Museum Art Recommender With Swipe Data

ISEF Category: Technology Enhances the Arts

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Most museum apps still recommend art like a bookstore recommends bestsellers. That misses what makes art personal. A piece that feels familiar to one visitor may feel random to another. Your project asks whether a smarter model can match people with art they actually want to keep exploring.

What Is It?

A recommender system is a tool that predicts what someone will like next. You already see this idea in movie and music apps. For this project, you would apply it to museum art. Instead of ranking only by popularity, your model would try to learn taste from swipe data, where visitors tap like or dislike on artworks.

The key idea is cultural proximity. That means artworks can feel closer to a visitor because of shared themes, style, region, time period, or visual patterns. Contrastive learning is a machine learning method that teaches a model to pull similar items closer together and push different items apart. Think of it like sorting photos into piles, then teaching the computer which pieces belong in the same pile and which do not.

Why This Is a Good Topic

This is a strong science fair topic because you can test a real prediction, not just build an app. You can compare a swipe-based model against a popularity baseline and measure whether people engage more with the recommendations. The project connects computer science, art access, and user behavior. You can learn data cleaning, feature design, model evaluation, and statistics without needing a wet lab.

Research Questions

• How does a swipe-based recommender compare with a popularity baseline for predicting which artworks visitors click or save?
• What is the effect of adding culture-related features, such as region, period, or medium, on recommendation accuracy?
• Does contrastive learning improve top-k engagement metrics compared with a standard collaborative filtering model?
• To what extent do recommendations based on visual features differ from recommendations based on metadata only?
• Which visitor signals, like like, dislike, dwell time, or repeat views, best predict later engagement with museum art?
• How does a culturally proximal recommendation list change diversity compared with a popularity-ranked list?

Basic Materials

• Laptop or desktop computer with internet access.
• Spreadsheet software for tracking labels and metrics.
• Python installed on a personal computer or school machine.
• Jupyter Notebook for cleaning data and testing models.
• Access to Met Open Access API or Rijksmuseum API documentation.
• Free image viewer or browser-based gallery for rating artwork samples.
• External storage or cloud folder for organizing downloaded metadata and images.

Advanced Materials

• Computer with a GPU, if available, for faster model training.
• Python environment with PyTorch or TensorFlow.
• API access to museum collection metadata and image records.
• Dataset storage for artwork images, metadata, and user interaction logs.
• Annotation tool for collecting swipe labels from test users.
• Statistics package for significance testing and confidence intervals.
• Version control system such as Git for tracking model changes.

Software & Tools

• Python: Cleans museum metadata, trains models, and calculates evaluation metrics.
• Jupyter Notebook: Lets you test ideas, plot results, and document code in one place.
• Pandas: Organizes artwork metadata and visitor interaction tables.
• scikit-learn: Builds baseline recommenders and scoring models.
• PyTorch: Trains a contrastive learning model on artwork features and swipe labels.

Experiment Steps

1. Define the recommendation task, the user action you will predict, and the baseline you will challenge.
2. Gather a museum artwork set with images, metadata, and enough labeled interactions to train and test your model.
3. Decide which features matter most, such as visual similarity, period, geography, medium, or artist metadata.
4. Build a baseline recommender first, then design a contrastive model that learns from like and dislike pairs.
5. Plan your evaluation metrics, including engagement, ranking quality, and diversity, so you can compare models fairly.
6. Test whether culturally proximal recommendations outperform popularity-based rankings for your chosen user group.

Common Pitfalls

• Using only popularity as a benchmark, which makes the new model look better or worse without a fair comparison.
• Mixing museum metadata from different sources without cleaning labels, which creates duplicate artists, periods, or object records.
• Training and testing on the same visitors, which inflates performance and hides overfitting.
• Collecting too few swipe labels, which makes the model chase noise instead of taste patterns.
• Measuring success only with accuracy, which misses whether the recommendations actually feel engaging or culturally closer.

What Makes This Competitive

A stronger project will do more than prove that a recommender can run. You should compare several model choices, use a clean baseline, and report more than one metric. If you test whether cultural proximity helps different visitor groups in different ways, your project becomes much stronger. Clear evaluation, careful labeling, and a smart fairness check can move this from a demo to real research.

Project Variations

• Use school gallery data or student survey swipes instead of museum API images to test the same model on a smaller, local art set.
• Compare visual-feature recommendations with text-only metadata recommendations to see which signal better predicts engagement.
• Test whether recommendations change when you group artworks by region, medium, or historical period instead of by popularity.

Learn More

• Met Open Access: Search the Metropolitan Museum of Art's collection API and open-access image records for artwork metadata and images.
• Rijksmuseum API: Search the Rijksmuseum developer documentation for collection data, object fields, and image access rules.
• TensorFlow Recommenders: A free library for building and testing recommender systems.
• PyTorch Tutorials: Official guides for building neural networks and contrastive learning models.
• scikit-learn User Guide: A free reference for baseline models, train-test splits, and evaluation metrics.
• PubMed: Search for review articles on recommender systems, user preference modeling, and cultural analytics.