mklim / steamsuggest

Experimental recommendation engine for Steam.

GameRecommender

This is an experimental videogame recommendation program using free data collected from Steam.

There are two engines, one in Hive, and one in Pig. Each is in its own subdirectory. The wrapper functions are in Ruby.

Methods

Getting the Data

Steam has a free RESTful API that allows anyone to access user data, as long as they know the user's ID and the user in question has their player data publicly available.

A ruby script reads a seed ID and a recursion depth from a config.json file, and then collects the games a user has played starting with the seed user and then branching out n=depth levels for each of that user's friends. The script was only saves each user once, regardless of how many people in the social network are friends with that person.

For example: Bob is my seed ID. Bob is friends with Joe, and Joe is friends with Thomas. Thomas and Bob are not friends. If I set config.json to have a depth of 2, getData.rb will pull all of Bob's player data and all of Joe's player data. If I set config.json to have a depth of 3, I will also pull Thomas' data along with Bob and Joe. 4 would pull all of Thomas' friends as well, etc. A depth of 1 only pulls data for the seed. A depth of 0 pulls nothing.

Warning: each level of depth exponentially increases the amount of data pulled, the time it takes to retrieve it, and the time it takes to build recommendations off of it.

Data files are saved in JSON format for pig, and CSV for hive:

/data/<SEED-ID>-<RECURSION-DEPTH>.json
/data/<SEED-ID>-<RECURSION-DEPTH>.csv

General Recommendation Algorithm

1. Create a list of game pairs for every user. Only include pairs where the seed user has played at least 1 of the two games.

   Example pair: uid:123, id1:1, name1:Skyrim, hours1:30, userHrsPlayed: 5 id2:2, name2:TF2, hours2:60

2. Calculate a related score for this pair based on hours played for each game.

    score = max_hours_played - 
            (max_hours_played - min_hours_played)
   

3. Get average score for each game pair across all users.

4. For each pair, multiply the score by hours the seed user played the first game in the pair. (Add one to both to keep games played for less than one hour from shrinking in value compared to games played for over an hour.)

    recommendation = (1+score) * (1+userHrsPlayed)
   

5. Split the pairs, keeping the final score for both games, and filter out any games the user has already played.

6. Save a list of the unplayed games sorted by score descending.

Usage

1. Install Ruby, Pig and Hive.

2. Create a config.json containing a steam64 ID and a depth.

    {"seed": "76561197968575517", "depth": 2}
   

3. Build a dataset: ruby populateBaseData.rb

4. Run both engines. You can use the included script ruby runEngines.rb or run them from the command line with your own specific options.

5. ruby printResults.rb to output the results to the Terminal.

Limitations

Games are considered related if a single user played both of them. Games are considered to be more highly related if a user plays them both for a high number of hours. This is built on top of the (generally faulty) assumption that users only like one type of game, and puts short games at a disadvantage compared to long ones.

Each of these scripts was only tested on one laptop, and may or may not be compatible with other OS's or Pig/Hive setups.