Feature selection method for mental health detection model
Through complex function model and feature selection recurrent units, combined with deep convolutional automatic encoder and variational inference, the feature set is optimized, which solves the problem of insufficient feature selection in mental health detection by traditional methods, and improves the accuracy and generalization of the model ability.
Patent Information
- Application Number
- CN202410975452.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2024-11-15
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional feature selection methods are difficult to fully capture the complex relationships and interactions between features in mental health detection, resulting in insufficient model performance and relying on expert subjective judgment or simple statistical methods to effectively process high-dimensional data.
The complex function model and feature selection cycle units are adopted, and the feature set is optimized through data preprocessing, preliminary screening, feature scoring functions and model training, and the feature collection is extracted and screened with deep convolutional automatic encoder and variational inference. Take into account the importance of features, interactions and model prediction capabilities in a comprehensive way.
It significantly improves the accuracy and generalization ability of mental health detection models, can better capture the complex relationships between data and the prediction ability of models, and select a more predictive feature set.
Smart Images

Figure CN118969270A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of mental health detection, and in particular to a feature selection method for a mental health detection model. Background Art
[0002] Mental health refers to all aspects of the mind and the process of activities being in a good or normal state. The ideal state of mental health is to maintain a healthy character, normal intelligence, correct cognition, appropriate emotions, reasonable will, positive attitude, appropriate behavior, and good adaptation. In mental health testing, feature selection is a key step, which directly affects the performance and accuracy of the model.
[0003] However, in mental health testing, feature selection of data is an important factor affecting model performance. Traditional feature selection methods usually rely on a single model or simple statistical methods, which cannot fully capture the complex relationships and interactions between features. In addition, when processing high-dimensional data, existing methods often find it difficult to find the optimal feature subset, resulting in insufficient model performance. At the same time, existing feature selection methods mostly rely on the subjective judgment of experts or simple statistical methods, which cannot fully consider the complex relationships between features and the predictive ability of the model. Therefore, the present invention proposes a feature selection method for complex function models and complex formulas, which optimizes the feature set through repeated detection cycle steps to improve the accuracy and generalization ability of the mental health detection model. Summary of the invention
[0004] To solve the above problems, the present invention provides the following technical solutions:
[0005] A feature selection method for a mental health detection model, characterized in that it includes the following steps:
[0006] Step 1: a data acquisition unit, wherein the data acquisition unit acquires original data;
[0007] Step 2: a data processing unit, wherein the data preprocessing unit performs data preprocessing on the original data set collected by the data collection unit, including missing value filling, outlier processing and normalization;
[0008] Step 3: Perform a preliminary screening of features based on statistics or simple machine learning models to remove features that are irrelevant or have low relevance to mental health status;
[0009] Step 4, feature selection loop unit, loops through the following loop steps:
[0010] S1: Construct a candidate feature set: randomly select a part of the features from the remaining features based on a specific strategy to form a candidate feature set;
[0011] S2: Feature scoring function: Define a complex feature scoring function F(X,θ) that takes into account the importance of features, interactions between features, and model prediction capabilities;
[0012] S3: Build and train the model: Build a mental health detection model based on the candidate feature set and train the model using training data;
[0013] S4: Evaluate the performance of the main model: Use the validation data to evaluate the trained main model and calculate the evaluation indicators;
[0014] S5: Update the candidate feature set: Update the candidate feature set according to the model performance evaluation results and feature scoring function, remove features with low scores, and add new potential features;
[0015] S6: Repeat the above process: repeat steps S1-S5 until the stop condition is met;
[0016] Step 5, output unit: comprehensively consider the importance of features, interactions between features, and model prediction capabilities.
[0017] Step 6, an early warning unit, wherein the early warning module compares the calculated freight rate with the actual freight rate data and issues an early warning according to the pre-warning.
[0018] Furthermore, the data acquisition unit includes a storage device, X = {X1, X2..., X n}, where X n It represents the nth piece of data, where n is a positive integer. The collected data is then sent to the data preprocessing unit for processing.
[0019] Furthermore, the data preprocessing unit includes a processor, which is used to analyze the data of the data acquisition unit, wherein the data set preprocessing includes normalizing the features to ensure the scale consistency between different features, wherein the normalization formula is:
[0020] Where μ is the mean and σ is the standard deviation.
