Multi-view clustering method based on feature contrast learning

Abstract

The invention discloses a simple and effective multi-view clustering method framework based on feature contrast learning, and data of each view is fused in an original data stage, so that models are prevented from being customized for different views, and the risk of losing important information is reduced. In order to cope with noise and unavailable data in real data, two data enhancement modes are designed for fused data, and the model is helped to better adapt to an imperfect data environment. A deep shared encoder is constructed in the framework, comparative learning is performed through two types of data enhanced features, representative features are extracted, and the clustering effect is further improved. Through effective fusion and feature extraction of different view data, the accuracy and reliability of multi-view clustering can be remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of multi-view learning, and more specifically, relates to an unsupervised multi-view clustering method based on feature contrast learning. Background Art

[0002] In the field of multi-view clustering, significant progress has been made by leveraging the consistency and complementarity of multi-view data. Multi-view data usually contains information from different feature spaces or modalities, such as images, text, and structured data, providing diverse perspectives on the same object. Therefore, it has become an important field of unsupervised clustering analysis. Learning representative features from multiple views plays a decisive role in clustering performance. With the emergence of deep learning models, multi-view clustering methods based on deep learning have gradually become mainstream. Existing deep multi-view clustering methods can be summarized into two categories: feature fusion and feature consistency. Feature fusion first uses a non-shared encoder to learn features from different views, and then performs feature fusion. On the fused features, early studies directly applied traditional single-view clustering methods to multi-view clustering, such as subspace methods and spectral methods. Subsequent studies have incorporated weighting or attention strategies into the fusion module to quantify the importance of views. Summary of the invention

[0003] The purpose of this invention is to improve the existing deep multi-view clustering algorithm, provide a new multi-view clustering framework based on feature contrast learning, and propose data fusion and data enhancement methods for most graph data. With a small number of parameters, it not only retains the important information in the original data, but also improves the generalization ability of the model.

[0004] The unsupervised multi-view clustering method based on feature contrast learning in this paper transforms the information fusion of each view from the original feature-level fusion to the original data-level fusion, which not only retains the important information in the original data, but also reduces the influence of deep model redundancy. Noise addition and masking data enhancement techniques for most graph data are proposed to increase the robustness of the model against multi-view data noise and data unavailability in reality.

[0005] The present invention has the following beneficial effects:

[0006] 1) This paper proposes a novel deep multi-view clustering framework to process multi-view data by data-level fusion, solving the network customization and model redundancy problems in existing methods.

[0007] 2) A noise and mask enhancement technique for multi-view data is proposed, which enables the deep encoder to learn useful information from noisy and missing data.

[0008] 3) Carry out consistency contrast learning for the features of two types of data enhancement. No need to design a decoder, build reconstruction loss, reduce training parameters, and avoid model collapse. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a flowchart of a specific implementation method of the present invention based on feature contrast learning;

[0010] Figure 2 It is a framework diagram of the present invention based on feature contrast learning; DETAILED DESCRIPTION

[0011] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0012] Figure 1 FIG. 1 is a flowchart of a specific implementation of the multi-view clustering method based on feature contrast learning of the present invention. Figure 1 As shown in FIG. 1 , the specific steps of the multi-view clustering method based on feature contrast learning include:

[0013] S101: Data initialization:

[0014] The present invention needs to fuse the data of each view at the original data level, needs to normalize and scale each view row by row, and unify the scale of each view to facilitate fusion.

[0015] S102: Multi-view data enhancement:

[0016] In order to train the system to cope with real noisy data and unavailable data, the present invention proposes noise data enhancement and mask data enhancement for multi-view data. These two types of data enhancement can set a data enhancement ratio threshold.

[0017] Noisy multi-view data enhancement is designed by adding noise to certain views of each sample. The multi-view data noise enhancement formula is:

[0018]

[0019] Where p is the threshold parameter that controls the threshold of the noise data enhancement ratio in multi-view data, and σ is the parameter for the standard deviation in adding Gaussian noise.

[0020] Noise augmented data for the i-th sample of the v-th view The generation method is:

[0021]

[0022] δ i v is a random sampling of uniform distribution, and ò ... 2 ) is the noise randomly sampled in .

[0023] The missing multi-view data enhancement method is designed by setting the values ​​on some views of each sample to null. The multi-view data masking data enhancement formula is:

[0024]

[0025] Where r is the threshold parameter controlling the threshold of the missing data enhancement ratio in multi-view data. The generation method is:

[0026]

[0027] Let A∈{0,1} N×V A is a random indicator matrix designed to ensure that for each sample, data from different views can be selectively set to zero while ensuring that at least one view’s data point is not masked. Multiply the original data with the indicator matrix A to simulate the state of missing data.

