Geometric enhanced molecular representation learning method for attribute prediction

Abstract

The invention relates to a geometric enhanced molecular representation learning method for attribute prediction, and belongs to the field of deep learning. According to the method, firstly, molecular data are modeled into voxel data, secondly, the number of voxel data channels is expanded through traditional convolution operation, so that information of all the channels can be interacted, then, an improved 3D convolutional neural network is used for multi-scale feature extraction, and finally, extracted geometric features are used for predicting molecular attributes. The deep learning method provided by the invention optimizes the performance of the convolutional neural network in molecular representation learning tasks on the premise of not introducing a great amount of calculation. Compared with other algorithms of the same type, the method shows better performance on a plurality of data sets.

Description

Technical Field

[0001] The present invention relates to a geometric enhanced molecular representation learning method for property prediction, belonging to the field of deep learning. Background Art

[0002] In the forefront of scientific research, the three-dimensional representation of molecules plays a central and indispensable role, which accurately and comprehensively depicts the spatial geometric configuration of molecules, laying a solid foundation for in-depth exploration of the physical and chemical properties of molecules. Given the crucial role of the three-dimensional representation of molecules in revealing the microscopic world of molecules, it is particularly urgent and important to efficiently and accurately extract the three-dimensional features of molecules. This process is not only a key step in understanding the relationship between molecular structure and function but also a prerequisite for making breakthroughs in fields such as molecular design, drug discovery, and material development.

[0003] In the research field of molecular representation learning, many outstanding models have been developed, among which methods such as DeepDrug3D, ATOM3D, and Drug3D-Net are particularly prominent. These methods use voxelization technology to successfully convert biomolecules and small molecule drugs into three-dimensional data forms and skillfully use 3D convolutional neural networks (3DCNNs) to extract features and predict properties of molecules. Research results show that considering the three-dimensional information of molecules can significantly improve the prediction accuracy and the generalization ability of the model in the task of predicting the properties of complex molecules. 3DCNNs play an indispensable and crucial role in deeply mining the three-dimensional structural features of molecules and accurately predicting molecular properties, having a profound impact on the progress of the field of molecular representation learning.

[0004] Regarding the improvement of convolutional neural networks (CNNs), recent research has mainly focused on two strategies: one is to build deeper models by increasing the number of network layers, and the other is to use larger convolutional kernels to expand the receptive field. For example, the ConvNeXt model has achieved excellent results with 7×7 dependency convolutions in image processing tasks. In addition, classic studies such as AlexNet and Inception v1 have also adopted large convolutional kernels as large as 11×11 and 7×7 respectively. Although stacking multiple layers of convolutional neural networks can gradually abstract the features of the input data layer by layer and build an increasingly complex representation system, excessive convolutional operations often lead to the loss of original information and bring high computational costs. Some studies have shown that retaining some original information is of positive significance for improving model performance in certain specific tasks. On the other hand, although expanding the size of the convolutional kernel can expand the receptive field of the model and thus capture more extensive context information, this method has the problem of a single feature extraction angle, and due to the density of convolutional calculations, expanding the convolutional kernel also leads to a huge consumption of computing resources. To address these challenges, InceptionNeXt has improved the 2D CNN. This improvement strategy decomposes the large convolutional kernel into four parallel branches along the channel dimension, including small square convolutional kernels, two orthogonal strip-shaped large convolutional kernels, and an identity mapping. This design not only effectively constructs a network structure with high throughput but also maintains excellent performance, providing a new idea for solving the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a geometric augmented molecular representation learning method for property prediction, aiming to optimize the model's ability to extract molecular data features by improving the traditional 3D CNN and improve the performance of convolutional neural networks in molecular property prediction.

[0006] To achieve the above object, the technical solution of the present invention is: a geometric augmented molecular representation learning method for property prediction, and the specific steps are as follows:

[0007] Step1: Model the 3D coordinate data of the molecule to obtain voxel data containing molecular geometric information;

[0008] Step2: Expand the number of channels of the molecular voxel data and fuse the features of each channel;

[0009] Step3: Extract the geometric features of the processed molecular voxel data;

[0010] Step4: Use MLP to perform final molecular property prediction on the extracted geometric features, including regression and classification tasks.

