Neural networks have been criticised for their inability to perform continual learning due to catastrophic forgetting and rapid unlearning of a past concept when a new concept is introduced. Catastrophic forgetting can be alleviated by specifically designed models and training techniques. This paper outlines a novel Spline Additive Model (SAM). SAM exhibits intrinsic memory retention with sufficient expressive power for many practical tasks, but is not a universal function approximator. SAM is extended with the Kolmogorov-Arnold representation theorem to a novel universal function approximator, called the Kolmogorov-Arnold Spline Additive Model - KASAM. The memory retention, expressive power and limitations of SAM and KASAM are illustrated analytically and empirically. SAM exhibited robust but imperfect memory retention, with small regions of overlapping interference in sequential learning tasks. KASAM exhibited greater susceptibility to catastrophic forgetting. KASAM in combination with pseudo-rehearsal training techniques exhibited superior performance in regression tasks and memory retention.
This paper proposes a one-covariate-at-a-time multiple testing (OCMT) approach to choose significant variables in high-dimensional nonparametric additive regression models. Similarly to Chudik, Kapetanios and Pesaran (2018), we consider the statistical significance of individual nonparametric additive components one at a time and take into account the multiple testing nature of the problem. One-stage and multiple-stage procedures are both considered. The former works well in terms of the true positive rate only if the marginal effects of all signals are strong enough; the latter helps to pick up hidden signals that have weak marginal effects. Simulations demonstrate the good finite sample performance of the proposed procedures. As an empirical application, we use the OCMT procedure on a dataset we extracted from the Longitudinal Survey on Rural Urban Migration in China. We find that our procedure works well in terms of the out-of-sample forecast root mean square errors, compared with competing methods.
The number of information systems (IS) studies dealing with explainable artificial intelligence (XAI) is currently exploding as the field demands more transparency about the internal decision logic of machine learning (ML) models. However, most techniques subsumed under XAI provide post-hoc-analytical explanations, which have to be considered with caution as they only use approximations of the underlying ML model. Therefore, our paper investigates a series of intrinsically interpretable ML models and discusses their suitability for the IS community. More specifically, our focus is on advanced extensions of generalized additive models (GAM) in which predictors are modeled independently in a non-linear way to generate shape functions that can capture arbitrary patterns but remain fully interpretable. In our study, we evaluate the prediction qualities of five GAMs as compared to six traditional ML models and assess their visual outputs for model interpretability. On this basis, we investigate their merits and limitations and derive design implications for further improvements.
Estimating treatment effects is of great importance for many biomedical applications with observational data. Particularly, interpretability of the treatment effects is preferable for many biomedical researchers. In this paper, we first provide a theoretical analysis and derive an upper bound for the bias of average treatment effect (ATE) estimation under the strong ignorability assumption. Derived by leveraging appealing properties of the Weighted Energy Distance, our upper bound is tighter than what has been reported in the literature. Motivated by the theoretical analysis, we propose a novel objective function for estimating the ATE that uses the energy distance balancing score and hence does not require correct specification of the propensity score model. We also leverage recently developed neural additive models to improve interpretability of deep learning models used for potential outcome prediction. We further enhance our proposed model with an energy distance balancing score weighted regularization. The superiority of our proposed model over current state-of-the-art methods is demonstrated in semi-synthetic experiments using two benchmark datasets, namely, IHDP and ACIC.
This paper covers predicting high-resolution electricity peak demand features given lower-resolution data. This is a relevant setup as it answers whether limited higher-resolution monitoring helps to estimate future high-resolution peak loads when the high-resolution data is no longer available. That question is particularly interesting for network operators considering replacing high-resolution monitoring predictive models due to economic considerations. We propose models to predict half-hourly minima and maxima of high-resolution (every minute) electricity load data while model inputs are of a lower resolution (30 minutes). We combine predictions of generalized additive models (GAM) and deep artificial neural networks (DNN), which are popular in load forecasting. We extensively analyze the prediction models, including the input parameters' importance, focusing on load, weather, and seasonal effects. The proposed method won a data competition organized by Western Power Distribution, a British distribution network operator. In addition, we provide a rigorous evaluation study that goes beyond the competition frame to analyze the models' robustness. The results show that the proposed methods are superior to the competition benchmark concerning the out-of-sample root mean squared error (RMSE). This holds regarding the competition month and the supplementary evaluation study, which covers an additional eleven months. Overall, our proposed model combination reduces the out-of-sample RMSE by 57.4\% compared to the benchmark.
