Social media content hotness prediction method fusing multi-scale cascade and time sequence features

Abstract

The application discloses a social media content heat prediction method fusing multi-scale cascades and time sequence features, comprising the following steps: using early propagation features of social media content, processing different size time windows to obtain multi-scale cascade graphs and heat trend sequences respectively; applying a graph convolutional neural network to embed the multi-scale cascade graph sequence and using a one-dimensional convolution to extract multi-scale trend short-term dynamic features, extracting multi-scale cascade graph and trend dynamic features through different Transformer Encoders; using a time window selection network to realize multi-scale feature fusion: after the network model is trained, inputting early multi-scale cascade graphs and heat trend sequences of social media to be predicted, and outputting social media content heat growth prediction values in a future period of time. Through effective fusion of multi-scale cascade graph features and trend features, the application has better prediction effect compared with a single scale method, and can obtain more optimal global prediction results and better generalization performance.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of social media content heat prediction, and particularly relates to a social media content heat prediction method fusing multi-scale cascade and trend time sequence features. BACKGROUND

[0002] The social media heat prediction task is quite widely applied in many fields including advertisement, recommendation system, information retrieval, etc. The existing heat prediction models generally utilize the cascade graph features and trend change features used by the social media content in the early propagation process. The main methods used include traditional machine learning methods, deep learning methods, point process methods and graph neural network methods, which have become the research focus. The above methods consider the dynamics in the social media propagation, and improve the accuracy of the microblog heat prediction to a higher position. However, when the existing methods study the dynamics of information propagation, a single scale sequence is often used, so part of the information of other scales is lost. When a relatively large scale is used, the overall trend of the heat trend can be grasped, but the local heat fluctuation change information in a small time period is ignored. If the multi-scale dynamic features can be combined and the time window selection network is used to fully and effectively fuse them, the heat prediction accuracy will be improved. The gating method is an effective method for fusing different features. SUMMARY

[0003] In order to effectively use the multi-scale features ignored by the existing technology, different size time windows are needed for data preprocessing, and the preprocessed multi-scale data needs a suitable method for modeling, extracting related multi-scale features, and an effective fusion strategy is needed for the multi-scale features, so as to complete the final heat growth prediction output.

[0004] In order to solve the above technical problems, the application provides a social media content heat prediction method fusing multi-scale cascade and time sequence features, and the steps are as follows:

[0005] Step 1, data preprocessing: using the early propagation features of the social media content, different size scale time windows are used to process to obtain multi-scale cascade graph sequences and multi-scale heat trend sequences respectively;

[0006] Step 2, multi-scale cascade graph dynamic feature modeling: the graph in the multi-scale cascade graph sequence obtained in step 1 is converted into an adjacency matrix sequence, the Laplacian matrix of the adjacency matrix is calculated, and the initial feature matrix of the nodes in each subgraph in the subgraph sequence is obtained through the PageRank algorithm; for the multi-scale cascade graph sequence of different scales, the multi-scale cascade graph sequence is embedded by applying a graph convolutional neural network, the Laplacian matrix and the initial feature matrix are input, and the node features are obtained; the dynamic correlation between the multi-scale cascade graph sequences is established, and the multi-scale cascade graph dynamic features are extracted by a Transformer Encoder;

[0007] Step 3, multi-scale trend dynamic feature modeling: the multi-scale trend sequence obtained in step 1 is used to extract multi-scale trend short-term dynamic features by one-dimensional convolution, and long-term dynamic change features, i.e., multi-scale trend dynamic features, are extracted by a Transformer Encoder;

[0008] Step 4, multi-scale feature fusion by using a time window selection network: the multi-scale cascade graph dynamic features and the multi-scale trend dynamic features are filtered, the filtered multi-scale cascade graph dynamic features and the multi-scale trend dynamic features are spliced, and then the multi-scale fusion features are obtained by using the time window selection network;

[0009] Step 5, training of the network model and output of the heat prediction result: the multi-scale fusion features are input into a full connection layer, and the output of the full connection layer is a heat growth prediction value; in the network model training process, the MSLE loss function is used in a back propagation mode to optimize the weights of each layer, and the feedback result is output to the input end of each layer of the entire network model; after the network model is trained, the early multi-scale cascade graph sequence and the multi-scale heat trend sequence of the microblog to be predicted are input, and the output of the network model is a heat growth prediction value in a future period of time, so that the heat prediction of the social media content is realized.