[0011] The specific content of Step1 is as follows:

[0012] Step 1.1: For the molecular data providing the 3D coordinates of the molecule, model it as voxel data;

[0013] Step 1.2: For the molecular data not providing the 3D coordinates of the molecule, simulate its 3D coordinates using molecular dynamics technology and then model it as voxel data.

[0014] The specific content of Step 2 is as follows:

[0015] Step 2.1: Using the obtained voxel data as input, expand the number of channels using ordinary 3D CNN and fuse the information of each channel to obtain the output after 3D CNN processing;

[0016] Step 2.2: Pass the obtained output as input to the CBAM attention module, and the CBAM attention module assigns different weights according to the importance of the channels by analyzing the input feature map.

[0017] The specific content of the CBAM attention module is as follows:

[0018] It includes a channel attention module and a spatial attention module;

[0019] The channel attention module first performs global information aggregation on each channel of the input 3D feature map through global average pooling operation, and then compresses and restores the channel features through two layers of 3D convolutional layers. Among them, the first convolutional layer compresses the number of channels from the original dimension to the number of channels of the output data of this layer, and the second convolutional layer restores the number of channels to the original dimension. Finally, a channel weight map is generated through the Sigmoid activation function, and the channel weight map is multiplied element-wise with the input feature map to achieve weighting of the channel dimension of the feature map;

[0020] The spatial attention module first generates a feature map with a depth of 1 through 3D convolutional operation to reflect the importance of each spatial position, then refines the spatial features through subsequent convolutional layers, then generates a spatial weight map through the Sigmoid activation function, and finally the spatial weight map is multiplied element-wise with the input feature map to achieve weighting of the spatial dimension of the feature map.

[0021] The specific content of Step 3 is as follows:

[0022] Step 3.1: Divide the input data into five parts according to the number of channels;

[0023] Step 3.2: Do not perform any convolutional operation on the first part after division and perform an identity mapping;

[0024] Step3.3: For the second part after division, perform multi-dimensional feature extraction using a 3×3×3 three-dimensional convolutional kernel to achieve three-dimensional feature representation in three orthogonal spatial dimensions of height, width, and depth;

[0025] Step3.4: For the third part after division, perform a convolution operation using a 1×5×1 strip-shaped convolutional kernel to construct a feature map in two dimensions of height and depth, while retaining the global context information of the width;

[0026] Step3.5: In the fourth part after division, deploy a 1×1×5 strip-shaped convolutional kernel for convolution processing to extract local feature patterns in the height and width dimensions, while maintaining the continuity of the depth dimension;

[0027] Step3.6: For the fifth part after division, perform a convolution operation using a 5×1×1 strip-shaped convolutional kernel to construct a feature correlation in the width and depth dimensions, while maintaining the structural information of the height dimension;

[0028] Step3.7: Concatenate the outputs of the above five parts as the extracted features;

[0029] Step3.8: Perform a pooling operation on the extracted features;

[0030] Step3.9: Use the CBAM attention module to assign different weights to the obtained features according to their importance to achieve weight allocation.

[0031] The specific content of Step4 is as follows:

[0032] Flatten the extracted geometric features as the input data of the multi-layer perceptron MLP;

[0033] When dealing with a regression task, configure the output layer as a single neuron, and the neuron directly outputs the predicted value without an additional activation function;

[0034] When dealing with a binary classification task, configure the output layer as a single neuron and use the Sigmoid activation function. This function maps the linear output to the interval (0,1) to generate a probability value P ∈ (0,1). This probability value represents the confidence that the input sample belongs to the target attribute class, and then compare the size of P with a preset threshold to complete the attribute prediction.

[0035] The beneficial effects of the present invention are as follows:

[0036] (1) Compared with many models based on convolutional neural networks, the present invention can extract geometric features of molecular voxel data at multiple scales, is applicable to both regression and classification tasks, and has strong universality;

[0037] (2) The present invention realizes the improvement of the traditional 3DCNN, enhancing the processing ability of the traditional 3D convolutional neural network for molecular voxel data;

[0038] (3) A modified 3DCNN deep learning method proposed by the present invention is used to enhance the performance of the molecular representation learning task. By optimizing the network structure, it effectively captures the three-dimensional features of molecules and is verified on multiple data sets with and without molecular geometric information, significantly improving the performance of the convolutional neural network in molecular representation. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is the flowchart of the steps of the present invention;

[0040] Figure 2 is a schematic diagram of modeling the doxorubicin molecular data as voxels in the present invention;

[0041] Figure 3 is the overall architecture diagram of the deep learning model proposed by the present invention;

[0042] Figure 4 is a schematic diagram of the working principle of the improved convolution module used in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] The present invention will be further described below in conjunction with the drawings and specific embodiments on the basis of the description.

