A few parameters are made clear through Steam's Interactive Recommender, which can be found in the Steam Labs:
Its description mentions how:
This experiment looks at how much you've played each game in your Steam library, and uses the magic of machine learning to recommend games you might like. Filter your results by picking games that are popular or niche, and drill down by release date and tags.
So it seems to consider playtime as a large factor (as it logically determines one's favourite games), and magic! ...of machine learning.
One the Steam page introducing this then new feature, it reads (my emphasis):
How It Works
The Interactive Recommender uses a machine learning model that is trained based on the playtime histories of millions of Steam users. It's not directly affected by tags or reviews — it instead learns about the games on Steam by looking at what users actually play. The basic idea is that if there are other players with similar play habits to you, who also play a game that you haven't tried yet, then that game is likely to be one you'll enjoy too.
We're also starting to apply the underlying model in other parts of the Steam store, where we think it can help players see the most relevant content or make more informed choices. For example, when viewing the page for a particular game, you may sometimes see "Players like you love this game" shown as a reason why the game is relevant to you, alongside other factors.
Not particularly clear, but a good start.