[0021] Furthermore, the preliminary feature screening can use a composite model based on deep convolutional autoencoder and variational inference to achieve automatic feature extraction and screening of high-dimensional image data. The specific formula is as follows:
[0022] X filttered =CAE(X)+VI(X)
[0023] Where CAE(X) represents the feature extraction of input data X by the deep convolutional autoencoder, VI(X) is the process of variational inference model used to optimize feature screening, and the composite model is based on the ability of the autoencoder to extract meaningful feature representations and further optimizes and selects the most relevant features through variational inference.
[0024] Furthermore, the deep convolutional autoencoder includes two parts, an encoder and a decoder, which are used to learn the compressed representation Z of the input data X and reconstruct the input data.
[0025] Furthermore, the encoder f encoder (X), the encoder usually consists of multiple convolutional layers and pooling layers, and its output is the feature representation Z of the latent space. The specific formula is:
[0026] Z=ConvPool L (...ConvPool l (X)
[0027] ConvPool l Represents the convolution and pooling operations of layer l.
[0028] The decoder CAE -1 (Z), decode the latent representation Z into a reconstructed output using deconvolution and upsampling techniques
[0029] The specific decoding process can be expressed as:
[0030]
[0031] Where DeconvUnpool l Represents the deconvolution and upsampling operations of layer l
[0032] Furthermore, the variational inference is used to estimate the posterior distribution of the latent variables, and the model is trained by maximizing the variational lower bound, including the variational encoder, decoder and variational lower bound (ELBO). The variational encoder is specifically analyzed as follows:
[0033] The variational encoder maps the input X to a distribution q of latent variables Z Φ (Z|X), usually assuming that Z is normally distributed The formula is as follows:
[0034]
[0035] where μ Φ (X),σ 2 Φ (X) is the mean and variance of the encoder network output.
[0036] The decoder p θ (X|Z), the conditional distribution p that generates X from the latent variable Z θ (X|Z), it is usually assumed that X follows another normal distribution. The specific formula is as follows:
[0037]
[0038] Among them, μ θ (X),σ 2 θ (X) is the mean and variance of the decoder network output.
[0039] The variational inference training optimizes the model by maximizing the variational lower bound, (ELBO), which is a lower bound on the log marginal probability of observing data X:
[0040]
[0041] in, is the reconstruction term (loss function), and the second term is the KL divergence, which measures the potential distribution q after encoding Φ The difference between (Z|X) and the prior validation distribution p(Z).
[0042] Furthermore, the specific formula of the feature scoring function is as follows:
[0043]
[0044] Among them G i , H i It is a complex deep neural network layer or recurrent neural network model for advanced computation of feature importance evaluation and subset interactions, α i , β j is the adaptive learning rate and model complexity parameter, n is the size of the candidate feature set, and m is the number of subsets.
[0045] Furthermore, the G i , H i The specific analysis is as follows:
[0046]
[0047]
[0048] Overview: The specific formula of the feature scoring function is as follows:
[0049]
[0050] Among them, x i ,θ i 、x subsetj ,θsubsetj is the input feature vector, W G1 , W G2 are the weight matrices of the first and second layers of the G model, respectively, and W H1 , W H2 are the weight matrices of the first and second layers of the H model; b G1 , b G2 are the bias vectors of the first and second layers of the G model, and b H1 , b H2 are the bias vectors of the first and second layers of the H model, respectively, and η is the activation function, G i (x i ,θ i ) and H j (x subsetj ,θ subsetj ) is the feature vector after being processed by the deep neural network.
[0051] Furthermore, even if the evaluation function reaches an evaluation index improvement less than a preset threshold, the feature selection loop unit stops looping, that is, F(X,θ)≤F 标准阀值 (X,θ).
[0052] Furthermore, the warning module includes a warning light, F(X,θ)>F 标准阀值 (X,θ), the early warning module will issue an early warning, and the feature selection loop unit will continue to cycle.
[0053] Compared with the prior art, the present invention has the following beneficial effects:
[0054] 1. In the present invention, firstly, a data collection unit is used to collect original data; a data processing unit is used, and a data preprocessing unit preprocesses the data collected by the data collection unit to preprocess the original data set, including missing value filling, outlier processing and normalization; a method based on statistics or simple machine learning models is used to preliminarily screen features, and remove features that are irrelevant to or have low correlation with the mental health status; a feature selection loop unit is used to loop the loop steps, and the output unit: comprehensively considers the importance of features, the interaction between features and the model prediction ability; an early warning unit, an early warning module compares the calculated freight rate with the actual freight rate data, and issues an early warning according to the pre-warning, and the above-mentioned units repeatedly detect the loop steps to optimize the feature set and improve the accuracy and generalization ability of the mental health detection model.
