Internet marketing activity key user participation degree trend prediction method

The trend sequence of key user participation in Internet marketing activities is decomposed through particle swarm optimization algorithm and variational modal model, and the DenseNet network is used to improve the time domain convolution network, solving the problem of nonlinear non-stationary trend sequence and intimate context connection. Achieve more accurate user engagement trend forecasts.

CN119067712AActive Publication Date: 2024-12-03ZHEJIANG SCI-TECH UNIV
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411580482.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-11-07
Publication Date
2024-12-03
Estimated Expiration
2044-11-07

AI Technical Summary

Technical Problem

There are problems in the trend prediction of trend sequences that are nonlinear and non-stationary and intimate context-related problems in existing Internet marketing activities, which limit the accuracy of the prediction.

Method used

The particle swarm optimization algorithm is used to optimize the variational modal model, and the key user participation trend sequence is decomposed into several subsequences, and the time domain convolution network is improved through the DenseNet network to perform convolution operations to analyze the key user participation trend.

Benefits of technology

Adaptively process nonlinear and non-stationary signals, split into multiple relatively stationary modes, avoid gradient vanishing and explosion problems, strengthen the ability to reuse features, and improve the accuracy of prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067712A_ABST
    Figure CN119067712A_ABST
Patent Text Reader

Abstract

The invention relates to an internet marketing activity key user participation degree trend prediction method, and belongs to the field of internet trend prediction, and the method comprises the following steps: S1, collecting marketing data, and constructing a key user participation degree trend sequence according to the classification weight and the interactive behavior score of key users in the marketing data; s2, constructing a variational modal model, optimizing an optimal parameter combination of the variational modal model by using a particle swarm optimization algorithm, and decomposing the key user participation degree trend sequence into a plurality of key user participation degree trend subsequences through the variational modal model; and S3, improving the time domain convolutional network by adopting a dense connection mode of a DenseNet network, respectively sending each key user participation degree trend subsequence into the improved time domain convolutional network for convolution operation, and analyzing the key user participation trend. According to the method, the problem that the networking marketing trend prediction trend sequence is non-linear and non-stable is solved, the problems of gradient disappearance and explosion are avoided, and the feature reuse capability is enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of Internet trend prediction, and in particular relates to a method for predicting the trend of key user participation in Internet marketing activities. Background Art

[0002] In recent years, with the vigorous development of information technology, there is more and more information on the Internet, the speed of information dissemination is getting faster and faster, and the channels for information dissemination are getting wider and wider. With the explosion of data volume and the emergence of endless websites, people are unable to quickly and intuitively perceive the dynamic development of the things they are concerned about. Although all major well-known websites currently have hot rankings or search rankings, their functions are completely facing all users. However, the information that different groups of people pay attention to is completely different, and the dissemination trend of information is also unpredictable. Therefore, there is a need for a technical solution that can perform hot spot ranking and hot spot prediction based on the Internet information of user-defined sites. Based on the same time slice and the number of reprints of Internet information, the Internet information at the current time can be compared with the Internet information at the same time slice to obtain the changes in the Internet information at the current time.

[0003] Recurrent Neural Networks (RNNs) are able to store information reasonably due to their unique recurrent architecture, which enables them to effectively capture the temporal relationship between data when processing time series data, and are widely used in tasks such as trend prediction. However, when processing long time series data, RNNs face the challenge of gradient vanishing or gradient exploding, which limits their ability to learn long-distance dependencies. LSTM introduced the concepts of forget gates and input gates to alleviate this problem by fine-tuning the flow of information. Subsequently, GRU improved computational efficiency by merging the forget gate and input gate into a single update gate and adding a reset gate. TCN, introduced in 2018, is an extension of the traditional convolutional neural network (CNN). It gradually expands its receptive field by stacking more convolutional layers and using larger dilation factors or increasing the size of the filter, thereby achieving a richer representation of information.

[0004] At present, the research on trend prediction of Internet marketing activities mainly relies on two technologies: autoregressive model and recurrent neural network. There are problems such as nonlinear and non-stationary trend sequences and loose contextual connection of trend sequences. Summary of the invention

[0005] The present invention provides a method for predicting the trend of key user participation in an Internet marketing activity, so as to solve the problems of nonlinear and non-stationary trend sequences and loose contextual connection of trend sequences in existing Internet marketing trend prediction.

