Comparison of Bayesian predictive methods for model selection
Abstract
The goal of this paper is to compare several widely used Bayesian model selection methods in practical model selection problems, highlight their differences and give recommendations about the preferred approaches. We focus on the variable subset selection for regression and classification and perform several numerical experiments using both simulated and real world data. The results show that the optimization of a utility estimate such as the cross-validation (CV) score is liable to finding overfitted models due to relatively high variance in the utility estimates when the data is scarce. This can also lead to substantial selection induced bias and optimism in the performance evaluation for the selected model. From a predictive viewpoint, best results are obtained by accounting for model uncertainty by forming the full encompassing model, such as the Bayesian model averaging solution over the candidate models. If the encompassing model is too complex, it can be robustly simplified by the projection method, in which the information of the full model is projected onto the submodels. This approach is substantially less prone to overfitting than selection based on CV-score. Overall, the projection method appears to outperform also the maximum a posteriori model and the selection of the most probable variables. The study also demonstrates that the model selection can greatly benefit from using cross-validation outside the searching process both for guiding the model size selection and assessing the predictive performance of the finally selected model.
Keywords
Bayesian model selection Cross-validation Reference model Projection Selection bias1 Introduction
Model selection is one of the fundamental problems in statistical modeling. The often adopted view for model selection is not to identify the true underlying model but rather to find a model which is useful. Typically the usefulness of a model is measured by its ability to make predictions about future or yet unseen observations. Even though the prediction would not be the most important part concerning the modelling problem at hand, the predictive ability is still a natural measure for figuring out whether the model makes sense or not. If the model gives poor predictions, there is not much point in trying to interpret the model parameters. We refer to the model selection based on assessing the predictive ability of the candidate models as predictive model selection.
Numerous methods for Bayesian predictive model selection and assessment have been proposed and the various approaches and their theoretical properties have been extensively reviewed by Vehtari and Ojanen (2012). This paper is a follow-up to their work. The review of Vehtari and Ojanen (2012) being qualitative, our contribution is to compare many of the different methods quantitatively in practical model selection problems, discuss the differences, and give recommendations about the preferred approaches. We believe this study will give useful insights because the comparisons to the existing techniques are often inadequate in the original articles presenting new methods. Our experiments will focus on variable subset selection on linear models for regression and classification, but the discussion is general and the concepts apply also to other model selection problems.
A fairly popular method in Bayesian literature is to select the maximum a posteriori (MAP) model which, in the case of a uniform prior on the model space, reduces to maximizing the marginal likelihood and the model selection can be performed using Bayes factors (e.g., Kass and Raftery 1995; Han and Carlin 2001). In the input variable selection context, also the marginal probabilities of the variables have been used (e.g., Brown et al. 1998; Barbieri and Berger 2004; Narisetty and He 2014). In fact, often in the Bayesian variable selection literature, the selection is assumed to be performed using one of these approaches, and the studies focus on choosing priors for the model space and parameters of the individual models (see the review by, O’Hara and Sillanpää 2009). These studies stem from the fact that sampling the high dimensional model space is highly nontrivial, and because it is well known that both the Bayes factors and the marginal probabilities are sensitive to the prior choices (e.g., Kass and Raftery 1995; O’Hara and Sillanpää 2009).
However, we want to make a distinction between prior and model selection. More specifically, we want to address the question, given our prior beliefs, how should we choose the model? In our view, the prior choice should reflect our genuine beliefs about the unknown quantities, such as the number of relevant input variables. For this reason our goal is not to compare and review the vast literature on the priors and computation strategies, which is already done by O’Hara and Sillanpää (2009). Still, we feel this literature is very essential, giving tools for formulating different prior beliefs and performing the necessary posterior computations.
