the weighted sum of the i-th dimension data in the multidimensional data, w T 、w H 、w S and w W are the time information obtained by entity recognition processing, the additional information obtained by entity recognition processing, the weight coefficients of sentiment classification and sentiment intensity, and e jrepresents the weighted sum of the j-th dimension data in the multi-dimensional data, and n represents the number of dimensions of the multi-dimensional data.

[0037] Furthermore, according to the time graph network and the attention weight corresponding to each data dimension, preliminary account follower number trend results are obtained, including:

[0038]

[0039]

[0040]

[0041]

[0042]

[0043] In the formula, g(V i ,W ij ) is a convolution-based feature aggregation function, N(i) represents the neighbor nodes of node i

[0044] Point, c ij is the normalization coefficient, deg() represents the degree of the node, that is, the number of edges connected to node i, represents the number of fans at the initial prediction time point t+1, W ij (t) represents the weight of the edge between node i and node j at time point t, I ij (t) represents the number of interactions between node i and node j at time point t, represents the total number of interactions between node i and all nodes at time point t, α t represents the weight at time point t, F predicted (t) represents the predicted number of fans at time point t+1.

[0045] Furthermore, the preliminary account follower number trend result is corrected using causal inference analysis and a federated learning model to obtain a corrected account follower number trend result, including:

[0046] The causal inference analysis is as follows:

[0047] F(t+1)=β1·F(t)+β2·P′(t)+ε

[0048] In the formula, β1 represents the influence weight of the number of fans at time point t, β2 represents the influence weight of external factors, P′(t) represents the external factor data in the preprocessed multidimensional data, and ε represents the error;

[0049] Calculate β2 using historical data as follows:

[0050]

[0052] In the formula, P and F represent the historical average values ​​of external factor data and the number of fans, respectively;

[0053] The prediction formula corrected by the causal inference is:

[0054] F local (t+1)=F predicted (t+1)+β2·P′(t)

[0055] In the formula, F local (t+1) represents the number of fans at time point t+1 predicted after causal inference correction;

[0056] The federated learning model is as follows:

[0057]

[0058] In the formula, F global (t+1) is the number of global fans at time point t+1 predicted by federated learning aggregation, represents the predicted number of fans at time point t+1 on the ith platform after causal inference correction, N represents the number of platforms, represents the fan growth rate on the ith platform, represents the interactive activity on the ith platform, which is the ratio of the total amount of interactions to the number of fans. represents the impact factor of popular works on the ith platform, which is calculated by weighting according to the interaction type. represents the weight of negative events on the ith platform, taking the mean of (negative comments - positive comments) / (positive comments + negative comments) of the corresponding platform posts in the time period, ω (i) represents the platform characteristic weight on the ith platform, which is calculated based on the user scale, number of active users, content format, etc. of different platforms;

[0059] The final prediction result of combining causal inference analysis and federated learning model is:

[0060]

[0061] In the formula, is the final predicted number of global fans at time point t+1 on the i-th platform, and α represents the adjustment parameter.

[0062] Furthermore, it also includes:

[0063] Compare the corrected result of the number of followers of the account with the actual number of followers and calculate the error;

[0064] Based on the errors, the model parameters are iteratively optimized.

[0065] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0066] The present invention proposes a method for correcting the trend of the number of account fans. First, multi-dimensional data related to the number of account fans is collected and pre-processed to ensure the accuracy and consistency of the data, so as to provide high-quality data input for subsequent analysis; the dependency relationship between account fans and interactive behaviors is constructed by using a time graph network (TGN), and the trend of future fan number changes is predicted by combining time dependency and interactive behaviors. Through the time graph network, the time series changes of the number of fans and the influence of KOL interactive behaviors on the number of fans can be captured more accurately; the key events are weighted by using an attention mechanism to further improve the accuracy of trend prediction, so as to ensure that the model can fully consider the event time points that have a greater impact on the fluctuation of the number of fans. The present invention can significantly improve the accuracy of the prediction of the number of account fans by introducing technologies such as the collection and processing of multi-dimensional data, the construction of a time graph network, the weighting of an attention mechanism, and causal inference, while correcting the influence of abnormal fluctuations, and providing users with a more accurate trend of the number of fans. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A flowchart of a method for correcting a trend of the number of followers of an account provided by an embodiment of the present invention;