[0044] The 3DCNN model has achieved remarkable results in molecular feature extraction and property prediction. However, the current convolutional neural network still faces challenges such as information loss, high computational cost, and single feature extraction angle when processing molecular data. To address these problems, the present invention proposes an innovative geometric enhanced molecular representation learning model, aiming to improve the traditional 3DCNN to optimize the feature extraction process, thereby improving the performance of the convolutional neural network in the molecular property prediction task.

[0045] Example 1: As Figure 1 shown, a geometric enhanced molecular representation learning method for property prediction, the specific steps are as follows:

[0046] Step1: Model the 3D coordinate data of the molecule to obtain voxel data containing molecular geometric information.

[0047] Step1.1: For the molecular data providing the 3D coordinates of the molecule, model it as voxel data;

[0048] Step1.2: For the molecular data not providing the 3D coordinates of the molecule, use molecular dynamics technology to simulate its 3D coordinates and then model it as voxel data.

[0049] Specifically, for a dataset that already contains molecular three-dimensional coordinate information (such as QM9, etc.), it is directly converted into voxel data for representation. For a dataset that does not provide molecular three-dimensional information (such as ESOL, etc.), the three-dimensional coordinates are predicted and constructed through molecular dynamics simulation methods, and then converted into voxel data. Subsequently, the voxel data is used as the input of the deep learning model.

[0050] Furthermore, Figure 2 Taking doxorubicin as an example, it intuitively shows the modeled style from the SMILES (Simplified Molecular Input Line Entry System) sequence to the voxel data model. Starting from the SMILES string of doxorubicin, the present invention uses professional functions in an advanced cheminformatics software library (such as RDKit) to accurately parse the string, and then constructs the corresponding molecular object. This parsing step deeply mines the SMILES string, accurately identifies and reconstructs various atoms, chemical bonds, and their intricate connection relationships in the molecule. Subsequently, to obtain the three-dimensional structure of the doxorubicin molecule, the present invention implements a molecular modeling process, which includes two core steps: molecular embedding and molecular optimization. In the molecular embedding stage, the present invention assigns an initial three-dimensional coordinate to each atom in the molecule, laying the foundation for the molecular spatial structure. Immediately afterwards, in the molecular optimization stage, the present invention uses an energy minimization algorithm to adjust these coordinates, aiming to obtain a molecular model with a more reasonable structure and higher stability. In this process, the present invention adopts a force field method, especially MMFF (Merck Molecular Force Field), to accurately simulate and calculate the internal interaction forces of the molecule. Finally, Figure 2 part (a) in shows the doxorubicin model represented by voxel data with a resolution of 16×16×16, and part (b) shows the doxorubicin model represented by voxel data with a resolution increased to 48×48×48. These two models with different resolutions not only intuitively reflect the spatial structure characteristics of the doxorubicin molecule, but also reveal the important influence of voxel resolution on the molecular representation accuracy and detail capture ability.

[0051] Step2: Expand the number of channels of the molecular voxel data and fuse the features of each channel.

[0052] Step2.1: Using the obtained voxel data as the input, a common 3DCNN is used to expand the number of channels and fuse the information of each channel to obtain the output after 3DCNN processing. This design can effectively fuse the atomic information in different channels. Through convolutional operations, the number of output channels can be expanded, thereby providing appropriate feature inputs for subsequent modules. It can be expressed by formula (1) based on the above theoretical basis.

[0053] (1)

[0054] Wherein, X represents the input voxel data, 64 represents the number of output channels of this layer, 3x3x3 represents the size of the convolutional kernel of this convolutional layer, Conv3D represents the convolution operation, represents the output obtained after passing through this convolutional layer.

[0055] Step2.2: Use the obtained output as the input and pass it to the CBAM attention module. The CBAM attention module analyzes the input feature map and assigns different weights according to the importance of the channels.

[0056] The CBAM attention module includes a channel attention module and a spatial attention module.

