PlayerEmbed: Learning Behavioral Fingerprint Representations from Raw Input Sequences

Tobias Schneider; Lorenz Sparrenberg; Priya Tomar; Shahzeb Qamar; Tobias Deußer; Rafet Sifa

cs.LGOct 23, 2026

PlayerEmbed: Learning Behavioral Fingerprint Representations from Raw Input Sequences

Authors: Tobias Schneider, Lorenz Sparrenberg, Priya Tomar, Shahzeb Qamar, Tobias Deußer, Rafet Sifa

Organizations: Fraunhofer IAIS, Sankt Augustin, Germany · University of Bonn, Germany · Lamarr Institute for Machine Learning and Artificial Intelligence, Germany · PhenoRob Cluster of Excellence, University of Bonn, Germany

Abstract

Raw player input sequences in video games contain fine-grained behavior, but are high-dimensional and noisy, which makes large-scale analytics challenging. To address this, we introduce PlayerEmbed, a framework that learns compact and discriminative player embeddings using deep metric learning. Our approach uses a hybrid CNN-LSTM encoder trained on raw input sequences from 1,000 players. We evaluate the model on held-out sequences from these players and on sequences from 1,000 players entirely unseen during training to assess generalization to unseen identities. For player identification, a nearest-centroid classifier built from reference embeddings achieves strong performance, reaching balanced accuracies of 0.950 for players seen during training and 0.879 for players unseen during training. Systematically increasing the number of training players substantially improves generalization to unseen players, indicating a transition from simple fingerprinting to learning transferable behavioral patterns.

Figures & tables

Fig. 1: Flowchart of how the PlayerEmbed framework can be used for player identification using nearest-centroid classification.

Config	$M$	Centroid	Medoid	1-NN	5-NN	9-NN
K $\to$ K	1	0.790 $\pm$ .007	0.720 $\pm$ .010	0.742 $\pm$ .008	0.743 $\pm$ .008	0.740 $\pm$ .008
K $\to$ K	3	0.931 $\pm$ .004	0.889 $\pm$ .005	0.889 $\pm$ .004	0.886 $\pm$ .004	0.881 $\pm$ .004
K $\to$ K	6	0.950 $\pm$ .003	0.920 $\pm$ .004	0.917 $\pm$ .004	0.913 $\pm$ .004	0.908 $\pm$ .004
KU $\to$ U	1	0.621 $\pm$ .008	0.520 $\pm$ .008	0.599 $\pm$ .007	0.598 $\pm$ .007	0.601 $\pm$ .007
KU $\to$ U	3	0.833 $\pm$ .004	0.728 $\pm$ .006	0.775 $\pm$ .005	0.782 $\pm$ .005	0.783 $\pm$ .004
KU $\to$ U	6	0.879 $\pm$ .004	0.792 $\pm$ .007	0.827 $\pm$ .005	0.832 $\pm$ .005	0.831 $\pm$ .005

TABLE I: Balanced accuracy of classification methods across evaluation configs and averaging levels

M

for

n=1000

		Balanced Accuracy		Macro F1
Config	$n$	$M=1$	$M=6$	$M=1$	$M=6$
K $\to$ K	100	0.8990 $\pm$ .009	0.9998 $\pm$ .000	0.8985 $\pm$ .009	0.9998 $\pm$ .000
K $\to$ K	200	0.8993 $\pm$ .004	0.9996 $\pm$ .000	0.8977 $\pm$ .004	0.9996 $\pm$ .000
K $\to$ K	400	0.8900 $\pm$ .006	0.9967 $\pm$ .001	0.8825 $\pm$ .006	0.9964 $\pm$ .001
K $\to$ K	600	0.8564 $\pm$ .007	0.9851 $\pm$ .004	0.8407 $\pm$ .008	0.9826 $\pm$ .005
K $\to$ K	800	0.8281 $\pm$ .009	0.9704 $\pm$ .005	0.8015 $\pm$ .009	0.9641 $\pm$ .006

