Counterfactual explanation is a technique that is used in the field of AI to provide explanations on why the model has made a particular choice. In the context of recommender systems, it can be described as “a set of minimal actions performed by the user that, if removed, changes the recommendation to a different item”1. Generating a counterfactual sequence, which also means changing the minimal number of items in order to have a different prediction, can hint to which are the items that caused the prediction in the first place, effectively providing an explanation.
Counterfactual reasoning has been used not only to provide explanation, but also to do data augmentation2 to counteract the high sparseness that recommendation datasets have; to reduce popularity bias34; or as a data augmentation technique for contrastive learning5. The first and last cases can be applied because counterfactual data is data that the user might have interacted with, and so it can be thought as sampled from the same distribution of the real data, meaning that it’s better than doing some other type of augmentations.
Counterfactual sequences has been applied both sequential and non-sequential recommender systems, and they have been generated by the usage of perturbation models, such as VAE6; with the help of reinforcement learning techniques2; or with graph-based techniques1.
Techniques can be model-agnostic (in a black-box) fashion if they do not require to know anything about the model, but just having the model as a black-box; but they can also be gray-box, if some information about the model has to be knows, such as its gradients; or it can be white-box, where the model has to be completely open, and the architecture has to be changed in order to provide an explanation.
Zhou et al.7 implemented all of these types of models, using attention (even though Attention is not Explanation!8) for the white-box, adversarial perturbation for the gray-box and counterfactual perturbation for the black-box.
Counterfactual sequences are evaluated in two ways:
- Model Fidelity, which expresses the percentage of data that the model is able to explain, if the technique is used for explanation;
- Standard performance evaluation of the recommendation model with and without the counterfactual sequences augmentation, if counterfactual reasoning is used as an augmentation technique.
tags:#ai-explainability/counterfactual#recommender-systems
Footnotes
-
Ghazimatin et al. - PRINCE Provider-side Interpretability with Counterfactual Explanations in Recommender Systems ↩ ↩2
-
Chen et al. - Data Augmented Sequential Recommendation Based on Counterfactual Thinking ↩ ↩2
-
Wei et al. - Model-agnostic counterfactual reasoning for eliminating popularity bias in recommender system ↩
-
Ren et al. - Disentangled Counterfactual Reasoning for Unbiased Sequential Recommendation ↩
-
Wang et al. - Explanation Guided Contrastive Learning for Sequential Recommendation ↩
-
: Xu et al. - Learning Causal Explanations for Recommendation ↩
-
: Zhou et al. - 2021 - From Intrinsic to Counterfactual On the Explainability of Contextualized Recommender Systems ↩