Categorization of the different model selection methods discussed in this paper
View | Methods | |
---|---|---|
Section 2.2 | Generalization utility estimation (\( \mathscr {M} \)-open view) | Cross-validation, WAIC, DIC |
Section 2.3 | Self/posterior predictive criteria (Mixed view) | \(L^2\)-, \(L^2_ \text {cv} \)- and \(L^2_k\)-criteria |
Section 2.4 | Reference model approach (\( \mathscr {M} \)-completed view) | Reference predictive method, projection predictive method |
Section 2.5 | Model space approach (\( \mathscr {M} \)-closed view) | Maximum a posteriori model, median probability model |
Yet an alternative approach is to construct a full encompassing reference model which is believed to best represent the uncertainties regarding the modelling task and choose a simpler submodel that gives nearly similar predictions as the reference model. This approach was pioneered by Lindley (1968) who considered input variable selection for linear regression and used the model with all the variables as the reference model. The idea was extended by Brown et al. (1999, 2002). A related method is due to San Martini and Spezzaferri (1984) who used the Bayesian model averaging (BMA) solution as the reference model and Kullback–Leibler divergence for measuring the difference between the predictions of the reference model and the candidate model instead of the squared error loss used by Lindley. Another related method is the reference model projection by Goutis and Robert (1998) and Dupuis and Robert (2003) in which the information contained in the posterior of the reference model is projected onto the candidate models. The variations of this method include heuristic Lasso-type searching (Nott and Leng 2010) and approximative projection with different cost functions (Tran et al. 2012; Hahn and Carvalho 2015).
Although LOO-CV and WAIC can be used to obtain a nearly unbiased estimate of the predictive ability of a given model, both of these estimates contain a stochastic error term whose variance can be substantial when the dataset is not very large. This variance in the estimate may lead to overfitting in the selection process causing nonoptimal model selection and inducing bias in the performance estimate for the selected model (e.g., Ambroise and McLachlan 2002; Reunanen 2003; Cawley and Talbot 2010). The overfitting in the selection may be negligible if only a few models are being compared but, as we will demonstrate, may become a problem for a larger number of candidate models, such as in variable selection.
Our results show that the optimization of CV or WAIC may yield highly varying results and lead to selecting a model with predictive performance far from optimal. From the predictive point of view, best results are generally obtained by accounting for the model uncertainty and forming the full BMA solution over the candidate models, and one should not expect to do better by selection. Our results agree with what is known about the good performance of the BMA (Hoeting et al. 1999; Raftery and Zheng 2003). If the full model is too complex or the cost for observing all the variables is too high, the model can be simplified most robustly by the projection method which is considerably less vulnerable to the overfitting in the selection. The advantage of the projection approach comes from taking into account the model uncertainty in forming the full encompassing model and reduced variance in the performance evaluation of the candidate models. Overall, the projection framework outperforms also the selection of the most probable inputs or the MAP model, while both of these typically perform better than optimization based on CV or WAIC.
Despite its advantages, the projection approach has suffered from a difficulty in deciding how many variables should be selected in order to get predictions close to the reference model (Peltola et al. 2014; Robert 2014). Our study shows that the model selection can highly benefit from using cross-validation outside the variable searching process both for guiding the model size selection and assessing the predictive performance of the finally selected model. Using cross-validation for choosing only the model size rather than the variable combination introduces substantially less overfitting due to greatly reduced number of model comparisons (see Sect. 4.4 for discussion).
The paper is organized as follows. In Sect. 2 we shortly go through the model selection methods discussed in this paper. This part of the paper somewhat overlaps with the previous studies (especially with Vehtari and Ojanen 2012), but is included for maintaining easier readability as a standalone paper. Section 3 discusses and illustrates the overfitting in model selection and the consequent selection induced bias in the performance evaluation of the chosen model. Section 4 is devoted to the numerical experiments and forms the core of the paper. The discussion in Sect. 5 concludes the paper.
2 Approaches for Bayesian model selection
This section discusses shortly the proposed methods for Bayesian model selection relevant for this paper. We do not attempt anything close to a comprehensive review but summarize the methods under comparison. See the review by Vehtari and Ojanen (2012) for a more complete discussion.
The section is organized as follows. Section 2.1 discusses how the predictive ability of a model is defined in terms of an expected utility and Sects. 2.2–2.5 shortly review the methods. Following Bernardo and Smith (1994) and Vehtari and Ojanen (2012) we have categorized the methods into \( \mathscr {M} \)-closed, \( \mathscr {M} \)-completed, and \( \mathscr {M} \)-open views, see Table 1. \( \mathscr {M} \)-closed means assuming that one of the candidate models is the true data generating model. Under this assumption, one can set prior probabilities for each model and form the Bayesian model averaging solution (see Sect. 2.5). The \( \mathscr {M} \)-completed view abandons the idea of a true model, but still forms a reference model which is believed to be the best available description of the future observations. In the \( \mathscr {M} \)-open view one does not assume one of the models being true and also rejects the idea of constructing the reference model.