Anomaly detection (AD) plays an important role in numerous applications. We focus on two understudied aspects of AD that are critical for integration into real-world applications. First, most AD methods cannot incorporate labeled data that are often available in practice in small quantities and can be crucial to achieve high AD accuracy. Second, most AD methods are not interpretable, a bottleneck that prevents stakeholders from understanding the reason behind the anomalies. In this paper, we propose a novel AD framework that adapts a white-box model class, Generalized Additive Models, to detect anomalies using a partial identification objective which naturally handles noisy or heterogeneous features. In addition, the proposed framework, DIAD, can incorporate a small amount of labeled data to further boost anomaly detection performances in semi-supervised settings. We demonstrate the superiority of our framework compared to previous work in both unsupervised and semi-supervised settings using diverse tabular datasets. For example, under 5 labeled anomalies DIAD improves from 86.2\% to 89.4\% AUC by learning AD from unlabeled data. We also present insightful interpretations that explain why DIAD deems certain samples as anomalies.
Interpretable machine learning has demonstrated impressive performance while preserving explainability. In particular, neural additive models (NAM) offer the interpretability to the black-box deep learning and achieve state-of-the-art accuracy among the large family of generalized additive models. In order to empower NAM with feature selection and improve the generalization, we propose the sparse neural additive models (SNAM) that employ the group sparsity regularization (e.g. Group LASSO), where each feature is learned by a sub-network whose trainable parameters are clustered as a group. We study the theoretical properties for SNAM with novel techniques to tackle the non-parametric truth, thus extending from classical sparse linear models such as the LASSO, which only works on the parametric truth. Specifically, we show that SNAM with subgradient and proximal gradient descents provably converges to zero training loss as $t\to\infty$, and that the estimation error of SNAM vanishes asymptotically as $n\to\infty$. We also prove that SNAM, similar to LASSO, can have exact support recovery, i.e. perfect feature selection, with appropriate regularization. Moreover, we show that the SNAM can generalize well and preserve the `identifiability', recovering each feature's effect. We validate our theories via extensive experiments and further testify to the good accuracy and efficiency of SNAM.
We present fast classification techniques for sparse generalized linear and additive models. These techniques can handle thousands of features and thousands of observations in minutes, even in the presence of many highly correlated features. For fast sparse logistic regression, our computational speed-up over other best-subset search techniques owes to linear and quadratic surrogate cuts for the logistic loss that allow us to efficiently screen features for elimination, as well as use of a priority queue that favors a more uniform exploration of features. As an alternative to the logistic loss, we propose the exponential loss, which permits an analytical solution to the line search at each iteration. Our algorithms are generally 2 to 5 times faster than previous approaches. They produce interpretable models that have accuracy comparable to black box models on challenging datasets.
Though there is a strong consensus that word length and frequency are the most important single-word features determining visual-orthographic access to the mental lexicon, there is less agreement as how to best capture syntactic and semantic factors. The traditional approach in cognitive reading research assumes that word predictability from sentence context is best captured by cloze completion probability (CCP) derived from human performance data. We review recent research suggesting that probabilistic language models provide deeper explanations for syntactic and semantic effects than CCP. Then we compare CCP with (1) Symbolic n-gram models consolidate syntactic and semantic short-range relations by computing the probability of a word to occur, given two preceding words. (2) Topic models rely on subsymbolic representations to capture long-range semantic similarity by word co-occurrence counts in documents. (3) In recurrent neural networks (RNNs), the subsymbolic units are trained to predict the next word, given all preceding words in the sentences. To examine lexical retrieval, these models were used to predict single fixation durations and gaze durations to capture rapidly successful and standard lexical access, and total viewing time to capture late semantic integration. The linear item-level analyses showed greater correlations of all language models with all eye-movement measures than CCP. Then we examined non-linear relations between the different types of predictability and the reading times using generalized additive models. N-gram and RNN probabilities of the present word more consistently predicted reading performance compared with topic models or CCP.
Modern machine learning has achieved impressive prediction performance, but often sacrifices interpretability, a critical consideration in high-stakes domains such as medicine. In such settings, practitioners often use highly interpretable decision tree models, but these suffer from inductive bias against additive structure. To overcome this bias, we propose Fast Interpretable Greedy-Tree Sums (FIGS), which generalizes the CART algorithm to simultaneously grow a flexible number of trees in summation. By combining logical rules with addition, FIGS is able to adapt to additive structure while remaining highly interpretable. Extensive experiments on real-world datasets show that FIGS achieves state-of-the-art prediction performance. To demonstrate the usefulness of FIGS in high-stakes domains, we adapt FIGS to learn clinical decision instruments (CDIs), which are tools for guiding clinical decision-making. Specifically, we introduce a variant of FIGS known as G-FIGS that accounts for the heterogeneity in medical data. G-FIGS derives CDIs that reflect domain knowledge and enjoy improved specificity (by up to 20% over CART) without sacrificing sensitivity or interpretability. To provide further insight into FIGS, we prove that FIGS learns components of additive models, a property we refer to as disentanglement. Further, we show (under oracle conditions) that unconstrained tree-sum models leverage disentanglement to generalize more efficiently than single decision tree models when fitted to additive regression functions. Finally, to avoid overfitting with an unconstrained number of splits, we develop Bagging-FIGS, an ensemble version of FIGS that borrows the variance reduction techniques of random forests. Bagging-FIGS enjoys competitive performance with random forests and XGBoost on real-world datasets.
