Ferrivore

A website for idea discovery. Hosted at https://ferrivore.gerardkeiser.com.

Motivation

Ferrivore helps you find ideas that are interesting. Other services promise this, but their recommendation algorithms tend to show you things that you already agree with, from people you already follow. Ferrivore, on the other hand, has no user likes or follows; liking a post by a user does not automatically prioritize that user's other posts. It also encourages you to like ideas regardless of how much you agree with them, so that the algorithm can recommend all interesting ideas and not just sort you into a filter bubble.

Features and Usage

Upon login, users view two main pages: the homepage and a page with previously rated ideas.

On the homepage, users can view one idea at a time, and can either rate their agreement with the idea on a 0-7 scale, or dismiss the idea as uninteresting. Thereupon they can see how other users rated the idea, and request another idea. Users can select whether they want the next idea to be one they will probably agree or disagree with.

The first idea on the homepage will always be one predicted to cause disagreement.

On the ideas page, users can see a list of all ideas they have previously rated, and can also switch to ideas they have posted themselves. From there they can view additional details for a post, and also change their ratings or, if they have posted it, delete the post.

They can also edit their username, email, and password, as well as post their own ideas.

There is no simple way to check another user's posts. This is deliberate.

Implementation

The backend is a RESTful API written with the Flask framework and connected a Neo4j database.

The frontend is a React app using React Router and written with TypeScript. It authenticates with the backend using JWTs.

Currently there are only four options for a website visitor to get a new idea.

1. Random: Gets a random idea that the user has not interacted with already.
2. Popular: Gets the idea that the most global users have interacted positively with. It is not affected by the visitor's history.
3. Disagreeable: Takes all users adjacent to the visitor, calculates their Pearson similarity score, then finds all the ideas those users interacted with but the visitor hasn't, multiplies that user's similarity score by that user's agreement score to get a measure of how likely the visitor is to agree with that idea, adds them all together, and returns the idea with the lowest score (usually negative). For more explanation, the algorithm is very similar to the one used here. Internally, disagreement is measured with a negative number, so that a user with negative similarity score to the visitor who disagrees with an idea will increase the idea's agreement score with the visitor.
4. Agreeable: The same algorithm as the disagreeable option, but returning the highest score.

Install/Deploy

1. Start a Neo4j database. The default login details for testing and are in __init__.py in the backend, and the necessary constraints and indices can be found in set_db_properties in the seed.py file.
2. If deploying, make a .env file and edit the Dockerfile for the backend. This should be unnecessary if running locally with flask run.
3. The frontend connects by default to https://ferrivore.gerardkeiser.com/api; set the REACT_APP_BASE_URL to change this.

Upcoming Features

• Many improvements to the collaborative filtering system

• System to commend posts for being particularly rational or well-written, which boosts them in the recommendation system

• Reporting system to remove abusive posts

• Password recovery system

• Better styling, as always