Abstract Local model-agnostic additive explanation techniques decompose the predicted output of a black-box model into additive feature importance scores. Questions have been raised about the accuracy of the produced local additive explanations. We investigate this by studying whether some of the most popular explanation techniques can accurately explain the decisions of linear additive models. We show that even though the explanations generated by these techniques are linear additives, they can fail to provide accurate explanations when explaining linear additive models. In the experiments, we measure the accuracy of additive explanations, as produced by, e.g., LIME and SHAP, along with the non-additive explanations of Local Permutation Importance (LPI) when explaining Linear and Logistic Regression and Gaussian naive Bayes models over 40 tabular datasets. We also investigate the degree to which different factors, such as the number of numerical or categorical or correlated features, the predictive performance of the black-box model, explanation sample size, similarity metric, and the pre-processing technique used on the dataset can directly affect the accuracy of local explanations.

Interpretability is recognized as a key property of trustworthy predictive models. Only interpretable models make it straightforward to explain individual predictions, and allow inspection and analysis of the model itself. In real-world scenarios, these explanations and insights are often needed for a specific batch of predictions, i.e., a production set. If the input vectors for this production set are available when generating the predictive model, a methodology called oracle coaching can be used to produce highly accurate and interpretable models optimized for the specific production set. In this paper, oracle coaching is, for the first time, combined with the conformal prediction framework for predictive regression. A conformal regressor, which is built on top of a standard regression model, outputs valid prediction intervals, i.e., the error rate on novel data is bounded by a preset significance level, as long as the labeled data used for calibration is exchangeable with production data. Since validity is guaranteed for all conformal predictors, the key performance metric is the size of the prediction intervals, where tighter (more efficient) intervals are preferred. The efficiency of a conformal model depends on several factors, but more accurate underlying models will generally also lead to improved efficiency in the corresponding conformal predictor. A key contribution in this paper is the design of setups ensuring that when oracle coached regression trees, that per definition utilize knowledge about production data, are used as underlying models for conformal regressors, these remain valid. The experiments, using 20 publicly available regression data sets, demonstrate the validity of the suggested setups. Results also show that utilizing oracle-coached underlying models will generally lead to significantly more efficient conformal regressors, compared to when these are built on top of models induced using only training data.

Class labels output by standard decision trees are not very useful for making informed decisions, e.g., when comparing the expected utility of various alternatives. In contrast, probability estimation trees (PETs) output class probability distributions rather than single class labels. It is well known that estimating class probabilities in PETs by relative frequencies often lead to extreme probability estimates, and a number of approaches to provide more well-calibrated estimates have been proposed. In this study, a recent model-agnostic calibration approach, called Venn-Abers predictors is, for the first time, considered in the context of decision trees. Results from a large-scale empirical investigation are presented, comparing the novel approach to previous calibration techniques with respect to several different performance metrics, targeting both predictive performance and reliability of the estimates. All approaches are considered both with and without Laplace correction. The results show that using Venn-Abers predictors for calibration is a highly competitive approach, significantly outperforming Platt scaling, Isotonic regression and no calibration, with respect to almost all performance metrics used, independently of whether Laplace correction is applied or not. The only exception is AUC, where using non-calibrated PETs together with Laplace correction, actually is the best option, which can be explained by the fact that AUC is not affected by the absolute, but only relative, values of the probability estimates.

The Gradient Boosted Tree (GBT) algorithm is one of the most popular machine learning algorithms used in production, for tasks that include Click-Through Rate (CTR) prediction and learning-to-rank. To deal with the massive datasets available today, many distributed GBT methods have been proposed. However, they all assume a row-distributed dataset, addressing scalability only with respect to the number of data points and not the number of features, and increasing communication cost for high-dimensional data. In order to allow for scalability across both the data point and feature dimensions, and reduce communication cost, we propose block-distributed GBTs. We achieve communication efficiency by making full use of the data sparsity and adapting the Quickscorer algorithm to the block-distributed setting. We evaluate our approach using datasets with millions of features, and demonstrate that we are able to achieve multiple orders of magnitude reduction in communication cost for sparse data, with no loss in accuracy, while providing a more scalable design. As a result, we are able to reduce the training time for high-dimensional data, and allow more cost-effective scale-out without the need for expensive network communication.

