This project is a cython open-source implementation of the Compact Prediction Tree algorithm using multithreading.
CPT is a sequence prediction algorithm. It is a highly explainable model and good at predicting, in a finite alphabet, next value of a sequence. However, given a sequence, CPT cannot predict an element already present in this sequence. CPT needs a tuning.
This implementation is based on the following research papers
http://www.philippe-fournier-viger.com/ADMA2013_Compact_Prediction_trees.pdf
http://www.philippe-fournier-viger.com/spmf/PAKDD2015_Compact_Prediction_tree+.pdf
You can simply use pip install cpt.
For unix users, no wheels are published (problem with auditwheel which cannot repair wheels to "manylinux"), you should install cython then cpt: pip install cython cpt.
However unix users can simply install from sources: pip install cython && python setup.py install.
For osx users, do not forget to install brew's llvm and libomp. You can follow the directives of this issue: bluesheeptoken#68
You can test the model with the following code
from cpt import Cpt
model = Cpt()
model.fit([['hello', 'world'],
['hello', 'this', 'is', 'me'],
['hello', 'me']
])
model.predict([['hello'], ['hello', 'this']])
# Output: ['me', 'is']For an example with the compatibility with sklearn, you should check the documentation.
The model can be trained with the fit method.
If needed the model can be retrained with the same method. It adds new sequences to the model and do not remove the old ones.
The predictions are launched by default with multithreading with OpenMP.
The predictions can also be launched in a single thread with the option multithread=False in the predict method.
You can control the number of threads by setting the following environment variable OMP_NUM_THREADS.
You can pickle the model to save it, and load it later via pickle library.
from cpt import Cpt
import pickle
model = Cpt()
model.fit([['hello', 'world']])
dumped = pickle.dumps(model)
unpickled_model = pickle.loads(dumped)
print(model == unpickled_model)The CPT class has several methods to explain the predictions.
You can see which elements are considered as noise (with a low presence in sequences) with model.compute_noisy_items(noise_ratio).
You can retrieve trained sequences with model.retrieve_sequence(id).
You can find similar sequences with find_similar_sequences(sequence).
You can not yet retrieve automatically all similar sequences with the noise reduction technique.
CPT has 3 meta parameters that need to be tuned. You can check how to tune them in the documentation. To tune you can use the model_selection module from sklearn, you can find an example here on how to.
The benchmark has been made on the FIFA dataset, the data can be find on the spmf website.
Using multithreading CPT has been able to perform around 5000 predictions per second.
Without multithreading, CPT predicted around 1650 sequences per second.
Details on the benchmark can be found here.