Throughout Sect. 2, the notation assumes a model M which predicts an output y given an input variable x. The same notation is used both for scalars and vectors. We denote the future observations by \( \tilde{y} \) and the model parameters by \( \theta \). To make the notation simpler, we denote the training data as \(D= \{(x_i,y_i)\}_{i=1}^n\).
2.1 Predictive ability as an expected utility
2.2 Generalization utility estimation
2.2.1 Cross-validation
2.2.2 Information criteria
2.3 Mixed self and posterior predictive criteria
There exists a few criteria that are not unbiased estimates of the true generalization utility (2) but have been proposed for model selection. These criteria do not fit to the \( \mathscr {M} \)-open view since the candidate models are partially assessed based on their own predictive properties and therefore these criteria resemble \( \mathscr {M} \)-closed/\( \mathscr {M} \)-completed view (for a detailed discussion, see Vehtari and Ojanen 2012).
2.4 Reference model approach
Section 2.2 reviewed methods for utility estimation that are based on sample reuse without any assumptions on the true model (\( \mathscr {M} \)-open view). An alternative way is to construct a full encompassing reference model \(M_*\), which is believed to best describe our knowledge about the future observations, and perform the utility estimation almost as if it was the true data generating distribution (\( \mathscr {M} \)-completed view). We refer to this as the reference model approach. There are basically two somewhat different but related approaches that fit into this category, namely the reference and projection predictive methods, which will be discussed separately.
2.4.1 Reference predictive method
The reference predictive approach is inherently a less straightforward approach to model selection than the methods presented in Sect. 2.2, because it requires the construction of the reference model and it is not obvious how it should be done. San Martini and Spezzaferri (1984) proposed using the Bayesian model average (20) as the reference model (see Sect. 2.5). In the variable selection context, the model averaging corresponds to a spike-and-slab type prior (Mitchell and Beauchamp 1988) which is often considered as the “gold standard” and has been extensively used for linear models (see, e.g., George and McCulloch 1993, 1997; Raftery et al. 1997; Brown et al. 2002; Lee et al. 2003; O’Hara and Sillanpää 2009; Narisetty and He 2014) and extended and applied to regression for over one million variables (Peltola et al. 2012a, b). However, we emphasize that any other model or prior could be used as long as we believe it reflects our best knowledge of the problem and allows convenient computation. For instance the Horseshoe prior (Carvalho et al. 2009, 2010) has been shown to have desirable properties empirically and theoretically assuming a properly chosen shrinkage factor (Datta and Ghosh 2013; Pas et al. 2014).
2.4.2 Projection predictive method
A related but somewhat different alternative to the reference predictive method (previous section) is the projection approach. The idea is to project the information in the posterior of the reference model \(M_*\) onto the candidate models M so that the predictive distribution of the candidate model remains as close to the reference model as possible. Thus the key aspect is that the candidate model parameters are determined by the fit of the reference model, not by the data. Therefore also the prior needs to be specified only for the reference model. The idea behind the projection is quite generic and Vehtari and Ojanen (2012) discuss the general framework in more detail.
The obvious drawback in this approach are the increased computations (as the selection and reference model fitting is repeated K times), but in Sect. 4.3 we demonstrate that this approach may be very useful when choosing the final model size.
2.5 Model space approach
From a model selection point of view, one may choose the model maximizing (19) ending up with the maximum a posteriori (MAP) model. Assuming the true data generating model belongs to the set of the candidate models, MAP model can be shown to be the optimal choice under the zero-one utility function (utility being one if the true model is found, and zero otherwise). If the models are given equal prior probabilities, \(p(M)\propto 1\), finding the MAP model reduces to maximizing the marginal likelihood, also referred to as the type-II maximum likelihood.
3 Overfitting and selection induced bias
As discussed in Sect. 2.1, the performance of a model is usually defined in terms of the expected utility (2). Many of the proposed selection criteria reviewed in Sects. 2.2–2.5 can be thought of as estimates of this quantity even if not designed directly for this purpose.
The right plot shows a biased utility estimation method that either under or overestimates the ability of most of the models. However, due to smaller variance, the probability of choosing a model with better true performance is significantly increased (the maxima of the estimates focus closer to the true optimum). This example demonstrates that even though the unbiasedness is beneficial for the performance evaluation of a particular model, it is not necessarily important for model selection. For the selection, it is more important to be able to rank the competing models in an approximately correct order with a low variability.
