For years now, Valve has been testing new approaches to filter the glut of steam games to applications that individual users are most likely to show their interest in. To this end, the company today deploys an "Interactive Reviewer" based on machine learning and trained on "billions of gaming sessions" from the Steam user base.
In the past, Steam relied heavily on crowd metadata, such as user-provided tags, user-driven lists, global review scores, and sales data to drive its recommendation algorithms. . But the new Interactive Reviewer is different, says Valve, because it works without any initial internal or external information about the games themselves (with the exception of the release date). "Instead, the model discovers the games themselves during the training process," says Valve. "The model deduces the properties of games by learning what users do, not by looking at other extrinsic data."
Your own reading history is an essential part of this model driven by a neural network. The number of hours you put in each game in your library is compared to that of millions of other Steam users. The neural network can therefore make "informed suggestions" about the types of games that might interest you. "The idea is that if players with very similar play habits to yours also tend to play another game that you have not tried yet, this game will probably be a good recommendation for you" , writes Valve.
This, in turn, should avoid the problems experienced by developers trying to play with the system by choosing popular tags or relying on positive reviews, as they did with previous recommendation algorithms. "The best way for a developer to optimize this model is to create a game that people love to play with," writes Valve.
Individual users can change the settings of these suggestions based on the AI in order to privilege games released during a certain period or those that fall on one side or the other on a gradient of "popularity". "We have found that, especially for people playing a lot of games, digging into the" niche "of the lineup can be a very effective way to find hidden treasures," writes Valve.
Valve admits that this machine learning system is not ideal for brand new games, which do not have enough players to collect data, resulting in a cold start effect between chicken and egg . Existing systems such as Steam's Discovery Queue should get these titles in front of their first readers, writes the company.
The new recommendation engine is accompanied by two other experimental products from the recently launched "Steam Labs" brand. A Micro Trailers page automatically generates six-second video thumbnails to represent the titles, sorted by genre, while The Automated Show will collect 30 minutes of footage from the latest versions of Steam for easy viewing.