[0055] 2. The efficient feature selection method proposed in the present invention can significantly improve the accuracy and stability of the mental health detection model. Compared with the traditional feature selection method, the method of the present invention can better capture the complex relationship between the data and the predictive ability of the model, thereby selecting a feature set with more predictive ability. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] In order to facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings;
[0057] Figure 1 A schematic diagram of the steps of a feature selection method for a mental health detection model according to the present invention. DETAILED DESCRIPTION
[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0059] like Figure 1 As shown, a feature selection method for a mental health detection model includes the following steps:
[0060] The following steps are involved:
[0061] Step 1: a data acquisition unit, wherein the data acquisition unit acquires original data;
[0062] Step 2: a data processing unit, wherein the data preprocessing unit performs data preprocessing on the original data set collected by the data collection unit, including missing value filling, outlier processing and normalization;
[0063] Step 3: Perform a preliminary screening of features based on statistics or simple machine learning models to remove features that are irrelevant or have low relevance to mental health status;
[0064] Step 4, feature selection loop unit, loops through the following loop steps:
[0065] S1: Construct a candidate feature set: randomly select a part of the features from the remaining features based on a specific strategy to form a candidate feature set;
[0066] S2: Feature scoring function: Define a complex feature scoring function F(X,θ) that takes into account the importance of features, interactions between features, and model prediction capabilities;
[0067] S3: Build and train the model: Build a mental health detection model based on the candidate feature set and train the model using training data;
[0068] S4: Evaluate the performance of the main model: Use the validation data to evaluate the trained main model and calculate the evaluation indicators;
[0069] S5: Update the candidate feature set: Update the candidate feature set according to the model performance evaluation results and feature scoring function, remove features with low scores, and add new potential features;
[0070] S6: Repeat the above process: repeat steps S1-S5 until the stop condition is met;
[0071] Step 5, output unit: comprehensively consider the importance of features, interactions between features, and model prediction capabilities.
[0072] Step 6, an early warning unit, wherein the early warning module compares the calculated freight rate with the actual freight rate data and issues an early warning according to the pre-warning.
[0073] The data acquisition unit includes a storage, X={X1, X2..., X n}, where X n It represents the nth piece of data, where n is a positive integer. The collected data is then sent to the data preprocessing unit for processing.
[0074] The data preprocessing unit includes a processor, and the processor is used to analyze the data of the data acquisition unit, wherein the data set preprocessing includes normalizing the features to ensure the scale consistency between different features, wherein the normalization formula is:
[0075] Where μ is the mean and σ is the standard deviation.
[0076] The preliminary feature screening can use a composite model based on deep convolutional autoencoder and variational inference to achieve automatic feature extraction and screening of high-dimensional image data. The specific formula is as follows:
[0077] X filttered =CAE(X)+VI(X)
[0078] Where CAE(X) represents the feature extraction of input data X by the deep convolutional autoencoder, VI(X) is the process of variational inference model used to optimize feature screening, and the composite model is based on the ability of the autoencoder to extract meaningful feature representations and further optimizes and selects the most relevant features through variational inference.
[0079] The deep convolutional autoencoder includes two parts, an encoder and a decoder, which are used to learn the compressed representation Z of input data X and reconstruct the input data.
[0080] The encoder f encoder (X), the encoder usually consists of multiple convolutional layers and pooling layers, and its output is the feature representation Z of the latent space. The specific formula is:
[0081] Z=ConvPool L (...ConvPool l (X)
[0082] ConvPool l Represents the convolution and pooling operations of layer l.
[0083] The decoder CAE -1 (Z), decode the latent representation Z into a reconstructed output using deconvolution and upsampling techniques
[0084] The specific decoding process can be expressed as:
[0085]
[0086] Where DeconvUnpool l Represents the deconvolution and upsampling operations of layer l.
[0087] The variational inference is used to estimate the posterior distribution of the latent variables and train the model by maximizing the variational lower bound, including the variational encoder, decoder and variational lower bound (ELBO). The variational encoder is specifically analyzed as follows:
[0088] The variational encoder maps the input X to a distribution q of latent variables Z Φ (Z|X), usually assuming that Z is normally distributed The formula is as follows:
[0089]
[0090] where μ Φ (X),σ 2 Φ (X) is the mean and variance of the encoder network output.