[0068] Figure 2 A schematic diagram comparing the original trend and the revised trend of the number of fans provided in an embodiment of the present invention;

[0069] Figure 3 The embodiment of the present invention provides Figure 2 A magnified schematic diagram of the details. DETAILED DESCRIPTION

[0070] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0071] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;

[0072] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0073] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0074] Example 1

[0075] This implementation provides a method for correcting the trend of the number of followers of an account, such as Figure 1 As shown, the following steps are included:

[0076] Collect multi-dimensional data of the account on the social platform and pre-process the multi-dimensional data;

[0077] Constructing a time graph network according to the preprocessed multi-dimensional data, wherein each node in the time graph network represents a data dimension in the multi-dimensional data, and the edges in the time graph network represent the degree of influence between data of different dimensions;

[0078] Calculating the attention weight corresponding to each data dimension in the multi-dimensional data;

[0079] According to the time graph network and the attention weights corresponding to each data dimension, a preliminary result of the trend of the number of followers of the account is obtained;

[0080] The preliminary account follower number trend result is corrected using causal inference analysis and a federated learning model to obtain a corrected account follower number trend result.

[0081] The embodiments of the present invention provide a solution for accurately predicting the changing trend of the number of fans of an account by comprehensively collecting multi-dimensional data of the account, combining technical means such as data preprocessing, abnormal fluctuation detection, trend prediction and correction, etc., to solve the problems of data anomaly and insufficient accuracy in the current trend analysis of the number of fans of an account, improve the accuracy of data fitting and the reliability of prediction, thereby more accurately reflecting the real changing trend of the number of KOL fans and providing more powerful support for marketing decisions.

[0082] Example 2

[0083] This embodiment further discloses the following contents based on the embodiment 1:

[0084] The multi-dimensional data includes historical data on the number of fans, account work interaction data, account popular work data, account negative information, account hot search list data and external factor data.

[0085] In a specific embodiment, the historical data of the number of fans records the time series changes of the number of fans of an account, reflecting the overall trend of the number of fans over time;

[0086] Account work interaction data includes fan interaction data after the account publishes content, such as likes, comments, and reposts. This data reflects the intensity of interaction between the account's works and fans, and directly affects the fluctuation of the number of fans;

[0087] The account's popular works data identifies the high-interaction works released by the account, which can explain the rapid growth or decline of the number of fans in a certain period of time;

[0088] Negative account information includes negative news or public opinion events involving the account, which usually lead to a significant drop in the number of followers;

[0089] The account hot search list data records the time, topic and ranking of the account on the hot search list. Hot search list events have a significant impact on the fluctuation of the number of fans in the short term.

[0090] External factor data refers to adjustments to the platform's recommendation algorithm, changes in the content distribution mechanism, and changes in the review mechanism. These factors will lead to fan loss and affect short-term fluctuations in the number of fans.

[0091] In a further embodiment, preprocessing the multi-dimensional data includes:

[0092] Cleaning the multi-dimensional data;

[0093] In a specific embodiment, cleaning the multi-dimensional data includes removing duplicate records, null values, and data with inconsistent formats to ensure data integrity and consistency. D′=Clean(D) where D represents the set of all collected data sources, including historical data on the number of fans, interactive data, popular works, negative events, and hot search data, and D′ is the cleaned data set to ensure high data quality without noise and redundant information;

[0094] Align the cleaned multi-dimensional data in the time dimension;

[0095] In a specific embodiment, the timestamps of different data sources may be inconsistent, so all data are aligned in the time dimension to ensure that the number of fans, interactive data, and hot search list data have a unified time reference. Formula: D″=AlignTime(D′) Where D′ is the cleaned data, and D′′ is the time-aligned data set, ensuring that the data is synchronized on the same timeline, which is suitable for time series analysis;

[0096] Monitor the historical data of the number of fans in the aligned multi-dimensional data for abnormal data points, and correct the abnormal data points;

[0097] In a specific embodiment, when analyzing the change in the number of followers of an account, abnormal fluctuations in the number of followers (such as sudden growth or decline) may hide important trend information. Therefore, it is necessary to identify these fluctuations by calculating the rate of change of the number of followers. The rate of change calculation is used to capture the relative change of the number of followers in the time series to determine whether there is an abnormal fluctuation.

