Estimation and Interpretation of Machine Learning Models with Customized Surrogate Model

Machine learning has the potential to predict unseen data and thus improve the productivity and processes of daily life activities. Notwithstanding its adaptiveness, several sensitive applications based on such technology cannot compromise our trust in them; thus, highly accurate machine learning models require reason. Such models are black boxes for end-users. Therefore, the concept of interpretability plays the role if assisting users in a couple of ways. Interpretable models are models that possess the quality of explaining predictions. Different strategies have been proposed for the aforementioned concept but some of these require an excessive amount of effort, lack generalization, are not agnostic and are computationally expensive. Thus, in this work, we propose a strategy that can tackle the aforementioned issues. A surrogate model assisted us in building interpretable models. Moreover, it helped us achieve results with accuracy close to that of the black box model but with less processing time. Thus, the proposed technique is computationally cheaper than traditional methods. The significance of such a novel technique is that data science developers will not have to perform strenuous hands-on activities to undertake feature engineering tasks and end-users will have the graphical-based explanation of complex models in a comprehensive way—consequently building trust in a machine.

Reconstruction of pairwise interactions using energy-based models*

Pairwise models like the Ising model or the generalized Potts model have found many successful applications in fields like physics, biology, and economics. Closely connected is the problem of inverse statistical mechanics, where the goal is to infer the parameters of such models given observed data. An open problem in this field is the question of how to train these models in the case where the data contain additional higher-order interactions that are not present in the pairwise model. In this work, we propose an approach based on energy-based models and pseudolikelihood maximization to address these complications: we show that hybrid models, which combine a pairwise model and a neural network, can lead to significant improvements in the reconstruction of pairwise interactions. We show these improvements to hold consistently when compared to a standard approach using only the pairwise model and to an approach using only a neural network. This is in line with the general idea that simple interpretable models and complex black-box models are not necessarily a dichotomy: interpolating these two classes of models can allow to keep some advantages of both.

Deep Learning Predicts EBV Status in Gastric Cancer Based on Spatial Patterns of Lymphocyte Infiltration

EBV infection occurs in around 10% of gastric cancer cases and represents a distinct subtype, characterized by a unique mutation profile, hypermethylation, and overexpression of PD-L1. Moreover, EBV positive gastric cancer tends to have higher immune infiltration and a better prognosis. EBV infection status in gastric cancer is most commonly determined using PCR and in situ hybridization, but such a method requires good nucleic acid preservation. Detection of EBV status with histopathology images may complement PCR and in situ hybridization as a first step of EBV infection assessment. Here, we developed a deep learning-based algorithm to directly predict EBV infection in gastric cancer from H&E stained histopathology slides. Our model can not only predict EBV infection in gastric cancers from tumor regions but also from normal regions with potential changes induced by adjacent EBV+ regions within each H&E slide. Furthermore, in cohorts with zero EBV abundances, a significant difference of immune infiltration between high and low EBV score samples was observed, consistent with the immune infiltration difference observed between EBV positive and negative samples. Therefore, we hypothesized that our model’s prediction of EBV infection is partially driven by the spatial information of immune cell composition, which was supported by mostly positive local correlations between the EBV score and immune infiltration in both tumor and normal regions across all H&E slides. Finally, EBV scores calculated from our model were found to be significantly associated with prognosis. This framework can be readily applied to develop interpretable models for prediction of virus infection across cancers.

