Feature Weight Learning for Subjective Hierarchical Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional clustering techniques fail to reflect subjective similarity in multimedia content effectively, especially in hierarchical clustering where the number of clusters is not predetermined, and they require significant training data with invalid entries, making them inefficient for emotional content analysis.
Innovation Solution
An information processing apparatus that learns the weights of physical features using triplets with label information to determine the hierarchical structure, updating weights based on agreement with user-specified similarity, and iteratively refining the clustering process to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional constrained clustering techniques are used, then training data can be prepared with ML/CL pairs, but the technique cannot be applied to hierarchical clustering where the number of clusters is not determined beforehand
Solution Approach 1:
The patent changes the parameter of training data format from ML/CL pairs to triplet data with label information indicating the highest similarity pair. This parameter change enables the technique to work with hierarchical clustering where the number of clusters is not predetermined, as the triplet format can be processed at any level of hierarchy without requiring prior knowledge of the final cluster number.
2Adaptability or versatility
If semi-supervised clustering with XAB-type similarity data is used, then training data can be created without determining the number of clusters beforehand, but about 1/3 of the training data becomes invalid for evaluating clustering results
Solution Approach 1:
The patent extracts and utilizes only the valid portion of the triplet data where the label information specifies the highest similarity pair. By focusing on this extracted valid information, the method eliminates the 1/3 invalid data problem while maintaining the flexibility to handle any number of clusters in hierarchical structure.
Solution Approach 2:
The patent introduces feedback through label information that indicates the correct highest similarity pair in each triplet. This feedback mechanism allows the system to learn from correct answers and adjust the clustering algorithm to preferentially group similar items, thereby improving reliability without sacrificing flexibility.
3Measurement precision
If the number of training data is increased to improve clustering accuracy, then more training data can be processed, but the proportion of invalid data increases making learning more complicated
Solution Approach 1:
The patent converts the potential harm of large datasets with invalid data into a benefit by using the triplet format with label information. The label information acts as a guide that transforms potentially invalid comparisons into useful learning signals, allowing the system to learn from the structure of the data itself rather than requiring complex filtering or validation processes.
Data Source
AI summary
An information processing apparatus determines a weight of each physical feature for hierarchical clustering by acquiring training data of multiple pieces of content in triplets with label information indicating a pair specified by a user as having a highest degree of similarity among three contents of the triplet and executing hierarchical clustering using a feature vector of each piece of content of the training data and the weight of each feature to determine the hierarchical structure of the training data. The information processing apparatus updates the weight of each feature so that the degree of agreement between a pair combined first as being the same clusters among three contents of the triplet in a determined hierarchical structure and a pair indicated by label information corresponding to the triplet increases.