Steam recommender system using implicitly inferred ratings

How it works

This is just another recommender system. In fact, the main piece of work takes place before the recommender system is even initialised. In the data preprocessing segment, a dataset with users, game name, and hours played is imported. The idea was to convert this dataset into a dataset with users, games (or more generally items), and ratings which the recommender systems generally work with. This is why I used the hours played feature in order to deduce the implicit rating of a particular user for a particular game.

Implicit ratings

The implicit ratings are calculated using a formula that consists of two parts:

• calculating the joy_coefficient
• feeding the joy_coefficient to an activation function.

Joy coefficient

Joy coefficient of user u for the game g:

joy_coefficient_u_g = (hours_played_u_g - average_hours_played_g) / average_hours_played_g

Activation

The main idea of this part is to convert a continous value of the joy_coefficient into a value that ranges from 0 to 5.

A perfect function for that is the Sigmoid function, whose value approaches 1 as x approaches infinity, and also its value approaches 0 as x approaches negative infinity - it has a range of (0, 1). By multiplying the Sigmoid by 5, a function with a range of (0, 5) is obtained.

So finally, the rating of user u for a game g is obtained as:

rating_u_g = 5 * Sigmoid(joy_coefficient_u_g)

Example

From the dataset, a user with id 98649241 has played a game called Warframe for 83 hours. The average playtime for Warframe is 31.96 hours.

Calculating the joy_coefficient:

joy_coefficient = (83 - 31.96) / 31.96 = 1.596

The rating is then:

rating = 5 * Sigmoid(1.596) = 4.158

The user is playing the game more than average, and as a consequence the rating is quite high. This behaviour can be understood more clearly by taking a look at the plot below: https://drive.google.com/file/d/12krBswoap1c35m9Pp8pZfqxb1gWg5xIV/view?usp=sharing

In the above illustration you can also see that if the joy_coefficient is negative, the ratings will be in the range of (0, 2.5), and if the joy_coefficient is positive, the rating will be in the range of (2.5, 5).

Benefits of implicit ratings

• The system is not dependent on direct user feedback (ratings)
• Bias is generally decreased
• More ratings since the system is instantly aware of each users playtime for each game that they have bought, whereas with direct/explictit ratings the system is only aware of the user's rating for a game once they decide to actually rate it. Most of the users have played a lot of games, but have only left reviews for a handful.