In the signal processing and statistics literature, the minimum description length (MDL) principle is a popular tool for choosing model complexity. Successful examples include signal denoising and variable selection in linear regression, for which the corresponding MDL solutions often enjoy consistent properties and produce very promising empirical results. This paper demonstrates that MDL can be extended naturally to the high-dimensional setting, where the number of predictors $p$ is larger than the number of observations $n$. It first considers the case of linear regression, then allows for outliers in the data, and lastly extends to the robust fitting of nonparametric additive models. Results from numerical experiments are presented to demonstrate the efficiency and effectiveness of the MDL approach.
Suppose we observe a random vector $X$ from some distribution $P$ in a known family with unknown parameters. We ask the following question: when is it possible to split $X$ into two parts $f(X)$ and $g(X)$ such that neither part is sufficient to reconstruct $X$ by itself, but both together can recover $X$ fully, and the joint distribution of $(f(X),g(X))$ is tractable? As one example, if $X=(X_1,\dots,X_n)$ and $P$ is a product distribution, then for any $m<n$, we can split the sample to define $f(X)=(X_1,\dots,X_m)$ and $g(X)=(X_{m+1},\dots,X_n)$. Rasines and Young (2022) offers an alternative approach that uses additive Gaussian noise -- this enables post-selection inference in finite samples for Gaussian distributed data and asymptotically when errors are non-Gaussian. In this paper, we offer a more general methodology for achieving such a split in finite samples by borrowing ideas from Bayesian inference to yield a (frequentist) solution that can be viewed as a continuous analog of data splitting. We call our method data fission, as an alternative to data splitting, data carving and p-value masking. We exemplify the method on a few prototypical applications, such as post-selection inference for trend filtering and other regression problems.
In the context of smart grids and load balancing, daily peak load forecasting has become a critical activity for stakeholders of the energy industry. An understanding of peak magnitude and timing is paramount for the implementation of smart grid strategies such as peak shaving. The modelling approach proposed in this paper leverages high-resolution and low-resolution information to forecast daily peak demand size and timing. The resulting multi-resolution modelling framework can be adapted to different model classes. The key contributions of this paper are a) a general and formal introduction to the multi-resolution modelling approach, b) a discussion on modelling approaches at different resolutions implemented via Generalised Additive Models and Neural Networks and c) experimental results on real data from the UK electricity market. The results confirm that the predictive performance of the proposed modelling approach is competitive with that of low- and high-resolution alternatives.
Sparse shrunk additive models and sparse random feature models have been developed separately as methods to learn low-order functions, where there are few interactions between variables, but neither offers computational efficiency. On the other hand, $\ell_2$-based shrunk additive models are efficient but do not offer feature selection as the resulting coefficient vectors are dense. Inspired by the success of the iterative magnitude pruning technique in finding lottery tickets of neural networks, we propose a new method -- Sparser Random Feature Models via IMP (ShRIMP) -- to efficiently fit high-dimensional data with inherent low-dimensional structure in the form of sparse variable dependencies. Our method can be viewed as a combined process to construct and find sparse lottery tickets for two-layer dense networks. We explain the observed benefit of SHRIMP through a refined analysis on the generalization error for thresholded Basis Pursuit and resulting bounds on eigenvalues. From function approximation experiments on both synthetic data and real-world benchmark datasets, we show that SHRIMP obtains better than or competitive test accuracy compared to state-of-art sparse feature and additive methods such as SRFE-S, SSAM, and SALSA. Meanwhile, SHRIMP performs feature selection with low computational complexity and is robust to the pruning rate, indicating a robustness in the structure of the obtained subnetworks. We gain insight into the lottery ticket hypothesis through SHRIMP by noting a correspondence between our model and weight/neuron subnetworks.