Flexible Bayesian Nonlinear Model Configuration

Regression models are used in a wide range of applications providing a powerful scientific tool for researchers from different fields. Linear, or simple parametric, models are often not sufficient to describe complex relationships between input variables and a response. Such relationships can be better described through flexible approaches such as neural networks, but this results in less interpretable models and potential overfitting. Alternatively, specific parametric nonlinear functions can be used, but the specification of such functions is in general complicated. In this paper, we introduce a flexible approach for the construction and selection of highly flexible nonlinear parametric regression models. Nonlinear features are generated hierarchically, similarly to deep learning, but have additional flexibility on the possible types of features to be considered. This flexibility, combined with variable selection, allows us to find a small set of important features and thereby more interpretable models. Within the space of possible functions, a Bayesian approach, introducing priors for functions based on their complexity, is considered. A genetically modi ed mode jumping Markov chain Monte Carlo algorithm is adopted to perform Bayesian inference and estimate posterior probabilities for model averaging. In various applications, we illustrate how our approach is used to obtain meaningful nonlinear models. Additionally, we compare its predictive performance with several machine learning algorithms.

Random Forest Similarity Maps: A Scalable Visual Representation for Global and Local Interpretation

Machine Learning prediction algorithms have made significant contributions in today’s world, leading to increased usage in various domains. However, as ML algorithms surge, the need for transparent and interpretable models becomes essential. Visual representations have shown to be instrumental in addressing such an issue, allowing users to grasp models’ inner workings. Despite their popularity, visualization techniques still present visual scalability limitations, mainly when applied to analyze popular and complex models, such as Random Forests (RF). In this work, we propose Random Forest Similarity Map (RFMap), a scalable interactive visual analytics tool designed to analyze RF ensemble models. RFMap focuses on explaining the inner working mechanism of models through different views describing individual data instance predictions, providing an overview of the entire forest of trees, and highlighting instance input feature values. The interactive nature of RFMap allows users to visually interpret model errors and decisions, establishing the necessary confidence and user trust in RF models and improving performance.

Neurocomputational Models of Language Processing

Efforts to understand the brain bases of language face the Mapping Problem: At what level do linguistic computations and representations connect to human neurobiology? We review one approach to this problem that relies on rigorously defined computational models to specify the links between linguistic features and neural signals. Such tools can be used to estimate linguistic predictions, model linguistic features, and specify a sequence of processing steps that may be quantitatively fit to neural signals collected while participants use language. Progress has been helped by advances in machine learning, attention to linguistically interpretable models, and openly shared data sets that allow researchers to compare and contrast a variety of models. We describe one such data set in detail in the Supplementary Appendix. Expected final online publication date for the Annual Review of Linguistics, Volume 8 is January 2022. Please see http://www.annualreviews.org/page/journal/pubdates for revised estimates.

The Resolved Mutual Information Function as a Structural Fingerprint of Biomolecular Sequences for Interpretable Machine Learning Classifiers

In the present article we propose the application of variants of the mutual information function as characteristic fingerprints of biomolecular sequences for classification analysis. In particular, we consider the resolved mutual information functions based on Shannon-, Rényi-, and Tsallis-entropy. In combination with interpretable machine learning classifier models based on generalized learning vector quantization, a powerful methodology for sequence classification is achieved which allows substantial knowledge extraction in addition to the high classification ability due to the model-inherent robustness. Any potential (slightly) inferior performance of the used classifier is compensated by the additional knowledge provided by interpretable models. This knowledge may assist the user in the analysis and understanding of the used data and considered task. After theoretical justification of the concepts, we demonstrate the approach for various example data sets covering different areas in biomolecular sequence analysis.

Credit scoring development in the light of the new Brazilian General Data Protection Law

With the advent of the new Brazilian General Data Protection Law (LGPD) which determines the right to the explanation of automated decisions, the use of non-interpretable models for human beings, known as black boxes, for the purposes of credit risk assessment may remain unfeasible. Thus, three different methods commonly applied to credit scoring – logistic regression, decision tree, and support vector machine (SVM) – were adjusted to an anonymized sample of a consumer credit portfolio from a credit union. Their results were compared and the adequacy of the explanation achieved for each classifier was assessed. Particularly for the SVM, which generated a black box model, a local interpretation method – the SHapley Additive exPlanation (SHAP) – was incorporated, enabling this machine learning classifier to fulfill the requirements imposed by the new LGPD, in equivalence to the inherent comprehensibility of the white box models.