[0098] In a further embodiment, monitoring the historical data of the number of fans in the aligned multi-dimensional data for abnormal data points and correcting the abnormal data points includes:

[0099] Calculate the rate of change of the number of fans:

[0100]

[0101] Wherein, R(t) represents the change rate of the number of fans, F(t) represents the number of fans at time point t, and F(t - 1) represents the number of fans at time point t - 1;

[0102] Use an unsupervised learning model (such as the Isolation Forest algorithm) to monitor the time series data of the change rate of the number of fans and obtain abnormal data points. The purpose of this step is to automatically identify abnormal points in the change of the number of fans, such as a large number of fans lost due to negative events or a sharp increase in the number of fans due to major exposure. This step can filter out normal fluctuations and focus on significantly abnormal changes in the number of fans;

[0103] After detecting abnormal points, there may be situations of discontinuous data or drastic fluctuations. Therefore, these abnormal points need to be corrected. The correction method usually adopts the linear interpolation method to smooth the fan number data of the abnormal points to ensure the continuity of the time series data. The corrected data will be used for subsequent trend prediction to avoid excessive interference of abnormal points on the prediction results. The formula for correcting the abnormal data points is as follows:

[0104]

[0105] Wherein, F′(t) is the number of fans at the corrected time point t, and F(t + 1) represents the number of fans at time point t + 1. This step ensures the smoothness and continuity of the time series.

[0106] In a further embodiment, the preprocessing of the multi-dimensional data further includes:

[0107] Perform entity recognition processing on the text in the account hot search list data to obtain account entities and additional information. The additional information includes the account's time on the list, topics, and rankings;

[0108] In a specific embodiment, the processing of the text in the account hot search list includes word segmentation, removing stop words (such as "de", "shi") and special symbols, etc. Text preprocessing can convert complex natural language into an input form suitable for natural language processing models (such as BERT). Formula: T′ = Preprocess(T), where T represents the original hot search list text data, and T′ is the preprocessed text data, which is suitable for the entity recognition model;

[0109] Use the named entity recognition (NER) technology based on the BERT model to identify account entities from the preprocessed text, such as account names, related topics, key events, etc. The BERT model can accurately extract key information by capturing the context semantics in the text. Formula: E i = NER(t i ′), where t i′ represents a single piece of text data after preprocessing, E i Represents a set of entities extracted from the text, including KOL names, topics, etc.

[0110] By matching the identified entities with the known account list, the most likely account entities are screened out. Using the confidence score, the extracted entities are ranked and the most credible results are screened out. Formula: E′ i ={e|e∈E i ∩K}, where K represents a list of known account entities, E i ' is the set of account entities after matching, which ensures the accuracy of recognition results;

[0111] In order to fully understand the performance of an account, in addition to identifying the account entity, it is also necessary to extract additional information, including the time, topic, and ranking of the account. This information will be helpful for subsequent analysis of the number of fans, especially for explaining the key moments of fan growth or loss. The additional information extraction is expressed as: T i ,H i =ExtractTimeTopic(t′ i ) Note: Among them, T i represents the time information extracted from the text, H i Indicates the topic content. This information can provide richer context support for subsequent sentiment analysis and anomaly detection;

[0112] Use the BERT model to perform sentiment recognition and classification on the text in the interactive data of the account works to obtain sentiment classification and sentiment intensity;

[0113] In a specific embodiment, the comments of fans contain a lot of emotional information, so the comment data needs to be preprocessed. This operation includes word segmentation, removal of stop words and invalid symbols, and ensuring that the comment text is suitable for sentiment analysis models (such as BERT). The preprocessed comment text is represented as: C′=Preprocess(C), C is the original comment data, C′ is the preprocessed comment text, which removes irrelevant information and makes the text structure more suitable for sentiment analysis by the BERT model;

[0114] The BERT model is used to classify the sentiment of fan comments and identify whether the comments are positive, negative, or neutral. BERT can accurately capture the sentiment in comments through its understanding of context, which is an important step in analyzing the changes in fan sentiment. The sentiment classification result is expressed as: S i =BERT(C′ i )S i Indicates fan comments C i ′’s sentiment classification results. Through the BERT model, fan comments are classified into positive, negative or neutral;