[0010] Further, the social media content heat prediction method provided by the application has the following characteristics:

[0011] The content of step 1 is: the early propagation features of the social media content include a user forwarding cascade graph and a heat change value in each time period; for the propagation data in the first T time periods of the social media content, K different scale time windows are given, corresponding to K different scales S=[s1, s2...s k ], and the obtained multi-scale sequence is P=[p t 1 ,p t 2 ...p t N], where t∈S represents one of the scales, N=T / t, N represents the length of the sequence; the multi-scale cascade graph sequence is G=[g t 1 ,g t 2 ...g t N ], the multi-scale heat trend sequence is F=[f t 1 ,f t 2 ...f t N ].

[0012] In step 2, comprising:

[0013] 2-1) for the multi-scale cascade graph sequence G=[g t 1 ,g t 2 ...g t N ], first, the cascade graph sequence G is converted into the adjacency matrix sequence A=[a t 1 ,a t 2 ...a t N ], then the Laplacian matrix of the adjacency matrix is calculated At the same time, the PageRank algorithm is applied to each subgraph in the adjacency matrix sequence to obtain the initial feature matrix H 0 =[h t 1 ,h t 2 ...h t N ] of the nodes;

[0014] 2-2) the final node feature H (l+1) is obtained by formula (1), then the node feature is represented into the graph feature by the readout function, that is, the average of the sum of the features of each node; the multi-scale cascade graph feature sequence is obtained;

[0015]

[0016] In formula (1), H (l+1) is the node feature of the next layer, σ is the activation function, is the Laplacian matrix of the subgraph, H(l) is the node feature of the previous layer, W (l) is the parameter of the layer; when l=0, that is, in the first layer convolution, H (l)The initial feature matrix H of the node 0 = [h t 1 , h t 2 ... h t N ].

[0017] 2-3) Multi-scale cascading graph dynamic feature modeling is performed, for multi-scale cascading graph features, a TransformerEncoder is used to extract multi-scale cascading graph dynamic change features, and a multi-scale cascading graph dynamic feature representation is obtained through formula (2),

[0018] x0 = x cas + x p 0 s

[0019] y k = x k-1 + MSA (LN (x k-1 )) (2)

[0020] r cas = y cas + FFN (LN (y k ))

[0021] In formula (2), x cas is a multi-scale cascading graph feature, x pos is a position encoding, x0 is an initial input, MSA is a multi-head attention mechanism, LN is a layer regularization, and FFN represents a fully connected layer.

[0022] In step 3, first, a one-dimensional convolution is used to preliminarily extract multi-scale trend features through formula (3),

[0023] x trend = CNN (F = [f t 1 , f t 2 ... f t N ]) (3)

[0024] In formula (3), CNN is a one-dimensional convolution formula, F = [f t 1 , f t 2 ... f t N ] is a multi-scale heat trend sequence, and x trend is a multi-scale trend feature.

[0025] Then, multi-scale trend dynamic feature modeling is performed, and for the multi-scale trend feature, a TransformerEncoder is used to extract the multi-scale trend dynamic change feature, and the multi-scale trend dynamic feature representation r is obtained through formula (4) trend ,

[0026] x0=x trend +x pos

[0027] y k =x k-1 +MSA(LN(x k-1 )) (4)

[0028] r trend =y trend +FFN(LN(y k ))

[0029] In formula (4), x trend is a multi-scale trend feature, x pos is a position encoding, x0 is an initial input, MSA is a multi-head attention mechanism, LN is a layer normalization, and FFN represents a fully connected layer.

