0
$\begingroup$

Explainable AI can be achieved through intrinsically explainable models, like logistic and linear regression, or post-hoc explanations, like SHAP.

I want to use an intrinsically explainable model on tabular data for a classification task. However, logistic and linear regression have poor performance.

Are there any other intrinsically explainable models that have higher performance?

$\endgroup$

2 Answers 2

2
$\begingroup$

To add a bit more to @noe 's answer: when you have a small number of features, explainable models can do a lot for you because they usually operate by making a prediction directly using the input features, without any intermediate features. When the data is structured and the number of features is small there isn't much value in choosing a more complex model while losing explainability.

With a large number of features that changes. You have two issues. Models that make predictions directly from input data no longer have simple explanatory values. Absent feature engineering, it is probably best to use a more powerful model that can use a smaller number of secondary features in its decision. For example, a two-layer neural network is essentially two layers of logistic regression. So you can still analyze which secondary features are useful for the final layer. Then you can analyze to see what those features correlate within your data set.

$\endgroup$
9
  • $\begingroup$ Thank you! So you're not aware of any new methods that are intrinsically explainable? Is that because all the research is currently focused on powerful methodologies? $\endgroup$
    – Connor
    Feb 24, 2023 at 22:28
  • $\begingroup$ @Connor If you are willing to remove the words "intrinsically explainable" and replace with "has software tools to explain inferences" then you open yourself up to a lot more models. Also, are you suggesting that your tabular data has many features? That also determines which models to try. Edit: I am out at the moment. I can list models in a bit. $\endgroup$ Feb 24, 2023 at 22:37
  • 1
    $\begingroup$ @Connor. Well I disagree on classifying them all as post-hoc, many are firmly rooted in math and stats. But if you just need something basic to start with on 50 features then I would recommend Gradient Boosted Trees (GBTs), Generalized Linear Models (GLMs), Support Vector Machines (SVMs) or Naive Bayes. They have a lot of diversity in mechanism & theory between them so one or more of them should give you a decent baseline to build on. But GLMs and Naive Bayes allow more straightforward computation of confidence values. So maybe that is what you want. $\endgroup$ Feb 24, 2023 at 23:17
  • 1
    $\begingroup$ And yea, there are specifically new models that are explainable as far as I am concerned but it takes a better understanding of how the probing is done to feel comfortable with them. For some things like modern versions of Variational Bayes there is some good theory to it. But how happy you are with the explanations really depends on your application. $\endgroup$ Feb 24, 2023 at 23:22
  • 1
    $\begingroup$ They are new classes of deep variational methods that are still well understood in terms of bayesian statistics and have decent explainability (via latent variables) over unstructured data like images, sound, video etc. Other methods include understanding the high dimensional dynamics of a deep neural net via tools such as Topological Data Analysis, Diverse Counterfactual Explanations, Activation Maximization, etc. Then there is bleeding edge stuff like GFlow nets that aims to infer the underlying bayesian network of latent variables that determines decisions. $\endgroup$ Feb 25, 2023 at 0:45
1
$\begingroup$

I think there are two issues with the formulation of your question:

  • Model performance highly depends on the data.
  • Explainability is not a black/white concept.

That said, one may understand that gradient-boosted trees tend to give good results in general and can, to some degree, be considered interpretable if they are small enough.

$\endgroup$
3
  • $\begingroup$ Please could you explain what you mean by "Explainability is not a black/white concept". I see you have or are doing a PhD in neural machine translation! I imagine you're deep in the explainable AI literature. Are there any new intrinsically-explainable ideas or models that you think could out-perform logistic and linear regression, and trees? I'd like to use boosted trees, but unfortunately they're considered too complex for this use case. $\endgroup$
    – Connor
    Feb 24, 2023 at 17:47
  • 1
    $\begingroup$ Explainability is not a black/white concept because a type of model may be theoretically explainable but, in practice, a model of such a type it may not. For instance, decision trees are explainable...until they are too large, because the explanation is not useful for a human to understand. $\endgroup$
    – noe
    Feb 24, 2023 at 18:17
  • 2
    $\begingroup$ Because of the reason I described above, intrinsically explainable methods are simple methods by definition. However, in IA one does not normally renounce powerful known blackbox methods (i.e. deep learning) for the sake of intrinsic explainability. Therefore, currently, it is more common to see attempts to understand or probe black box models. The Blackbox NLP workshop is a paradigmatical example. $\endgroup$
    – noe
    Feb 24, 2023 at 18:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.