Discoverability has been Steam’s biggest problem for longer than anything else. Ever since Steam Greenlight ushered in an era of Lord of the Flies-esque mob law, an ever-increasing slice of the New Releases pie is dominated by early 90s edutainment software and full-on pornography. If you’re not in the market for children’s point and click games and/or anime sex simulators, finding a game that suits your taste can take more time and effort than most are willing to put in. Even yours truly only ends up on the Steam storefront when linked from Twitter or Discord, or when there’s a large sale happening. And in that second case, I’m mostly interested in what’s already on my wishlist, not adding new things to it.
Valve’s newest solution for this problem is machine learning, embodied in a new tool called the Interactive Recommender. Each user has their own Interactive Recommender, which can be found in the Steam Labs tab at the top of the store’s homepage. Once you’ve adjusted sliders for how “popular” or “niche” you want the results to be, and how old/new you want the results to be, the Interactive Recommender builds a list of 30 games for your perusal. You can also specify tags that the listed games must/must not have, which is helpful if you’re looking for a specific genre, or know that you’re definitely not in the market for, say, virtual reality games.
“Underlying this new recommender is a neural-network model that is trained to recommend games based on a user’s playtime history, along with other salient data,” Valve said. “We train the model based on data from many millions of Steam users and many billions of play sessions.”
The neural network ignores tags and review scores while learning, according to Valve; the only information that is explicitly fed to the machine is a game’s release date. “It turns out that using release date as part of the model training process yields better quality results than simply applying it as filter on the output.” The machine learns everything else on its own, “[disregarding] most of the usual data about a game, like genre or price point,” instead focusing on “what games you play and what games other people play” before making ” suggestions based on the decisions of other people playing games on Steam.” This, Valve hopes, will prevent the Interactive Recommender from becoming too insular, which might happen if it only recommended games similar to what the user was already playing.
Because the machine ignores metadata like tags, description text, and the like, developers need not optimize anything for the algorithm. “The best way for a developer to optimize for this model is to make a game that people enjoy playing,” says Valve. In theory, this prevents anyone from being able to game the system by optimizing their game’s storefront, description, achievements, or other malleable features to the network’s liking. The network only cares about who is playing what games, and how two things correlate.
Valve sees the Interactive Recommender as a supplemental discovery tool, rather than a replacement for existing systems, due to the “cold start” problem. Since the neural network only learns about games that people are already playing, brand-new games are ignored. “The model can’t recommend games that don’t have players yet, because it has no data about them,” Valve says. “It can react quite quickly, and when re-trained it picks up on new releases with just a few days of data. That said, it can’t fill the role played by the Discovery Queue in surfacing brand new content, and so we view this tool to be additive to existing mechanisms rather than a replacement for them.”
It’s not unlike Valve to take a “computers will solve all our problems” approach to an issue, and I think having a machine pick games is a lot better than offloading that labor onto the community. Under its default settings, my Interactive Recommender showed me a list of games that I have already heard of and, ultimately, passed on. This was not very useful. After moving both sliders as far into “new” and “niche” as they’d go, however, I was presented with a list of 30 games that I’d never heard of, and most of them genuinely seemed like they were my sorta deal. A couple of them now rest dormant in my wishlist, waiting for me to spend some fun money during a sale or moment of particular weakness. Mission accomplished, I guess!