Aesthetics-Guided Graph Clustering With Absent Modalities Imputation.

Area of Science:

• Computer Science
• Data Mining
• Image Analysis

Background:

• Clustering internet users by aesthetic styles is crucial for image modeling and data mining.
• Existing methods may struggle with multi-channel, potentially absent, user data.

Purpose of the Study:

• To develop a novel, partially supervised model for accurately clustering internet users based on photo aesthetics.
• To capture sparse representations of photo aesthetics by fusing multi-channel features.

Main Methods:

• A partially supervised model that fuses multi-channel features (gaze, quality scores, semantic tags), allowing for absent data.
• Utilizing KL-divergence to construct a large-scale graph representing user aesthetic correlations.
• Employing dense subgraph mining to detect aesthetic communities, accommodating outliers.

Main Results:

• The proposed method demonstrates superior performance on a million-scale Flickr image dataset.
• Successfully clusters internet users into distinct aesthetic communities.
• Identified aesthetic communities significantly enhance photo retargeting and video summarization.

Conclusions:

• The novel model effectively clusters users by aesthetic styles, outperforming existing approaches.
• The method's ability to handle missing data and outliers makes it robust for large-scale applications.
• Discovered aesthetic communities offer practical benefits for image and video processing tasks.