[0006] In order to solve the above technical problems, the technical solution provided by the present invention is: The present invention relates to a method for predicting the trend of key user participation in an Internet marketing activity, which comprises the following steps: S1. Collect marketing data and build a key user engagement trend sequence based on the classification weights and interactive behavior scores of key users in the marketing data; S2. Construct a variational modal model, use the particle swarm optimization algorithm to optimize the optimal parameter combination of the variational modal model, and decompose the key user engagement trend sequence into several key user engagement trend subsequences through the variational modal model; S3. The dense connection method of the DenseNet network is used to improve the time-domain convolutional network, and each key user participation trend subsequence is sent into the improved time-domain convolutional network for convolution operation to analyze the participation trend of key users.

[0007] Preferably, the specific steps of constructing the key user engagement trend sequence according to the classification weights and interactive behavior scores of the key users in the marketing data in S1 are: S1.1. Construct a half-life function. The expression of the half-life function is: (1), in, represents the half-life, t0 represents the start time of the marketing campaign, t represents the current time, and θ is the decay factor; S1.2. Construct a marketing activity heat function with a half-life function. The formula of the marketing activity heat function is: (2), Among them, Heat(t) represents the heat of the marketing activity, and Num(t) represents the number of participants in the activity on that day; S1.3. According to the classification weights, interactive behavior scores and marketing activity popularity of key users in the marketing data, a daily average engagement function is constructed. The expression of the daily average engagement function is: (3), Among them, P(t) represents the average daily participation, u i represents the i-th key user, u i ∈U, U represents the key user group, S(u i , t) represents the key user u i The interactive behavior score at time t, W(u i , t) represents the key user u at time t i The weight determined by the classification, N represents the total number of key user groups at time t; S1.4. Calculate the average daily engagement for each time period, sort the average daily engagement in chronological order, and construct a trend sequence of key user engagement.

[0008] Preferably, after calculating the average daily engagement of each time period, S1.4 further normalizes the average daily engagement of each time period so that data with different characteristics are scaled to the same scale, and sorts the normalized average daily engagement in chronological order to construct a key user engagement trend sequence; The formula for the normalization process is: (4), Among them, Norm(t) represents the average daily participation after normalization, P max and P min Respectively represent the maximum and minimum participation in a period of time.

[0009] Preferably, S2 introduces a variational modal model constructed by a quadratic penalty term and Lagrange multipliers, and the variational modal model formula is: (7), Where L represents the Lagrangian function, IMF k represents the kth key user engagement trend subsequence, K represents the number of key user engagement trend subsequences, {IMF k} represents the set of K key user engagement trend subsequences, that is, each intrinsic mode function corresponds to a key user engagement trend subsequence, ω k represents the center frequency of the kth modal component after decomposition, {ω k} represents the set of center frequencies of the K modal components after decomposition, X(t) represents the original user engagement time series, represents the partial derivative with respect to time t, δ(t) is expressed as a Dirac distribution, represents the convolution operation, α represents the quadratic penalty factor, λ represents the Lagrange multiplier, and r represents an imaginary number.

[0010] Preferably, the variational modal model includes the following constraints: (5), (6).

[0011] Preferably, the specific steps of S2 using a particle swarm optimization algorithm to optimize the optimal parameter combination of the variational modal model are: S2.1. Construct a particle swarm algorithm. The formula of the particle swarm algorithm is: (12), (13), Among them, v i represents the velocity of particle i, xi represents the position of particle i, represents the inertia weight, c1 and c2 represent the learning factors, r1 and r2 represent random numbers between (0,1), and p best,i represents the optimal position found by particle i so far, g best,i represents the optimal position found by the particle swarm so far; S2.2. The particle swarm optimization algorithm is used to optimize the total number K of key user participation trend subsequences and the quadratic penalty factor α with envelope entropy as the fitness function. The fitness function is: (14), (15), Where j represents the key user participation trend subsequence index, m(j) is the normalized form of the envelope signal obtained by Hilbert demodulation of each key user participation trend subsequence, and p j is the probability distribution, M represents the total number of particles, E p represents the fitness function; The specific process of optimizing the total number K of key user participation trend subsequences and the secondary penalty factor α is as follows: (1) Initialize a set of particles, each particle represents a candidate solution for the total number K of key user engagement trend subsequences and the quadratic penalty factor α, and set the initial position and initial velocity for each particle; (2) The performance of each particle is evaluated using the envelope entropy. That is, for a given total number K of key user engagement trend subsequences and a quadratic penalty factor α, VMD decomposition is performed and the envelope entropy of the result is calculated. The speed and position of each particle, the individual local optimum, and the population global optimum are all updated. (3) Repeat the iteration until the preset maximum number of iterations is reached, and finally output the global optimal solution.