[0057] The channel attention module is expressed as:

[0058] Perform global average pooling on the input feature map to generate a feature map with a shape of (in_channels, 1, 1, 1), where in_channels represents the number of input data channels. This operation can be represented by formula (2).

[0059] (2)

[0060] Where , where B represents the batch size and C represents the number of channels. AvgPool3D represents the global average pooling operation, represents the output obtained after the global average pooling operation.

[0061] Then, through the convolution operation, compress the number of channels from in_channels to in_channels / / reduction, where reduction represents the compression ratio of the channels, and " / / " represents integer division. Use the non-linear activation function ReLU to increase the expression ability of the model. This process can be described by formula (3).

[0062] (3)

[0063] Where, represents the output after non-linear activation by the ReLU function, and in_channels / / reduction represents the number of output channels of this layer.

[0064] Restore the number of channels to the original in_channels, and the size of the output feature map is the same as the input. Generate a weight coefficient in the range of [0, 1] through the Sigmoid function, which is used to weight each channel of the input feature map. This process can be represented by formula (4), and the final channel attention map can be obtained after this process .

[0065] (4)

[0066] Among them, represents the channel attention map obtained through the above operations.

[0067] Multiply the obtained channel attention map element-wise with the input feature map to weight the importance of each channel, and finally obtain the feature map weighted by channels , and this process can be expressed as formula (5).

[0068] (5)

[0069] The spatial attention module is expressed as:

[0070] The spatial attention mechanism weights the spatial dimension of the input feature map to highlight important spatial regions and suppress irrelevant regions. Specifically, first, a feature map with a depth of 1 is generated through a 3D convolution operation to reflect the importance of each spatial position, then the spatial features are refined through subsequent convolutional layers, and finally a spatial weight map is generated through the Sigmoid activation function. This weight map is multiplied element-wise with the input feature map to achieve weighting of the spatial dimension of the feature map, thereby increasing the model's attention to key spatial regions. This process can be expressed as formula (6).

[0071] (6)

[0072] In the formula, A represents the finally obtained spatial attention map, ReLU and Sigmoid represent activation functions, Conv3D represents a 3D convolution operation, and 3×3×3 represents the convolution kernel size.

[0073] Multiply the feature map that has been weighted by channels element-wise with the spatial attention map to obtain the final weighted feature map , and this process can be expressed as formula (7).

[0074] (7)

[0075] Specifically, Figure 3Shows the schematic architecture of the deep learning model of the present invention. In this model, the input voxel data first flows through a traditional Convolutional Neural Network (CNN). The number of output channels of this layer is set to 64, aiming to fully capture the key information in the data. It should be noted that in the process of converting molecular data into voxel representation, the present invention encodes specific atomic types as independent channels. Therefore, this CNN layer not only performs the function of feature extraction, but also effectively mixes and integrates the information from different atomic channels, thereby enhancing the model's ability to understand the complexity of molecular structures. After preliminary feature extraction, the data further passes through a pooling layer to reduce the dimension and highlight the key features, and then enters a CBAM attention layer. The CBAM layer can adaptively focus on more critical information regions by introducing attention mechanisms in both channel and spatial dimensions, further improving the feature expression ability and the discriminative power of the model. The processed data is then fed into the core component of the feature extraction part, the multi-scale convolution module, that is Figure 3 the BMSC module in

[0076] Step3: Extract the geometric features of the processed molecular voxel data.

[0077] Step3.1: The input data is divided by channels and split into several parts. Specifically, the input tensor is divided into five parts. This process can be expressed by formula (8).

[0078] (8)

[0079] In the formula, Split represents the division function, represents the input of this layer, where, are the depth, height, and width of the input data respectively. respectively represent the five parts after dividing

[0080] Step3.2: For ​​​​The partial input does not undergo any convolution operation, aiming to protect the original information from being completely lost due to multiple convolutions.

[0081] Step3.3: For the partial input, perform a convolution operation using a convolution kernel of size . This is a standard 3D convolution operation, and the input data of this part is filtered in three dimensions: depth, width, and height. A cubic convolution kernel is used in this part. This process can be expressed by formula (9).

[0082] (9)

[0083] In the formula, represents the output obtained after convolving the partial input. In the present invention, has a value of 3.

