IronDome

Why it exists

I wanted to learn machine learning and pattern recognition on a problem that mattered to someone, and a friend who buys Roblox games had one.

Studios look for “rockets”: games that already have everything Roblox’s recommendation algorithm looks for but haven’t been promoted by it yet. Buying one, or a share of one, cheaply before Roblox starts promoting it can return thousands of dollars. Finding them by hand doesn’t scale, and Roblox doesn’t publish a list of games to go through.

How it worked

Mapping games with “worms”

Roblox game ids aren’t numbered in order, so you can’t just count through them. IronDome started several crawlers, which I called worms, on random games it already knew about. Each worm asked Roblox for that game’s recommended games, and started a new worm on every one that wasn’t in the database yet. That mapped over 101,000 games.

Checking every game every hour

Once the map existed, several copies of the tracker recorded how each game changed over each hour. They spread their requests out and slowed down when Roblox pushed back, so they stayed within Roblox’s limits. Each game’s history went into a database.

Filling in history

A tracker that starts today knows nothing about last month. I filled in the past from a third-party site that had been tracking Roblox game stats for a while, so the model had long trends to learn from on day one.

Ranking

The model picked out the games whose player counts looked likely to get picked up by the algorithm soon.

What happened

My friend used its picks to decide what to buy, and it informed 3 purchases. When I was done, I handed the code over to them. Roblox has since changed its recommendation algorithm completely, partly because studios were doing this.

The code was transferred and is private, so everything here is from memory.