[0012] Preferably, the specific steps of decomposing the key user engagement trend sequence into a number of key user engagement trend subsequences by using a variational modal model in S2 are: S2.3. Initialize the key user engagement trend subsequence, modal component center frequency and Lagrange multiplier, and set the number of iterations n to 0; S2.4. Let the sequence number k of the key user engagement trend subsequence be 1; S2.5. Iterate the point user engagement trend subsequence and modal component center frequency by alternating direction multiplier method. The calculation formula of the key user participation trend subsequence is: (8), in, represents the trend subsequence of the key user engagement after iteration, ω represents the center frequency of the modal component, represents the original signal after iteration, z represents the subsequence index of the key user engagement trend after iteration, represents the iterated Lagrange multiplier, α represents the quadratic penalty factor; The calculation formula of the modal component center frequency is: (9), in, represents the center frequency of the modal component after iteration, d ω represents the calculus increment, n represents the number of iterations; S2.6. Is the sequence number k of the key user engagement trend subsequence the same as the total number K of the point user engagement trend subsequences? If not, return to S2.5. If so, update the Lagrange multiplier λ. The calculation formula of the Lagrange multiplier λ is: (10), in, and They represent the Lagrange multipliers λ before and after iteration, and τ represents the noise tolerance level; S2.7. Calculate the convergence accuracy after iteration and compare it with the accuracy threshold. If the accuracy is not less than the accuracy threshold, increase the number of iterations n by 1 and return to S2.4. If the accuracy is less than the accuracy threshold, stop iteration and output all key user engagement trend subsequences.

[0013] Preferably, the calculation formula of the S2.7 convergence accuracy is: (11).

[0014] Preferably, the S3 uses the dense connection mode of the DenseNet network to improve the time domain convolutional network. Specifically, the residual network structure in the time domain convolutional network is replaced by the DenseNet network structure. The dense block of the DenseNet network structure adopts a combination structure of a BN layer, a ReLU activation function, and a Conv converter. The expression of the ReLU activation function is: (16), Among them, x l represents the output of the first layer, H l Represents the nonlinear transformation of the lth layer.

[0015] Preferably, S3 sends each key user participation trend subsequence into the improved time domain convolutional network for convolution operation. The specific steps of analyzing the key user participation trend are: S3.1. Weight normalization is performed on the key user engagement trend subsequences obtained by decomposing the variational modal model, and the ReLU activation function is used to increase the nonlinear relationship between the networks; S3.2. Perform the dilated causal convolution operation of the time domain convolutional network on the key user engagement trend subsequence. The operation formula is: (17), Among them, F is the convolution kernel, f represents the convolution layer, m represents the convolution layer index, and x s represents the time step currently considered, , d is the dilation factor, c is the convolution kernel size, is the convolution operator, is historical data, which means that only convolution operations are performed on past states.

[0016] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects: 1. The method for predicting the trend of key user participation in Internet marketing activities involved in the present invention uses a particle swarm optimization algorithm to optimize the optimal parameter combination of a variational modal model, and decomposes a key user participation trend sequence into a number of key user participation trend subsequences through a variational modal model, adaptively processes nonlinear and non-stationary signals, and splits the key user participation trend sequence into multiple relatively stable modes, each mode having a unique central frequency, thereby solving the problem of nonlinearity and non-stationarity of the trend sequence for Internet marketing trend prediction.

