TLDR: Medium Article

Inspiration

That may sound silly, I am very afraid of the Government/BigTechCompany controlling us with news recommendation. Alphabet donating insane amounts of money to the DNC, CNN and FOX being the fake news specifically targeted on us. Living in an authoritarian state Russia is, I see people being misinformed about the evergoing war in Ukraine and omnipresent corruption. But is the U.S. Any different? The way people want to give up their freedoms in exchange for government safety is alarming and lightweight decentralized approaches may help with the social awareness of the current world situation. RL is lightweight and can be implemented into smart contracts, which drastically reduces government/bigtech involvement into our newsfeed.

What it does

Reinforcement learning uses a fraction of computation power that other recommendation methods (matrix factorization) require. I am the first to popularize reinforcement learning for news recommendation on Medium and the first to create a simple yet powerful Pytorch API for Models, Datasets, and primarily Algorithms.

How I built it

It took 3 months and 2 articles to write this thing. I have an in-depth (7k words) explanation of this: https://towardsdatascience.com/reinforcement-learning-ddpg-and-td3-for-news-recommendation-d3cddec26011

Accomplishments that I'm proud of

Recently I celebrated 50 Github stars. The docs have launched today. Medum is about to get 600 claps in total.

What I learned

Natural Language Processing, Graph Networks, Reinforcement Learning, State Representation, Temporal Convolution, Echo State networks, Chaos Free RNNs, Probabilistic Programming with Pyro

What's next for recnn

Finish the library, add more algorithms, models, state representation, fully integrate with pyro, add different VAEs for batch constrained representation (FFJORD, Glow, Hyperspherical Variational Auto-Encoders) Write a colab demo, continue working on documentation, maybe integrate it with smart contracts in a different library,

Built With

Share this project:

Updates