[0030] In step 4, the following is included:

[0031] 4-1) The concatenated graph dynamic features of the same scale are spliced with the trend dynamic features, and the spliced multi-scale feature representation R is obtained through formula (6): R={r(t0),r(t1)...r(t N-1 )}, wherein r cas is a multi-scale concatenated graph dynamic feature, and r trend is a multi-scale trend dynamic feature.

[0032] 4-2) The time window selection network includes a gated residual network and a time window weight weighting, wherein:

[0033] The gated residual network uses a linear layer plus an activation function as a gate, and adds a residual connection, to preliminarily select and filter the multi-scale features; specifically, the multi-scale features are preliminarily selected and filtered through formula (5):

[0034] GRN w (r(t i ))=LayerNorm(r(t i )+GLU w (η1))

[0035] η1=W 1,w η2+b 1,w (5)

[0036] η2 = ELU(W 2,w r(t i )+b 2,w )

[0037] r(t i ) is a multi-scale dynamic feature, LayerNorm represents layer regularization, η1 is a linear layer, W 1,w ,b 1,w are linear layer parameters, GLU w is a gating basic unit; ELU w is a gating activation function, W 2,w ,b 2,w are also linear layer parameters, GRN w (r(t i )) is a multi-scale feature after gating residual selection;

[0038] The time window weight weights: the multi-scale feature after gating residual selection is normalized by softmax to obtain the weight coefficient of each scale time window, and then multiplied by the sequence feature to obtain the preliminary selected multi-scale feature;

[0039] The feature obtained by formula (5) is normalized by formula (6), and the final multi-scale fusion feature is obtained by multiplication;

[0040]

[0041] In formula (6), is a multi-scale feature after gating residual selection, and Softmax is a weight normalization activation function. is the multiplication of the weight and the multi-scale feature after gating residual selection, is the final multi-scale fusion feature for prediction.

[0042] Compared with the prior art, the beneficial effects of the present application are:

[0043] Compared with the existing heat prediction method, the multi-scale dynamic features ignored can be fully utilized. And through effective fusion, the fused multi-scale dynamic features are obtained. The model is more robust when facing different data, so that the social media content heat value obtained by the final prediction is more accurate. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is the flow chart of the social media content heat prediction method of the present application which fuses multi-scale cascade and time sequence features. DETAILED DESCRIPTION

[0045] The application will be further described below in conjunction with the accompanying drawings and specific embodiments, but the following embodiments are by no means limiting to the application.

[0046] The design concept of the social media content heat prediction method based on graph neural network fusion of text and time sequence information is: mainly including multi-scale cascading graph sequence embedding to realize multi-scale cascading graph sequence dynamic feature modeling, multi-scale heat trend dynamic feature modeling method, multi-scale dynamic feature fusion modeling, network model training and heat prediction result output; through multi-layer graph convolutional neural network to extract cascading graph features, and Transformer Encoder to extract multi-scale cascading graph dynamic features, use one-dimensional convolution and Transformer Encoder to extract multi-scale trend dynamic features, use the time window selection network composed of the time window weight weighted by the gate residual network and the time window weight to realize the effective fusion of multi-scale features, and improve the training effect of the model. The final heat is predicted by using a multi-layer perceptron. The advantage of the application is that the multi-scale cascading graph features and trend features are effectively fused, and compared with the single scale method, the prediction effect is better, and the prediction result with better globality and generalization performance can be obtained. Figure 1 As shown in the figure, the steps are as follows:

[0047] Step 1, data preprocessing:

[0048] The early propagation characteristics of social media content, such as microblogs published by users in Sina microblog, are collected, and the user forwarding cascading graph and the heat change value of each time period in the early propagation of microblog are collected. According to different size scales of time window, multi-scale cascading graph sequence and multi-scale heat trend sequence are obtained. For the propagation data of social media content in the first T time period, K different scale time windows are given, corresponding to K different scales S=[s1,s2...s k ], the obtained multi-scale sequence is P=[p t 1 ,p t 2 ...p t N ], wherein t∈S represents one of the scales, N=T / t, and N represents the length of the sequence. The multi-scale cascading graph sequence is G=[g t 1 ,g t 2 ...g t N ], and the multi-scale heat trend sequence is F=[f t 1 ,f t 2 ...f t N ].

