A Whole New Map
We built the thing! Again! But better this time!
After several months of study, experimentation, and late nights we have finally forged a new Steam map that we feel is worthy of celebrating and we've seen some nice improvements in our core metrics! In this devlog I'm going to give a brief overview of the changes made. But first, for those who are new here, let's quickly provide a recap for those asking "what on earth is a Steam map?"
Steam Map? What are you on about?
For those new to Game Oracle let me provide a little TLDR - we built a map of the Steam marketplace. Using all available public information we've used some maths to chart how each game on Steam relates to each other, creating a 'map' of game similarity. Our map exists in a few different forms:
- The "full map" is a deep mathematical model that we provide our customers to search the market for existing games - our search tools are available in Data Explorer
- We use a condensed version to create clusters of market segments using an in-house clustering algorithm - this creates the market segments and ready-made analysis we serve up in Game Gap
- Our most condensed version is used to create a 2D map of market saturation, which is what you see in most of our imagery since it is the simplest to understand - this is the map you can explore in Saturation Map
Our most most complex map (left) is perfect for searching the market for games but is hard to visualise and is a bit too complex for analysing market segments. Our slightly more simplified map (middle) is perfect for clustering similar games together, enabling complex market analysis that can surface opportunities. The most simplified form of our map (right) lets us visualise market saturation in a single image.
As you move from left to right, from our deep map to our simplified map, we lose some information. This is the sacrifice we must make to enable visualising the market, but it unlocks an incredibly powerful tool for understanding the market, which is why we have been hard at work trying to improve the map - specifically, we have been developing a more reliable method of squeezing our deep mathematical map down into these more usable forms.
What's he building in there?!
We've been in the lab cooking, using something called "Deep Manifold Learning" to create the most reliable low-dimensional representations of the Steam marketplace. But you might ask, how on earth can you teach a model to map similar games on Steam? And how do you know if your model is correct? Well, we have been putting in the hours to create the most reliable data possible. Through careful curation, we have created a dataset of triplet pairs. What is a triplet pair? A triplet consists of an 'anchor', a 'positive' example, and a 'negative' example - we define an 'anchor' game that we compare to a similar game and a dissimilar game; we call these comparison games the 'positive' and 'negative', respectively.
The Sims 4 is similar to Paralives but is nothing like Phasmaphobia. A good model should understand this and place The Sims 4 closer to Paralives on the map and place Phasmaphobia far away from The Sims 4. Additionally, we would expect Paralives and The Sims 4 to live in the same cluster (market segment) in our Game Gap model, whereas Phasmaphobia should be in a completely different cluster.
If our model is performing well we can expect two things:
- The 'positive' game will be closer to the 'anchor' game on our map than the 'negative' game
- When we cluster games together on the map we expect the 'positive' game to be in the same cluster as the 'anchor' and the 'negative' game to be in a different cluster to the 'anchor'
We have used a variety of techniques to define our triplets of games:
- Using complex combinations of user defined tags we can find what games should be close to one another and what games should be far apart
- We can use information about popular video game franchises to define similar and dissimilar games
- Hours and hours of careful manual curation to ensure we have high quality data
The result was over 100k examples that we can use to train and evaluate the performance of our model.
So, how good is this model of yours?
Pretty good! After a lot of hard work we have a model that correctly arranges positive and negative pairs on the map for around 92% of our examples and that is when using our most simplified model that we use in Saturation Map. On top of that, our clustering results are extra impressive - when examining the clusters of anchor games, none of them are contaminated with negative examples and 70% contain the positive examples we classify as 'similar' to the anchor game.
As we transition from our old (left) map to our new shiny detailed map (right) we get an incredibly detailed view of the entire Steam market place - here in our Saturation Map each small square shows a collection similar games - bright red squares indicate many games and a high saturation, whereas blue squares mean there are fewer games and less competition.
The market "geography"
Before we wrap up this devlog, let's take a run around the new map and show you some of the cool sections of the market we can find on our Saturation Map.
Moving clockwise here is what we find:
- Mount Horror: this spooky mound consists of a massive bunch of games that all have the general theme of jump scares, psychological trauma, and outright terrifying experiences. In here you're going to find the likes of Layers of Fear, Alan Awake, and The Hat Man.
- Cat Island: a lot more wholesome on this little island of games! This cluster consists of cat-themed adventures and platformers. Here you will find everything from Stray to Super Catboy.
- Racing Archipelago: a large collection of different types of racing games with everything from the intense racing simulators of Assetto Corsa on one side of the archipelago, to the more comical Big Mutha Truckers 2 on the other.
- Spaceship Ridge: across this ridge of games expect to find all your space exploration games with the likes of Star Traders: Frontiers and Starship EVO.
- Tower Defence Peak: clumped together are your classic tower defence games like Bloons TD and Plants Vs Zombies, but also innovative newcomers like Thronefall and Nordhold.
- Dice Bay: in this corner we see a 'bay' of uniquely fun games all based around dice mechanics. In here, expect to find games like Dice With Death, Die In The Dungeon, and Astrea: Six-Sided Oracles.
There are so many more niches, more than we can invent geographical names for! However, we hope this offers a refreshing take on market research and an opportunity to literally get your hands on the data and discover what makes the Steam market tick.
