Due to the technological advancements across all aspects of modern society many of our routine daily decisions are now either delegated to, or driven by, algorithms. These algorithms, for example, decide which emails are considered spam, if our credit loan application is going to be approved, and whether our commute between locations increases the probability of avoiding a traffic jam or an accident. These ubiquitous algorithms, however, not only provide answers but also raise questions that our society needs to address. For instance, if an algorithm denies an applicant credit, that applicant would most certainly want to be informed about why that decision made to understand the adjustments they should make in order to receive a better outcome for their next application, or to challenge the decision. Similarly, a patient receiving medical treatment would appreciate understanding the evidence for a diagnosis, by what degree the diagnostic process is automated, and how much the clinician trusts the algorithmic process.
The algorithms we are focused on in this article are supervised machine learning algorithms that are used to build predictive models [Kelleher, Mac Namee, and D’Arcy2015]. Fundamentally, supervised machine learning algorithms learn to distinguish patterns in input space by associating inputs, , to outputs, . During the process of training a model, an algorithm strives to optimize a performance criterion using example data or past experience (for a more formal definition we point the reader to [Mitchell1997]). Since predictive models have already penetrated critical areas of society such as healthcare, justice systems, and the financial industry, it is necessary to understand how they arrive at decisions, to verify the integrity of the decision making process, and to ensure that processes are in accordance with the ethical and legal requirements. More succinctly, models need to be interpretable but this requirement poses a series of recent challenges and trends regarding their design and implementation details [Holzinger et al.2018].
As we have already mentioned, interpretability is an important factor for predictive models and the broader machine learning (ML) community, but its importance derives from the scenarios in which machine learning models are applied. For instance, it is crucial to have interpretable explanations from predictive models assisting with diagnosis in the medical domain to provide additional information to domain experts (doctors) and guide their decisions, but one could argue that the importance of interpretability fades when considering customers being recommended products while browsing online shops.
To discuss interpretability, first we need to define it. Throughout the rest of the article we will use the definition by Miller [Miller2017]: “interpretability is the degree to which a human can understand the cause of a decision”. Lipton [Lipton2016] goes further and divides the scope of interpretability into two categories: transparency and post-hoc explanations. Transparency refers to the intrinsic underlying mechanism of a model’s inner workings. According to Lipton [Lipton2016] this could potentially describe mechanisms at the level of the model, component, or algorithm. On the other hand post-hoc explanations aim to provide further information about the model by uncovering the importance of its parameters. A schematic representation of this categorization is shown in Figure 1.
Although it might not be evident at first sight we can visualize the two categories as being interconnected instead of separate, as shown in Figure 2. This is possible due to the large variety of post-hoc methods that are available and that can be integrated at different levels of the life-cycle of a model.
In this article we are concerned with the post-hoc explanation category of interpretability. This category can be further subdivided into their own groups. For the purpose of this study we will call them Group A and Group B. Approaches in Group A address the question of what has the model learned, either on a holistic or modular level. Approaches in Group B are concerned with why the model produced a specific behaviour, either for a single prediction or a group of predictions. Each of them can be further subdivided into two additional subgroups: model specific that are tightly intertwined with a specific modelling algorithm and model agnostic approaches that can be applied to models trained using any algorithm. A schematic of these subdivisions is presented in Figure 3.
The remainder of this article tests this categorisation by describing key approaches in Group A and Group B, as just described. We briefly describe these approaches, draw connections and distinctions between them, and identify promising areas for future work.
Group A: What has the model learned?
Approaches that we categorise under Group A, provide additional information about a trained predictive model, either holistically or at a modular level. We first review model agnostic approaches and then focus on model specific ones.
Model agnostic approaches from Group A can provide overall interpretations of any black box model. These interpretations could be in regards to the overall model or to its individual components. The most common type of approach in this category is rule induction, in which interpretable rules are extracted from a trained black box model to describe its important characteristics. Yang et al [Yang, Rangarajan, and Ranka2018], for example, introduced the use of compact binary trees to represent important decision rules that are implicitly contained within black box models. To retrieve the binary tree of explanations the authors recursively partition the input variable space maximizing the difference in contribution of input variables averaged from local explanations between the divided spaces to retrieve the most important rules that represent the magnitude of changes in the output variable due to one unit of change in the input variable. This is also known as sensitivity analysis in fields such as physics or biology.