[0049] Step 2, multi-scale cascade graph dynamic feature modeling:

[0050] Convert the graph in the multi-scale cascade graph sequence obtained in step 1 into an adjacency matrix sequence, calculate the Laplacian matrix of the adjacency matrix, and at the same time, obtain the initial feature matrix of the nodes of each subgraph in the subgraph sequence through the PageRank algorithm; for the multi-scale cascade graph sequence G = [g t 1 ,g t 2 ...g t N ], first, convert the cascade graph sequence G into an adjacency matrix sequence A = [a t 1 ,a t 2 ...a t N ], then calculate the Laplacian matrix of the adjacency matrix At the same time, apply the PageRank algorithm to each subgraph in the adjacency matrix sequence to obtain the initial feature matrix H 0 = [h t 1 ,h t 2 ...h t N ];

[0051] For multi-scale cascade graph sequences of different scales, graph convolutional neural networks are applied to embed the multi-scale cascade graph sequence, the Laplacian matrix and the initial feature matrix are input, and the node features are obtained; the dynamic correlation between the multi-scale cascade graph sequences is established, and the multi-scale cascade graph dynamic features are extracted through the Transformer Encoder.

[0052] First, the node features are extracted through the graph convolutional neural network, and the final node features H (l +1) are obtained through formula (1), and the node features are represented into graph features through the readout function, that is, the features of each node are added and then averaged; the multi-scale cascade graph feature sequence x cas is obtained.

[0053]

[0054] In formula (1), H (l+1) is the node feature of the next layer, σ is an activation function, is the Laplacian matrix of the subgraph, H (l) is the node feature of the previous layer, W (l) is the parameter of the layer; H 0= [h t 1 ,h t 2 ...h t N ] is the node feature of the first layer.

[0055] Then, the multi-scale cascading graph dynamic feature modeling is performed, and for the multi-scale cascading graph feature, the multi-scale cascading graph dynamic change feature is extracted by using the Transformer Encoder, and the multi-scale cascading graph dynamic feature representation r cas ,

[0056] x0 = x cas +x pos

[0057] y k =x k-1 +MSA(LN(x k-1 )) (2)

[0058] r cas =y cas +FFN(LN(y k ))

[0059] In formula (2), x cas is the multi-scale cascading graph feature, x pos is the position encoding, x0 is the initial input, MSA is the multi-head attention mechanism, LN is the layer regularization, and FFN represents the fully connected layer.

[0060] Step 3, multi-scale trend dynamic feature modeling:

[0061] The multi-scale trend sequence obtained in step 1 is used to extract the multi-scale trend short-term dynamic feature by using one-dimensional convolution, and the long-term dynamic change feature, i.e. the multi-scale trend dynamic feature, is extracted by using the Transformer Encoder.

[0062] First, the multi-scale trend feature is preliminarily extracted by one-dimensional convolution according to formula (3),

[0063] x trend =CNN(F = [f t 1 ,f t 2 ...f t N ]) (3)

[0064] In formula (3), CNN is the one-dimensional convolution formula, F = [f t 1 ,f t 2...f t N ]is a multi-scale trend sequence, x trend is a multi-scale trend feature;

[0065] Then, multi-scale trend dynamic feature modeling is performed, and for the multi-scale trend feature, a TransformerEncoder is used to extract the multi-scale trend dynamic change feature, and a multi-scale trend dynamic feature representation r trend ,

[0066] x0=x trend +x pos

[0067] y k =x k-1 +MSA(LN(x k-1 )) (4)

[0068] r trend =y trend +FFN(LN(y k ))

[0069] In formula (4), x trend is a multi-scale trend feature, x pos is a position encoding, x0 is an initial input, MSA is a multi-head attention mechanism, LN is a layer normalization, and FFN represents a fully connected layer.

