(Summary from Perplexity Deep Research)
Summary:
Unveiling the Factors of Aesthetic Preferences with Explainable AI https://bpspsychub.onlinelibrary.wiley.com/doi/full/10.1111/bjop.12707
This paper explores the complex factors influencing aesthetic preferences in images using machine learning (ML) models and explainable artificial intelligence (XAI) techniques. The researchers pioneered a novel approach by utilizing various ML models to predict aesthetic scores based on aesthetic attributes, then analyzed these predictions using SHapley Additive exPlanations (SHAP) to understand which attributes most influence aesthetic judgments.
Research Methodology
The study employed four different ML models for regression tasks:
- Random Forest (ensemble learning with decision trees)
- XGBoost (eXtreme Gradient Boosting)
- Support Vector Regression (SVR) with Radial Basis Function kernel
- Multilayer Perceptron (neural network approach)
These models were tested on three different image aesthetic datasets:
- Aesthetics with Attributes Database (AADB): 10,000 images with 11 attributes
- Explainable Visual Aesthetics (EVA): 4,070 images with 4 attributes
- Personalized image Aesthetics database with Rich Attributes (PARA): 31,220 images with 7 attributes
The models were trained to predict the overall aesthetic scores using attribute ratings as inputs. The researchers then applied SHAP analysis to interpret the importance of different attributes in the predictions.
Key Findings
Model Performance
Support Vector Regression (SVR) consistently performed slightly better than other models across all three datasets, though all models demonstrated good predictive capability[1].
Influential Attributes
The SHAP analysis revealed different influential attributes across datasets:
- AADB: "Content" emerged as the most influential attribute, followed by "object emphasis" and "color harmony"[1]
- EVA: "Semantics" was most important, followed by "composition and depth"[1]
- PARA: "Quality" was the most influential attribute[1]
Attribute Interactions
The study identified interesting interactions between attributes. For example, in the AADB dataset, there were positive interactions between "balancing elements" and "content," as well as between "color harmony" and "content"[1].
Comparison with Linear Regression
The researchers compared their ML models with traditional linear regression and found that while linear regression performed well on some metrics, it wasn't always appropriate due to violated assumptions. For instance, the AADB dataset didn't meet the linearity assumption required for linear regression, making ML models more suitable[1].
Contributions and Implications
This study makes several significant contributions:
- It introduces ML models for regression to gain insights into aesthetic preferences in images[1]
- It provides the first detailed comparative analysis of various ML models within computational aesthetics[1]
- It pioneers the utilization of attribute information in image aesthetic benchmarks through a data mining approach[1]
- It presents the first application of the SHAP method in understanding image aesthetics[1]
The researchers argue that aesthetics is an excellent case study for AI applications in psychology because aesthetic preferences involve complex interactions between perception, cognition, and emotion. The study demonstrates how computational methods and XAI can enhance our understanding of psychological processes underlying aesthetic judgments[1].
Conclusion
This research offers a novel perspective on understanding image aesthetics by focusing on attribute scores and using ML models with XAI techniques. It emphasizes the importance of using appropriate models based on data characteristics and the value of examining multiple models to ensure consistent results. The findings contribute significantly to the field of computational aesthetics and demonstrate the potential of AI methods in psychological research[1].
- https://bpspsychub.onlinelibrary.wiley.com/doi/full/10.1111/bjop.12707
- Soydaner, D., & Wagemans, J. (2024). Unveiling the factors of aesthetic preferences with explainable AI. British Journal of Psychology. https://doi.org/10.1111/bjop.12707
Below is a detailed explanation of key terms and concepts from the summary:
Core Concepts in Aesthetic Preference Research
Explainable AI (XAI) refers to artificial intelligence systems designed to make their decision-making processes transparent and interpretable to humans.
In this study, XAI techniques like SHapley Additive exPlanations (SHAP) were used to identify which image attributes (e.g., color harmony, composition) most influenced aesthetic judgments.
SHAP quantifies the contribution of each input feature (attribute) to a model’s prediction, enabling researchers to understand complex relationships between variables.
Machine Learning (ML) Models are computational algorithms that learn patterns from data without explicit programming. The study tested four types:
- Random Forest: An ensemble method that combines multiple decision trees to improve prediction accuracy.
- XGBoost: A scalable gradient-boosting framework optimized for speed and performance.
- Support Vector Regression (SVR): A regression model that maps data into high-dimensional space to find optimal boundaries.
- Multilayer Perceptron (MLP): A neural network with interconnected layers of nodes, capable of learning non-linear relationships.
Datasets and Attributes
Aesthetics with Attributes Database (AADB) contains 10,000 images annotated with 11 attributes, such as "content" (subject matter) and "object emphasis" (visual prominence of key elements).
Explainable Visual Aesthetics (EVA) includes 4,070 images rated for four attributes like "semantics" (meaningfulness) and "composition".
The Personalized image Aesthetics database with Rich Attributes (PARA) comprises 31,220 images with seven attributes, including "quality" (technical excellence).
Statistical and Methodological Terms
Regression Tasks involve predicting continuous numerical outcomes (e.g., aesthetic scores) rather than categorical labels.
Linear Regression assumes a straight-line relationship between variables, which was contrasted with ML models in this study.
Attribute Interactions describe how combinations of features (e.g., "content" and "color harmony") jointly influence predictions, revealing synergistic effects.
Psychological and Computational Frameworks
Computational Aesthetics is an interdisciplinary field combining psychology, computer science, and art to quantify aesthetic experiences.
Aesthetic Scores are numerical ratings representing perceived beauty or appeal, often aggregated from human evaluators.
Psychological Processes in this context refer to cognitive and emotional mechanisms (e.g., perception, valuation) that underlie aesthetic preferences.
This glossary provides foundational knowledge to navigate the study’s exploration of how machine learning and XAI elucidate the drivers of aesthetic preferences. Each term is integral to understanding the methodological rigor and theoretical implications of the research.