The Black Box Explanation through Transparent Approximations (BETA) [Lakkaraju et al.2017], approach is another rule induction method, closely connected to earlier work presented by [Lakkaraju, Bach, and Jure2016]. BETA learns a compact two-level decision set in which each rule explains part of the model behaviour unambiguously. BETA includes a novel objective function so that the learning process is optimized for high fidelity (increased agreement between explanations and the model), low unambiguity (less overlaps between decision rules in the explanation), and high interpretability (the explainable decision set is lightweight and small). These aspects are combined into one objective function that is optimized.
The approaches described so far can be applied to predictive models, but there are also rule induction approaches for more complex models. Penkox [Penkov and Ramamoorthy2017], for example, introduced the
-machine, which can extract LISP-like programs from observed data traces to explain the behaviour of agents in dynamical systems whose behaviour is driven by deep Q network (DQN) models trained using reinforcement learning. The proposed method utilizes an optimisation procedure based on backpropagation, gradient descent, andsearch in order to learn programs and explain various phenomena.
Model specific methods from Group A are
designed specifically to derive overall interpretations of what models
trained using specific algorithms, or families thereof, have
learned. One compelling example is maximum variance total
maximum variance total variation denoising(MVTV) [Tansey, Thomason, and Scott2017]
which is a regression approach that explicitly gives interpretability of the model which is of high priority and is useful when simple linear or additive models fail to provide adequate performance. Recently there has been an additional effort to extend a class of algorithms with high classification accuracy rate such as Deep Neural Networks (DNN) in order to extract insightful explanations in the form of decision diagram rules. An example of such an effort is described inlearning customized and optimized lists of rules with mathematical programming [Rudin and Ertekin2018]
, providing a formal approach in order to derive interpretable lists of rules. The methodology empowers solving a custom objective function along with user defined constrains utilizing mixed integer programming (MIP). This provides two immediate benefits, fist, the pruning of the large space of the derived rules and second, the guaranteed of selecting the most optimal rules that closely describe the classifier’s inner process.
Another approach which strives to extract interpretations from DNN is
presented in this looks looks like that: deep learning for
this looks looks like that: deep learning for interpretable image recognition[Chen et al.2018]. Here the emphasis is based on defining a reasoning process similar to humans when deciding if two objects belong to the same group based on common characteristics exhibited among them. Thus, the methodology is trying to discover a combination of exemplar patches from an image along with its prototypes during the classification process. This ensures that the final classification process adheres to principles closely related to human decision making, therefore the final classification process is interpretable by design. One such a example is showing and image of a cat to a human and asking them to describe the process by which they came to the final conclusion. Most probably they would focus on particular areas of the image while making the final decision based on previous experience, and excluding non similar objects. In addition, bayesian patchworks: an approach to case-based reasoning [Moghaddass and Rudin2018]
can be thought of as an extension of the previous methodology by designing models that can mimic logical processes resembled in individuals while reasoning about previous examples. The main idea is based on deriving a generative hierarchical model who’s final classification decision is based on examining similar examples exhibiting common features to identify a common ancestor. Going back to our earlier example of identifying an image of a cat, that would translate into a process of, (1) finding previous examples with common feature vectors (i.e. its neighbours), (2) deciding how many of the neighbours share important properties, (3) identifying those neighbours that are connected in terms of their features (i.e. share a common ancestor), (4) assign a label to the new example based on the influence of its neighbours and ancestor (e.g. similar to a voting approach but from a hierarchical perspective).
In summary one can interpret Group A as being the focus of investigating interpretable techniques providing either a holistic or modular view of the behaviour of a black-box model. The main idea underlying these efforts is to provide a coherent narrative through the process of a logical sequence which the end user can interpret as part of the overall process. Thus, approaches in this direction quite often utilize methods that either can provide this narrative such as decision rules, decision sets or inherently transparent methods such as regression variants. Similar efforts can be found in the works of lakkaraju2016 lakkaraju2016, condry2016 condry2016 and letham2015 letham2015.
Group B: Why does the model produce this behaviour?
In contrast to those in Group A that seek to explain a global view of what a model has learned, approaches in Group B seek to explain the reasons behind a specific prediction for a specific test instance. Again we summarise both model agnostic and model specific methods.