[0070] Step 4, multi-scale feature fusion is realized by using a time window selection network:

[0071] The filtered multi-scale concatenated graph dynamic features and multi-scale trend dynamic features are spliced, and then the multi-scale fusion features are obtained by using the time window selection network.

[0072] 4-1) Splicing of concatenated graph features and trend change features, splicing the concatenated graph dynamic features and the trend dynamic features of the same scale, through the spliced multi-scale feature representation R={r(t0),r(t1)...r(t N-1 )} is obtained, where r cas is a multi-scale concatenated graph dynamic feature, and r trend is a multi-scale trend dynamic feature.

[0073] 4-2) Multi-scale feature fusion is performed by using a time window selection network, and the time window selection network includes a gated residual network and a time window weight weighting, wherein:

[0074] The gating residual network: using a linear layer plus an activation function as a gate, while adding a residual connection, preliminarily selects and filters the multi-scale features; specifically: preliminarily selecting and filtering the multi-scale features through formula (5);

[0075] GRN w (r(t i ))=LayerNorm(r(t i )+GLU w (η1))

[0076] η1=W 1,w η2+b 1,w (5)

[0077] η2=ELU(W 2,w r(t i )+b 2,w )

[0078] Formula (5), wherein r(t i ) is a multi-scale dynamic feature, LayerNorm represents layer regularization, η1 is a linear layer, W 1,w ,b 1,w are linear layer parameters, GLU w is a gating basic unit; ELU w is a gating activation function, W 2,w ,b 2,w are also linear layer parameters, GRN w (r(t i )) is a multi-scale feature after gating residual selection.

[0079] The time window weight weighting: the multi-scale features after gating residual selection are normalized by using softmax to obtain the weight coefficients of each scale time window, and then multiplied by the sequence features to obtain the preliminarily selected multi-scale features.

[0080] The features obtained by formula (5) are normalized by formula (6), and then multiplied to obtain the final multi-scale fusion features;

[0081]

[0082] In formula (6), is a multi-scale feature after gating residual selection, and Softmax is a weight normalization activation function. is multiplied by the weight and the multi-scale feature after gating residual selection, is the final multi-scale fusion feature used for prediction.

[0083] Step 5, training of the network model and output of the heat prediction result:

[0084] The multi-scale fusion features are input into a full connection layer to output a heat growth prediction value; during the network model training process, each level weight value is optimized through the MSLE loss function reverse propagation mode, and the feedback result is output to the input end of each level of the entire network model; after the network model training is completed, only the early multi-scale cascade graph sequence and the multi-scale heat trend sequence of the social media to be predicted, such as Weibo, need to be input, and the output of the network model is the Weibo content heat growth prediction value in the future period of time, so as to realize the social media content heat prediction.

[0085] Table 1 is the prediction MSLE (mean square logarithmic error) result comparison of the method (denoted as MCTNN in the table) proposed in the application and other social media content heat prediction methods in the prior art, including two main data sets, Weibo2016 and Weibo2020, respectively, representing the real Weibo data of Sina Weibo in 2016 and 2020, and the observation time of the two data sets is set to 1 hour, 2 hours and 3 hours. The experimental methods compared include Features-deep, Deepcas, Deep-hawkes, and CasCN, which are all well-known and effective baseline methods in the field of social media content heat prediction, and all use a single scale cascade combined with time sequence information and deep learning method for heat prediction.

[0086] Table 1

[0087]

[0088] From the experimental results in Table 1, it can be seen that the prediction effect of the method of the application on the two real data sets is obviously better than that of the baseline method based on a single scale. The effectiveness of the application of the method of fusing multi-scale cascade and time sequence features is proved.