TABLE II: Balanced accuracy and macro F1 for nearest-centroid classification across evaluation configurations, numbers of known training players

n

, and averaging levels

M

	$n=100$		$n=1000$
Method	$M=1$	$M=6$	$M=1$	$M=6$
MLP	0.5522 $\pm$ .0060	0.7633 $\pm$ .0068	0.2513 $\pm$ .0015	0.3946 $\pm$ .0082
SVM	0.4407 $\pm$ .0024	0.7951 $\pm$ .0077	0.1332 $\pm$ .0013	0.3549 $\pm$ .0042
Centroid	0.3197 $\pm$ .0040	0.7566 $\pm$ .0072	0.1226 $\pm$ .0006	0.3701 $\pm$ .0056
XGBoost	0.5418 $\pm$ .0049	0.7124 $\pm$ .0055	0.2200 $\pm$ .0026	0.3148 $\pm$ .0053
PlayerEmbed	0.8990 $\pm$ .0090	0.9998 $\pm$ .0000	0.7899 $\pm$ .0070	0.9503 $\pm$ .0030

TABLE III: Balanced accuracy of traditional baselines compared to PlayerEmbed . Bold marks the best baseline.

Model	K $\to$ K		KU $\to$ U
	$M=1$	$M=6$	$M=1$	$M=6$
CNN-LSTM	0.790 $\pm$ .007	0.950 $\pm$ .003	0.621 $\pm$ .008	0.879 $\pm$ .004
Transformer L=3	0.698 $\pm$ .005	0.931 $\pm$ .002	0.385 $\pm$ .010	0.785 $\pm$ .006
CNN-LSTM ${}_{\text{no head}}$	0.702 $\pm$ .017	0.921 $\pm$ .009	0.524 $\pm$ .018	0.846 $\pm$ .009
Transformer L=2	0.681 $\pm$ .004	0.921 $\pm$ .002	0.444 $\pm$ .009	0.803 $\pm$ .003
LSTM	0.680 $\pm$ .027	0.914 $\pm$ .009	0.506 $\pm$ .020	0.830 $\pm$ .012

TABLE IV: Balanced accuracy for

n=1000

on known and unknown players across averaging levels

M

. Rows are ordered by known-player performance at

M=6

Fig. 2: UMAP visualization of player centroids for $K_{Test}$ and $U_{Test}$ with HDBSCAN cluster assignments.

Representation Learning

Representation learning develops data transformations that expose useful structure for prediction, transfer, generation, and control.

See all 799 papers

BehaviorFormer: Discovering Behavioral Patterns via Clusterable Next-Token Prediction
Oct 12, 2026Tobias Schneider, Danial Fahrner, Soren Antebi +4
Fraunhofer IAIS, Sankt Augustin, Germany · University of Bonn, Bonn, Germany
Beyond a Single Crop: Contextualized Embeddings for Player Re-Identification
Oct 9, 2026Ge Luo, Changhuo Yang, Weisi Fan +2Person Re-Identification Sports Video Analysis
Central South University · Iowa State University · Shandong University
Beyond Asking: A Pipeline for Personalized Game Generation that Reads Players from Behavior
Aug 17, 2026Yifan Lu, Xiaopeng Yuan, Haohan WangUser Modeling Personality Modeling in Language Models
Imperial College London · University of Illinois Urbana-Champaign

Oct 12, 2026J/K move · Enter open · S save

BehaviorFormer: Discovering Behavioral Patterns via Clusterable Next-Token Prediction

Tobias Schneider, Danial Fahrner, Soren Antebi, Daniel Uedelhoven+3

Fraunhofer IAIS, Sankt Augustin, Germany · University of Bonn, Bonn, Germany

Open paper AskPDF arXiv

Discovering behavioral patterns in open-ended games is challenging due to the lack of clear objectives or ground-truth annotations. In this work, we extend a next-token prediction Transformer with the deep clustering algorithm DCSS to learn meaningful representations with well-separated clusters. In a controlled Minecraft session with 26 players, the method discovers 23 distinct behavioral patterns that generalize across the studied players, further refining structures present in self-supervised representations into clearly separable clusters. We group the clusters into seven interpretable categories covering stationary, roaming, travel, mining, base work, logging, and farming behavior.

Abstract

Representation Learning

Explore similar work

BehaviorFormer: Discovering Behavioral Patterns via Clusterable Next-Token Prediction

Beyond a Single Crop: Contextualized Embeddings for Player Re-Identification

Beyond Asking: A Pipeline for Personalized Game Generation that Reads Players from Behavior