The local interpretable model agnostic explanations (LIME) [Ribeiro, Singh, and Guestrin2016] approach is a seminal example of a model agnostic Group B method. LIME is based on two simple ideas: perturbation and locally linear approximation. By repeatedly perturbing the values of the input and observing the effect this has on the output an understanding can be developed of how different inputs relate to the original output. LIME performs a perturbation of a specific input and then trains a linear model to describe the relationships between the (perturbed) inputs and the predicted outputs . The simple linear model approximates the more complex original black box model locally in the vicinity of the prediction that is to be explained. The linear model is used as an explanation for the original prediction.
Approaches similar to LIME are also presented in [Baehrens et al.2010] and [Robnik-Šikonja and Kononenko2008]. Anchors [Ribeiro, Singh, and Guestrin2018] is a particularly interesting extension to LIME that adds rule induction to make the explanations more compelling. The ASTRID method [Henelius, Puolamäki, and Ukkonen2017] provides a form of post-hoc interpretability to gain insight into how a model reaches specific predictions by examining attribute interactions in the dataset.
It can be more fruitful to build approaches that only apply to specific model types as this allows the inherent structure and learning mechanisms of these models to be utilised in generating explanations. This section describes interesting examples taking this approach.
The recent resurgence in interest in deep neural network approaches [Goodfellow et al.2016] has seen parallel interest in developing explanation approaches for these deep network models. Montavon et al [Montavon et al.2017], for example, propose a novel methodology for interpreting generic multilayer neural networks by decomposing the network classification into contributions of its input elements. Their methodology is applicable to a broad set of input data, learning tasks and network architectures. Explanations are generated as heatmaps overlaid on input artefacts (such as images) that indicate the areas of those artefacts that contributed most strongly to the model output. LSTMVis [Strobelt et al.2017]
, is a visual analysis tool for recurrent neural networks with a focus on understanding hidden dynamic states learned by these models. LSTMVis also relies heavily on the use of visualisations to communicate explanations.
Ensemble methods can also be difficult to interpret and specific approaches for these types of models have been developed. SHAP (SHapely Additive exPlanation) [Lundberg, Erion, and Lee2017]
, for example, addresses a fundamental problem with feature attribution when interpreting ensemble predictive models such as gradient boosting and random forests, due to the fact that they are inconsistent and can lower a feature’s assigned importance. SHAP uses approaches form game theory to recognise genuinely important features.
Overall the aim of Group B, in contrast to Group A, is to provide explanations regarding the behaviour of a black-box model based on the predictions emitted by the model (single or multiple predictions). The main idea is based on examining features or interactions thereof in order to understand which pose the highest shift in regards to the black-box model’s behaviour. This approach is known as sensitivity analysis in fields such as physics and biology and adheres to the study of uncertainty in the output of a mathematical model or system related to different sources of uncertainty in its inputs. Similar approaches can be found in [Chen et al.2018], [Ancona et al.2017] [Kahng et al.2017] [Samek et al.2017] [Qi and Li2017] [Liu, Luan, and Sun2017] [Lapuschkin et al.2016] [Zhou et al.2016] [Mahendran and Vedaldi2016] and [Li et al.2016].
In summary in this article we are trying to provide an overview of the multifaceted aspects of interpretability. Overall we proposed a hierarchical view of interpretability adopted from [Lipton2016] with minor modifications. Despite the fact that the emphasis has been on introducing and disseminating post-hoc methodologies we have always strove for a balance either by pointing the reader to the equivalent bibliography or by introducing methodologies that could potentially fall under the transparency group as well.
A recurring theme we have noticed throughout the existing literature is the lack of clear distinction between the notions of interpretability and debug(ability). It seems that most of the existing methodologies provide a way of debugging a predictive model and implicitly deriving the necessary interpretations. This constitutes an additional confusion to the reader and overall for the community as well.
Finally, we would like to point the reader to an interesting direction of research on constructing bug free predictive models, [Selsam, Liang, and Dill2017], falling under the group of transparency methodologies, which potentially holds a promise on possibly alleviating the need for interpretability. Furthermore, work towards an axiomatic treatment of interpretability [Bodenhofer and Bauer2000] has already paved the path in which the community should be invested, in order to alleviate the multifaceted properties of interpretability exhibited until recently. We hope that this article can serve as a guide in navigating the different aspects of interpretability to the interested reader.