[0017] 2. The method for predicting the trend of key user participation in Internet marketing activities involved in the present invention adopts the dense connection method of the DenseNet network to improve the time domain convolutional network, and sends each key user participation trend subsequence into the improved time domain convolutional network for convolution operation. The time domain convolutional network realizes effective processing of sequence data through causal convolution, avoids the gradient disappearance and explosion problems, and enhances the ability of feature reuse by introducing the dense connection method of DenseNet. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is the architecture diagram of the key user engagement trend prediction model for Internet marketing activities.

[0019] Figure 2 This is the flow chart of VMD algorithm.

[0020] Figure 3 To improve the DTCN structure diagram.

[0021] Figure 4 This is a series chart showing the trend of key user engagement.

[0022] Figure 5This is the VMD decomposition result diagram of key user time series data.

[0023] Figure 6 Error box plot of the key user engagement trend prediction method.

[0024] Figure 7 This is a result analysis chart of the key user engagement trend prediction method.

[0025] Figure 8 This is a diagram of the empirical application of the key user engagement trend prediction method. DETAILED DESCRIPTION

[0026] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with examples. The following examples are used to illustrate the present invention but are not used to limit the scope of the present invention.

[0027] See attached Figure 1 As shown, the present invention relates to a method for predicting the trend of key user participation in Internet marketing activities, which comprises the following steps: S1. Collect marketing data, focus on key user groups, and divide key users into three categories, namely loyal users, valuable users and opinion leaders. Loyal users refer to those consumer groups that interact with the brand for a long time and continue to show engagement, defined by indicators such as the cumulative number of likes, reposts, favorites, and the cumulative viewing time of live videos; valuable users represent those groups that have a decisive influence on the brand or event in terms of contribution, divided by the cumulative amount of consumption of peripheral products and services, and the number of transactions with large purchases and frequent transactions; opinion leaders refer to groups with significant influence in the exclusive online community of the brand or event, defined by the cumulative number of fans and the number of comment support in the event discussion. By assigning specific weights to the participation scores of different types of key user groups and calculating their weighted averages, a key user engagement trend sequence is constructed based on the classification weights and interactive behavior scores of key users in the marketing data. The specific steps are as follows: S1.1. Construct a half-life function. The expression of the half-life function is: (1), in, represents the half-life, t0 represents the start time of the marketing campaign, and t represents the current time. is the attenuation factor; S1.2. Construct a marketing activity heat function with a half-life function. The formula of the marketing activity heat function is: (2), Among them, Heat(t) represents the heat of the marketing activity, and Num(t) represents the number of participants in the activity on that day; S1.3. According to the classification weights, interactive behavior scores and marketing activity popularity of key users in the marketing data, a daily average engagement function is constructed. The expression of the daily average engagement function is: (3), Among them, P(t) represents the average daily participation, u i represents the i-th key user, u i ∈U, U represents the key user group, S(u i , t) represents the key user u i The interactive behavior score at time t, W(u i , t) represents the key user u at time t i The weight determined by the classification, N represents the total number of key user groups at time t; S1.4. Calculate the average daily engagement of each time period, normalize the average daily engagement of each time period, so that data with different characteristics are on the same scale, improve the efficiency and accuracy of model training, sort the normalized average daily engagement in chronological order, and construct a trend sequence of key user engagement; The formula for the normalization process is: (4), Among them, Norm(t) represents the average daily participation after normalization, P max and P min Respectively represent the maximum and minimum participation in a period of time.

[0028] S2. Constructing variational modal model: The variational problem constrained by the variational modal model needs to meet two conditions: the sum of the bandwidth of the center frequency of each modal vector is minimum and the sum of all modal vectors is equal to the original signal. Therefore, the variational modal model includes the following constraints: (5), (6); Among them, IMF k represents the kth key user engagement trend subsequence, K represents the number of key user engagement trend subsequences, {IMF k} represents the set of K key user engagement trend subsequences, that is, each intrinsic mode function corresponds to a key user engagement trend subsequence, ω k represents the center frequency of the kth modal component after decomposition, {ω k} represents the set of center frequencies of the K modal components after decomposition, X(t) represents the original user engagement time series, represents the partial derivative with respect to time t, δ(t) is expressed as a Dirac distribution, represents the convolution operation, α represents the quadratic penalty factor, λ represents the Lagrange multiplier, and r represents an imaginary number.