[0084] Step3.4: For the partial input, perform a convolution using a strip-shaped convolution kernel of size . This process can be expressed by formula (10).

[0085] (10)

[0086] In the formula, represents the output obtained after convolving the partial input. In the present invention, has a value of 5.

[0087] Step3.5: For the partial input, perform a convolution using a strip-shaped convolution kernel of size . This process can be expressed by formula (11).

[0088] (11)

[0089] In the formula, represents the output obtained after convolving the partial input.

[0090] Step3.6: For the partial input, perform a convolution using a strip-shaped convolution kernel of size . This process can be expressed by formula (12).

[0091] (12)

[0092] In the formula, represents the output obtained after convolving the partial input.

[0093] Step 3.7: Concatenate the outputs of each branch to obtain the final output tensor. The concatenation operation is performed in the channel dimension. This process can be represented by Equation (13).

[0094] (13)

[0095] In the formula, is the output obtained by separately performing convolution on each input part and then concatenating them, represents the identity mapping of the original input information, and Concat represents concatenating the outputs of the above five parts in the channel dimension.

[0096] Step 3.8: Perform a max pooling operation on the output to reduce the number of parameters.

[0097] Step 3.9: Use the CBAM attention module to enhance the expression of important features by weighting the input feature maps, thereby optimizing the performance of the model.

[0098] Specifically, Figure 4 depicts the workflow of the improved 3D convolutional neural network, which aims to enhance the understanding ability of complex 3D molecular data through multi-scale feature extraction. After the data is input into this network layer, it is first divided into five independent parts according to the channel characteristics of the data. Immediately afterwards, for these five parts, the present invention respectively applies convolution operations with different-sized convolutional kernels independently. The core of this multi-scale convolution strategy lies in that convolutional kernels of different sizes can capture spatial features in different ranges, thereby more comprehensively revealing the diversity and complexity of the internal structure of molecules. In Figure 4 the next level, the sizes of the convolutional kernels used for each part are clearly marked. The NULL part in the figure represents the identity mapping, and no convolution operation is performed on the input of this part. This reflects the fine control of the feature extraction scale by the present invention. After each part undergoes independent convolution processing, respective feature map outputs will be generated. These outputs not only retain the key information of the original data but also incorporate the local spatial features extracted by the convolutional kernels. Finally, the present invention concatenates these feature map outputs from different parts in the channel dimension to form a composite output that combines multi-scale feature information. This concatenation operation not only integrates features of different scales but also retains the relative positions and correlations between them, providing a richer and more comprehensive feature representation for subsequent network layers.

[0099] Step 4: Use an MLP to perform regression and classification tasks for the final molecular properties on the extracted geometric features.

[0100] Specifically, the geometric features extracted in the above steps are flattened and used as the input data for the multi-layer perceptron (MLP). When dealing with regression tasks, the output layer is configured with a single neuron that directly outputs the predicted value without the need for an additional activation function. For classification tasks, the output layer is also set to a single neuron, and the sigmoid activation function is applied to convert the output into a probability value, thereby reflecting the likelihood of belonging to a certain class. Such a design enables the MLP to flexibly adapt to different prediction requirements. In this embodiment, the tasks involved include ESOL for predicting the solubility of organic compounds; FreeSolv for predicting the free energy of molecules in solution; QM9 for predicting quantum chemical properties based on molecular structure, such as molecular energy, polarizability, molecular orbitals, etc.; these three tasks are regression tasks. For the Tox21 compound toxicity prediction, this task belongs to the classification task. Table 1 shows the basic information of the datasets involved in the present invention. Tables 2, 3, 4, and 5 respectively show the performance of the deep learning method proposed in the present invention on each experimental dataset. The "-" in the table indicates that the model did not use this evaluation metric.

[0101] Table 1 Basic Information of the Datasets Used in the Present Invention

[0102] Table 2 Prediction Performance of the QM9 Dataset with MAE (Mean Absolute Error, the smaller the better)

[0103] Table 3 Prediction Performance of the ESOL Dataset with MAE and RMSE (Mean Absolute Error and Root Mean Square Error, the smaller the better)

[0104] Table 4 Prediction Performance of the FreeSolv Dataset with MAE and RMSE (Mean Absolute Error and Root Mean Square Error, the smaller the better)

[0105] Table 5 Prediction Performance of Each Task in the Tox21 Dataset

[0106] The specific implementation manners of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above implementation manners, and various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those of ordinary skill in the art.

