💻 computer science

Aesthetics Prediction of Websites using Federated Learning

This paper proposes a Federated Learning-based approach using EfficientNet-B0 to predict website interface aesthetics by incorporating individual user preferences across five demographic groups, achieving a promising Mean Squared Error of 0.0534 while overcoming the limitations of centralized data collection and average ratings.

Original authors: Ranjan Maity, Jahnavi Roy, Rahul Son Boro, Deepali Borgayary

Published 2026-10-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ranjan Maity, Jahnavi Roy, Rahul Son Boro, Deepali Borgayary

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

We live in a world where the look of a digital screen matters as much as what it does. Before a person even reads a single word on a website or app, their brain has already formed an opinion on whether it is pleasing or off-putting. This split-second judgment, often made in less than a blink, can determine if a user stays to explore or leaves immediately. For designers, this creates a difficult challenge: beauty is deeply personal. What one person finds elegant, another might find cluttered or boring. Traditionally, to understand these preferences, researchers have gathered thousands of people to rate images, then averaged the scores to find a "general" opinion. But this approach has a flaw. By blending everyone's tastes into a single number, it erases the unique preferences of individuals. It assumes that a student, a working professional, and a retiree all see the world in the same way, which is rarely true. Furthermore, collecting all these personal ratings in one central database raises concerns about privacy, as it requires users to hand over their private opinions to a central authority.

A team of researchers at the Central Institute of Technology in Kokrajhar, India, set out to solve these problems by building a new kind of system that respects individual taste without compromising privacy. They proposed a method called Federated Learning, a technique that allows computers to learn from data without ever moving that data from the user's device. Instead of gathering all the ratings into one giant pile, the system sends a smart computer program to the users' devices. The program learns from the ratings stored locally on each device, and only the lessons learned—the mathematical adjustments to the program—are sent back to a central server. The raw ratings, the personal opinions, never leave the user's phone or computer. To test this idea, the researchers created a dataset of one hundred website interfaces, capturing both desktop and mobile versions of fifty different sites ranging from government portals to shopping apps. They then asked one hundred people to rate these screens on a scale of one to ten. The participants represented a mix of ages, genders, and occupations, ensuring a wide variety of perspectives.

The researchers first needed to find the best computer brain to do the heavy lifting of recognizing visual beauty. They tested three different deep learning models, which are complex systems designed to mimic how the human eye processes images. After running extensive trials, they found that a model called EfficientNet-B0 was the most effective. It was fast, required fewer computing resources, and produced the most accurate predictions. However, the real innovation lay in how they handled the human element. The team realized that people do not use rating scales the same way. Some people tend to give high scores to everything, while others are naturally more critical. To fix this, they developed a way to adjust each person's ratings before the computer learned from them. They calculated how a specific person usually rates things compared to the group average and shifted their scores accordingly. This allowed the computer to learn the true shape of a person's preference, rather than just their tendency to be generous or harsh.

To see if this approach worked in a real-world scenario, the researchers divided the one hundred participants into five distinct groups based on their gender and occupation: female non-working, female students, female working, male students, and male working. Each group acted as a separate "client" in the Federated Learning system. The central server sent the initial EfficientNet-B0 model to each of these five groups. Each group trained the model on their own local data, learning what their specific demographic found beautiful. Once the training was complete, the groups sent their updated model weights back to the server. The server then combined these updates to create a new, smarter global model that understood the nuances of all five groups. This cycle repeated for thirty-three rounds. Throughout this process, the system never saw the raw ratings; it only saw the mathematical improvements made by each group.

The results showed that this privacy-preserving approach was highly effective. The final model achieved a very low error rate, meaning its predictions of how much a person would like a website were extremely close to the actual ratings given by the users. The system performed particularly well for groups with more similar tastes, such as female non-working participants, while still maintaining strong accuracy for groups with more diverse opinions. Crucially, the researchers found that the mobile versions of websites were rated almost identically to their desktop counterparts, suggesting that the principles of visual appeal hold true across different screen sizes. The study demonstrated that it is possible to build a system that learns from a wide variety of people without ever needing to collect their private data in a central location. By using a single global model that is adapted through a clever mathematical adjustment of ratings, the system can offer personalized aesthetic predictions to anyone, regardless of their background. This work suggests that the future of user interface design could rely on systems that are not only smart enough to understand beauty but also respectful enough to keep personal preferences private.

Technical Summary: Aesthetics Prediction of Websites using Federated Learning

Problem Statement
The visual aesthetics of user interfaces (UI) are critical to user satisfaction and engagement, yet aesthetic perception is inherently subjective and varies significantly across individuals due to factors such as age, gender, and occupation. Existing computational approaches to aesthetic prediction typically rely on centralized data collection and aggregate population-level ratings to produce a single "average" aesthetic score. This methodology fails to account for inter-individual variability and ignores personal preferences. Furthermore, traditional methods like surveys and heuristic evaluations are time-consuming, difficult to scale, and prone to human bias. While deep learning models have shown promise in image aesthetic assessment, they generally lack mechanisms to handle the non-IID (Non-Independent and Identically Distributed) nature of aesthetic data across diverse user groups without compromising privacy or requiring massive centralized datasets.

