Python-based implementation of PSyKE, i.e. a Platform for Symbolic Knowledge Extraction
Project description
PSyKE
Some quick links:
Reference paper
Federico Sabbatini, Giovanni Ciatto, Roberta Calegari, Andrea Omicini. "On the Design of PSyKE: A Platform for Symbolic Knowledge Extraction", in: WOA 2021 – 22nd Workshop “From Objects to Agents”, Aachen, Sun SITE Central Europe, RWTH Aachen University, 2021, 2963, pp. 29 - 48.
Bibtex:
@inproceedings{psyke-woa2021,
articleno = 3,
author = {Sabbatini, Federico and Ciatto, Giovanni and Calegari, Roberta and Omicini, Andrea},
booktitle = {WOA 2021 -- 22nd Workshop ``From Objects to Agents''},
editor = {Calegari, Roberta and Ciatto, Giovanni and Denti, Enrico and Omicini, Andrea and Sartor, Giovanni},
issn = {1613-0073},
keywords = {explainable AI, knowledge extraction, interpretable prediction, PSyKE},
location = {Bologna, Italy},
month = oct,
note = {22nd Workshop ``From Objects to Agents'' (WOA 2021), Bologna, Italy, 1--3~} # sep # {~2021. Proceedings},
numpages = 20,
pages = {29--48},
publisher = {Sun SITE Central Europe, RWTH Aachen University},
series = {CEUR Workshop Proceedings},
subseries = {AI*IA Series},
title = {On the Design of {PSyKE}: A Platform for Symbolic Knowledge Extraction},
url = {http://ceur-ws.org/Vol-2963/paper14.pdf},
volume = 2963,
year = 2021
}
Intro
PSyKE (Platform for Symbolic Knowledge Extraction) is intended as a library for extracting symbolic knowledge (in the form of logic rules) out of sub-symbolic predictors.
More precisely, PSyKE offers a general purpose API for knowledge extraction, and a number of different algorithms implementing it, supporting both classification and regression problems. The extracted knowledge consists of a Prolog theory (i.e. a list of Horn clauses).
PSyKE relies on 2ppy (tuProlog in Python) for logic support, which is in turn based on the 2p-Kt logic ecosystem.
Class diagram overview:
PSyKE is designed around the notion of extractor.
More precisely, an Extractor
is any object capable of extracting a logic Theory
out of a trained sub-symbolic regressor or classifier.
Accordingly, any Extractor
is composed of
(i) a trained predictor (i.e., black-box used as an oracle) and
(ii) a set of discrete feature descriptors (as some algorithms require a discrete dataset), and it provides two methods:
extract
: given a dataset it returns a logic theory;predict
: predicts a value using the extracted rules instead of the original predictor.
Currently, the supported extraction algorithms are:
- CART, straightforward extracts rules from both classification and regression decision trees;
- Classification:
- Regression:
Users
End users
PSyKE is deployed as a library on Pypi, and it can therefore be installed as Python package by running:
pip install psyke
Requirements
- numpy 1.21.3+
- pandas 1.3.4+
- scikit-learn 1.0.1+
- 2ppy 0.3.3+
Test requirements
- skl2onnx 1.10.0+
- onnxruntime 1.9.0+
- parameterized 0.8.1+
Once installed, one can create an extractor from a predictor (e.g. Neural Networks, Support Vector Machines, K-Nearest Neighbor, Random Forests, etc.) and from the dataset used to train the predictor.
Note: the predictor must expose a method named
predict
to be properly used as oracle.
End users
A brief example is presented in demo.py
script.
Using sklearn iris dataset we train a K-Nearest Neighbor to predict the correct iris class.
Before training, we make the dataset discrete.
After that we create two different extractors: REAL and Trepan.
We output the extracted theory for both extractors.
REAL extracted rules:
iris(PetalLength_0, PetalWidth_0, SepalLength_0, SepalWidth_0, setosa) :- '=<'(PetalWidth_0, 0.65).
iris(PetalLength_1, PetalWidth_1, SepalLength_1, SepalWidth_1, versicolor) :- ('>'(PetalLength_1, 4.87), '>'(SepalLength_1, 6.26)).
iris(PetalLength_2, PetalWidth_2, SepalLength_2, SepalWidth_2, versicolor) :- '>'(PetalWidth_2, 1.64).
iris(PetalLength_3, PetalWidth_3, SepalLength_3, SepalWidth_3, virginica) :- '=<'(SepalWidth_3, 2.87).
iris(PetalLength_4, PetalWidth_4, SepalLength_4, SepalWidth_4, virginica) :- in(SepalLength_4, [5.39, 6.26]).
iris(PetalLength_5, PetalWidth_5, SepalLength_5, SepalWidth_5, virginica) :- in(PetalWidth_5, [0.65, 1.64]).
Trepan extracted rules:
iris(PetalLength_6, PetalWidth_6, SepalLength_6, SepalWidth_6, virginica) :- ('>'(PetalLength_6, 2.28), in(PetalLength_6, [2.28, 4.87])).
iris(PetalLength_7, PetalWidth_7, SepalLength_7, SepalWidth_7, versicolor) :- '>'(PetalLength_7, 2.28).
iris(PetalLength_8, PetalWidth_8, SepalLength_8, SepalWidth_8, setosa) :- true.
Developers
Working with PSyKE codebase requires a number of tools to be installed:
- Python 3.9+
- JDK 11+ (please ensure the
JAVA_HOME
environment variable is properly configured) - Git 2.20+
Develop PSyKE with PyCharm
To participate in the development of PSyKE, we suggest the PyCharm IDE.
Importing the project
- Clone this repository in a folder of your preference using
git_clone
appropriately - Open PyCharm
- Select
Open
- Navigate your file system and find the folder where you cloned the repository
- Click
Open
Developing the project
Contributions to this project are welcome. Just some rules:
- We use git flow, so if you write new features, please do so in a separate
feature/
branch - We recommend forking the project, developing your stuff, then contributing back vie pull request
- Commit often
- Stay in sync with the
develop
(ormaster
) branch (pull frequently if the build passes) - Do not introduce low quality or untested code
Issue tracking
If you meet some problem in using or developing PSyKE, you are encouraged to signal it through the project "Issues" section on GitHub.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for psyke-0.1.0.dev28-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 712d1c0cfab518b2661448029fbb285e9a16e9b4eddc7c41cdc1a07eb90ef7f2 |
|
MD5 | 59590ddcc33cc165697f9ba55cd848c1 |
|
BLAKE2b-256 | 24a120eb872c8ede025a1a0bd39a2ddc66f84324c59e84998d4030e45ee26bcb |