A global, model-agnostic, contrastive explainer for any text classifier
- TOC {:toc}
Imagine we have a text classifier, let's say M1, that is retrained with new data generating M2.
- Can we estimate to what extent M2 classifies new data coherently to the past predictions made by the previous model M1?
- Why do the criteria used by M1 result in class c, but M2 does not use the same criteria to classify as c anymore?
- Can we use natural language to explain the differences between models making them more comprehensible to final users?
ContrXT is about explaining how a classifier changed its predictions through time. Also, it can be used to explain the differences in the classification behaviours of two distinct classifiers at a time.
ContrXT takes as input the training data and predictions of two distinct classifiers M1 and M2. Then, it traces the decision criteria of both classifiers by encoding the changes in the decision logic through Binary Decision Diagrams. Then (ii) it provides "global, model-agnostic, time-contrastive (T-contrast) "explanations in natural language, estimating why -and to what extent- the model has modified its behaviour over time.
Details on how ContrXT works can be found in this paper Below a link to the demo video (4K)
ContrXT.Demo.1.mp4
ContrXT takes as input the "feature data" (can be training or test, labelled or unlabelled) and the corresponding "labels" predicted by the classifier. This means you don't need to wrap ContrXT within your code at all! As optional parameters, the user can specify:
- the coverage of the dataset to be used (default is 100%); otherwise, a sampling procedure is used;
- to obtain explanations either for the multiclass case (default: one class vs all) or the two-class case (class vs class, by restricting the surrogate generation to those classes);
ContrXT provides two outputs:
A picture estimating the differences among the classification criteria of both classifiers M1 and M2, that are Add and Del values. To estimate the degree of changes among classes, ContrXT also provides add_global and del_global. In the case of a multiclass classifier, ContrXT suggests focusing on classes that went through major alterations from M1 to M2, distinguishing between three groups according to their Add and Del values being above or below the 75th percentile.
The picture above is generated by ContrXT automatically. It provides indicators of changes in classification criteria from model M1 to M2 for each 20newsgroup class, using a DT as a surrogate (0.8 fidelity) to explain two BERT classifiers.
The indicators allow the user to concentrate on classes that have changed more. For example, one might closely look at class atheism as for this class, the number of deleted paths is higher than the added ones
The Natural Language Explanation (NLE) for atheism reveals the presence of the word bill leads the retrained classifier M2 to assign the label atheism to a specific record, whilst the presence of such a feature was not a criterion for the previous classifier M1. Conversely, the explanation shows that M1 used the feature keith to assign the label, whilst M2 discarded this rule.
Both terms refer to the name of the posts' authors: Bill's posts are only contained within the dataset used to retrain whilst Keith's ones are more frequent in the initial dataset rather than the second one (dataset taken from Jin, P., Zhang, Y., Chen, X., & Xia, Y. Bag-of-embeddings for text classification. In IJCAI-2016).
Finally, M2 discarded the rule having political atheist that was sufficient for M1 for classifying the instance.
The NLE shows the differences between the two models. However, a user might also wish to see example instances in the datasets where these rules apply.
To do so, ContrXT provides the get_rule_examples function, which requires the user to specify a rule to be applied and the number of examples to show.
Our experiments reveal Add/Del is not correlated with the models' accuracy. For example, we assessed a correlation between Add/Del and the change in performance of the classifiers in terms of F1-score on 20newsgroup. To this end, we computed the Spearman's \rho between the Add of every class and its change in f1-score between the two classifiers. The correlation values are not significant, p=0.91$, ro=-0.11 for the Add and p=0.65, ro=-0.04 for Del indicator. This confirms that Add/Del are not related to the f1-score of the trained model. Instead, they estimate its behaviour change handling new data, considering which rules have been added or deleted to the past.
ContrXT is available on PyPi. Simply run:
pip install contrxt
Or clone the repository and run:
pip install .
The PyEDA package is required but has not been added to the dependencies. This is due to installation errors on Windows. If you are on Linux or Mac, you should be able to install it by running:
pip3 install pyeda
However, if you are on Windows, we found that the best way to install is through Christophe Gohlke's pythonlibs page. For further information, please consult the official PyEDA installation documentation.
To produce the PDF files, a Graphviz installation is also required. Full documentation on how to install Graphviz on any platform is available here.
A complete example of ContrXT usage is provided in the notebook "ContrXT Demo" inside of the main repository folder. A notebook example with ML model training is also available in this repository, which can also be accessed in this Google Colab notebook. Complete documentation of the package functions will also be available shortly.
Basic unit tests are provided. To run them, after installation, execute the following command while in the main directory:
python -m unittest discover
ContrXT provides REST-API to generate explanations for any text classifier. Our API enables users to get the outcome of ContrXT as discussed above (i.e., Indicators and Natural language explanations) with no need to install or configure the tool locally. The required input from a user are (i) the training data and (ii) the predicted labels by the classifier of their choice. Users are required to upload two csv files for two datasets for which the schema is shown in the following JSON.
schema = {
"type" : "csv",
"columns" : {
"corpus" : {"type" : "string"},
"predicted" : {"type" : "string"},
},
}
The ContrXT's API can be invoked using a few lines code shown in Code below.
import requests, io
from zipfile import ZipFile
files = {
'time_1': open(t1_csv_path, 'rb'),
'time_2': open(t2_csv_path, 'rb')
}
r = requests.post('[URLandPort]', files=files)
result = ZipFile(io.BytesIO(r.content))
Notice. To avoid improper use of ContrXT'server resources, we ask users to ask for free credentials through this link.
Currently, we support explaining predictions for text classifiers only, but we are working to extend it to deal with tabular data.
To cite ContrXT please refer to the following paper
@article{ContrXT,
author = {Lorenzo Malandri and Fabio Mercorio and Mario Mezzanzanica and Andrea Seveso and Navid Nobani},
title = {ContrXT: Generating Contrastive Explanations from any Text Classifier},
year = {2021},
publisher = {Elsevier},
issn = {1566-2535},
year = {2021},
issn = {1566-2535},
doi = {https://doi.org/10.1016/j.inffus.2021.11.016},
url = {https://www.sciencedirect.com/science/article/pii/S1566253521002426},
journal = {Information Fusion}
}
@inproceedings{ContrXTdemo,
author = {Lorenzo Malandri and
Fabio Mercorio and
Mario Mezzanzanica and
Navid Nobani and
Andrea Seveso},
title = {The Good, the Bad, and the Explainer: {A} Tool for Contrastive Explanations
of Text Classifiers},
booktitle = {Proceedings of the Thirty-First International Joint Conference on
Artificial Intelligence, {IJCAI} 2022, Vienna, Austria, 23-29 July
2022},
pages = {5936--5939},
year = {2022},
url = {https://doi.org/10.24963/ijcai.2022/858},
doi = {10.24963/ijcai.2022/858}
}