[0089] Although the application has been described above with reference to the drawings, the application is not limited to the specific embodiments described above, and the specific embodiments described above are only illustrative, not limiting, and those of ordinary skill in the art can make many modifications under the inspiration of the application without departing from the purpose of the application, and these all belong to the protection of the application.

Claims

1. A method for predicting the popularity of social media content by integrating multi-scale cascade and temporal features, characterized by the following steps: include: Step 1, Data Preprocessing: Utilizing the early dissemination characteristics of social media content, the data is processed according to time windows of different sizes to obtain multi-scale cascaded graph sequences and multi-scale popularity trend sequences. Step 2, Dynamic Feature Modeling of Multi-Scale Cascaded Graphs: Transform the graphs in the multi-scale cascaded graph sequence obtained in Step 1 into an adjacency matrix sequence, calculate the Laplacian matrix of the adjacency matrix, and obtain the initial feature matrix of the nodes for each subgraph in the subgraph sequence using the PageRank algorithm. For multi-scale cascaded graph sequences of different scales, a graph convolutional neural network is applied to embed the multi-scale cascaded graph sequences. The node features are obtained by taking the Laplacian matrix and the initial feature matrix as input. Establish dynamic correlations between multi-scale cascaded graph sequences and extract dynamic features of multi-scale cascaded graphs using Transformer Encoder; Step 3: Multi-scale trend dynamic feature modeling: The multi-scale heat trend sequence obtained in Step 1 is used to extract the short-term dynamic features of the multi-scale trend using one-dimensional convolution, and the long-term dynamic change features, i.e., the multi-scale trend dynamic features, are extracted using Transformer Encoder. Step 4: Use the time window selection network to achieve multi-scale feature fusion: Filter the dynamic features of the multi-scale cascaded graph and the dynamic features of the multi-scale trend, concatenate the filtered dynamic features of the multi-scale cascaded graph and the dynamic features of the multi-scale trend, and then use the time window selection network to obtain the multi-scale fused features. Step 5: Training the network model and outputting the popularity prediction results: The multi-scale fusion features are used as the input to the fully connected layer, and the output of the fully connected layer is the predicted value of popularity growth. During the training of the network model, the weights of each layer are optimized by backpropagation using the MSLE loss function, and the feedback results are output to the input terminals of each layer of the entire network model. After the network model is trained, the early multi-scale cascade graph sequence and multi-scale popularity trend sequence of the social media to be predicted are input. The output of the network model is the predicted value of the popularity growth of social media content in the future, thereby realizing the prediction of social media content popularity.

2. The social media content popularity prediction method according to claim 1, characterized in that, Step 1 is as follows: The early dissemination characteristics of social media content include user forwarding cascade graphs and popularity changes for each time period; for the dissemination data of social media content within the first T time periods, given K different scale time windows, corresponding to K different scales S = [s1, s2...s... k The resulting multi-scale sequence is P = [p] t 1 ,p t 2 ...p t N ], where t∈S represents one of the scales, N=T / t, and N represents the length of the sequence; the multi-scale cascaded graph sequence is G=[g t 1 ,g t 2 ...g t N The multi-scale heat trend sequence is F = [f t 1 ,f t 2 ...f t N ].