This work has been supported by a research grant by Science Foundation Ireland under grant number SFI/15/CDA/3520. We would also like to thank Nvidia for their generours donation of GPU hardware in support of our research.
- [Ancona et al.2017] Ancona, M.; Ceolini, E.; Öztireli, A. C.; and Gross, M. H. 2017. A unified view of gradient-based attribution methods for Deep Neural Networks. NIPS Workshop on Interpreting, Explaining and Visualizing Deep Learning - Now What? 31.
- [Baehrens et al.2010] Baehrens, D.; Schroeter, T.; Harmeling, S.; Kawanabe, M.; Hansen, K.; and Müller, K.-R. 2010. How to Explain Individual Classification Decisions. JMLR 11:1803–1831.
- [Bodenhofer and Bauer2000] Bodenhofer, U., and Bauer, P. 2000. Towards an axiomatic treatment of ”interpretability”. SCCH 13.
- [Chen et al.2018] Chen, C.; Li, O.; Barnett, A.; Su, J.; and Rudin, C. 2018. This looks like that: Deep learning for interpretable image recognition. ICML 35.
- [Condry2016] Condry, N. 2016. Meaningful Models: Utilizing Conceptual Structure to Improve Machine Learning Interpretability. ICML Workshop on Human Interpretability in Machine Learning (WHI 2016), New York, NY 33.
- [Goodfellow et al.2016] Goodfellow, I.; Bengio, Y.; Courville, A.; and Bengio, Y. 2016. Deep learning, volume 1. MIT press Cambridge.
- [Henelius, Puolamäki, and Ukkonen2017] Henelius, A.; Puolamäki, K.; and Ukkonen, A. 2017. Interpreting Classifiers through Attribute Interactions in Datasets. ICML Workshop on Human Interpretability in Machine Learning (WHI 2017), Sydney, NSW, Australia 34.
- [Holzinger et al.2018] Holzinger, A.; Kieseberg, P.; Weippl, E.; and Tjoa, A. M. 2018. Current advances, trends and challenges of machine learning and knowledge extraction: From machine learning to explainable a.i. CD-MAKE, Lecture Notes in Computer Science 11015, Springer 1–8.
- [Kahng et al.2017] Kahng, M.; Andrews, P. Y.; Kalro, A.; and Chau, D. H. 2017. ActiVis: Visual Exploration of Industry-Scale Deep Neural Network Models. IEEE Transactions on Visualization and Computer Graphics 24(1):88–97.
- [Kelleher, Mac Namee, and D’Arcy2015] Kelleher, J. D.; Mac Namee, B.; and D’Arcy, A. 2015. Fundamentals of machine learning for predictive data analytics: algorithms, worked examples, and case studies. MIT Press.
- [Lakkaraju, Bach, and Jure2016] Lakkaraju, H.; Bach, S. H.; and Jure, L. 2016. Interpretable Decision Sets: A Joint Framework for Description and Prediction. KDD.
- [Lakkaraju et al.2017] Lakkaraju, H.; Kamar, E.; Caruana, R.; and Leskovec, J. 2017. Interpretable & Explorable Approximations of Black Box Models. FATML.
- [Lapuschkin et al.2016] Lapuschkin, S.; Binder, A.; Montavon, G.; Müller, K.-R.; and Samek, W. 2016. Analyzing Classifiers: Fisher Vectors and Deep Neural Networks. CVPR.
- [Letham et al.2015] Letham, B.; Rudin, C.; McCormick, T. H.; and Madigan, D. 2015. Interpretable classifiers using rules and Bayesian analysis: Building a better stroke prediction model. The Annals of Applied Statistics 9(3):1350–1371.
- [Li et al.2016] Li, J.; Chen, X.; Hovy, E.; and Jurafsky, D. 2016. Visualizing and Understanding Neural Models in NLP. HLT-NAACL.
- [Lipton2016] Lipton, Z. C. 2016. The Mythos of Model Interpretability. ICML Workshop on Human Interpretability in Machine Learning (WHI 2016), New York, NY.