The overfitting in model selection and the selection induced bias are important concepts that have received relatively little attention compared to the vast literature on model selection in general. However, the topic has been discussed for example by Rencher and Pun (1980), Ambroise and McLachlan (2002), Reunanen (2003), Varma and Simon (2006), and Cawley and Talbot (2010). These authors discuss mainly the model selection using cross-validation, but the ideas apply also to other utility estimation methods. As discussed in Sect. 2.2, cross-validation gives a nearly unbiased estimate of the generalization performance of any given model, but the selection process may overfit when the variance in the utility estimates is high (as depicted in the left plot of Fig. 1). This will be demonstrated empirically in Sect. 4. The variance in the utility estimate is different for different estimation methods but may generally be considerable for small datasets. The amount of overfitting in selection increases with the number of models being compared, and may become a problem for example in variable selection where the number of candidate models grows quickly with the number of variables.
4 Numerical experiments
This section compares the methods presented in Sect. 2 in practical variable selection problems. Section 4.1 discusses the used models and Sects. 4.2 and 4.3 show illustrative examples using simulated and real world data, respectively. The reader is encouraged to go through the simulated examples first as they illustrate most of the important concepts with more detailed discussion. In Sect. 4.4 we then discuss the use of cross-validation for guiding the model size selection and for performance evaluation of the finally selected model. Finally, Sect. 4.5 provides a short note on the computational aspects.
4.1 Models
4.2 Simulated data
Compared model selection methods for the experiments
Abbreviation | Method |
---|---|
CV-10 | Tenfold cross-validation optimization (3) |
WAIC | WAIC optimization (4) |
DIC | DIC optimization (6) |
L2 | \(L^2\)-criterion optimization (8) |
L2-CV | \(L^2_ \text {cv} \)-criterion optimization (9) |
L2-k | \(L^2_k\)-criterion optimization with \(k=1\) (10) |
MAP | Maximum a posteriori model |
MPP/Median | Sort the variables according to their marginal posterior probabilities (MPP), choose all with probability 0.5 or more (Median) (22) |
BMA-ref | Posterior predictive discrepancy minimization from BMA (12), choose smallest model having 95 % explanatory power (17) |
BMA-proj | Projection of BMA to submodels (15), choose smallest model having 95 % explanatory power (17) |
The experiments were carried out by varying the training set size \(n = 100,200,400\) and the correlation coefficient \(\rho =0,0.5,0.9\). We used the regression model (24) with prior parameters \(\alpha _\tau = \beta _\tau = \alpha _\sigma = \beta _\sigma = 0.5\). The posterior inference did not seem to be sensitive to these choices. As the reference model \(M_*\), we used the BMA (20) over the different input combinations with prior \(a=1\), \(b = 10\) for the number of inputs (26). For each combination of \((n,\rho )\), we performed the variable selection with each method listed in Table 2 for 50 different data realizations. As a search heuristic, we used the standard forward search, also known as the stepwise regression. In other words, starting from the empty model, at each step we select the variable that increases the utility estimate (like CV, WAIC, DIC, etc.) the most. The Median and MAP model where estimated from the RJMCMC samples that were drawn to form the BMA (as discussed in Sect. 4.1).
Figure 2 shows the average number of selected variables and the test utilities of the selected models in each data setting with respect to the BMA. First, a striking observation is that none of the methods is able to find a model with better predictive performance than the BMA. From the predictive point of view, model averaging yields generally the best results on expectation, and one should not expect to do better by selection. This result is in perfect accordance with what is known about the good performance of the BMA (Hoeting et al. 1999; Raftery and Zheng 2003). Thus, the primary motivation for selection should be the simplification of the model without substantially compromising the predictive accuracy, rather than trying to improve over the predictions obtained by taking into account the model uncertainty.
