Data Gathering
Our modeling dataset was built from the Steam Kaggle dataset, enhanced with external sources including SteamSpy, SteamCharts, Backloggd, and Metacritic.
In particular, using SteamSpy API we updated the following attributes:
- positive : number of postive reviews from Steam
- negative : number of negative reviews from Steam
- owners : number of estimated owners for each Steam game
- average_forever : average playtime in minutes per game
- median_forever : median playtime in minutes per game
- price : Steam current price
- initialprice : Steam launch price
- discount : discount percentage on Steam
- ccu : maximum number of concurrent player in the last 24 hours
- languages : list of languages in which each game has been translated
- genre : Steam genres associated to each game
- tags : Steam tags assigned to each game
From SteamChart, using Beautiful Soup, we scraped all-time peak which represents the maximum number of concurrent players for a game starting from the release date.
From Backloggd, using also Beautiful Soup, we scraped the following data:
- Played : number of times a game has been played
- Playing : number of current players
- Backlogs : number of times in which a game has been bought but not played even once
- Wishlist : number of times in which a game has been put in a wishlist
- Lists : number of times in which a game appears in user-made list
- Reviews : number of reviews for each game on Backloggd
- Likes : likes number
- Platforms : list of platforms on which a game has been released
Finally, from Metacritic, using Selenium, we scraped for each game
- Metascore : average critics rating
- User Score : average users rating
- Total number of reviews for critics and users
- Reviews from critic and users (with their score, date and platform)
The final dataset contained more than 90,000 games; however, after filtering it by considering only games with a Metacritic page, its size was reduced to around 5,000 games. This choice allowed us to consider only games with a certain market visibility, also taking care of a large amount of missing values present in the Kaggle dataset for less renowned games. Finally, considering we only had partial data available for 2024, we removed all games released in that year from the dataset. We also removed from the dataset games that, despite having a Metacritic page, did not have any user reviews on the platform (this reduced the datased to 4878 games).