1. Introduction

Authors:
@LinxiaGong(巩琳霞), @XiaochuanFeng(冯小川), @DezhiYe(叶徳志)*, @LiHao(李浩)
@RunzeWu(吴润泽), @JianrongTao(陶建容), @ChangjieFan(范长杰), @PengCui(崔鹏)
*The work was done during the internship.

Keywords: Game Matchmaking, User Engagement, Graph Embedding

Long Story Short

This paper gives:

1. An generalized iterative two stage data-driven matchmaking framework, namely OptMatch, which has minimal product knowledge and data required, by utilizing only win/lose/score results of matches;

2. (stage 1) offline learning:
   • extracts two interpersonal relations for representing and understanding tacit coordination interactions among players;
   • learns the representation vectors to incorporate the high-order interactions;
   • trains a model(i.e., OptMatch-Net) to encode team-up effect and predict the match outcome;
3. (stage 2) online planning:
   • leverages the representation vectors of players and OptMatch-Net model to maximize the (predicted) gross utilities for the queuing players

Advantages:

• applicable to most of gaming products, fast and easy to implement
• minimal knowledge about the products and data required
• robust to data drift

Limitations:

• (for online games and e-sports) a hero can appear at maximun once in a team

It’s worth noting:

• this method is complementary to user portrait

2. Matchmaking System: OptMatch

2.1 System Overview

2.2 Offline Learning

2.2.1 Relationship Mining

2.2.2 OptMatch-Net

Secondly, a neural network (i.e. OptMatch-Net) is used to learn the intra-team interactions with a self-attention based Team2Vec Layer, and learn the inter-team interactions with a TeamComparison Layer.

Intra-team interaction (Team-up effect): each player can be beneficial/ disadvantageous to other team members

Inter-team interaction (Rock-paper-scissors effect): there might be no absolute measure to rank strengths of the teams

2.3 Online Planning

3. Experiments Details (Updating)

3.1 Datasets

» View the details and analysis of the datasets or view the detail page of the dataset through the dataset name (in the following table).

Dataset	Matches	Heroes	Players
Dota2 (5v5)	50,000	113	10,815
LOL (5v5)	623,263	145	/
LOL Championship (5v5)	187,588	139	43,706
NBA	3,342	/	949
Fever Basketball (3v3)	851,648	40	33,873

Dataset Split: Matches are sorted by the time for each dataset. Then we take the first 80% matches as the training set and the remaining 20% matches as the test set. Why? This ensures no leak of result information from the test set.

3.2 Codes structure

src
├── data_analysis
|   └── hero_combination_analysis.py
├── experiments
├── graph_embedding
│   └── ge_optmatch.py
├── models
│   └── optMatch-Net.py
├── requirements.txt
└── utils

3.3 Experiment Example

3.4 Online A/B Test Result

4. Discussion

4.1 Players’ Utility (Players’ Engagement)

Matchmaking aims to maximize the gross utility (i.e. satisfaction/engagement) of players.