[0091] The decoder p θ (X|Z), the conditional distribution p that generates X from the latent variable Z θ (X|Z), it is usually assumed that X follows another normal distribution. The specific formula is as follows:
[0092]
[0093] Among them, μ θ (X),σ 2 θ (X) is the mean and variance of the decoder network output.
[0094] The variational inference training optimizes the model by maximizing the variational lower bound, (ELBO), which is a lower bound on the log marginal probability of observing data X:
[0095]
[0096] in, is the reconstruction term (loss function), and the second term is the KL divergence, which measures the potential distribution q after encoding Φ The difference between (Z|X) and the prior validation distribution p(Z).
[0097] The specific formula of the feature scoring function is as follows:
[0098]
[0099] Among them G i , H i It is a complex deep neural network layer or recurrent neural network model for advanced computation of feature importance evaluation and subset interactions, α i , β j is the adaptive learning rate and model complexity parameter, n is the size of the candidate feature set, and m is the number of subsets.
[0100] The G i , H i The specific analysis is as follows:
[0101] G i (x i ,θ i )=η(W G2 .η(W G1 .[x i ,θ i ]+b G1 +b G2 )
[0102] H j (x subsetj ,θ subsetj )=η(W H2 .η(W H1 .[x subsetj ,θ subsetj ]+b H1 +b H2 )
[0103] Overview: The specific formula of the feature scoring function is as follows:
[0104]
[0105] Among them, x i ,θ i 、x subsetj ,θ subsetj is the input feature vector, W G1 , W G2are the weight matrices of the first and second layers of the G model, respectively, and W H1 , W H2 are the weight matrices of the first and second layers of the H model; b G1 , b G2 are the bias vectors of the first and second layers of the G model, and b H1 , b H2 are the bias vectors of the first and second layers of the H model, respectively, and η is the activation function, G i (x i ,θ i ) and H j (x subsetj ,θ subsetj ) is the feature vector after being processed by the deep neural network.
[0106] Inter-layer transformation: First layer (hidden layer) transformation:
[0107] Where W G1 This is the weight matrix connecting the input features to the hidden layer, which determines the linear combination of the input features, b G1 This is the bias term of the hidden layer. η is used to adjust the activation of each neuron, introducing nonlinearity, enabling the network to learn complex patterns. The formula is:
[0108] Z G1 =W G1 .[x i ,θ i ]+b G1
[0109] H G1 =η(Z G1 )
[0110] Where Z G1 =W G1 .[x i ,θ i ]+b G1 is the result of linear transformation, H G1 =η(Z G1 ) is the result after applying the activation function and serves as the input of the next layer.
[0111] The second layer (output layer) transformation:
[0112] W G2 This is the weight matrix connecting the hidden layer to the output layer, b G2 is the bias term of the output layer,
[0113] The formula is:
[0114] Z G2 =W G2 .HG1 +b G2
[0115] G(x i ,θ i )=η(Z G2 )
[0116] Where Z G2 =W G2 .H G1 +b G2 is the linear transformation result of the second layer, G(x i ,θ i )=η(Z G2 ) is the final output after the activation function.
[0117] By analyzing the weight W of the first layer (hidden layer) G1 , the importance of the feature can be evaluated. The larger the absolute value of the weight, the greater the contribution of the feature to the model. i ,θ i The derivative of to evaluate the importance of features;
[0118] The subset interactions can be analyzed by analyzing the activations of the hidden layer H G1 Different combinations of features will have different effects on the activation of the hidden layer, which can reveal the interaction between features. The specific pattern of feature interaction can be identified by further analyzing the weights and activation patterns of the hidden layer.
[0119] Even if the evaluation function reaches an evaluation index improvement less than a preset threshold, the feature selection loop unit stops looping, that is, F(X,θ)≤F 标准阀值 (X,θ).
[0120] The warning module includes a warning light, F(X,θ)>F 标准阀值 (X,θ), the early warning module will issue an early warning, and the feature selection loop unit will continue to cycle.
