Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Across-Game Engagement Modelling via Few-Shot Learning

Kosmas Pinitas, Konstantinos Makantasis, Georgios N. Yannakakis

TL;DR

This work tackles domain generalisation for user engagement modelling in video games under limited data by proposing a multidomain few-shot learning framework. The method decomposes cross-game prediction into game-specific subproblems, using a relabelling function $R_Y(y,n)$ to create non-overlapping domain-specific labels and a representation backbone that feeds into episodic $N$-way, $K$-shot learning. Evaluated on the GameVibe Few-Shot (GVFS) dataset, built from 30 FPS games, the approach compares Prototypical Network, Matching Network, and Supervised Contrastive losses against a traditional end-to-end baseline, across multiple backbones and settings. Results show few-shot learners generally outperform the baseline, with SC often delivering the best accuracy, demonstrating that few-shot learning can robustly generalise engagement predictions across diverse games while reducing labelled data requirements. This framework has broad implications for scalable, domain-general affect modelling beyond games and can be extended to additional modalities and domains.

Abstract

Domain generalisation involves learning artificial intelligence (AI) models that can maintain high performance across diverse domains within a specific task. In video games, for instance, such AI models can supposedly learn to detect player actions across different games. Despite recent advancements in AI, domain generalisation for modelling the users' experience remains largely unexplored. While video games present unique challenges and opportunities for the analysis of user experience -- due to their dynamic and rich contextual nature -- modelling such experiences is limited by generally small datasets. As a result, conventional modelling methods often struggle to bridge the domain gap between users and games due to their reliance on large labelled training data and assumptions of common distributions of user experience. In this paper, we tackle this challenge by introducing a framework that decomposes the general domain-agnostic modelling of user experience into several domain-specific and game-dependent tasks that can be solved via few-shot learning. We test our framework on a variation of the publicly available GameVibe corpus, designed specifically to test a model's ability to predict user engagement across different first-person shooter games. Our findings demonstrate the superior performance of few-shot learners over traditional modelling methods and thus showcase the potential of few-shot learning for robust experience modelling in video games and beyond.

Across-Game Engagement Modelling via Few-Shot Learning

TL;DR

This work tackles domain generalisation for user engagement modelling in video games under limited data by proposing a multidomain few-shot learning framework. The method decomposes cross-game prediction into game-specific subproblems, using a relabelling function to create non-overlapping domain-specific labels and a representation backbone that feeds into episodic -way, -shot learning. Evaluated on the GameVibe Few-Shot (GVFS) dataset, built from 30 FPS games, the approach compares Prototypical Network, Matching Network, and Supervised Contrastive losses against a traditional end-to-end baseline, across multiple backbones and settings. Results show few-shot learners generally outperform the baseline, with SC often delivering the best accuracy, demonstrating that few-shot learning can robustly generalise engagement predictions across diverse games while reducing labelled data requirements. This framework has broad implications for scalable, domain-general affect modelling beyond games and can be extended to additional modalities and domains.

Abstract

Domain generalisation involves learning artificial intelligence (AI) models that can maintain high performance across diverse domains within a specific task. In video games, for instance, such AI models can supposedly learn to detect player actions across different games. Despite recent advancements in AI, domain generalisation for modelling the users' experience remains largely unexplored. While video games present unique challenges and opportunities for the analysis of user experience -- due to their dynamic and rich contextual nature -- modelling such experiences is limited by generally small datasets. As a result, conventional modelling methods often struggle to bridge the domain gap between users and games due to their reliance on large labelled training data and assumptions of common distributions of user experience. In this paper, we tackle this challenge by introducing a framework that decomposes the general domain-agnostic modelling of user experience into several domain-specific and game-dependent tasks that can be solved via few-shot learning. We test our framework on a variation of the publicly available GameVibe corpus, designed specifically to test a model's ability to predict user engagement across different first-person shooter games. Our findings demonstrate the superior performance of few-shot learners over traditional modelling methods and thus showcase the potential of few-shot learning for robust experience modelling in video games and beyond.
Paper Structure (12 sections, 6 equations, 3 figures, 3 tables)

This paper contains 12 sections, 6 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Method
  4. Problem Setting
  5. Representation Components
  6. Learning Objectives
  7. Experiments
  8. The GameVibe Few-Shot Dataset
  9. Experiment Protocol
  10. Results
  11. Discussion
  12. Conclusion

Figures (3)

  • Figure 1: Illustration of a classification problem---viewed as 2D plots of projected embeddings---containing three domains $D$ (yellow, magenta, blue) and two classes $C$ (green, pink). Such a problem is the prediction of user engagement classes (high vs. low) across several games (domains). Plot (a) showcases that the two classes cannot be easily separated across the entire dataset of domains. Plot (b) shows that points within the same domain are clustered together independently of their class. The rightmost plot (c) illustrates the method introduced in this paper combining both class and domain information: the method results in isolated domains allowing for classification within these more homogeneous groups.
  • Figure 2: A problem with 3 domains (yellow, red, green) and 2 classes ($C_0$ and $C_1$) per domain. $S$ and $Q$ represent the support and query sets, respectively. We first extract embeddings using a pre-trained frozen feature extractor. Following this step, we pass the extracted embeddings through a trainable projection layer and perform $L2$ normalisation. Finally, we optimise the few-shot losses using the resulting $S$ and $Q$.
  • Figure 3: Screenshots from the 30 different FPS games annotated for engagement. List of game titles: (1) Apex Legends; (2) Battlefield 1942; (3) Blitz Brigade; (4) Borderlands 3; (5) Corridor 7; (6) Counter-Strike 2016; (7) Counter-Strike 2018; (8) Counter-Strike 2019; (9) Counter-Strike: Global Offensive; (10) Doom; (11) Dusk; (12) Far Cry 1; (13) Fortnite; (14) Heretic; (15) Hrot; (16) Insurgency; (17) Modern Combat: Sandstorm; (18) Medal of Honor 2010; (19) Medal of Honor 1999; (20) Medal of Honor: Pacific Assault; (21) Operation Bodycount; (22) Outlaws; (23) Overwatch 2; (24) PUBG; (25) Superhot; (26) Team Fortress 2; (27) Void Bastards; (28) Wolfenstein 3D; (29) Wolfenstein New Order; (30) Wolfram Wolfenstein.