hoangdang09/game-telemetry-data

Citation

If you use this dataset, please cite:

@article{dang2026effect,
  title={Effect-Level Validation for Causal Discovery},
  author={Dang, Hoang and Pham, Luan and Nguyen, Minh},
  journal={arXiv preprint arXiv:2602.08340},
  year={2026}
}

Game Telemetry Dataset

Description

This dataset contains real behavioral telemetry from a deployed Web3 role-playing game, released for causal inference and observational analysis of early PvP exposure and Day-1 retention.

The public file contains 1,280 de-identified, user-level observations and six variables: one treatment, one outcome, and four pre-treatment covariates.

Variables

Variable	Type	Definition
web3	Binary	Wallet connection status, used as a proxy for platform commitment
total_session	Numeric	Total login sessions before treatment time, used as a proxy for engagement intensity
time_play_level1	Numeric	Time spent on Level 1 before treatment time, used as a proxy for onboarding friction
total_pve_battle	Numeric	Number of PvE battles before treatment time, used as a proxy for baseline progression
pvp	Binary	Treatment indicator: early exposure to or participation in optional Arena/PvP mode
r1	Binary	Outcome indicator: Day-1 retention, equal to 1 if the user returned during hours 24-48

Collection And Provenance

The dataset is derived from real behavioral telemetry from a deployed Web3 RPG. Raw operational telemetry was collected through Firebase during normal gameplay over a five-month worldwide collection period in 2024. Exact calendar months are withheld to reduce re-identification and business-sensitivity risks.

Users consented to gameplay telemetry collection under the game's Terms of Service and/or Privacy Policy. No prospective intervention, crowdsourcing task, or human-subject experiment was conducted for this public dataset release.

The raw telemetry contained 77,698 users and 26 telemetry-derived features covering sessions, progression milestones, economic transactions, and system interactions. The analytic cohort was constructed by restricting to paying players, applying quality filters and missingness removal, and retaining the complete-case public analytic sample of 1,280 observations.

For final effect estimates in the accompanying paper, overlap trimming further restricts the common-support sample to 1,170 users. The hosted public dataset contains the 1,280-row complete-case analytic sample.

De-identification And Sensitive Information

The public dataset contains no direct identifiers. It does not include user IDs, wallet addresses, device IDs, IP addresses, email addresses, raw timestamps, raw event logs, or raw transaction logs.

The dataset includes de-identified user-level behavioral aggregates and a binary Web3 wallet-connection indicator. Although these fields are not direct identifiers, behavioral telemetry can still be sensitive in some contexts. Users should not attempt to re-identify players or link the data with external information.

Intended Use And Construct Validity

This dataset is intended for causal inference, observational analysis, and evaluation of adjustment, overlap, refutation, and causal-discovery workflows for early PvP exposure and Day-1 retention.

The main constructs are defined as follows:

• pvp measures early exposure to the optional competitive gameplay loop.
• r1 measures return during hours 24-48.
• web3 proxies platform commitment.
• time_play_level1 proxies onboarding friction.
• total_pve_battle proxies baseline progression.
• total_session proxies engagement intensity.

Validity has been established only for the accompanying game-telemetry analysis context. The dataset is not validated for demographic fairness analysis, individual player scoring, general-purpose retention benchmarking, or policy decisions about other games or player populations.

Limitations

The dataset is not representative of all players, all Web3 users, all RPGs, or the full deployed-game population. The analytic cohort is restricted to paying players and users active until PvP eligibility.

PvP exposure is self-selected rather than randomized. Causal conclusions require adjustment assumptions and remain vulnerable to residual or unmeasured confounding.

The public release contains only six aggregate analysis variables. It omits the raw 26-feature telemetry matrix, raw event logs, timestamps, user identifiers, and detailed acquisition context.

The dataset does not include demographics such as age, gender, language, country, or socioeconomic status, so demographic representativeness and fairness claims cannot be evaluated.

Exact collection months within 2024 are intentionally withheld. The dataset should not be used for time-series seasonality analysis or claims requiring exact calendar timing.

Biases

PvP exposure is optional, so treated users are self-selected. In the analytic sample, PvP-exposed users are more active and more likely to have Web3 wallet connection than non-PvP users.

The cohort restriction to paying players underrepresents non-paying players and may overrepresent more committed or higher-value users.

Users who churn before PvP eligibility are excluded, which underrepresents the earliest churn population.

The data were collected worldwide, but the public release does not include country or regional fields; regional coverage and regional imbalance cannot be assessed from the released file.

Synthetic Data

This dataset contains real telemetry data.

hasSyntheticData: false

Social Impact

Potential benefits include transparent evaluation of causal-discovery and effect-validation methods on real behavioral telemetry, helping researchers and product teams avoid overinterpreting unstable graph structures.

Potential risks include misuse for engagement optimization, player targeting, or retention manipulation without adequate user-welfare, privacy, or fairness review. The release mitigates some privacy risks by excluding direct identifiers, raw timestamps, wallet addresses, and raw event logs, but users should still treat behavioral telemetry as sensitive and avoid re-identification or individual-level profiling.

Croissant Metadata

For NeurIPS 2026 Evaluations and Datasets Track submission, use the completed Croissant metadata file with core and Responsible AI metadata fields, then validate it with the Croissant validator before submission.

Required RAI/provenance fields included in the companion Croissant file:

• rai:dataLimitations
• rai:dataBiases
• rai:personalSensitiveInformation
• rai:dataUseCases
• rai:dataSocialImpact
• rai:hasSyntheticData
• prov:wasDerivedFrom
• prov:wasGeneratedBy

License

This dataset is released under the Creative Commons Attribution 4.0 International license (CC BY 4.0).