[0121] The working principle of the present invention is as follows: the original data is collected by using a data collection unit; the collected data is X={X1, X2..., X n}, where X n Represents the nth data, n is a positive integer, and then the collected data is sent to the data preprocessing unit for processing. The data processing unit performs data preprocessing on the original data set collected by the data collection unit, including missing value filling, outlier processing and normalization. That is, data set preprocessing includes normalizing the features to ensure the scale consistency between different features. The normalization formula is: Where μ is the mean and σ is the standard deviation. The features are preliminarily screened based on statistics or simple machine learning models to remove features that are irrelevant or less relevant to mental health status. That is, the initial feature screening can use a composite model based on deep convolutional autoencoders and variational inference to achieve automatic feature extraction and screening of high-dimensional image data. The specific formula is as follows:
[0122] X filttered =CAE(X)+VI(X)
[0123] Where CAE(X) represents the feature extraction of input data X by the deep convolutional autoencoder, VI(X) is the process of optimizing feature screening by the variational inference model, and the composite model extracts meaningful feature representations based on the ability of the autoencoder, and further optimizes and selects the most relevant features through variational inference. The specific formula of the feature scoring function is as follows:
[0124]
[0125] The data is evaluated in a concluding manner. Even if the evaluation function reaches an evaluation index improvement less than the preset threshold, the feature selection loop unit stops looping, that is, F(X,θ)≤F 标准阀值 (X,θ), otherwise the warning module includes a warning light, F(X,θ)>F 标准阀值 (X,θ), the warning module will issue a warning, and the feature selection loop unit will continue to loop. The loop unit is: looping and executing the following loop steps:
[0126] S1: Construct a candidate feature set: randomly select a part of the features from the remaining features based on a specific strategy to form a candidate feature set;
[0127] S2: Feature scoring function: Define a complex feature scoring function F(X,θ) that takes into account the importance of features, interactions between features, and model prediction capabilities;
[0128] S3: Build and train the model: Build a mental health detection model based on the candidate feature set and train the model using training data;
[0129] S4: Evaluate the performance of the main model: Use the validation data to evaluate the trained main model and calculate the evaluation indicators;
[0130] S5: Update the candidate feature set: Update the candidate feature set according to the model performance evaluation results and feature scoring function, remove features with low scores, and add new potential features;
[0131] S6: Repeat the above process: Repeat steps S1-S5 until the stop condition is met.
[0132] Through this technical solution: the above-mentioned unit repeatedly detects the cycle steps to optimize the feature set and improve the accuracy and generalization ability of the mental health detection model. Compared with the traditional feature selection method, the method of the present invention can better capture the complex relationship between the data and the predictive ability of the model, thereby selecting a feature set with better predictive ability.
[0133] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.
Claims
1. A feature selection method for a mental health detection model, characterized in that: The following steps are involved: Step 1: a data acquisition unit, wherein the data acquisition unit acquires original data; Step 2: a data processing unit, wherein the data preprocessing unit performs data preprocessing on the original data set collected by the data collection unit, including missing value filling, outlier processing and normalization; Step 3: Perform a preliminary screening of features based on statistics or simple machine learning models to remove features that are irrelevant or have low relevance to mental health status; Step 4, feature selection loop unit, loops through the following loop steps: S1: Construct a candidate feature set: randomly select a part of the features from the remaining features based on a specific strategy to form a candidate feature set; S2: Feature scoring function: Define a complex feature scoring function F(X,θ), taking into account the importance of features, interactions between features, and model prediction capabilities; S3: Build and train the model: Build a mental health detection model based on the candidate feature set and train the model using training data; S4: Evaluate the performance of the main model: Use the validation data to evaluate the trained main model and calculate the evaluation indicators; S5: Update the candidate feature set: Update the candidate feature set according to the model performance evaluation results and feature scoring function, remove features with low scores, and add new potential features; S6: Repeat the above process: repeat steps S1-S5 until the stop condition is met; Step 5, output unit: comprehensively consider the importance of features, interactions between features, and model prediction capabilities. Step 6, an early warning unit, wherein the early warning module compares the calculated freight rate with the actual freight rate data and issues an early warning according to the pre-warning.
2. A feature selection method for a mental health detection model according to claim 1, characterized in that: The data acquisition unit includes a storage, X={X1, X2..., X n }, where X n It represents the nth piece of data, where n is a positive integer. The collected data is then sent to the data preprocessing unit for processing.
3. A feature selection method for a mental health detection model according to claim 2, characterized in that: The data preprocessing unit includes a processor, and the processor is used to analyze the data of the data acquisition unit, wherein the data set preprocessing includes normalizing the features to ensure the scale consistency between different features, wherein the normalization formula is: Where μ is the mean and σ is the standard deviation.