[0115] In addition to identifying the emotion category, it is also necessary to further quantify the intensity of the emotion. By weighted analysis of the emotion keywords, the amplitude of the fans' emotion fluctuation can be evaluated. The emotion intensity can reflect the degree of influence of the comments on the fans' emotion changes. The result of the emotion intensity analysis is expressed as: W i =Weight(S i ), emotional intensity W i is the classification result S i The results of the weighted analysis reflect the strength of the fans' comments. This step provides quantitative support for subsequent analysis;

[0116] Combine the sentiment classification and sentiment intensity data with the timestamp to output complete sentiment fluctuation data for subsequent anomaly detection and fan number trend analysis. These sentiment data can explain the sentiment driving factors behind the fan number fluctuation. The sentiment data output is expressed as: i ={S i ,W i ,T i},R i is a set of sentiment data, including sentiment classification S i , emotional intensity W i , and timestamp t i These data provide an emotional basis for subsequent anomaly detection and fan trend prediction.

[0117] In a further embodiment, in the fan number trend prediction, multi-dimensional data input is first combined, and the dependency relationship between the fan number change and the interactive behavior is constructed through a time graph network, and the time graph network includes:

[0118] G=(V,E)

[0119] V={F′(t),I′(t),H′(t),N′(t),T i ,H i ,R i}

[0120] E={(i,j,w ij )}

[0121] In the formula, G represents the time graph network, V is the node set, which represents the preprocessed multi-dimensional data, and E is the edge set, which represents the dependency relationship between nodes; I′(t) represents the preprocessed account work interaction data, H′(t) represents the preprocessed account popular work data, N′(t) represents the preprocessed account negative information, and T i represents the time information obtained by entity recognition processing, H i represents the additional information obtained by entity recognition processing, R iRepresents sentiment classification S i and emotional intensity W i ;(i,j,w ij ) represents the weight w of the edge between node i and node j ij .

[0122] In a further embodiment, calculating the attention weight corresponding to each data dimension in the multi-dimensional data includes:

[0123]

[0124] e i =w T ·T i +w H ·H i +w S ·S i +w W ·W i

[0125] In the formula, a i represents the attention weight of the i-th dimension data in the multi-dimensional data, e i is the weighted sum of the i-th dimension data in the multidimensional data, w T 、w H 、w S and w W are the time information obtained by entity recognition processing, the additional information obtained by entity recognition processing, the weight coefficients of sentiment classification and sentiment intensity, and e j represents the weighted sum of the j-th dimension data in the multi-dimensional data, and n represents the number of dimensions of the multi-dimensional data.

[0126] In a specific embodiment, the attention weighting mechanism is used to weight various factors of the change in the number of fans (such as historical data, interaction, emotional fluctuations, hot events, etc.) to determine their importance at different time points. This can help capture the short-term or long-term impact of key events and emotional fluctuations, enhance the accuracy of predictions, and measure the contribution of each feature to the prediction of the number of fans at the current moment by generating a similarity measure of the features.

[0127] In a further embodiment, according to the time graph network and the attention weight corresponding to each data dimension, a preliminary account follower number trend result is obtained, including:

[0128]

[0129]

[0130]

[0131]

[0132]

[0133] In the formula, g(V i ,W ij ) is a convolution-based feature aggregation function, N(i) represents the neighbor nodes of node i, c ij is the normalization coefficient, deg() represents the degree of the node, that is, the number of edges connected to node i, represents the number of fans at the initial prediction time point t+1, W ij (t) represents the weight of the edge between node i and node j at time point t, I ij (t) represents the number of interactions between node i and node j at time point t, represents the total number of interactions between node i and all nodes at time point t, α t represents the weight at time point t, F predicted (t) represents the predicted number of fans at time point t+1.

[0134] In a specific embodiment, the above formula represents that in the time graph network TGN, the feature V i The temporal dependence of u , dynamically adjust the contribution of each feature while capturing the complex dependencies in time series data, thereby affecting the final impact on the trend of the number of fans.