[0029] The variational mode model constructed by introducing quadratic penalty terms and Lagrange multipliers is formulated as follows: (7), Where L represents the Lagrangian function, IMF k represents the kth modal component, i.e., the key user engagement trend subsequence, K represents the number of key user engagement trend subsequences, {IMF k} represents the set of K key user engagement trend subsequences, that is, each intrinsic mode function corresponds to a key user engagement trend subsequence, ω k represents the center frequency of the kth modal component after decomposition, {ω k} represents the set of center frequencies of the K modal components after decomposition, X(t) represents the original user engagement time series, represents the partial derivative with respect to time t, δ(t) is expressed as a Dirac distribution, represents the convolution operation, α represents the quadratic penalty factor, λ represents the Lagrange multiplier, and r represents an imaginary number.

[0030] Reference Figure 1 and attached Figure 2 As shown in the figure, the particle swarm optimization algorithm is used to optimize the optimal parameter combination of the variational mode model (VMD). The specific steps are: S2.1. Construct a particle swarm algorithm. The formula of the particle swarm algorithm is: (12), (13), Among them, v i represents the velocity of particle i, x i represents the position of particle i, ω represents the inertia weight, c1 and c2 represent learning factors, r1 and r2 represent random numbers between (0,1), and p best,i represents the optimal position found by particle i so far, g best,i represents the optimal position found by the particle swarm so far; S2.2. The particle swarm optimization algorithm is used to optimize the total number K of key user participation trend subsequences and the quadratic penalty factor α with envelope entropy as the fitness function. The fitness function is: (14), (15), Where j represents the key user participation trend subsequence index, m(j) is the normalized form of the envelope signal obtained by Hilbert demodulation of each key user participation trend subsequence, and p j is the probability distribution, M represents the total number of particles, E p represents the fitness function; The specific process of optimizing the total number K of key user participation trend subsequences and the secondary penalty factor α is as follows: (1) Initialize a set of particles, each particle represents a candidate solution for the total number K of key user engagement trend subsequences and the quadratic penalty factor α, and set the initial position and initial velocity for each particle; (2) The performance of each particle is evaluated using the envelope entropy. That is, for a given total number K of key user engagement trend subsequences and a quadratic penalty factor α, VMD decomposition is performed and the envelope entropy of the result is calculated. The speed and position of each particle, the individual local optimum, and the population global optimum are all updated. (3) Repeat the iteration until the preset maximum number of iterations is reached, and finally output the global optimal solution.

[0031] The optimal parameter combination is obtained by optimizing the particle swarm optimization algorithm, and the key user participation trend sequence is decomposed into K key user participation trend subsequences IMF through the variational mode model. k (t), k=1,2,…,K, such as Figure 2 As shown, the specific steps are: S2.3. Initialize the key user engagement trend subsequence , modal component center frequency and Lagrange multipliers , let the number of iterations n be 0; S2.4. Let the sequence number k of the key user engagement trend subsequence be 1; S2.5. Iterate the point user engagement trend subsequence and modal component center frequency by alternating direction multiplier method. The calculation formula of the key user participation trend subsequence is: (8), in, represents the trend subsequence of the key user engagement after iteration, ω represents the center frequency of the modal component, represents the original signal after iteration, z represents the subsequence index of the key user engagement trend after iteration, represents the iterated Lagrange multiplier, α represents the quadratic penalty factor; The calculation formula of the modal component center frequency is: (9), in, represents the center frequency of the modal component after iteration, d ω represents the calculus increment, and n represents the number of iterations; S2.6. Is the sequence number k of the key user engagement trend subsequence the same as the total number K of the point user engagement trend subsequences? If not, return to S2.5. If so, update the Lagrange multiplier λ. The calculation formula of the Lagrange multiplier λ is: (10), in, and They represent the Lagrange multipliers λ before and after iteration, and τ represents the noise tolerance level; S2.7. Calculate the convergence accuracy after iteration. The calculation formula of convergence accuracy ε is: (11), Compare the convergence accuracy with the accuracy threshold ε0. If the accuracy is not less than the accuracy threshold, that is, ε≥ε0, increase the number of iterations n by 1 and return to S2.4. If the accuracy is less than the accuracy threshold, that is, ε<ε0, stop the iteration and output all key user engagement trend subsequences.

