Personalized design aesthetic preference modeling: a variational autoencoder and meta-learning approach for...

Area of Science:

• Computational aesthetics
• Artificial intelligence in design
• Human-computer interaction

Background:

• Traditional design evaluation methods lack personalization and generalization.
• Existing approaches struggle with capturing nuanced individual aesthetic preferences.
• Need for adaptive systems in design and content curation.

Purpose of the Study:

• To develop a comprehensive framework for personalized design aesthetic preference modeling.
• To integrate variational autoencoders (VAE) with meta-learning for enhanced preference prediction.
• To enable robust generalization across diverse design domains.

Main Methods:

• Utilized variational autoencoders (VAE) for probabilistic modeling of aesthetic uncertainty.
• Employed meta-learning for rapid adaptation to individual user preferences with minimal data.
• Integrated multi-modal feature extraction (visual, textual, behavioral) using attention-based fusion.

Main Results:

• Achieved 84.7% accuracy in aesthetic preference prediction.
• Demonstrated a 70% reduction in adaptation requirements compared to traditional transfer learning.
• Validated superior performance across six diverse design datasets.

Conclusions:

• The proposed framework effectively models personalized design aesthetics.
• It successfully addresses challenges in personalization and cross-domain generalization.
• Provides practical foundations for intelligent design tools and automated content curation.