Clustering Method Using Reference Data Initialization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing clustering methods face challenges in accurately determining the number of classes and initial cluster placement, especially when the number of classes is unknown or when clusters are partially overlapped, leading to potential empty classes and suboptimal clustering results.
Innovation Solution
The method determines an initial value of model parameters for an input data set based on a similar reference data set, which has been previously clustered, and modifies these parameters to match the input data set, using a combination of reference parameters from multiple similar data sets to improve clustering accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional clustering methods are used to classify data sets, then the clustering process can be performed, but the accuracy is reduced due to empty classes and suboptimal initial cluster placement
Solution Approach 1:
The patent applies preliminary action by using a reference data set to pre-determine optimal model parameters before clustering the input data set. The system stores previously clustered reference data and its optimal parameters, then retrieves and applies these parameters as initial values for the input data set, avoiding the need to start clustering from random initializations that may lead to empty classes or local optima.
Solution Approach 2:
The patent uses copying by replicating the successful clustering configuration from a reference data set to the input data set. The model parameters (number of classes, initial cluster centers, covariance matrices) that proved effective for the reference data are copied and adapted for the input data, providing a reliable starting point that improves convergence and accuracy.
2Adaptability or versatility
If the number of classes is unknown or clusters are partially overlapped, then clustering can still be performed, but the initial cluster placement becomes suboptimal leading to empty classes
Solution Approach 1:
The system performs preliminary determination of the number of classes and initial cluster configuration using a reference data set with similar characteristics. By analyzing the reference data's optimal clustering structure, the system pre-calculates appropriate model parameters that adapt to unknown or overlapping classes in the input data, avoiding random initialization issues.
Solution Approach 2:
The patent applies parameter changes by dynamically adjusting model parameters based on the reference data set's characteristics. The system modifies the number of classes, initial cluster centers, and covariance matrices to match the reference data's optimal configuration, thereby adapting to unknown or overlapping classes while maintaining accurate cluster placement.
3Ease of operation
If random initial values are used for model parameters, then the clustering algorithm can start, but it may converge to local optimum solutions
Solution Approach 1:
The patent eliminates random initialization by performing preliminary determination of model parameters using a reference data set. The system pre-calculates optimal initial values for cluster centers, covariances, and mixing coefficients based on the reference data's clustering structure, ensuring the algorithm starts from a globally optimal or near-optimal configuration rather than a random one.
Solution Approach 2:
The system implements feedback by using the known optimal parameters from the reference data set to guide the initialization of the input data set's clustering. The reference parameters serve as feedback information that informs and improves the initial configuration, creating a closed-loop approach where past successful configurations inform future clustering operations.
Data Source
AI summary
An information processing apparatus, a clustering method, and a clustering program stored on a recording medium, each of which determines an initial value of model parameter of an input data set based on a model parameter of a reference data set that is similar to the input data set and is previously clustered, modifies the initial value so as to match the input data set, and to obtain a clustering result of the input data set using the updated initial value of model parameter.