giuseppec / beta3_IRT

Source code of $\beta^3$-IRT(https://arxiv.org/abs/1903.04016)

beta3_IRT

Source code of the paper $\beta^3$-IRT: A New Item Response Model and its Applications

Requirements

The source code was originally developed on:

1. Python 2.7.12
2. Tensorflow 1.2.0
3. Edward 1.3.4

It was also tested on:

1. Python 3.6.6
2. Tensorflow 1.10.0
3. Edward 1.3.5

which requires manually fixing the compatible issues of Tensorflow (> 1.2.0) in Edward. Please find a fixed version from this fork https://github.com/yc14600/edward and install Edward from the source.

Usage

There are two steps to run experiments:

1. Train classifiers and generate response data for the $\beta^3$ IRT model, for example, run the following command:

   python gen_irt_data.py --dataset moons --data_size 400 --noise_fraction 0.2 --seed 42
   

2. The first step will automatically generate data files for the second step, and the file named with "irt_data_*.csv" is the input parameter of the command to run $\beta^3$ IRT model, i.e.:

   python betairt_test.py --IRT_dfile irt_data_moons_s400_f20_sd42_m12.csv --a_prior_mean 1. --a_prior_std 1.  
   

Citing $\beta^3$-IRT

Biblatex entry:

@inproceedings{chen2019beta,
  title={$\beta^3$-IRT: A New Item Response Model and its Applications},
  author={Chen, Yu and Filho, Telmo Silva and Prud{\^e}ncio, Ricardo BC and Diethe, Tom and Flach, Peter},
  booktitle={Proceedings of the 22nd International Conference on Artificial Intelligence and Statistics (AISTATS) },
  year={2019}
  }