The output of a machine learning algorithm can usually be represented by one or more multivariate functions of its input variables. Knowing the global properties of such functions can help in understanding the system that produced the data as well as interpreting and explaining corresponding model predictions. A method is presented for representing a general multivariate function as a tree of simpler functions. This tree exposes the global internal structure of the function by uncovering and describing the combined joint influences of subsets of its input variables. Given the inputs and corresponding function values, a function tree is constructed that can be used to rapidly identify and compute all of the function's main and interaction effects up to high order. Interaction effects involving up to four variables are graphically visualized.

The estimation of nested functions (i.e., functions of functions) is one of the central reasons for the success and popularity of machine learning. Today, artificial neural networks are the predominant class of algorithms in this area, known as representational learning. Here, we introduce Representational Gradient Boosting (RGB), a nonparametric algorithm that estimates functions with multi-layer architectures obtained using backpropagation in the space of functions. RGB does not need to assume a functional form in the nodes or output (e.g., linear models or rectified linear units), but rather estimates these transformations. RGB can be seen as an optimized stacking procedure where a meta algorithm learns how to combine different classes of functions (e.g., Neural Networks (NN) and Gradient Boosting (GB)), while building and optimizing them jointly in an attempt to compensate each other's weaknesses. This highlights a stark difference with current approaches to meta-learning that combine models only after they have been built independently. We showed that providing optimized stacking is one of the main advantages of RGB over current approaches. Additionally, due to the nested nature of RGB we also showed how it improves over GB in problems that have several high-order interactions. Finally, we investigate both theoretically and in practice the problem of recovering nested functions and the value of prior knowledge.

Many regression and classification procedures fit a parameterized function $f(x;w)$ of predictor variables $x$ to data $\{x_{i},y_{i}\}_1^N$ based on some loss criterion $L(y,f)$. Often, regularization is applied to improve accuracy by placing a constraint $P(w)\leq t$ on the values of the parameters $w$. Although efficient methods exist for finding solutions to these constrained optimization problems for all values of $t\geq0$ in the special case when $f$ is a linear function, none are available when $f$ is non-linear (e.g. Neural Networks). Here we present a fast algorithm that provides all such solutions for any differentiable function $f$ and loss $L$, and any constraint $P$ that is an increasing monotone function of the absolute value of each parameter. Applications involving sparsity inducing regularization of arbitrary Neural Networks are discussed. Empirical results indicate that these sparse solutions are usually superior to their dense counterparts in both accuracy and interpretability. This improvement in accuracy can often make Neural Networks competitive with, and sometimes superior to, state-of-the-art methods in the analysis of tabular data.

We propose a new method for supervised learning, especially suited to wide data where the number of features is much greater than the number of observations. The method combines the lasso ($\ell_1$) sparsity penalty with a quadratic penalty that shrinks the coefficient vector toward the leading principal components of the feature matrix. We call the proposed method the "principal components lasso" ("pcLasso"). The method can be especially powerful if the features are pre-assigned to groups (such as cell-pathways, assays or protein interaction networks). In that case, pcLasso shrinks each group-wise component of the solution toward the leading principal components of that group. In the process, it also carries out selection of the feature groups. We provide some theory for this method and illustrate it on a number of simulated and real data examples.

The observation that decision trees are boosting algorithms, as cited in our work (1) and acknowledged by Nock and Nielsen (2), was first established by refs. 3 and 4. This was later used by refs. 5 and 6 to develop, to the best of our knowledge, the first decision tree algorithms based purely on boosting. This work, cited in our article, precedes refs. 7 and 8 cited by Nock and Nielsen (2). The original and important contributions of refs. 7 and 8 as they pertain to this discussion was to theoretically prove convergence rates for decision tree algorithms built with boosting, along with … [↵][1]1To whom correspondence may be addressed. Email: gilmer.valdes{at}ucsf.edu. [1]: #xref-corresp-1-1