The enormous amounts of data that are continuously recorded in electronic health record systems offer ample opportunities for data science applications to improve healthcare. There are, however, challenges involved in using such data for machine learning, such as high dimensionality and sparsity, as well as an inherent heterogeneity that does not allow the distinct types of clinical data to be treated in an identical manner. On the other hand, there are also similarities across data types that may be exploited, e.g., the possibility of representing some of them as sequences. Here, we apply the notions underlying distributional semantics, i.e., methods that model the meaning of words in semantic (vector) space on the basis of co-occurrence information, to four distinct types of clinical data: free-text notes, on the one hand, and clinical events, in the form of diagnosis codes, drug codes and measurements, on the other hand. Each semantic space contains continuous vector representations for every unique word and event, which can then be used to create representations of, e.g., care episodes that, in turn, can be exploited by the learning algorithm. This approach does not only reduce sparsity, but also takes into account, and explicitly models, similarities between various items, and it does so in an entirely data-driven fashion. Here, we report on a series of experiments using the random forest learning algorithm that demonstrate the effectiveness, in terms of accuracy and area under ROC curve, of the proposed representation form over the commonly used bag-of-items counterpart. The experiments are conducted on 27 real datasets that each involves the (binary) classification task of detecting a particular adverse drug event. It is also shown that combining structured and unstructured data leads to significant improvements over using only one of them.

The ability to detect adverse drug events (ADEs) in electronic health records (EHRs) is useful in many medical applications, such as alerting systems that indicate when an ADE-specific diagnosis code should be assigned. Automating the detection of ADEs can be attempted by applying machine learning to existing, labeled EHR data. How to do this in an effective manner is, however, an open question. The issues addressed in this study concern the granularity of the classification task: (1) If we wish to predict the occurrence of any ADE, is it advantageous to conflate the various ADE class labels prior to learning, or should they be merged post prediction? (2) If we wish to predict a family of ADEs or even a specific ADE, can the predictive performance be enhanced by dividing the classification task into a cascading scheme: predicting first, on a coarse level, whether there is an ADE or not, and, in the former case, followed by a more specific prediction on which family the ADE belongs to, and then finally a prediction on the specific ADE within that particular family? In this study, we conduct a series of experiments using a real, clinical dataset comprising healthcare episodes that have been assigned one of eight ADE-related diagnosis codes and a set of randomly extracted episodes that have not been assigned any ADE code. It is shown that, when distinguishing between ADEs and non-ADEs, merging the various ADE labels prior to learning leads to significantly higher predictive performance in terms of accuracy and area under ROC curve. A cascade of random forests is moreover constructed to determine either the family of ADEs or the specific class label; here, the performance is indeed enhanced compared to directly employing a one-step prediction. This study concludes that, if predictive performance is of primary importance, the cascading scheme should be the recommended approach over employing a one-step prediction for detecting ADEs in EHRs.

Online predictive modeling of streaming data is a key task for big data analytics. In this paper, a novel approach for efficient online learning of regression trees is proposed, which continuously updates, rather than retrains, the tree as more labeled data become available. A conformal predictor outputs prediction sets instead of point predictions; which for regression translates into prediction intervals. The key property of a conformal predictor is that it is always valid, i.e., the error rate, on novel data, is bounded by a preset significance level. Here, we suggest applying Mondrian conformal prediction on top of the resulting models, in order to obtain regression trees where not only the tree, but also each and every rule, corresponding to a path from the root node to a leaf, is valid. Using Mondrian conformal prediction, it becomes possible to analyze and explore the different rules separately, knowing that their accuracy, in the long run, will not be below the preset significance level. An empirical investigation, using 17 publicly available data sets, confirms that the resulting rules are independently valid, but also shows that the prediction intervals are smaller, on average, than when only the global model is required to be valid. All-in-all, the suggested method provides a data miner or a decision maker with highly informative predictive models of streaming data.

