Popularity is not the same thing as a good recommendation
A RAWG gaming-data project raises a product-design question: which signals describe a player's fit, not just a game's reach?
A catalogue is not a taste model
A RAWG-Gaming-Data-ML project connects machine-learning exploration with game data. That is a rich product space: a catalogue can describe titles, but a recommendation has to reason about a particular player, moment and reason to choose.
Popularity is an easy proxy because it is visible and plentiful. It can also make a recommender repeat the same familiar titles, bury niche interests and confuse exposure with preference. The right signal depends on the job: discovery, similarity, replay, or a recommendation that fits a constraint the player actually named.
Make the comparison fair
Before training a model, define what counts as a relevant match and what a user should be able to inspect. Separate features available at recommendation time from information that appears later; otherwise the evaluation can quietly learn from the answer key.
Compare a simple popularity baseline with content-based and behavioural approaches. Evaluate beyond one average score: examine coverage, novelty and how results change for players with different histories. A recommendation system should explain its trade-offs, not turn one ranking into an objective truth.
Let discovery stay personal
A useful interface gives people ways to steer: “more like this”, “surprise me”, or “show why it fits”. Those controls create better feedback than a silent click count and give the model a clearer relationship with the person using it.