What did you develop?

• We developed Terra Tinder, a Web App that matches users to suitable datasets. Our app will learn the users preferences through their likes and dislikes, meaning that the user will be shown datasets that fit their preferences.

How does it work?

• The backend is responsible for computing recommendations based on positive and negative feedback given by the user on previous datasets (i.e. swiping right or left). First, metadata on all ~30,000 NASA Earth Science Datasets was gathered via the associated API and saved in data/data.csv. Each dataset was mapped to a point in a 15-dimensional embedding space by using the word2vec package in R. This data was saved in a CSV file in data/embeddings.csv The backend server runs on Python. Upon startup, an adjacency matrix is calculated based on the cosine distance between all dataset embeddings. This adjacency matrix contains roughly 900 million entries and would be too large to commit to version control. Instead it is calculated once upon startup of the server and stored in memory. The server provides an API for the frontend for recommending datasets. Dataset recommendation is achieved using an algorithm that takes into account the previously liked datasets (which are supplied as part of the API request), and associated similarity scores to other datasets. Furthermore, the backend extracts keywords from the recommended dataset's title using NLP techniques in order to conduct a Bing image search to obtain a relevant image to display in the web app.
• The frontend is written in JavaScript using React. It establishes a connection to the backend server and uses the backend API to deliver information. 

Why is it important?

• Our application is important because it displays NASA Earth datasets in an accessible way so that all users (regardless of their scientific background) can access relevant data. This means that more users can be reached.

What do you hope to achieve?

• We hope that our application is helpful for users who are unsure of which dataset is for them. The NASA Earth datasets can be matched appropriately to users using our algorithms.
• We also hope that users can learn more about datasets as they swipe through different datasets.

Our project addresses this challenge in the following ways:

• Our application allows users todiscover new NASA Earth datasets, as the format of our app directs users towards datasets that are suitable for their needs. When a user "likes" a dataset, our Web App will then recommend similar datasets using similarity scores generated through word embedding. We are able to compare different datasets and assess their similarities. As users swipe, they can discover more about datasets with the 'See More' feature. They can also discover new datasets quickly and efficiently with the "like" button.
• Our application allows users tovisualize NASA Earth datasets in an innovative way. Our application uses dataset titles to fetch images from Bing Images. When users are "liking" or "disliking" datasets, the image search feature will allow them to visualize the NASA Earth datasets. We believe this is an accessible feature that will help users from a range of scientific backgrounds to understand and visualize NASA Earth datasets, as the images are clear and coherent with the dataset titles.
• Our application helps users find data relevant to their needs. We believe our application does this well, as we utilize natural language processing and word embedding to find new datasets based on those viewed by the user. This means that datasets shown are relevant to users needs.
• Our application allows users to analyse the data, but in a quick and efficient way. The format allows for quick analysis, as the 'See More' application allows for in depth understanding of each dataset, but the "like/dislike" format means that only datasets that are of interest of the user are looked into.
• Our application is user-friendly, as it is simple, accessible and and the format is efficient.