Methodology
The authors propose a privacy-preserving framework for personalized UI aesthetic prediction using Federated Learning (FL). The methodology proceeds through the following stages:

Dataset Construction:
- A custom dataset of 100 website interfaces was created, comprising 50 desktop and 50 mobile versions of the same sites across ten diverse domains (e.g., government, e-commerce, news).
- Screenshots were captured at standard resolutions (1920x1080 for desktop; 640x1136 for mobile).
- 100 participants rated these 100 interfaces using a 10-point Likert scale. The Latin Square method was employed to mitigate order bias.
- Statistical analysis (paired t-test) confirmed no significant difference in aesthetic ratings between desktop and mobile versions ( $p > 0.05$ ), allowing them to be treated as a unified dataset.
Base Model Selection:
- Three architectures were evaluated: EfficientNet-B0, ResNet-50, and Vision Transformer (ViT-B/16).
- EfficientNet-B0 was selected as the base model due to its superior performance, lower computational cost, and fewer parameters, making it suitable for edge deployment.
- The model was adapted for regression by replacing the final classification layer with a single linear unit to predict a continuous aesthetic score.
Data Preprocessing and Personalization:
- Image Preprocessing: Inputs were resized to 224x224, augmented with mild random flips, rotations, and color jitter, and normalized using ImageNet statistics.
- Rating Normalization (Key Contribution): To address subjectivity, a user-specific normalization strategy was implemented. Raw ratings were adjusted by removing the user's mean rating tendency and re-centering them around the global mean ( $r_{centered} = r_{raw} - \mu_u + \mu_g$ ). These values were then min-max scaled to the [0, 1] interval. This allows the model to learn relative preferences rather than absolute scores.
Federated Learning Architecture:
- Clients: The 100 users were grouped into five demographic clients based on the intersection of gender and occupation: Female Non-working, Female Student, Female Working, Male Student, and Male Working.
- Algorithm: The Federated Averaging (FedAvg) algorithm was used. A central server initializes the global EfficientNet-B0 model with ImageNet-1K pre-trained weights.
- Training Process: In each communication round, the server distributes the global model to the five clients. Each client trains locally for one epoch using its specific demographic data. Only the updated model weights (not raw data) are sent back to the server.
- Aggregation: The server computes a weighted average of the client weights, proportional to the number of training samples per client, to update the global model.
- Hyperparameters: Training utilized the AdamW optimizer ( $1 \times 10^{-4}$ learning rate), Mean Squared Error (MSE) loss, and Automatic Mixed Precision (AMP). Early stopping was triggered if validation loss did not improve for 10 consecutive rounds.

Key Results

Convergence: The global model converged after 33 communication rounds. The validation Mean Squared Error (MSE) decreased from 0.0702 in round 1 to a final 0.0534 on the normalized [0, 1] scale.
Performance Metrics: The final model achieved an RMSE of 0.2312 and an MAE of 0.1837.
Client-wise Variation: Performance varied by demographic group, reflecting internal preference diversity. The "Female Non-working" group achieved the lowest error (MSE = 0.0169), while the "Female Working" group showed the highest (MSE = 0.0890). The authors note that groups with higher internal diversity naturally yield higher error rates.
Error Distribution: Prediction errors were approximately normally distributed around zero, with the majority of errors falling within $\pm0.15$ on the normalized scale (roughly $\pm1.35$ points on the original 1–10 scale).

Significance and Claims
The paper presents a framework for personalized UI aesthetic prediction using Federated Averaging (FedAvg), an EfficientNet-B0 backbone, and a user-adaptive rating normalization pipeline. The system distributes the learning process across five demographic client groups, ensuring raw user rating data never leaves the local device. The authors demonstrate that a single global model can serve diverse users effectively without centralized data pooling. Personalization is achieved through the rating transformation pipeline (user-mean centering and demographic group bias re-injection) rather than maintaining separate local models.

However, the authors explicitly acknowledge a prominent limitation: the global model exhibits an inclination to predict scores concentrated around the dataset's average (approximately 0.5–0.6 on the normalized 0-to-1 scale). The current system relies exclusively on the rating transformation pipeline for personalization and does not perform client-specific fine-tuning or local head personalization after federated training.

Limitations
The authors modestly acknowledge a limitation: the global model tends to predict scores concentrated around the dataset's average (0.5–0.6 on the normalized scale). The current implementation relies solely on the rating transformation pipeline for personalization and does not perform client-specific fine-tuning or local head personalization after federated training.

Technical Summary: Aesthetics Prediction of Websites using Federated Learning

More like this