[0032] S3. Refer to the attached Figure 1 and 3 As shown in the figure, the time domain convolutional network is improved by using the dense connection method of the DenseNet network, that is, the residual network structure in the time domain convolutional network is replaced by the DenseNet network structure. The dense block Dense Block of the DenseNet network structure adopts a combination structure of BN layer, ReLU activation function, and Conv converter. Before inputting to the next convolutional layer, these features must be standardized through the BN layer to ensure the consistency and stability of the data, and then the convolution operation is performed. The Transition layer adjusts the number of channels by multiplying the convolution unit and uses average pooling to reduce the size of the feature map, integrates features between DenseBlocks and performs downsampling. The expression of the ReLU activation function is: (16), Among them, x l represents the output of the first layer, H l represents the nonlinear transformation of the lth layer, Each key user participation trend subsequence is sent to the improved time-domain convolutional network for convolution operation to analyze the key user participation trend. The specific steps are as follows: S3.1. Weight normalization is performed on the key user engagement trend subsequences obtained by decomposing the variational modal model, and the ReLU activation function is used to increase the nonlinear relationship between the networks; S3.2. The time domain convolution network (TCN) convolution layer performs the dilated causal convolution operation of the time domain convolution network on the key user engagement trend subsequence. The operation formula is: (17), Among them, F is the convolution kernel, f represents the convolution layer, m represents the convolution layer index, and x s represents the time step currently considered, , d is the expansion factor, is the convolution kernel size, is the convolution operator, is historical data, which means that only convolution operations are performed on past states.

[0033] Experimental example The original data of this experimental example comes from the marketing activity data of a large Internet marketing platform in 2023. This dataset is derived from a real Internet marketing activity scenario, has a large-scale and complete feature system, and covers various data. A total of 691 activities have been held, with a total of 4,242,713 participating users and a total of 1,358,331 members. The C dataset of the "Display Competition" marketing activity was processed accordingly. This dataset covers the basic interactive information of 62,967 users, and the event was held from February 1 to July 31. The data from February 1 to June 30 is used as the training set, and the data from July 1 to July 31 is used as the test set. The details of the key user division are shown in Table 1:

[0034] Table 1: Detailed rules for key user groups

[0035] The key user participation behavior standard table is shown in Table 2: Table 2: Key user participation behavior standards

[0036] Key user engagement trends such as Figure 4 As shown. The particle swarm algorithm is used to find the optimal parameters K and α of VMD, where the parameter optimization boundary ranges of K and α are set to [3,15] and [1000,2500] respectively, the number of algorithm iterations is 20 times, the population size is 10, the learning factors c1 and c2 are 2 and 0.5 respectively, and the inertia weight ω is 0.9. Based on the above parameter settings, the original time series is decomposed into 6 subsequences, and the decomposition results are shown in Figure 5 The mean absolute error (MAE), symmetric mean absolute percentage error (SMAPE), root mean square error (RMSE), and coefficient of determination (R 2) is used as the standard for evaluating the prediction accuracy of the model. The currently popular single model and combined model are compared with the method of the present invention. The experimental performance is shown in Table 3.

[0037] Table 3 Performance of all trend prediction models on the marketing campaign C dataset

[0038] In addition to the prediction error indicators, the prediction model error box plot and the prediction model result analysis chart are as follows Figure 6 and Figure 7 On a large Internet marketing platform, the model was empirically applied to the marketing activities. The visualization interface is shown as follows: Figure 8 shown.

[0039] The present invention is described in detail above in conjunction with the embodiments, but the contents described are only preferred embodiments of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of application of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for predicting the trend of key user participation in an Internet marketing activity, characterized by: It includes the following steps: S1. Collect marketing data and build a key user engagement trend sequence based on the classification weights and interactive behavior scores of key users in the marketing data; S2. Construct a variational modal model, use the particle swarm optimization algorithm to optimize the optimal parameter combination of the variational modal model, and decompose the key user engagement trend sequence into several key user engagement trend subsequences through the variational modal model; S3. The dense connection method of the DenseNet network is used to improve the time-domain convolutional network, and each key user participation trend subsequence is sent into the improved time-domain convolutional network for convolution operation to analyze the participation trend of key users.

