Sparse subspace clustering method based on selective coordinate descent optimization

Abstract

The invention discloses a sparse subspace clustering method based on selective coordinate descent optimization. The method comprises the following steps: 1, establishing a sparse subspace clustering model, substituting the sparse subspace clustering model into the Lasso formula, converting the sparse subspace clustering model into a quadratic programming problem, and solving a similarity matrix; 2, starting to solve the similarity matrix, traversing all features in the first time iteration by using a coordinate descent method, and using a calculated solution as an initial value; 3, starting from the second time iteration, traversing each feature, if the current solution item is non-zero, updating the feature of the coordinate position by using the coordinate descent method; otherwise, skipping the update of the feature of the coordinate position; repeating the process until the convergence of an objective function; and 4, after the similarity matrix is obtained, performing a spectral clustering process to obtain a classification serial number. According to the sparse subspace clustering method disclosed by the invention, whether the current solution item is zero or non-zero is quickly judged by using an infinite norm rule, thereby avoiding an unnecessary calculation process of a zero item result and shortening the time of optimization solution; and the effect on large-scale sparse problems is extremely obvious.

Description

technical field [0001] The invention belongs to the technical field of machine learning data processing, and in particular relates to a sparse subspace clustering method based on selective coordinate descent optimization. Background technique [0002] Clustering is one of the important issues in unsupervised machine learning research, and it has been widely studied and applied in image processing, data mining, social networks and other fields. However, in many practical problems, with the continuous and rapid increase of data dimensions, the so-called "curse of dimensionality" problem is becoming more and more prominent. If high-dimensional data is solved by traditional methods, the time complexity is often unbearable. Therefore, efficient modeling and calculation of high-dimensional data has become an important challenge and difficulty in data mining. [0003] Elhamifar & Vidal proposed a sparse subspace clustering model based on the self-expression property. The model us...

AI Technical Summary

Problems solved by technology

If the degree of sparsity of the solution is relatively high and the total dimension is relatively high, this way of traversing all the features obviously wastes a lot of time and calculation process

Images