4. A feature selection method for a mental health detection model according to claim 3, characterized in that: The preliminary feature screening can use a composite model based on deep convolutional autoencoder and variational inference to achieve automatic feature extraction and screening of high-dimensional image data. The specific formula is as follows: X filttered =CAE(X)+VI(X) Where CAE(X) represents the feature extraction of input data X by the deep convolutional autoencoder, VI(X) is the process of variational inference model used to optimize feature screening, and the composite model is based on the ability of the autoencoder to extract meaningful feature representations and further optimizes and selects the most relevant features through variational inference.
5. A feature selection method for a mental health detection model according to claim 4, characterized in that: The deep convolutional autoencoder includes two parts, an encoder and a decoder, which are used to learn the compressed representation Z of input data X and reconstruct the input data.
6. A feature selection method for a mental health detection model according to claim 5, characterized in that: The encoder f encoder (X), the encoder usually consists of multiple convolutional layers and pooling layers, and its output is the feature representation Z of the latent space. The specific formula is: Z=ConvPool L (...ConvPool l (X) ConvPool l Represents the convolution and pooling operations of layer l. The decoder CAE -1 (Z), decode the latent representation Z into a reconstructed output using deconvolution and upsampling techniques The specific decoding process can be expressed as: Where DeconvUnpool l Represents the deconvolution and upsampling operations of layer l.
7. A feature selection method for a mental health detection model according to claim 6, characterized in that: The variational inference is used to estimate the posterior distribution of the latent variables and train the model by maximizing the variational lower bound, including the variational encoder, decoder and variational lower bound (ELBO). The variational encoder is specifically analyzed as follows: The variational encoder maps the input X to a distribution q of latent variables Z Φ (ZX), usually assuming that Z is normally distributed The formula is as follows: where μ Φ (X),σ 2 Φ (X) is the mean and variance of the encoder network output. The decoder p θ (XZ), the conditional distribution p that generates X from the latent variable Z θ (XZ), it is usually assumed that X obeys another normal distribution. The specific formula is as follows: Among them, μ θ (X),σ 2 θ (X) is the mean and variance of the decoder network output. The variational inference training optimizes the model by maximizing the variational lower bound, (ELBO), which is a lower bound on the log marginal probability of observing data X: Among them, qΦ(ZX) [logp θ (XZ)] is the reconstruction term (loss function), and the second term is the KL divergence, which measures the potential distribution q after encoding Φ (ZX) and the difference between the prior validation distribution p(Z).
8. A feature selection method for a mental health detection model according to claim 7, characterized in that: The specific formula of the feature scoring function is as follows: Among them G i , H i It is a complex deep neural network layer or recurrent neural network model for advanced computation of feature importance evaluation and subset interactions, α i , β j is the adaptive learning rate and model complexity parameter, n is the size of the candidate feature set, and m is the number of subsets.
9. A feature selection method for a mental health detection model according to claim 8, characterized in that: The G i , H i The specific analysis is as follows: G i (x i ,i i )=η(W G2 .η(W G1 .[x i ,i i ]+b G1 +b G2 ) H j (x subsetj ,i subsetj )=η(W H2 .η(W H1 .[x subsetj ,i subsetj ]+b H1 +b H2 ) Overview: The specific formula of the feature scoring function is as follows: Among them, x i ,θ i 、x subsetj ,θ subsetj is the input feature vector, W G1 , W G2 are the weight matrices of the first and second layers of the G model, respectively, and W H1 , W H2 are the weight matrices of the first and second layers of the H model; b G1 , b G2 are the bias vectors of the first and second layers of the G model, and b H1 , b H2 are the bias vectors of the first and second layers of the H model, respectively, and η is the activation function, G i (x i ,θ i ) and H j (x subsetj ,θ subsetj ) is the feature vector after being processed by the deep neural network.
10. A feature selection method for a mental health detection model according to claim 9, characterized in that: Even if the evaluation function reaches an evaluation index improvement less than a preset threshold, the feature selection loop unit stops looping, that is, F(X,θ)≤F 标准阀值 (X,θ).
11. A feature selection method for a mental health detection model according to claim 10, characterized in that: The warning module includes a warning light, F(X,θ)>F 标准阀值 The (X,θ) warning module will issue a warning, and the feature selection loop unit will continue to cycle.
Citation Information
Patent Citations
Mental health early warning method and system based on ensemble learning
CN116092678A
College student mental health state assessment model and early warning method based on campus behavior data analysis
CN116825348A