2. The method for predicting the trend of key user participation in Internet marketing activities according to claim 1, characterized in that: The specific steps of S1 to construct a key user engagement trend sequence based on the classification weights and interactive behavior scores of key users in the marketing data are: S1.

1. Construct a half-life function. The expression of the half-life function is: (1), in, represents the half-life, t0 represents the start time of the marketing campaign, t represents the current time, and θ is the decay factor; S1.

2. Construct a marketing activity heat function with a half-life function. The formula of the marketing activity heat function is: (2), Among them, Heat(t) represents the heat of the marketing activity, and Num(t) represents the number of participants in the activity on that day; S1.

3. According to the classification weights, interactive behavior scores and marketing activity popularity of key users in the marketing data, a daily average engagement function is constructed. The expression of the daily average engagement function is: (3), Among them, P(t) represents the average daily participation, u i represents the i-th key user, u i ∈U, U represents the key user group, S(u i ,t) represents the key user u i The score of interactive behavior at time t, W(u i , t) represents the key user u at time t i The weight determined by the classification, N represents the total number of key user groups at time t; S1.

4. Calculate the average daily engagement for each time period, sort the average daily engagement in chronological order, and construct a trend sequence of key user engagement.

3. The method for predicting the trend of key user participation in Internet marketing activities according to claim 2, characterized in that: After calculating the average daily engagement of each time period, S1.4 also normalizes the average daily engagement of each time period so that data with different features are scaled to the same scale, and sorts the normalized average daily engagement in chronological order to construct a trend sequence of key user engagement; The formula for the normalization process is: (4), Among them, Norm(t) represents the average daily participation after normalization, P max and P min Respectively represent the maximum and minimum participation in a period of time.

4. The method for predicting the trend of key user participation in Internet marketing activities according to claim 1, characterized in that: The S2 introduces a variational modal model constructed by a quadratic penalty term and Lagrange multipliers. The variational modal model formula is: (7) Where L represents the Lagrangian function, IMF k represents the kth key user engagement trend subsequence, K represents the number of key user engagement trend subsequences, {IMF k } represents the set of K key user engagement trend subsequences, that is, each intrinsic mode function corresponds to a key user engagement trend subsequence, ω k represents the center frequency of the kth modal component after decomposition, {ω k } represents the set of center frequencies of the K modal components after decomposition, X(t) represents the original user engagement time series, represents the partial derivative with respect to time t, δ(t) is expressed as a Dirac distribution, represents the convolution operation, α represents the quadratic penalty factor, λ represents the Lagrange multiplier, and r represents an imaginary number.

5. The method for predicting the trend of key user participation in Internet marketing activities according to claim 4 is characterized by: The variational modal model includes the following constraints: (5), (6)。 6. The method for predicting the trend of key user participation in Internet marketing activities according to claim 4, characterized in that: The specific steps of S2 using the particle swarm optimization algorithm to optimize the optimal parameter combination of the variational modal model are: S2.

1. Construct a particle swarm algorithm. The formula of the particle swarm algorithm is: (12), (13), Where i represents the particle index, v i represents the velocity of particle i, x i represents the position of particle i, represents the inertia weight, c1 and c2 represent the learning factors, r1 and r2 represent random numbers between (0,1), and p best,i represents the optimal position found by particle i so far, g best,i represents the optimal position found by the particle swarm so far; S2.