[0135] In a further embodiment, since an account may experience abnormal fluctuations due to external factors such as changes in platform algorithms, the causal inference model analyzes the causal relationship of external interference factors and analyzes these abnormal fluctuations to make the prediction results more accurate. The preliminary account follower number trend results are corrected using causal inference analysis and a federated learning model to obtain the account follower number trend correction results, including:

[0136] The causal inference analysis is as follows:

[0137] F(t+1)=β1·F(t)+β2·P′(t)+ε

[0138] In the formula, β1 represents the influence weight of the number of fans at time point t, β2 represents the influence weight of external factors, P′(t) represents the external factor data in the preprocessed multidimensional data, and ε represents the error;

[0139] Calculate β2 using historical data as follows:

[0140]

[0141] In the formula, and They represent the historical average values ​​of external factor data and number of fans respectively;

[0142] The prediction formula corrected by the causal inference is:

[0143] F local (t+1)=f predicted (t+1)+β2·P′(t)

[0144] In the formula, F local (t+1) represents the number of fans at the time point t+1 predicted after causal inference correction. Through causal analysis, we can determine the impact of external factors P(t) on the number of fans, and use the weight β2 to correct the prediction result;

[0145] The main function of the federated learning model is to integrate data features from different platforms to ensure the consistency and accuracy of the fan number prediction results across platforms. The number of fans, interactive behaviors, popular content, and negative events on each platform may differ. Federated learning aggregates the model parameters of each platform to eliminate the deviation between platforms, making the fan number prediction model more widely applicable and generalizable. The federated learning model is as follows:

[0146]

[0147] In the formula, F global (t+1) is the number of global fans at time point t+1 predicted by federated learning aggregation, represents the predicted number of fans at time point t+1 on the ith platform after causal inference correction, N represents the number of platforms, represents the fan growth rate on the ith platform, specifically (number of fans in this period - number of fans in the previous period) / number of fans in the previous period x 100%; represents the interactive activity on the ith platform, which is the ratio of the total amount of interactions to the number of fans. It represents the impact factor of popular works on the ith platform, which is calculated by weighting according to the interaction type (number of views, number of likes, number of comments, etc.). represents the weight of negative events on the ith platform, taking the mean of (negative comments - positive comments) / (positive comments + negative comments) of the corresponding platform posts in the time period, ω (i) represents the platform characteristic weight on the ith platform, which is calculated based on the user scale, number of active users, content format, etc. of different platforms;

[0148] The federated learning model not only generates global prediction results, but also provides a basis for personalized prediction for each platform. Can be combined with the global model F global (t+1), generate the final prediction result of the platform;

[0149] The final prediction result of combining causal inference analysis and federated learning model is:

[0150]

[0151] In the formula, is the final predicted number of global fans at time point t+1 on the ith platform, and α represents the adjustment parameter;

[0152] By combining the global predictions generated by federated learning, the personalized predictions of local models, and the corrections of causal inference, personalized prediction results for each platform are finally generated.

[0153] Example 3

[0154] This embodiment further discloses the following contents based on Embodiment 1 and Embodiment 2:

[0155] Compare the corrected result of the number of followers of the account with the actual number of followers and calculate the error;

[0156] In a specific embodiment, after the number of fans is predicted, the model will evaluate the accuracy of the prediction through an error feedback mechanism. The model will compare the final prediction result with the actual number of fans, calculate the error E(t), and provide a basis for subsequent parameter adjustments.

[0157] The error feedback formula is as follows: E(t) = F actual (t)-F final (t), where F actual (t): The actual number of fans at time point t. final (t): The final prediction result at time point t (after correction). E(t): Error feedback, used to measure the prediction deviation of the model.

[0158] By analyzing the errors at different time points, the model can be helped to identify its shortcomings in specific scenarios or time periods, which can serve as a basis for iterative optimization.

[0159] iteratively optimizing model parameters according to the error;

[0160] Based on the error feedback results, the model adjusts parameters and performs iterative optimization. This process dynamically adjusts key parameters (such as the anomaly detection threshold δ outher , time weight ω(t) and external interference weight β2), to ensure that the model can adapt to different KOL data scenarios, gradually reduce errors and improve prediction accuracy.

[0161] The specific adjustment formula is: Among them: θ: represents the set of all parameters to be optimized, including v outher, ω(t), β2. E(t): prediction error at the current time point t. η: adjust the step size (learning rate) to control the update speed of the parameters. It represents the gradient of the prediction error with respect to the parameter θ, indicating the sensitivity of each parameter to the error.

