• Presented at the AAAI 2019 conference in Honolulu, Hawaii on January 27, 2019 [arXiv] [slides].
• Applied in the Best Paper Award of the EDM 2019 conference in Montreal, Canada on July 2, 2019.

@inproceedings{Vie2019,
  Author = {{Vie}, Jill-J{\^e}nn and {Kashima}, Hisashi},
  Booktitle = {Proceedings of the 33th {AAAI} Conference on Artificial Intelligence},
  Title = {{Knowledge Tracing Machines: Factorization Machines for Knowledge Tracing}},
  Pages = {750--757},
  Url = {https://arxiv.org/abs/1811.03388},
  Year = 2019
}

Authors: Jill-Jênn Vie, Hisashi Kashima

Presented at the Optimizing Human Learning workshop in Kingston, Jamaica on June 4, 2019.

Slides from the tutorial are available here. A Jupyter notebook will be available "soon" on Binder.

The tutorial makes you play with the models to assess weak generalization. To assess strong generalization and reproduce the experiments of the paper, you may want to use scikit-learn's GroupShuffleSplit.

python3 -m venv venv
. venv/bin/activate
pip install -r requirements.txt  # Will install numpy, scipy, pandas, scikit-learn, pywFM

If you also want to get the factorization machines running (KTM for d > 0), you should also do:

make libfm

Select a dataset and the features you want to include.

data/<dataset>/data.csv should contain the following columns:

user, item, skill, correct, wins, fails

where wins and fails are the number of successful and unsuccessful attempts at the corresponding skill.

data/<dataset>/needed.csv needs to contain:

user_id, item_id, correct

(Note the difference.)

And data/<dataset>/q_mat.npz should be a q-matrix under scipy.sparse format.

If you want to compute wins and fails like in PFA or DAS3H, you should run encode_tw.py instead of this file, with the --pfa option for PFA or --tw for DAS3H.

Are you excited? If so, check <sktm.py>.

pipe = Pipeline([
    ('onehot', OneHotEncoder(handle_unknown='ignore')),
    ('lr', LogisticRegression(solver='liblinear'))
])

# IRT
pipe.fit(df_train[['user', 'item']], df_train['correct'])
print(pipe.predict_proba(df_test[['user', 'item']]))

# PFA
pipe.fit(df_train[['skill', 'wins', 'fails']], df_train['correct'])
print(pipe.predict_proba(df_test[['skill', 'wins', 'fails']]))

sktm contains efficient parallel cross validation over 5 folds, stratified by group (i.e. strong generalization).

Usage:

mkdir data/assistments09
wget https://jiji.cat/weasel2018/data.csv -P data/assistments09
python sktm.py data/assistments09/data.csv --model (irt|pfa|sktm)  # Choose which model

For factorization machines, replace LogisticRegression with from fm import FMClassifier.

• Assistments 2009
• Our reformatted version of the Assistments 2009 dataset.
• Berkeley and Castor datasets are private.
• Datasets ECPE and TIMSS 2003 come from the CDM package:

> install.packages('CDM')
> library('CDM')
> dim(fraction.subtraction.data)
[1] 536  20
> dim(data.ecpe$data)
[1] 2922   29
> dim(data.timss03.G8.su$data)
[1] 757  25

python encode.py --users --items  # To get the encodings (npz)
python lr.py data/dummy/X-ui.npz  # To get results (txt)

You can also download the Assistments 2009 dataset into data/assistments09 and change the dataset:

python encode.py --dataset assistments09 --skills --wins --fails  # Will encode PFA sparse features into X-swf.npz

If you are lazy, you can also just do make and try to understand what is going on in the Makefile.

Choffin et al. proposed the DAS3H model, and we implemented it using queues. This code is faster than the original KTM encoding.

To prepare a dataset like Assistments, see examples in the data folder.
Skill information should be available either as skill_id, or skill_ids separated with ~~, or in a q-matrix q_mat.npz.

python encode_tw.py --dataset dummy_tw --tw  # Will encode DAS3H sparse features into X.npz

Then you can run lr.py or fm.py, see below.

If you want to encode PFA features:

python encode.py --skills --wins --fails  # Will create X-swf.npz

For logistic regression:

python lr.py data/dummy/X-swf.npz
# Will save weights in coef0.npy

For factorization machines of size d = 5:

python fm.py --d 5 data/dummy/X-swf.npz
# Will save weights in w.npy and V.npy

The following code does not work if you don't have user_id as column in CSV file.

NEW! For an online MIRT model:

python omirt.py --d 0 data/assist09/needed.csv  # Will load LR: coef0.npy
python omirt.py --d 5 data/assist09/needed.csv  # Will load FM: w.npy and V.npy

# Will train a IRT model on Fraction dataset with learning rate 0.01
python omirt.py --d 0 data/fraction/needed.csv --lr 0.01 --lr2 0.

NEW! For an IRT or deeper model with Keras, for batching and early stopping:

python dmirt.py data/assist09/needed.csv

It will also create a model.png file with the architecture (here just IRT with L2 regularization):

Those numbers may change according to your random state seed.

On the Assistments 2009 dataset:

Computation times are given for a i7 with 2.6 GHz, with 200 epochs of FM training.

On the Assistments 2009 dataset:

On the Duolingo French dataset:

We also showed that Knowledge Tracing Machines (Bayesian FMs) got better results than Deep Factorization Machines on the Duolingo dataset. See our article: Deep Factorization Machines for Knowledge Tracing and poster at the BEA workshop at New Orleans, LA on June 5, 2018.

@inproceedings{Vie2018,
  Author = {{Vie}, Jill-J{\^e}nn},
  Booktitle = {{Proceedings of the Thirteenth Workshop on Innovative Use of NLP for Building Educational Applications}},
  Pages = {370--373},
  Title = {{Deep Factorization Machines for Knowledge Tracing}},
  Url = {http://arxiv.org/abs/1805.00356},
  Year = 2018}