This repository contains the code and the dataset for building and evaluating prediction models for team victory in Dota 2.

The data was collected from the OpenDota platform and you can find the used dataset here:

In this project, we build victory prediction models for Dota 2 and use an interpretability technique (SHAP values) to explain the prediction by means of feature importance.

For more details on how we collected the data and extracted the model features, please check our Dota 2 paper accepted at the 16th AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE 2020):

The following dependencies are required to run the scripts of this project:

• Python 3.7

• Git 2.20.1.windows.1

• Numpy 1.18.4

  pip install numpy

• Pandas 0.24.2

  pip install pandas

• scikit-learn 0.20.3

  pip install scikit-learn

• xgboost 1.1.0

  pip install xgboost

• matplotlib 3.0.2

  pip install matplotlib

• shap 0.35.0

  pip install shap

data-collection

The data-collection directory contains the dataset-related code.

• The hero_changelog and hero_historical_stats directories contain the changelog and attribute values for each Dota 2 hero since Dota 2 first version.

• The identifiers of the matches that compose our dataset can be found in the match_ids files (available in csv and json formats), along with additional information.

• The file opendota-sql-query.txt contains the SQL query used to retrieve the match ids from the SQL functionality within the OpenDota platform.

• The file request_matches.py reads the match IDs from the csv file (must be in the same directory) and request the match file using the OpenDota API (you must create the output directory in the same directory level as the request_matches.py file.

data-analysis

The data-analysis directory contains the data analysis-related code, including the code to build and evaluate the prediction models and the code to apply SHAP values to obtain the predicion explanation by means of feature importance.

• The file feature_engineering.ipynb contains the code to compute the machine learning model features.

• The hero-attributes directory contains the script to collect the changelog for all (current) Dota 2 heroes and parse them to extract heroes' attributes for all the game versions.

• The prediction-models directory contains instructions on how to run the code to build and evaluate the prediction models (XGBoost, Random Forest, and Logistic Regression) for team victory in Dota 2.

• The prediction-explanation-SHAP directory contains instructions on how to run the script to apply the SHAP values technique to explain victory predictions for Dota 2 matches.

Note that the code within the prediction-models and prediction-explanation-SHAP directories only depend on the computed features, which are readily-available in the model_features_pre-match directory.

If you have any question or suggestion, please contact the repository developer at markosviggiato[at]gmail.com

https://markosviggiato.github.io/