To get more insight to the problem, let us examine more closely how the predictive performance of the submodels change when variables are selected. Figure 3 shows the CV and test utilities after sorting the variables, as a function of the number of variables selected along the forward search path when \(\rho =0.5\). The CV-utility (tenfold) is computed within the data used for selection (n points), and the test utility on independent data (note that computing the CV-curve for BMA-proj requires cross-validating the BMA and performing the projection for the submodels separately for each fold). The search paths for CV-10 (top row) demonstrate the overfitting in model selection; starting from the empty model and adding variables one at a time one finds models that have high CV utility but much worse test utility. In other words, the performance of the models at the search path is dramatically overestimated and the gap between the two curves denotes the selection induced bias. Yet in other words, after selection (sorting the variables) the CV utility is an optimistic estimate for the selected models. Note, however, that for the empty model and the model with all the variables the CV utility and the test utility are on average almost the same because these models do not involve any selection. The overfitting in the selection process decreases when the size of the training set grows because the variance of the error term in decomposition (23) becomes smaller, but the effect is still visible for \(n=400\). The behaviour is very similar also for WAIC, DIC, L2, L2-CV and L2-k (the results for the last three are left out to save space).
Ordering the variables according to their marginal posterior probabilities (MPP) and selecting the Median model works much better than CV-10, leading to selection of smaller models with good predictive performance. However, even better results are obtained by using the reference model approach, especially the projection (BMA-proj). The results clearly show that the projection approach is much less vulnerable to the overfitting than CV, WAIC and DIC, even though the CV utility is still a biased estimate of the true predictive ability for the chosen models. Even for the smallest dataset size, the projection is able to find models with predictive ability very close to the BMA with about 10–15 variables on average. Moreover, the projection has the inherent advantage over the other methods performing reasonably well (like MPP/Median) that when more variables are added, the submodels get ever closer to the reference model (BMA), thus avoiding the dip in the predictive accuracy apparent with the other methods around 10 variables. This is simply because the submodels are constructed to be similar than the model averaging solution which yields the best results (this point will be further discussed below, see discussion related to Fig. 6).
Figure 4 shows the variability in the performance of the selected models for the same selection methods as in Fig. 3. The grey lines denote the test utilities for the selected models as a function of number of selected variables for different data realizations and the black line denotes the average (same as in Fig. 3). For small training set sizes the variability in the predictive performance of the selected submodels is very high for CV-10, WAIC, DIC, and MPP. The reference model approach, especially the projection, reduces the variability substantially finding sparse submodels with predictive performance close to the BMA in all the data realizations. This is another property that makes the projection approach appealing. Figure 4 further emphasizes how difficult it is to improve over the BMA in predictive terms; most of the time the model averaging yields better predictions than any of the found submodels. Moreover, even when better submodels are found (the cases where the grey lines exceed the zero level), the difference in the predictive performance is relatively small.
Although our main focus is on the predictive ability of the chosen models, we also studied how the different methods are able to choose the truly relevant variables over the irrelevant ones. Here by “relevant” we mean those 15 variables that were used to generate the output y even though it might not be completely clear how the relevance should be defined when there are correlations between the variables (for example, should we select one or both of two correlating variables which both correlate with the output). Figure 5 shows the proportion of relevant variables chosen (vertical axis) versus proportion of irrelevant variables chosen (horizontal axis) on average (the larger the area under the curve, the better). In this aspect, ordering the variables according to their marginal probabilities seems to work best, slightly better than the projection. The other methods seem to perform somewhat worse. Interestingly, although the projection does not necessarily order the variables any better than the marginal posterior probability order, the predictive ability of the projected submodels is on average better and varies less as Fig. 4 demonstrates.
To explain this behaviour, we did one more experiment and studied the difference of learning the submodel parameters by the projection from the BMA compared to fitting the submodels to the data. We performed this analysis both when the selection was done by the marginal probabilities or by forward search minimizing the discrepancy to the BMA, see Fig. 6. The results show that constructing the submodels by projection improves the results regardless of which of the two methods is used to sort the variables, and in this example, there seems to be little difference in the final results as long as projection is used to construct the submodels.
This effect can be explained by transmission of information from the reference model. Recall that the BMA corresponds to setting the sparsifying spike-and-slab prior for the full model. Because the prior information is transmitted also to the submodels in the projection, it is natural that the submodels benefit from this [compared to the Gaussian prior in (24)] especially when some irrelevant variables have been included. Furthermore, the noise level is also learned from the full model [see Eq. (31)], which reduces overfitting of the submodels as the full model best represents the uncertainties related to the problem and best captures the correct the noise level. More detailed analysis of these effects would be useful, but we do not focus on it more in this paper and leave it to the future research.
