This is a visual recommendation system for movies. Movies are plotted on a graph, with more similar movies being plotted closer together.

Try searching for your favorite movie on the left side. Click on a search result to see nearby neighbors to your movie, or explore the graph on your own. You can also search by director or production company. Note that the movie dataset only goes up to 2017.

To move around, click and drag the graph. To zoom in, use the scroll wheel. To display movie info, click on a dot.

What is t-SNE?
t-SNE stands for t-distributed stochastic neighbor embedding, and is a technique for reducing high dimensional data to low dimensional data. In our original data, each movie is associated with 3,000 data points (rating data, high dimension), and is reduced via t-SNE to two cooridnates (x and y coordinates of the plot, low dimension).

The goal of this dimension reduction is to keep points which are close together and more similar together in a human-accessible plot. 3,000 dimensions are difficult for humans to comprehend; two are well-understandable.

What type of information is used to plot the movies?
Movie rating data by user is the only information which determines the coordinates of each point. If the same user rates two movies similarly, they become closer together, and if they are rated differently, they are bought further apart.

What makes this different from other movie recommendation engines?
Most movie reccomendation engines predict ratings, one value. Rather than relying on one value, this visualizer uses two numbers to try to capture movie rating and similarity. This allows the visualizer to capture more nuanced details than a simple predicted rating.

This visualizer was made by using rating data from the MovieLens dataset. Metadata is provided by The Movie Databse.

Two dimensional coordinates of movies were obtained by using t-distributed stochastic neighbor embedding (t-SNE) on rating data.

Source code is avalible here.

This visualizer is powered by the package scatter-gl.

Special thanks to the University of Minnesota, and The Movie Data (TMDb) base for providing rating data and metadata respectively.

This product uses the University of Minnesota's MovieLens dataset, but is not endorsed or certified by the University of Minnesota. This product uses the TMDb API but is not endorsed or certified by TMDb.

Each user index represents one user the original data used to build this model. Given a specific user, the graph is colored by what rating they gave a specific film, or grey if the film is unrated.
This information gives additional insight into the projections outputted by t-SNE, and the general trends which t-SNE captures.

We use an SVD model, one of the components of the BelKor solution to the Netflix Prize as a reference for our visualizer. This SVD model is optimized via RSME and has an RSME of 0.84 with a five-fold train-test split.
This allows the highlighting of differences between the t-SNE visualizer and conventional reccomendation engines. Numeric rating values are not neccessarily indicators of movie similarity - two dimensional movie projections allow the capturing of additional information, particularly similarity, for movie reccomendations.

An original space refers to the high dimensional data which a lower dimensional space reflects. The 3,000 rating values per movie is the original space, while the two x, y coordinates for each movie is the lower dimensional space. One of the goals of our dimension reduction is to keep movies in the original space close to their neighbors - nearest neighbors in the original space are provided for understanding this process.