Cross-Game Modeling of Player’s Behaviour in Free-To-Play Games
Andrej Vítek (Slovak University of Technology in Bratislava)
Introduction
While the free-to-play model allows players to try the game, it also creates a non-contractual setting between the game and a player. It is very common for players to leave the game early. This high amount of early churn causes problems for the player modelling tasks, as we have small amounts of data about a large portion of players.
In our work, we aim to explore how to model new players for which we have only a small amount of data available. We propose two approaches to reduce the problem of cold-start players. The first approach is focused on using information from multiple games. Our second approach is based on using additional information about other players of a given game.
Cross-game modelling
Almost every game is different from its design perspective, content and in play-style of players. As a result, created user models are often heavily context-dependent and it is difficult to create a general complex player’s model. Multiple studies have focused on the problem of cross-game modelling with various approaches as:
- Feature mapping – functions that transform feature set across games and are often combined with transfer learning
- Independent features – removing features that are dependent on the game
- Abstract features – using aggregation and transformation to create more general features
These techniques and their uses still show multiple problems and challenges. Most studies that focus on cross-game modelling still require extensive manual setup by an expert. Moreover, these approaches are still partially specific to a given group of games or don’t consider a player’s behaviour that is specific to a given game.
Feature mapping based on unsupervised translation
We propose an approach based on unsupervised transfer learning and translation. It is based on hypotheses, that sequences of actions are similar to sentences of words. Moreover, the behaviour patterns of users are similar to the language grammar and structure. As different games contain different actions, the problem of finding a mapping between them is similar to the problem of language translation. This approach is based on the following steps:
- Creation of action embedding for each game individually
- Using adversarial learning to learn the matrix, which can align action embeddings of different games into common subspace.
Combined group-based and individual modelling
This approach is based on future direction from a survey by Hooshyar et al. (2018) and consists of the following steps:
- Creation of group-based models representing new players
- Categorize player to a group based on his/her behaviour
- Update player models and reassign a player to better match his behaviour
As there are various segmentation strategies, in this approach we will explore which segmentation can represent the new players the best for the given task.
Initial results
Our initial work was focused on exploring the viability of cross-game approach by comparing game-dependent, game-independent and combined models on datasets of three different games. Used samples contained approximately 6500 players and 40 in-game actions per game. Models were compared by a churn prediction task.
For small amounts of data used (3 and 6 days of players history), the model with game-independent features has the best performance. Best results for all data (12 days of players history) are achieved by the model that combines game-dependent and game-independent features.