[Liu, Luan, and Sun2017]
Liu, Y. D. Y.; Luan, H.; and Sun, M.
Visualizing and Understanding Neural Machine Translation.55th AMACL 1150–1159.
- [Lundberg, Erion, and Lee2017] Lundberg, S. M.; Erion, G. G.; and Lee, S.-I. 2017. Consistent Individualized Feature Attribution for Tree Ensembles. ICML Workshop on Human Interpretability in Machine Learning (WHI 2017), Sydney, NSW, Australia 34.
[Mahendran and Vedaldi2016]
Mahendran, A., and Vedaldi, A.
Visualizing Deep Convolutional Neural Networks Using Natural Pre-Images.IJCV 120(3):233–255.
Explanation in Artificial Intelligence: Insights from the Social Sciences.Artificial Intelligence.
- [Mitchell1997] Mitchell, T. M. 1997. Machine Learning. The McGraw-Hill International Series in Engineering and Computer Science.
- [Moghaddass and Rudin2018] Moghaddass, R., and Rudin, C. 2018. Bayesian patchworks: an approach to case-based reasoning. CoRR abs/1809.03541.
- [Montavon et al.2017] Montavon, G.; Lapuschkin, S.; Binder, A.; Samek, W.; and Müller, K.-R. 2017. Explaining nonlinear classification decisions with deep Taylor decomposition. Pattern Recognition 65:211–222.
- [Penkov and Ramamoorthy2017] Penkov, S., and Ramamoorthy, S. 2017. Using Program Induction to Interpret Transition System Dynamics. ICML Workshop on Human Interpretability in Machine Learning (WHI 2017), Sydney, NSW, Australia.
- [Qi and Li2017] Qi, Z., and Li, F. 2017. Learning Explainable Embeddings for Deep Networks. NIPS Workshop on Interpretable Machine Learning 31.
- [Ribeiro, Singh, and Guestrin2016] Ribeiro, M. T.; Singh, S.; and Guestrin, C. 2016. ”Why Should I Trust You?”: Explaining the Predictions of Any Classifier. ACM KDD.
- [Ribeiro, Singh, and Guestrin2018] Ribeiro, M. T.; Singh, S.; and Guestrin, C. 2018. Anchors: High-Precision Model-Agnostic Explanations. AAAI Press 32:1527–1535.
- [Robnik-Šikonja and Kononenko2008] Robnik-Šikonja, M., and Kononenko, I. 2008. Explaining Classifications for Individual Instances. IEEE Trans. on Knowledge and Data Engineering 20(5):589–600.
- [Rudin and Ertekin2018] Rudin, C., and Ertekin, Ş. 2018. Learning customized and optimized lists of rules with mathematical programming. Mathematical Programming Computation 10(4):659–702.
- [Samek et al.2017] Samek, W.; Binder, A.; Montavon, G.; Bach, S.; and Müller, K.-R. 2017. Evaluating the visualization of what a Deep Neural Network has learned. IEEE Trans. on Neural Networks and Learning Systems.
- [Selsam, Liang, and Dill2017] Selsam, D.; Liang, P.; and Dill, D. L. 2017. Developing bug-free machine learning systems with formal mathematics. ICML 34.
- [Strobelt et al.2017] Strobelt, H.; Gehrmann, S.; Pfister, H.; and Rush, A. M. 2017. LSTMVis: A Tool for Visual Analysis of Hidden State Dynamics in Recurrent Neural Networks. IEEE Transactions on Visualization and Computer Graphics 24(1):667–676.
- [Tansey, Thomason, and Scott2017] Tansey, W.; Thomason, J.; and Scott, J. G. 2017. Interpretable Low-Dimensional Regression via Data-Adaptive Smoothing. ICML Workshop on Human Interpretability in Machine Learning (WHI 2017), Sydney, NSW, Australia.
[Yang, Rangarajan, and Ranka2018]
Yang, C.; Rangarajan, A.; and Ranka, S.
Global Model Interpretation via Recursive Partitioning.
IEEE International Conference on Data Science and Systems (DSS-2018)4.
[Zhou et al.2016]
Zhou, B.; Khosla, A.; Lapedriza, A.; Oliva, A.; and Torralba, A.
Learning Deep Features for Discriminative Localization.CVPR.