3. The social media content popularity prediction method according to claim 2, characterized in that, In step 2, 2-1) For multi-scale cascaded graph sequence level G = [g t 1 ,g t 2 ...g t N First, the cascaded graph sequence G is transformed into an adjacency matrix sequence A = [a t 1 ,a t 2 ...a t N Then, the Laplace matrix of the adjacency matrix is ​​calculated. Simultaneously, the PageRank algorithm is applied to each subgraph in the adjacency matrix sequence to obtain the initial feature matrix H of the nodes. 0 =[h t 1 ,h t 2 ...h t N ]; 2-2) The final node feature H is obtained through formula (1). (l+1) Then, the node features are represented as graph features using the readout function, which involves summing the feature representations of each node and then averaging them; this yields the multi-scale cascaded graph feature sequence x. cas ; In equation (1), H (l+1) The node features of the next layer are represented by σ, which is the activation function. H is the Laplacian matrix of the subgraph. (l) It is the node feature of the previous layer, W (l) For this layer parameter, when l = 0, i.e., in the first convolutional layer, H (l) H is the initial characteristic matrix of the node. 0 =[h t 1 ,h t 2 ...h t N ]; 2-3) Perform dynamic feature modeling of multi-scale cascaded graphs. For the features of multi-scale cascaded graphs, use TransformerEncoder to extract the dynamic change features of multi-scale cascaded graphs, and obtain the dynamic feature representation r of multi-scale cascaded graphs through formula (2). cas , x0 = x cas +x pos y k =x k-1 +MSA(LN(x k-1 )) (2) r cas =the cas +FFN(LN(y k )) In equation (2), x cas For multi-scale cascaded graph feature sequences, x pos For position encoding, x0 is the initial input, MSA is the multi-head attention mechanism, LN is the layer regularization, and FFN represents a fully connected layer.

4. The social media content popularity prediction method according to claim 3, characterized in that, In step 3, First, multi-scale trend features are initially extracted using one-dimensional convolution according to formula (3). x trend =CNN(F=[f t 1 ,f t 2 ...f t N ]) (3) In equation (3), CNN is a one-dimensional convolution formula, F = [f t 1 ,f t 2 ...f t N [x] is a multi-scale heat trend sequence. trend It exhibits multi-scale trend characteristics; Then, multi-scale trend dynamic feature modeling is performed. For the multi-scale trend features, the Transformer Encoder is used to extract the multi-scale trend dynamic change features, and the multi-scale trend dynamic feature representation r is obtained through formula (4). trend , x0=x trend +x pos y k =x k-1 +MSA(LN(x k-1 )) (4) r trend =the trend +FFN(LN(y k )) In equation (4), x trend For multi-scale trend characteristics, x pos For position encoding, x0 is the initial input, MSA is the multi-head attention mechanism, LN is the layer regularization, and FFN represents a fully connected layer.

5. The social media content popularity prediction method according to claim 4, characterized in that, Step 4 includes: 4-1) Concatenate the dynamic features of cascaded plots of the same scale with the dynamic features of trends, and then... The concatenated multi-scale feature representation is obtained as follows: R = {r(t0), r(t1)...r(t...} N-1 )}, where r cas For the dynamic features of multi-scale cascaded graphs, r trend It represents a multi-scale trend dynamic characteristic; 4-2) The time window selection network includes a gated residual network and a time window weighted network, wherein: The gated residual network described above uses a linear layer with an activation function as a gate, and adds residual connections to perform preliminary selection and filtering of multi-scale features; specifically, it performs preliminary selection and filtering of multi-scale features through equation (5); GRN w (r(t i ))=LayerNorm(r(t i )+GLU w (η1)) η1=W 1,w η2+b 1,w (5) η2=ELU(W 2,w r(t i )+b 2,w ) In equation (5), r(t) i ) represents multi-scale dynamic features, LayerNorm represents layer regularization, η1 is a linear layer, and W 1,w ,b 1,w For linear layer parameters, GLU w It is a basic gating unit; ELU w It is a gated activation function, W 2,w ,b 2,w Also with linear layer parameters, GRN w (r(t i ()) represents the multi-scale features after gated residual selection; The time window weighting is as follows: the weights of the multi-scale features selected by gated residuals are normalized by softmax to obtain the weight coefficients of each scale time window, and then multiplied with the sequence features to obtain the initially selected multi-scale features. The features obtained from equation (5) are weighted and normalized using equation (6), and then multiplied to obtain the final multi-scale fusion features. In equation (6), The features are multi-scale features after gated residual selection, and Softmax is the weighted normalized activation function. The weights are multiplied by the multi-scale features selected by the gated residuals. These are the final multi-scale fusion features used for prediction.

Images