2. The particle swarm optimization algorithm is used to optimize the total number K of key user participation trend subsequences and the quadratic penalty factor α with envelope entropy as the fitness function. The fitness function is: (14), (15), Where j represents the key user participation trend subsequence index, m(j) is the normalized form of the envelope signal obtained by Hilbert demodulation of each key user participation trend subsequence, and p j is the probability distribution, M represents the total number of particles, E p represents the fitness function; The specific process of optimizing the total number K of key user participation trend subsequences and the secondary penalty factor α is as follows: (1) Initialize a set of particles, each particle represents a candidate solution for the total number K of key user engagement trend subsequences and the quadratic penalty factor α, and set the initial position and initial velocity for each particle; (2) The performance of each particle is evaluated using the envelope entropy. That is, for a given total number K of key user engagement trend subsequences and a quadratic penalty factor α, VMD decomposition is performed and the envelope entropy of the result is calculated. The speed and position of each particle, the individual local optimum, and the population global optimum are all updated. (3) Repeat the iteration until the preset maximum number of iterations is reached, and finally output the global optimal solution.

7. The method for predicting the trend of key user participation in Internet marketing activities according to claim 6, characterized in that: The specific steps of S2 decomposing the key user engagement trend sequence into several key user engagement trend subsequences through the variational modal model are: S2.

3. Initialize the key user engagement trend subsequence, modal component center frequency and Lagrange multiplier, and set the number of iterations n to 0; S2.

4. Let the sequence number k of the key user engagement trend subsequence be 1; S2.

5. Iterate the point user engagement trend subsequence and modal component center frequency by alternating direction multiplier method. The calculation formula of the key user participation trend subsequence is: (8), in, represents the trend subsequence of the key user engagement after iteration, ω represents the center frequency of the modal component, represents the original signal after iteration, z represents the subsequence index of the key user engagement trend after iteration, represents the iterated Lagrange multiplier, α represents the quadratic penalty factor; The calculation formula of the modal component center frequency is: (9), in, represents the center frequency of the modal component after iteration, d ω represents the calculus increment, and n represents the number of iterations; S2.

6. Is the sequence number k of the key user engagement trend subsequence the same as the total number K of the point user engagement trend subsequences? If not, return to S2.

5. If so, update the Lagrange multiplier λ. The calculation formula of the Lagrange multiplier λ is: (10), in, and They represent the Lagrange multipliers λ before and after iteration, and τ represents the noise tolerance level; S2.

7. Calculate the convergence accuracy after iteration and compare it with the accuracy threshold. If the accuracy is not less than the accuracy threshold, increase the number of iterations n by 1 and return to S2.

4. If the accuracy is less than the accuracy threshold, stop iteration and output all key user engagement trend subsequences.

8. The method for predicting the trend of key user participation in Internet marketing activities according to claim 7, characterized in that: The calculation formula of the S2.7 convergence accuracy is: (11)。 9. The method for predicting the trend of key user participation in Internet marketing activities according to claim 1, characterized in that: The S3 uses the dense connection mode of the DenseNet network to improve the time domain convolution network. Specifically, the residual network structure in the time domain convolution network is replaced by the DenseNet network structure. The dense block of the DenseNet network structure adopts a combination structure of a BN layer, a ReLU activation function, and a Conv converter. The expression of the ReLU activation function is: (16), Among them, x l represents the output of layer l, H l Represents the nonlinear transformation of the lth layer.

10. The method for predicting the trend of key user participation in Internet marketing activities according to claim 9, characterized in that: S3 sends each key user participation trend subsequence into the improved time-domain convolutional network for convolution operation. The specific steps of analyzing the key user participation trend are as follows: S3.

1. Weight normalization is performed on the key user engagement trend subsequences obtained by decomposing the variational modal model, and the ReLU activation function is used to increase the nonlinear relationship between the networks; S3.

2. Perform the dilated causal convolution operation of the time domain convolutional network on the key user engagement trend subsequence. The operation formula is: (17), Among them, F is the convolution kernel, f represents the convolution layer, m represents the convolution layer index, and x s represents the time step currently considered, , d is the dilation factor, c is the convolution kernel size, is the convolution operator, is historical data, which means that only convolution operations are performed on past states.

Citation Information

Patent Citations

  • Bearing degradation state identification and predication method based on variation modal decomposition-transfer entropy

    CN109580222A

  • Rolling bearing fault diagnosis method based on VMD-DenseNet

    CN115114954A

  • Short-term power load prediction method based on CVMD-GRU-DenseNet hybrid model

    CN115511162A

  • Artificial intelligence identification of high-value audiences for marketing campaigns

    US11113707B1

  • AU2020104000A4