[0028] S103: Multi-view data fusion:

[0029] The present invention realizes data fusion by horizontally splicing the multi-view data after data enhancement in sequence.

[0030]

[0031]

[0032] S104: Extract features from two types of enhanced data:

[0033] The two types of fused enhanced data are input into the shared deep encoding for feature extraction to obtain the corresponding representative low-dimensional features. and missing fusion data Input shared deep encoder E θ , perform feature extraction to obtain high-level features and

[0034]

[0035] S105: Comparative learning of the extracted features:

[0036] Advanced features for noisy data and advanced features of missing data Perform consistency contrast learning. If the number of training samples in one iteration is B, the features can form 2|B| sample pairs, where Positive sample pairs. For each high-level feature Can form (2|B|-2) negative sample pairs, and is a negative sample pair set, and S - Represents. The loss objective function composed of positive sample set and negative sample set is:

[0037]

[0038] Where τ is a hyperparameter and the distance between two high-level features is calculated by cosine similarity:

[0039]

[0040] The overall objective loss function of a batch of training process is as follows:

[0041]

[0042] S106, S107: number of iterations t <t max When the time comes, a new round of training begins;

[0043] S108: Output characteristics:

[0044] When the number of iterations t = t max When , the features are output and the final labels are obtained through k-means.

[0045] In order to verify the effectiveness of the present invention, we compared it with four existing deep multi-view clustering algorithms based on feature contrast learning, and conducted experiments on 6 public datasets to verify the effectiveness of the present invention. The datasets include the Berkeley Drosophila Genome Project (BDGP), the Handwritten Digits Dataset (DIGIT), the News Dataset (NGs), the Visual Object Classes Dataset (VOC), the Deep Human Action Video Dataset (DHA), and the Columbia Object Image Library (COIL-20). The four comparison methods are CPSPAN (Cross-view Partial Sample and Prototype Alignment Network), CVCL (cross-view contrastive learning), DSIMVC (Deep Safe Incomplete Multi-View Clustering), and DSMVC (Deep safe multi-view clustering). The evaluation indicators include clustering accuracy (ACC), normalized mutual information (NMI), and adjusted Rand index (ARI).

[0046] The present invention is implemented by PyTorch and trained on NVIDIA RTX3060TI GPU. Table 1 shows the experimental results of the present invention and the comparative method in the present invention.

[0047]

[0048]

[0049] Table 1

[0050] As shown in Table 1, the average performance of the present invention on 6 open data sets is better than that of the other 4 methods, which proves the superiority of the present invention.

[0051] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A multi-view clustering unsupervised learning method based on feature contrast learning, characterized in that: The following steps are involved: S1: Initialize the multi-view data, enhance the multi-view data and perform data fusion on each view data: S1.1: In multi-view data X m In each view, the data is normalized and scaled row by row; S1.2: Perform random noise and masking on the multi-view data according to probability to obtain the corresponding enhanced data X noise and X missing : Noisy multi-view data enhancement is designed by adding noise to certain views of each sample. The multi-view data noise enhancement formula is: Where p is the threshold parameter that controls the ratio of noise data enhancement in multi-view data, and σ is the parameter for the standard deviation of Gaussian noise addition. Noise augmented data for the i-th sample of the v-th view The generation method is: is a random sampling of uniform distribution, and ò is a random sampling of uniform distribution N(0,σ 2 ) is the noise randomly sampled in . The missing multi-view data enhancement method is designed by setting the values ​​on some views of each sample to null. The multi-view data masking data enhancement formula is: Where r is the threshold parameter controlling the threshold of the missing data enhancement ratio in multi-view data. The generation method is: Let A∈{0,1} N×V A is a random indicator matrix designed to ensure that for each sample, data from different views can be selectively set to zero while ensuring that at least one view’s data point is not masked. Multiply the original data with the indicator matrix A to simulate the state of missing data. S1.3: X in view order noise and X missing Perform horizontal splicing and fusion to obtain fused data and S2: Fusion of noise into data and missing fusion data Input shared deep encoder E θ , perform feature extraction to obtain high-level features and S3: Advanced features for noisy data and advanced features of missing data Perform consistency contrast learning. If the number of training samples in one iteration is B, the features can form 2|B| sample pairs, where Positive sample pairs. For each high-level feature Can form (2|B|-2) negative sample pairs, and is a negative sample pair set, and S - Represents. The loss objective function composed of positive sample set and negative sample set is: Where τ is a hyperparameter and the distance between two high-level features is calculated by cosine similarity: The overall objective loss function of a batch of training process is as follows:

Images