[0162] Model optimization is an ongoing process. After each parameter adjustment, the model needs to undergo multiple rounds of iterative optimization to ensure that the performance in different data sets and scenarios can meet expectations. In each iteration, the model will be gradually adjusted based on the feedback results until the error is reduced to an acceptable range. Iterative optimization can help the model maintain efficient prediction capabilities when dealing with a variety of complex scenarios.

[0163] The iterative optimization formula is as follows: in, The model parameter set after the (n+1)th round of iterative optimization. The set of model parameters after the nth round of iterative optimization, including v outher , ω(t), β2 and other key parameters. η: learning rate, which controls the amplitude of model parameter update. represents the gradient of the prediction error of the nth round parameter set. E(t): prediction error at time t, through the formula E(t) = F actual (t)-F final (t) is calculated. represents the total prediction error at all times t.

[0164] After the iterative optimization is completed, the optimized model parameters and prediction results are used as output for subsequent practical applications or further testing. The output of this step can provide higher prediction accuracy and robustness for new data inputs, while recording the key adjustment points in the optimization process to provide a basis for future model updates.

[0165] The optimization result output formula is as follows: Among them: F finaloptimized (t+1): The final optimized fan number prediction result, which combines the global model prediction, local model prediction and causal inference correction, represents the predicted value at time t+1. The final model parameters after multiple rounds of iterations include δ outher ,ω(t),β2 and other key parameters.

[0166] In a specific embodiment, Figure 2 As shown in the figure, the original trend (red) and the corrected trend (blue) of the number of fans in a certain period of time are shown. The figure shows the difference between the trend before and after the data correction. The corrected trend is smoother and eliminates obvious abnormal fluctuations.

[0167] like Figure 3As shown, the red line represents the abnormal jumps and fluctuations in the original data, and the blue line represents the data corrected by the method of the embodiment of the present invention. The corrected trend shows a more reasonable and stable growth.

[0168] The same or similar reference numerals correspond to the same or similar components;

[0169] The terms used in the drawings to describe positional relationships are only used for illustrative purposes and should not be construed as limiting this patent;

[0170] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for correcting the trend of the number of followers of an account, characterized in that: The following steps are involved: Collecting multi-dimensional data of the account on the social platform and pre-processing the multi-dimensional data; Constructing a time graph network according to the preprocessed multi-dimensional data, wherein each node in the time graph network represents a data dimension in the multi-dimensional data, and the edges in the time graph network represent the degree of influence between data of different dimensions; Calculating the attention weight corresponding to each data dimension in the multi-dimensional data; According to the time graph network and the attention weights corresponding to each data dimension, a preliminary result of the trend of the number of followers of the account is obtained; The preliminary account follower number trend result is corrected using causal inference analysis and a federated learning model to obtain a corrected account follower number trend result.

2. The method for correcting the trend of the number of followers of an account according to claim 1, characterized in that: The multi-dimensional data includes historical data on the number of fans, account work interaction data, account popular work data, account negative information, account hot search list data and external factor data.

3. The method for correcting the trend of the number of followers of an account according to claim 2, characterized in that: Preprocessing the multi-dimensional data includes: Cleaning the multi-dimensional data; Align the cleaned multi-dimensional data in the time dimension; The historical data of the number of fans in the aligned multi-dimensional data is monitored for abnormal data points, and the abnormal data points are corrected.

4. The method for correcting the trend of the number of followers of an account according to claim 4, characterized in that: Monitoring abnormal data points of the historical data of the number of fans in the aligned multi-dimensional data and correcting the abnormal data points include: Calculate the rate of change of the number of fans: In the formula, R(t) represents the rate of change of the number of fans, F(t) represents the number of fans at time point t, and F(t-1) represents the number of fans at time point t-1; Use unsupervised learning models to monitor the time series data of the change rate of the number of fans and obtain abnormal data points; Correction of the anomalous data point: In the formula, F ′ (t) is the corrected number of fans at time point t, and F(t+1) is the number of fans at time point t+1.

5. The method for correcting the trend of the number of followers of an account according to claim 4, characterized in that: Preprocessing the multi-dimensional data further includes: Perform entity recognition processing on the text in the account hot search list data to obtain the account entity and additional information, wherein the additional information includes the time when the account was listed, the topic and the ranking; Use the BERT model to perform sentiment recognition and classification on the text in the interactive data of the account works to obtain sentiment classification and sentiment intensity.

