Depth subspace clustering method and system for realizing effective feature extraction

Abstract

The embodiment of the invention provides a depth subspace clustering method and system for realizing effective feature extraction, and the method comprises the steps: extracting an input sample feature expression through employing an auto-encoder, and obtaining an auto-encoder loss function through minimizing the error between an input sample and an output sample; realizing subspace learning by utilizing a self-expression module to obtain a self-expression module loss function; extracting features related to the clustering task from the potential representation through a supervised feature learning module to obtain a supervised feature learning module loss function; obtaining an optimization objective function according to the auto-encoder loss function, the self-expression module loss function and the supervised feature learning module loss function; and calculating an affinity matrix according to the obtained correlation coefficient matrix, and performing spectral clustering by utilizing the affinity matrix. According to the embodiment of the invention, features useful for clustering can be extracted, the influence of noise data is reduced, the clustering effect is improved, andthe problem of misalignment of labels is also avoided.

Description

technical field [0001] The present invention relates to the field of machine learning technology, in particular to a deep subspace clustering method for effective feature extraction Background technique [0002] Clustering is an important task in unsupervised learning for many practical applications. [0003] Among many models, subspace learning is one of the most commonly used methods. In the prior art, subspace clustering methods include: traditional methods such as low-rank representation clustering, deep subspace clustering based on autoencoder and self-supervision mechanism. [0004] However, feature extraction and clustering are performed separately in traditional methods such as low-rank representation clustering. Firstly, this will make the model unable to obtain an optimal solution, and secondly, there is no mutual promotion process between the two steps. The deep subspace clustering based on the autoencoder solves this problem to a certain extent, but the model s...

AI Technical Summary

Problems solved by technology

[0004] However, feature extraction and clustering are performed separately in traditional methods such as low-rank representation clustering. Firstly, this will prevent the model from obtaining an optimal solution. Secondly, there is no mutual promotion process between the two steps.

The deep subspace clustering based on the autoencoder solves this problem to a certain extent, but the model still cannot guarantee that the learned intermediate hidden layer representation is useful for the clustering task; the model based on the self-supervision mechanism will have inconsistent labels situation, not conducive to feature learning

Images