This is our graduation project for the Bachelor's degree in computer science from Cairo University, Faculty of computers and AI.
As gaming community is growing bigger and bigger every day, the need for matching players together in online Lobbies is evolving as well.
From that a big problem arose which is that players with different personalities get matched up together, which results in a bad gaming experience for all of them.
So our aim is to help those and others in having a better gaming experience through an efficient Player Recommendation system that would improve the quality of teams.
To solve this problem and offer a better gaming experience for all the players with different personalities, we decided to make a platform that contains all the Players and Match them with the most suitable players for them based on their personalities.
The Player when signing up must take a personality test that will help us define his behavior and personality.
Using a Machine Learning Model (Unsupervised Learning techniques), we match the player with other players that are close in personality and also play the same games as well.
The website allows players to do the following:
- Sign Up and Login to the system
- Edit profile information and profile photo
- Add games and their handles in them in their profile
- Find peers playing a specific game
- Send a peer up request
- Accept or Decline a peer up request
- Review a peer (Updates the player's personality scores)
- Unpeer a peer
When a player requests to find a peer playing a certain game, the back-end application receives the request, retrieves the user's ID and the game's ID, and makes a request for the ML Model server, which in turn runs the ML Model with the specified parameters and returns a list of suggested player IDs which is responded back to the user in the form of profile tokens.
When a player reviews a peer, a list of 10 questions is selected randomly from the list of personality questions (same questions in the personality test but re-phrased for reviews).
And when a player submits a review, the peer's personality scores get updated for those questions based on the average score for all persons who reviewed this player.
This allows for more realistic personality scores for all players in order to improve the quality of the service.
Used Java and Spring Boot for building the back-end server using Spring Data and Spring REST to allow communication with the front-end, also responsible for communication with the Machine Learning model server to expose the peers suggestion service.
Built in Angular and TypeScript and communicates with the back-end using RESTful APIs.
The model itself is built using SKLearn (K-Means Clustering and K-Nearest-Neighbors), while the REST API server is built using Flask to expose the service to the back-end application.
Database used is MySQL and the database schema is generated using Spring Data model entities, and the ML Model also uses views to the database to access personality-related data and apply the matching service.
The datasets we've worked on during this project in order to build the model and test its effectiveness are:
Contains about 1 million records of users’ answers to the Big 5 personality test, which gives scores for a person’s personality traits like extraversion and agreeableness using 50 questions with answers from 1 to 5.
Contains 200K records of Steam (an online gaming platform) users, each record contains a player ID and Game name.
We’ve adjusted the datasets we worked on to meet our project’s needs according to the following:
We’ve selected only the features (questions) that can be relevant to our purpose in the project in the personality test dataset, and we further adjusted them to match our needs in the gaming platform, and this is the list of features we’ll be working with in the personality dataset:
And we also adjusted the questions to allow players to review each others in order to have a more realistic view of the players personalities:
And in the Gaming dataset:
1783 rows in the personality test dataset had missing values, which is a small number compared to the size of the dataset, and in these records, the whole rows had missing values not just one cell, so we simply removed these records.
The personality test dataset should contain values only between 1 and 5 inclusive, while some of the values (exactly 203108) had a value of 0 which should not be valid, so we mapped these values to 1 instead, as they correspond to the least value.
We used random sampling to sample the dataset to preserve the distribution structure of features, we first worked with a small sample (about 10,000 records) to match the number of users in the steam dataset.
We have tried multiple ML unsupervised learning techniques in order to know which one will serve us best throughout the project, such as Agglomerative clustering, K-Means clustering and K-Nearest-Neighbors, and we decided to:
Use a combination of both K-Means clustering and K-Nearest-Neighbors.
We first applied K-Means clustering using the personality dataset on the whole dataset of users in the system, and this step needs to be done periodically (every day for example), and when a user chooses to play a certain game, KNN is applied to the user’s cluster in order to sort them in ascending order of distance, and then we match the player with the closest players in his cluster that also play the same game.
This flowchart shows the approach for finding peers:
This project was created by:
Under the supervision of Dr. Soha Makady, Cairo University, Faculty of Computers and AI, Computer Science department.