6. The method for correcting the trend of the number of followers of an account according to claim 5, characterized in that: The time graph network comprises: G=(V,E) v={F ′ (t),I ′ (t),H ′ (t),N ′ (t),T i ,H i ,R i } E={(i,j,w ij )} In the formula, G represents the time graph network, V is the node set, which represents the preprocessed multi-dimensional data, and E is the edge set, which represents the dependency relationship between nodes; I ′ (t) represents the pre-processed account work interaction data, H ′ (t) represents the pre-processed account popular works data, N ′ (t) represents the negative account information after preprocessing, T i represents the time information obtained by entity recognition processing, H i represents the additional information obtained by entity recognition processing, R i Represents sentiment classification S i and emotional intensity H i ;(i,j,w ij ) represents the weight w of the edge between node i and node j ij .

7. The method for correcting the trend of the number of followers of an account according to claim 6, characterized in that: Calculating the attention weight corresponding to each data dimension in the multi-dimensional data includes: e i =w T ·T i +w H ·T i +w S ·S i +w W ·W i In the formula, a i represents the attention weight of the i-th dimension data in the multi-dimensional data, e i is the weighted sum of the i-th dimension data in the multidimensional data, w T 、w H 、w S and w W are the time information obtained by entity recognition processing, the additional information obtained by entity recognition processing, the weight coefficients of sentiment classification and sentiment intensity, and e j represents the weighted sum of the j-th dimension data in the multi-dimensional data, and n represents the number of dimensions of the multi-dimensional data.

8. The method for correcting the trend of the number of followers of an account according to claim 7, characterized in that: According to the time graph network and the attention weights corresponding to each data dimension, preliminary account follower number trend results are obtained, including: In the formula, g(V i ,W ij ) is a convolution-based feature aggregation function, N(i) represents the neighbor nodes of node i, c ij is the normalization coefficient, deg() represents the degree of the node, that is, the number of edges connected to node i, represents the number of fans at the initial prediction time point t+1, W ij (t) represents the weight of the edge between node i and node j at time point t, I ij (t) represents the number of interactions between node i and node j at time point t, represents the total number of interactions between node i and all nodes at time point t, α t represents the weight at time point t, F predicted (t) represents the predicted number of fans at time point t+1.

9. The method for correcting the trend of the number of followers of an account according to claim 8, characterized in that: The preliminary account follower number trend result is corrected using causal inference analysis and a federated learning model to obtain a corrected account follower number trend result, including: The causal inference analysis is as follows: F(t+1)=β1·F(t)+β2·P′(t)+ε In the formula, β1 represents the influence weight of the number of fans at time point t, β2 represents the influence weight of external factors, P′(t) represents the external factor data in the preprocessed multidimensional data, and ε represents the error; Calculate β2 using historical data as follows: In the formula, and They represent the historical average values ​​of external factor data and number of fans respectively; The prediction formula corrected by the causal inference is: F local (t+1)=F predicted (t+1)+β2·P′(t) In the formula, F local (t+1) represents the number of fans at time point t+1 predicted after causal inference correction; The federated learning model is as follows: In the formula, F global (t+1) is the number of global fans at time point t+1 predicted by federated learning aggregation, represents the predicted number of fans at time point t+1 on the ith platform after causal inference correction, N represents the number of platforms, represents the fan growth rate on the ith platform, represents the interactive activity on the ith platform, which is the ratio of the total amount of interactions to the number of fans. represents the impact factor of popular works on the ith platform, which is calculated by weighting according to the interaction type. represents the weight of negative events on the ith platform, taking the mean of (negative comments - positive comments) / (positive comments + negative comments) of the corresponding platform posts in the time period, ω (i) represents the platform characteristic weight on the ith platform, which is calculated based on the user scale, number of active users, content format, etc. of different platforms; The final prediction result of combining causal inference analysis and federated learning model is: In the formula, is the final predicted number of global fans at time point t+1 on the i-th platform, and α represents the adjustment parameter.

10. The method for correcting the trend of the number of followers of an account according to any one of claims 1 to 9, characterized in that: Also includes: Compare the corrected result of the number of followers of the account with the actual number of followers and calculate the error; Based on the errors, the model parameters are iteratively optimized.

Images