Conformal prediction is a relatively new framework in which the predictive models output sets of predictions with a bound on the error rate, i.e., in a classification context, the probability of excluding the correct class label is lower than a predefined significance level. An investigation of the use of decision trees within the conformal prediction framework is presented, with the overall purpose to determine the effect of different algorithmic choices, including split criterion, pruning scheme and way to calculate the probability estimates. Since the error rate is bounded by the framework, the most important property of conformal predictors is efficiency, which concerns minimizing the number of elements in the output prediction sets. Results from one of the largest empirical investigations to date within the conformal prediction framework are presented, showing that in order to optimize efficiency, the decision trees should be induced using no pruning and with smoothed probability estimates. The choice of split criterion to use for the actual induction of the trees did not turn out to have any major impact on the efficiency. Finally, the experimentation also showed that when using decision trees, standard inductive conformal prediction was as efficient as the recently suggested method cross-conformal prediction. This is an encouraging results since cross-conformal prediction uses several decision trees, thus sacrificing the interpretability of a single decision tree.

Both theory and a wealth of empirical studies have established that ensembles are more accurate than single predictive models. Unfortunately, the problem of how to maximize ensemble accuracy is, especially for classification, far from solved. In essence, the key problem is to find a suitable criterion, typically based on training or selection set performance, highly correlated with ensemble accuracy on novel data. Several studies have, however, shown that it is difficult to come up with a single measure, such as ensemble or base classifier selection set accuracy, or some measure based on diversity, that is a good general predictor for ensemble test accuracy. This paper presents a novel technique that for each learning task searches for the most effective combination of given atomic measures, by means of a genetic algorithm. Ensembles built from either neural networks or random forests were empirically evaluated on 30 UCI datasets. The experimental results show that when using the generated combined optimization criteria to rank candidate ensembles, a higher test set accuracy for the top ranked ensemble was achieved, compared to using ensemble accuracy on selection data alone. Furthermore, when creating ensembles from a pool of neural networks, the use of the generated combined criteria was shown to generally outperform the use of estimated ensemble accuracy as the single optimization criterion.

Two strategies for fusing information from multiple sources when generating predictive models in the domain of pesticide classification are investigated: i) fusing different sets of features (molecular descriptors) before building a model and ii) fusing the classifiers built from the individual descriptor sets. An empirical investigation demonstrates that the choice of strategy can have a significant impact on the predictive performance. Furthermore, the experiment shows that the best strategy is dependent on the type of predictive model considered. When generating a decision tree for pesticide classification, a statistically significant difference in accuracy is observed in favor of combining predictions from the individual models compared to generating a single model from the fused set of molecular descriptors. On the other hand, when the model consists of an ensemble of decision trees, a statistically significant difference in accuracy is observed in favor of building the model from the fused set of descriptors compared to fusing ensemble models built from the individual sources.

Decision lists (or ordered rule sets) have two attractive properties compared to unordered rule sets: they require a simpler classification procedure and they allow for a more compact representation. However, it is an open question what effect these properties have on the area under the ROC curve (AUC). Two ways of forming decision lists are considered in this study: by generating a sequence of rules, with a default rule for one of the classes, and by imposing an order upon rules that have been generated for all classes. An empirical investigation shows that the latter method gives a significantly higher AUC than the former, demonstrating that the compactness obtained by using one of the classes as a default is indeed associated with a cost. Furthermore, by using all applicable rules rather than the first in an ordered set, an even further significant improvement in AUC is obtained, demonstrating that the simple classification procedure is also associated with a cost. The observed gains in AUC for unordered rule sets compared to decision lists can be explained by that learning rules for all classes as well as combining multiple rules allow for examples to be ranked according to a more fine-grained scale compared to when applying rules in a fixed order and providing a default rule for one of the classes.