Summary of the real world datasets and used priors
Dataset | Type | p | n | Prior parameters |
---|---|---|---|---|
Crime | Regression | 102 | 1992 | \({\alpha _\tau = \beta _\tau = 0.5}\), \({\alpha _\sigma = \beta _\sigma = 0.5}\), \({a = b = 2}\) |
Ionosphere | Classification | 33 | 351 | \({\alpha _\tau = \beta _\tau = 0.5},\, {a = b = 2}\) |
Sonar | Classification | 60 | 208 | \({\alpha _\tau = \beta _\tau = 0.5},\, {a = b = 2}\) |
Ovarian cancer | Classification | 1536 | 54 | \({\alpha _\tau = \beta _\tau = 2},\, {a = 1,\, b = 1200}\) |
Colon cancer | Classification | 2000 | 62 | \({\alpha _\tau = \beta _\tau = 2},\, {a = 1,\, b = 2000}\) |
4.3 Real world datasets
We also studied the performance of the different methods on several real world datasets. Five publicly available^{1} datasets were used and they are summarized in Table 3. One of the datasets deals with regression and the rest with binary classification. As a preprocessing, we normalized all the input variables to have zero mean and unit variance. For the Crime dataset we also log-normalized the original non-negative target variable (crimes per population) to get a real-valued and more Gaussian output. From this dataset we also removed some input variables and observations with missing values (the given p and n in Table 3 are after removing the missing values). We will not discuss the datasets in detail but refer to the sources for more information.
We then performed the variable selection using the methods in Table 2 except the ones based on the squared error (L2, L2-CV, L2-k) were not used for the classification problems. For Ovarian and Colon datasets, due to large number of variables, we also replaced the tenfold-CV by the importance sampling LOO-CV (IS-LOO-CV) to reduce the computation time. For these two datasets we also performed the forward searching only up to 10 variables.
Figure 8 summarizes the results. The left column shows the average number of variables selected and the right column the estimated out-of-sample utilities for the chosen models (out-of-sample utilities are estimated using hold-out samples not used for selection as explained above). The results are qualitatively very similar to those obtained for the simulated experiments (Sect. 4.2). Again we conclude that the BMA solution gives better predictions than any of the selection methods when measured on independent data. Moreover, the results demonstrate again that model selection using CV, WAIC, DIC, L2, L2-CV, or L2-k is liable to overfitting especially when the dataset is small compared to the number of variables. Overall, MAP and Median models tend to perform better but show non-desirable performance on some of the datasets. Especially the Median model performs badly for Ovarian and Colon datasets where almost all the variables have marginal posterior probability less than 0.5 (depending on the split into training and validation sets). The projection (BMA-proj) shows the most robust performance choosing models with predictive ability close to the BMA for all the datasets.
Figure 9 shows the CV (red) and out-of-sample (black) utilities for the chosen models as a function of number of chosen variables for the classification problems. CV-utilities (tenfold) are computed after sorting the variables within the same data used for selection, and the out-of-sample utilities are estimated on hold-out samples not used for selection as explained earlier. This figure is analogous to Fig. 3 showing the magnitude of the selection induced bias (the difference between the red and black lines). Especially for the last three datasets (Sonar, Ovarian, Colon) the selection induced bias is considerable for all the methods, which emphasizes the importance of validating the variable searching process in order to avoid bias in performance evaluation for the found models. Overall, the projection (BMA-proj) appears to find models with best out-of-sample accuracy for a given model complexity, albeit for Ovarian dataset choosing about five most probable inputs (MPP) would perform even better. Moreover, the uncertainty in the out-of-sample performance for a given number of variables is also the smallest for the projection over all the datasets.
The same applies to the Crime dataset, see Fig. 10. For any given number of variables the projection is able to find models with predictive ability closest to the BMA and also with the least variability, the difference to the other methods being the largest when the dataset size is small. For Crime data, some additional results using hierarchical shrinkage prior for the full model are presented in Appendix 2.2.
4.4 On choosing the final model size
Figure 11 shows the final expected utility on independent data for different values of U and \(\alpha \) for the different datasets. The results show that for a large enough credible level \(\alpha =0.95\), the applied selection rule appears to be safe in the sense that the final expected utility remains always above the level of U (the dots stay above the diagonal line). This means that there is no substantial amount of selection induced bias at this stage, and the second level of validation is not necessarily needed. However, this does not always hold for smaller values of \(\alpha \) (\(\alpha =0.5\) and 0.05).
