Machine learning andare helping automate an ever-increasing array of tasks, with ever-increasing accuracy. They are supported by the growing volume of data used to feed them, and the growing sophistication in algorithms.
The flip side of more complex algorithms, however, is less interpretability. In many cases, the ability to retrace and explain outcomes reached by machine learning models (ML) is crucial, as:
"Trust models based on responsible authorities are being replaced by algorithmic trust models to ensure privacy and security of data, source of assets and identity of individuals and things. Algorithmic trust helps to ensure that organizations will not be exposed to the risk and costs of losing the trust of their customers, employees and partners. Emerging technologies tied to algorithmic trust include secure access service edge, differential privacy, authenticated provenance, bring your own identity, responsible AI and explainable AI."
From the peak of inflated expectations to a deep dive in machine learning interpretability
The above quote is taken from Gartner's newly released 2020 Hype Cycle for Emerging Technologies. In it, explainable AI is placed at the peak of inflated expectations. In other words, we have reached peak hype for explainable AI. To put that into perspective, a recap may be useful.
As experts such as Gary Marcus point out, AI is probably not what you think it is. Many people today conflate AI with machine learning. While machine learning has made strides in recent years, it's not the only type of AI we have. Rule-based, symbolic AI has been around for years, and it has always been explainable.
Incidentally, that kind of AI, in the form of "Ontologies and Graphs" is also included in the same Gartner Hype Cycle, albeit in a different phase -- the trough of disillusionment. Incidentally, again, that's conflating. Ontologies are part of AI, while graphs, not necessarily.
That said: If you are interested in getting a better understanding of the state of the art in explainable AI machine learning, reading Christoph Molnar's book is a good place to start. Molnar is a data scientist and Ph.D. candidate in interpretable machine learning. Molnar has written the book Interpretable Machine Learning: A Guide for Making Black Box Models Explainable, in which he elaborates on the issue and examines methods for achieving explainability.
Recently, Molnar and a group of researchers attempted to addresses ML practitioners by raising awareness of pitfalls and pointing out solutions for correct model interpretation, as well as ML researchers by discussing open issues for further research. Their work was published as a research paper, titled Pitfalls to Avoid when Interpreting Machine Learning Models, by the ICML 2020 Workshop XXAI: Extending Explainable AI Beyond Deep Models and Classifiers.
Similar to Molnar's book, the paper is thorough. Admittedly, however, it's also more involved. Yet, Molnar has striven to make it more approachable by means of visualization, using what he dubs "poorly drawn comics" to highlight each pitfall. As with Molnar's book on interpretable machine learning, we summarize findings here, while encouraging readers to dive in for themselves.
The paper mainly focuses on the pitfalls of global interpretation techniques when the full functional relationship underlying the data is to be analyzed. Discussion of "local" interpretation methods, where individual predictions are to be explained, is out of scope. For a reference on global vs. local interpretations, you can refer to Molnar's book as previously covered on ZDNet.
Authors note that ML models usually contain non-linear effects and higher-order interactions. As interpretations are based on simplifying assumptions, the associated conclusions are only valid if we have checked that the assumptions underlying our simplifications are not substantially violated.
In classical statistics this process is called "model diagnostics," and the research claims that a similar process is necessary for interpretable ML (IML) based techniques. The research identifies and describes pitfalls to avoid when interpreting ML models, reviews (partial) solutions for practitioners, and discusses open issues that require further research.
Bad Model Generalization, Unnecessary Use of Complex Models
Under- or overfitting models will result in misleading interpretations regarding true feature effects and importance scores, as the model does not match the underlying data generating process well. Evaluation of training data should not be used for ML models due to the danger of overfitting. We have to resort to out-of-sample validation such as cross-validation procedures.
Formally, IML methods are designed to interpret the model instead of drawing inferences about the data generating process. In practice, however, the latter is the goal of the analysis, not the former. If a model approximates the data generating process well enough, its interpretation should reveal insights into the underlying process. Interpretations can only be as good as their underlying models. It is crucial to properly evaluate models using training and test splits -- ideally using a resampling scheme.
Flexible models should be part of the model selection process so that the true data-generating function is more likely to be discovered. This is important, as the Bayes error for most practical situations is unknown, and we cannot make absolute statements about whether a model already fits the data optimally.
Using opaque, complex ML models when an interpretable model would have been sufficient (i.e., having similar performance) is considered a common mistake. Starting with simple, interpretable models and gradually increasing complexity in a controlled, step-wise manner, where predictive performance is carefully measured and compared is recommended.
Measures of model complexity allow us to quantify the trade-off between complexity and performance and to automatically optimize for multiple objectives beyond performance. Some steps toward quantifying model complexity have been made. However, further research is required as there is no single perfect definition of interpretability but rather multiple, depending on the context.