Claims

1. A geometric enhanced molecular representation learning method for property prediction, characterized in that: Step1: Model the 3D coordinate data of the molecule to obtain voxel data containing molecular geometric information; Step2: Expand the number of channels of the molecular voxel data and fuse the features of each channel; Step3: Extract the geometric features of the processed molecular voxel data; Step4: Use MLP for the regression and classification tasks of the final molecular properties on the extracted geometric features.

2. The geometric enhanced molecular representation learning method for attribute prediction according to claim 1, wherein The specific content of Step1 is as follows: Step1.1: For the molecular data providing the 3D coordinates of the molecule, model it as voxel data; Step1.2: For the molecular data not providing the 3D coordinates of the molecule, use molecular dynamics technology to simulate its 3D coordinates and then model it as voxel data.

3. A geometric augmentation molecular representation learning method for attribute prediction according to claim 1, characterized in that, The specific content of Step2 is as follows: Step2.1: Use the obtained voxel data as input, expand the number of channels using ordinary 3D CNN and fuse the information of each channel to obtain the output after 3D CNN processing; Step2.2: Pass the obtained output as input to the CBAM attention module, and the CBAM attention module assigns different weights according to the importance of the channels by analyzing the input feature map.

4. A geometric augmented molecular representation learning method for property prediction according to claim 3, characterized in that, The specific content of the CBAM attention module is as follows: It includes a channel attention module and a spatial attention module; The channel attention module first performs global information aggregation on each channel of the input 3D feature map through global average pooling operation, and then compresses and restores the channel features through two layers of 3D convolutional layers. Among them, the first convolutional layer compresses the number of channels from the original dimension to the number of channels of the output data of this layer, and the second convolutional layer restores the number of channels to the original dimension. Finally, a channel weight map is generated through the Sigmoid activation function, and the channel weight map is multiplied element by element with the input feature map to achieve weighting of the channel dimension of the feature map; The spatial attention module first generates a feature map with a depth of 1 through 3D convolutional operation to reflect the importance of each spatial position, then refines the spatial features through subsequent convolutional layers, then generates a spatial weight map through the Sigmoid activation function, and finally the spatial weight map is multiplied element by element with the input feature map to achieve weighting of the spatial dimension of the feature map.

5. A geometric augmented molecular representation learning method for attribute prediction according to claim 1, characterized in that The specific content of Step3 is as follows: Step3.1: Divide the input data into five parts according to the number of channels; Step3.2: The first part after division does not perform any convolutional operation and performs an identity mapping; Step3.3: For the second part after division, use a 3×3×3 three-dimensional convolutional kernel for multi-dimensional feature extraction to achieve three-dimensional feature representation in three orthogonal spatial dimensions of height, width, and depth; Step3.4: For the third part after division, use a 1×5×1 strip convolutional kernel for convolution operation to construct a feature map in two dimensions of height and depth, while retaining the global context information of the width; Step3.5: In the fourth part after division, deploy a 1×1×5 strip-shaped convolutional kernel for convolution processing to extract local feature patterns in the height and width dimensions while maintaining the continuity of the depth dimension; Step3.6: For the fifth part after division, implement a 5×1×1 strip-shaped convolutional kernel for convolution operation to build feature associations in the width and depth dimensions while maintaining the structural information of the height dimension; Step3.7: Concatenate the outputs of the above five parts as the extracted features; Step3.8: Perform a pooling operation on the extracted features; Step3.9: Use the CBAM attention module to assign different weights to the obtained features according to their importance to achieve weight allocation.

6. A geometric augmentation molecular representation learning method for attribute prediction according to claim 1, characterized in that, The specific content of Step4 is as follows: Flatten the extracted geometric features as the input data of the multi-layer perceptron MLP; When dealing with regression tasks, configure the output layer as a single neuron that directly outputs the predicted value without an additional activation function; When dealing with binary classification tasks, configure the output layer as a single neuron and use the Sigmoid activation function. This function maps the linear output to the interval (0,1) to generate a probability value P ∈ (0,1). This probability value represents the confidence that the input sample belongs to the target attribute category. Then compare the size of P with a preset threshold to complete attribute prediction.

Images