We can make these results more intuitive by looking at an example. Consider the projection search path for the Sonar dataset in Fig. 9 (bottom row, second column). Assume now that \(U = -0.01\), meaning that we are willing to lose 0.01 in the MLPD compared to the BMA, which is about 5 % of the difference between the BMA and the empty model. Since the grey bars denote the central 95 % credible intervals, setting \(\alpha =0.975\) would correspond to choosing the smallest model for which the lower limit of the grey bar falls above \(U=-0.01\) (because then the performance of the submodel is greater than this with probability 0.975). This would lead to choosing about 35 variables, and we could be quite confident that the final expected utility would be within 0.01 from the BMA. On the other hand, setting \(\alpha =0.025\) corresponds to choosing the smallest model for which the upper limit of the grey bar falls above \(U=-0.01\) (because then the performance of the submodel is greater than this with probability 0.025). This would lead to choosing only three variables, but in this case we cannot expect confidently that the final expected utility would be within 0.01 from the BMA, and in fact it would be lower than this with high probability.
Following the example above, one can get an idea of the effect of \(\alpha \) and U to the final model size and the final expected utility. Generally speaking, U determines how much we are willing to compromise the predictive accuracy and \(\alpha \) determines how condident we want to be about the final utility estimate. It is application specific whether it is more important to have high predictive accuracy or to reduce the number of variables at the cost of losing predictive accuracy, but a reasonable recommendation might be to use \(\alpha \ge 0.95\) and U to be 5 % of the difference between the reference model and the empty model. Based on Fig. 11, this combination would appear to give predictive accuracy close to the BMA, and based on Fig. 9 yield quite effective reduction in the number of variables (choosing about 20 variables for Ionosphere, 35 for Sonar, and 5–10 variables for Ovarian and Colon by visual inspection).
As a final remark, one might wonder why the cross-validation works well if used only to decide the model size but poorly if used directly to optimize the variable combination (as depicted e.g. in Fig. 3)? As discussed in Sect. 3, the amount of overfitting in the selection and the consequent selection bias depends on the variance of the utility estimate for a given model (over the data realizations) and the number of candidate models being compared. For the latter reason, the overfitting is considerably smaller when cross-validation is used only to decide the model size by comparing only \(p+1\) models (given the variable ordering), in contrast to selecting variable combination in the forward search phase among \(O(p^2)\) models. Moreover, the utilities of two consecutive models at the search path are likely to be highly correlated, which further reduces the freedom of selection and therefore the overfitting at this point. It is for this same reason why cross-validation may yield reasonable results when used to choose a single hyperparameter among a small set of values, in which case essentially only a few different models are being compared.
Based on the results presented in this section, despite the increased computational effort, we believe the use of cross-validation on top of the variable searching is highly advisable both for choosing the final model size and giving a nearly unbiased estimate of the out-of-sample performance for the selected model. We emphasize the importance of this regardless of the method used for searching the variables, but generally we recommend using the projection given its overall superior performance in our experiments.
4.5 Computational considerations
We conclude with a remark on the computational aspects. The results in Sects. 4.2 and 4.3 emphasized that from the predictive point of view, the model averaging over the different submodels yields often better results than selection of any single model. Forming the model averaging solution over the variable combinations may be computationally challenging, but is quite feasible up to problems with a few thousand variables or less with, for instance, a straightforward implementation of the RJMCMC algorithm with simple proposal distributions. Scalability up to a million variables can be obtained with more sophisticated and efficient proposals (Peltola et al. 2012b). The computations could also be sped up by using a hierarchical shrinkage prior such as the horseshoe (see Appendix 2.1) for the regression weights in the full model, instead of the spike-and-slab (which is equivalent to the model averaging over the variable combinations).
After forming the full reference model (either the BMA or using some alternative prior), the subsequent computations needed for the actual selection are typically less burdensome. Computing the MAP and the Median models from the MCMC samples from the model space is easy, but these cannot be computed with alternative priors. Constructing the submodels by projection and performing a forward search through the variable combinations is somewhat more laborous, but takes still usually considerably less time than forming the reference model. The projection approach can also be used with any prior for the full model, as the posterior samples are all that is needed (see Appendix 2).