Ignoring Feature Dependence
This pitfall is further analyzed in three sub-categories: Interpretation with extrapolation, confusing correlation with dependence, and misunderstanding conditional interpretation.
Interpretation with Extrapolation refers to producing artificial data points that are used for model predictions with perturbations. These are aggregated to produce global interpretations. But if features are dependent, perturbation approaches produce unrealistic data points. In addition, even if features are independent, using an equidistant grid can produce unrealistic values for the feature of interest. Both issues can result in misleading interpretations.
Before applying interpretation methods, practitioners should check for dependencies between features in the data (e.g., via descriptive statistics or measures of dependence). When it is unavoidable to include dependent features in the model, which is usually the case in ML scenarios, additional information regarding the strength and shape of the dependence structure should be provided.
Confusing correlation with dependence is a typical error. The Pearson correlation coefficient (PCC) is a measure used to track dependency among ML features. But features with PCC close to zero can still be dependent and cause misleading model interpretations. While independence between two features implies that the PCC is zero, the converse is generally false.
Any type of dependence between features can have a strong impact on the interpretation of the results of IML methods. Thus, knowledge about (possibly non-linear) dependencies between features is crucial. Low-dimensional data can be visualized to detect dependence. For high-dimensional data, several other measures of dependence in addition to PCC can be used.
Misunderstanding conditional interpretation. Conditional variants to estimate feature effects and importance scores require a different interpretation. While conditional variants for feature effects avoid model extrapolations, these methods answer a different question. Interpretation methods that perturb features independently of others also yield an unconditional interpretation.
Conditional variants do not replace values independently of other features, but in such a way that they conform to the conditional distribution. This changes the interpretation as the effects of all dependent features become entangled. The safest option would be to remove dependent features, but this is usually infeasible in practice.
When features are highly dependent and conditional effects and importance scores are used, the practitioner has to be aware of the distinct interpretation. Currently, no approach allows us to simultaneously avoid model extrapolations and to allow a conditional interpretation of effects and importance scores for dependent features.
Misleading Effect due to Interactions, Ignoring Estimation Uncertainty, Ignoring Multiple Comparisons
Global interpretation methods can produce misleading interpretations when features interact. Many interpretation methods cannot separate interactions from main effects. Most methods that identify and visualize interactions are not able to identify higher-order interactions and interactions of dependent features.
There are some methods to deal with this, but further research is still warranted. Furthermore, solutions lack in automatic detection and ranking of all interactions of a model as well as specifying the type of modeled interaction.
Due to the variance in the estimation process, interpretations of ML models can become misleading. When sampling techniques are used to approximate expected values, estimates vary, depending on the data used for the estimation. Furthermore, the obtained ML model is also a random variable, as it is generated on randomly sampled data and the inducing algorithm might contain stochastic components as well.
Hence, the model variance has to be taken into account. The true effect of a feature may be flat, but purely by chance, especially on smaller data, an effect might algorithmically be detected. This effect could cancel out once averaged over multiple model fits. The researchers note the uncertainty in feature effect methods has not been studied in detail.
Simultaneously testing the importance of multiple features will result in false-positive interpretations if the multiple comparisons problem (MCP) is ignored. MCP is well known in significance tests for linear models and similarly exists in testing for feature importance in ML.
For example, when simultaneously testing the importance of 50 features, even if all features are unimportant, the probability of observing that at least one feature is significantly important is ≈ 0.923. Multiple comparisons will even be more problematic, the higher dimensional a dataset is. Since MCP is well known in statistics, the authors refer practitioners to existing overviews and discussions of alternative adjustment methods.
Unjustified Causal Interpretation
Practitioners are often interested in causal insights into the underlying data-generating mechanisms, which IML methods, in general, do not provide. Common causal questions include the identification of causes and effects, predicting the effects of interventions, and answering counterfactual questions. In the search for answers, researchers can be tempted to interpret the result of IML methods from a causal perspective.
However, a causal interpretation of predictive models is often not possible. Standard supervised ML models are not designed to model causal relationships but to merely exploit associations. A model may, therefore, rely on the causes and effects of the target variable as well as on variables that help to reconstruct unobserved influences.
Consequently, the question of whether a variable is relevant to a predictive model does not directly indicate whether a variable is a cause, an effect, or does not stand in any causal relation to the target variable.
As the researchers note, the challenge of causal discovery and inference remains an open key issue in the field of machine learning. Careful research is required to make explicit under which assumptions what insight about the underlying data generating mechanism can be gained by interpreting a machine learning model
Groundwork vs. Hype
Molnar et. al. offer an involved review of the pitfalls of global model-agnostic interpretation techniques for ML. Although as they note their list is far from complete, they cover common ones that pose a particularly high risk.
They aim to encourage a more cautious approach when interpreting ML models in practice, to point practitioners to already (partially) available solutions, and to stimulate further research.
Contrasting this highly involved and detailed groundwork to high-level hype and trends on explainable AI may be instructive.