Recent strides in interpretable machine learning (ML) research reveal that models exploit undesirable patterns in the data to make predictions, which potentially causes harms in deployment. However, it is unclear how we can fix these models. We present our ongoing work, GAM Changer, an open-source interactive system to help data scientists and domain experts easily and responsibly edit their Generalized Additive Models (GAMs). With novel visualization techniques, our tool puts interpretability into action -- empowering human users to analyze, validate, and align model behaviors with their knowledge and values. Built using modern web technologies, our tool runs locally in users' computational notebooks or web browsers without requiring extra compute resources, lowering the barrier to creating more responsible ML models. GAM Changer is available at https://interpret.ml/gam-changer.
The recent abundance of data on electricity consumption at different scales opens new challenges and highlights the need for new techniques to leverage information present at finer scales in order to improve forecasts at wider scales. In this work, we take advantage of the similarity between this hierarchical prediction problem and multi-scale transfer learning. We develop two methods for hierarchical transfer learning, based respectively on the stacking of generalized additive models and random forests, and on the use of aggregation of experts. We apply these methods to two problems of electricity load forecasting at national scale, using smart meter data in the first case, and regional data in the second case. For these two usecases, we compare the performances of our methods to that of benchmark algorithms, and we investigate their behaviour using variable importance analysis. Our results demonstrate the interest of both methods, which lead to a significant improvement of the predictions.
In this study an exploration of insurance risk transfer is undertaken for the cyber insurance industry in the United States of America, based on the leading industry dataset of cyber events provided by Advisen. We seek to address two core unresolved questions. First, what factors are the most significant covariates that may explain the frequency and severity of cyber loss events and are they heterogeneous over cyber risk categories? Second, is cyber risk insurable in regards to the required premiums, risk pool sizes and how would this decision vary with the insured companies industry sector and size? We address these questions through a combination of regression models based on the class of Generalised Additive Models for Location Shape and Scale (GAMLSS) and a class of ordinal regressions. These models will then form the basis for our analysis of frequency and severity of cyber risk loss processes. We investigate the viability of insurance for cyber risk using a utility modelling framework with premium calculated by classical certainty equivalence analysis utilising the developed regression models. Our results provide several new key insights into the nature of insurability of cyber risk and rigorously address the two insurance questions posed in a real data driven case study analysis.
Although reinforcement learning (RL) has tremendous success in many fields, applying RL to real-world settings such as healthcare is challenging when the reward is hard to specify and no exploration is allowed. In this work, we focus on recovering clinicians' rewards in treating patients. We incorporate the what-if reasoning to explain the clinician's treatments based on their potential future outcomes. We use generalized additive models (GAMs) - a class of accurate, interpretable models - to recover the reward. In both simulation and a real-world hospital dataset, we show our model outperforms baselines. Finally, our model's explanations match several clinical guidelines when treating patients while we found the commonly-used linear model often contradicts them.
There are proposals that extend the classical generalized additive models (GAMs) to accommodate high-dimensional data ($p>>n$) using group sparse regularization. However, the sparse regularization may induce excess shrinkage when estimating smooth functions, damaging predictive performance. Moreover, most of these GAMs consider an "all-in-all-out" approach for functional selection, rendering them difficult to answer if nonlinear effects are necessary. While some Bayesian models can address these shortcomings, using Markov chain Monte Carlo algorithms for model fitting creates a new challenge, scalability. Hence, we propose Bayesian hierarchical generalized additive models as a solution: we consider the smoothing penalty for proper shrinkage of curve interpolation via reparameterization. A novel two-part spike-and-slab LASSO prior for smooth functions is developed to address the sparsity of signals while providing extra flexibility to select the linear or nonlinear components of smooth functions. A scalable and deterministic algorithm, EM-Coordinate Descent, is implemented in an open-source R package BHAM. Simulation studies and metabolomics data analyses demonstrate improved predictive and computational performance against state-of-the-art models. Functional selection performance suggests trade-offs exist regarding the effect hierarchy assumption.
Post-hoc global/local feature attribution methods are progressively being employed to understand the decisions of complex machine learning models. Yet, because of limited amounts of data, it is possible to obtain a diversity of models with good empirical performance but that provide very different explanations for the same prediction, making it hard to derive insight from them. In this work, instead of aiming at reducing the under-specification of model explanations, we fully embrace it and extract logical statements about feature attributions that are consistent across all models with good empirical performance (i.e. all models in the Rashomon Set). We show that partial orders of local/global feature importance arise from this methodology enabling more nuanced interpretations by allowing pairs of features to be incomparable when there is no consensus on their relative importance. We prove that every relation among features present in these partial orders also holds in the rankings provided by existing approaches. Finally, we present three use cases employing hypothesis spaces with tractable Rashomon Sets (Additive models, Kernel Ridge, and Random Forests) and show that partial orders allow one to extract consistent local and global interpretations of models despite their under-specification.