The methods that do not rely on the construction of the full reference model (CV, WAIC, DIC, and the \(L^2\)-variants) are typically computationally easier as they avoid the burden of forming the reference model in the first place. However, if one has to go through a lot of models in the search, the procedure is fast only if the fitting of the submodels is fast. This is the case for instance for the Gaussian linear model for which the posterior computations are obtained analytically, but in other problems like in classification, sampling the parameters separately for each submodel may be very expensive, and one may be forced to use faster approximations (such as the Laplace method or expectation propagation). On the other hand, it must be kept in mind that the possible computational savings from avoiding the construction of the full model may come at the risk of overfitting and reduced predictive accuracy, as our results show.
5 Conclusions
In this paper we have shortly reviewed many of the proposed methods for Bayesian predictive model selection and illustrated their use and performance in practical variable selection problems for regression and binary classification, where the goal is to select a minimal subset of input variables with a good predictive accuracy. The experiments have been carried out using both simulated and several real world datasets.
The numerical experiments show that the overfitting in the selection may be a potential problem and hinder the model selection considerably. This is the case especially when the dataset is small (high variance in the utility estimates) and the number of models under comparison large (large number of variables). Especially vulnerable methods for this type of overfitting are CV, WAIC, DIC and other methods that rely on data reuse and have therefore relatively high variance in the utility estimates. From the predictive point of view, better results are generally obtained by accounting for the model uncertainty and forming the full encompassing (reference) model with all the variables and best possible prior information on the sparsity level. Our results showed that Bayesian model averaging (BMA) over the candidate models yields often the best results on expectation, and one should not expect to do better by selection. This agrees with what is known about the good performance of the BMA (Hoeting et al. 1999; Raftery and Zheng 2003).
If the full model is too complex or the cost for observing all the variables is too high, the model can be simplified most robustly by the projection method which is considerably less vulnerable to the overfitting in the selection. The advantage of the projection approach comes from taking into account the uncertainty in forming the full encompassing model and then finding a simpler model which gives similar answers as the full model. Overall, the projection framework outperforms also the selection of the most probable variables or variable combination (Median and MAP models) being able to best retain the predictive ability of the full model while effectively reducing the model complexity. The results also demonstrated that the projection does not only outperform the other methods on average but the variability over the different data realizations is also considerably smaller compared to the other methods. In addition, the numerical experiments showed that constructing the submodels by the projection from the full model may improve the predictive accuracy even when some other strategy, such as marginal probabilities, are used to rank the variables.
Despite its advantages, the projection method has the inherent challenge of forming the reference model in the first place. There is no automated way of coming up with a good reference model which emphasizes the model criticism. However, as already stressed, incorporating the best possible prior information into the full encompassing model is formally the correct Bayesian way of dealing with the model uncertainty and often seems to also provide the best predictions in practice. In this study we used the model averaging over the variable combinations as the reference model, but similar results are obtained also with the hierarchical shrinkage prior (see Appendix 2).
Another issue is that, even though the projection method seems the most robust way of searching for good submodels, the estimated discrepancy between the reference model and a submodel is in general an unreliable indicator of the predictive performance of the submodel. In variable selection, this property makes it problematic to decide how many variables should be selected to obtain predictive performance close to the reference model, even though the minimization of the discrepancy from the reference model typically finds a good search path through the model space.
However, the results show that this problem can be solved by using cross-validation outside the searching process, as this allows studying the tradeoff between the number of included variables and the predictive performance, which we believe is highly informative. Moreover, we demonstrated that selecting the number of variables this way does not produce considerable overfitting or selection induced bias in the utility estimate for the finally selected model, because the selection is conditioned on a greatly reduced number of models (see Sect. 4.4). While this still leaves the user the responsibility of deciding the final model size, we emphasize that this decision depends on the application and the costs of the inputs. Without any costs for the variables, we would simply recommend using them all and carrying out the full Bayesian inference on the model.
The first three datasets are available at the UCI Machine Learning repository https://archive.ics.uci.edu/ml/index.html.
Ovarian cancer dataset can be found at http://www.dcs.gla.ac.uk/~srogers/lpd/lpd.html and the Colon cancer data at http://genomics-pubs.princeton.edu/oncology/affydata/index.html.
Acknowledgments
We thank Arno Solin, Tomi Peltola and the anonymous reviewers for their valuable comments to improve the manuscript. We also acknowledge the computational resources provided by Aalto Science-IT project.
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.