Machine learning is proving invaluable across disciplines. However, its success is often limited by the quality and quantity of available data, while its adoption is limited by the level of trust afforded by given models. Human vs. machine performance is commonly compared empirically to decide whether a certain task should be performed by a computer or an expert. In reality, the optimal learning strategy may involve combining the complementary strengths of humans and machines. Here, we present expert-augmented machine learning (EAML), an automated method that guides the extraction of expert knowledge and its integration into machine-learned models. We used a large dataset of intensive-care patient data to derive 126 decision rules that predict hospital mortality. Using an online platform, we asked 15 clinicians to assess the relative risk of the subpopulation defined by each rule compared to the total sample. We compared the clinician-assessed risk to the empirical risk and found that, while clinicians agreed with the data in most cases, there were notable exceptions where they overestimated or underestimated the true risk. Studying the rules with greatest disagreement, we identified problems with the training data, including one miscoded variable and one hidden confounder. Filtering the rules based on the extent of disagreement between clinician-assessed risk and empirical risk, we improved performance on out-of-sample data and were able to train with less data. EAML provides a platform for automated creation of problem-specific priors, which help build robust and dependable machine-learning models in critical applications.

Professor Efron has presented us with a thought-provoking paper on the relationship between prediction, estimation, and attribution in the modern era of data science. While we appreciate many of his arguments, we see more of a continuum between the old and new methodology, and the opportunity for both to improve through their synergy.

We propose a generalization of the lasso that allows the model coefficients to vary as a function of a general set of modifying variables. These modifiers might be variables such as gender, age or time. The paradigm is quite general, with each lasso coefficient modified by a sparse linear function of the modifying variables $Z$. The model is estimated in a hierarchical fashion to control the degrees of freedom and avoid overfitting. The modifying variables may be observed, observed only in the training set, or unobserved overall. There are connections of our proposal to varying coefficient models and high-dimensional interaction models. We present a computationally efficient algorithm for its optimization, with exact screening rules to facilitate application to large numbers of predictors. The method is illustrated on a number of different simulated and real examples.

The goal of regression analysis is to predict the value of a numeric outcome variable y given a vector of joint values of other (predictor) variables x. Usually a particular x-vector does not specify a repeatable value for y, but rather a probability distribution of possible y--values, p(y|x). This distribution has a location, scale and shape, all of which can depend on x, and are needed to infer likely values for y given x. Regression methods usually assume that training data y-values are perfect numeric realizations from some well behaived p(y|x). Often actual training data y-values are discrete, truncated and/or arbitrary censored. Regression procedures based on an optimal transformation strategy are presented for estimating location, scale and shape of p(y|x) as general functions of x, in the possible presence of such imperfect training data. In addition, validation diagnostics are presented to ascertain the quality of the solutions.

Often machine learning methods are applied and results reported in cases where there is little to no information concerning accuracy of the output. Simply because a computer program returns a result does not insure its validity. If decisions are to be made based on such results it is important to have some notion of their veracity. Contrast trees represent a new approach for assessing the accuracy of many types of machine learning estimates that are not amenable to standard (cross) validation methods. In situations where inaccuracies are detected boosted contrast trees can often improve performance. A special case, distribution boosting, provides an assumption free method for estimating the full probability distribution of an outcome variable given any set of joint input predictor variable values.

The expansion of machine learning to high-stakes application domains such as medicine, finance, and criminal justice, where making informed decisions requires clear understanding of the model, has increased the interest in interpretable machine learning. The widely used Classification and Regression Trees (CART) have played a major role in health sciences, due to their simple and intuitive explanation of predictions. Ensemble methods like gradient boosting can improve the accuracy of decision trees, but at the expense of the interpretability of the generated model. Additive models, such as those produced by gradient boosting, and full interaction models, such as CART, have been investigated largely in isolation. We show that these models exist along a spectrum, revealing previously unseen connections between these approaches. This paper introduces a rigorous formalization for the additive tree, an empirically validated learning technique for creating a single decision tree, and shows that this method can produce models equivalent to CART or gradient boosted stumps at the extremes by varying a single parameter. Although the additive tree is designed primarily to provide both the model interpretability and predictive performance needed for high-stakes applications like medicine, it also can produce decision trees represented by hybrid models between CART and boosted stumps that can outperform either of these approaches.