An empirical comparison of supervised learning algorithms

R. CaruanaAlexandru Niculescu-Mizil

article2006ICML2,904 citations

Presents a comprehensive benchmark of ten supervised learning algorithms across nine performance metrics, revealing how post-training probability calibration enables boosted decision trees and random forests to outperform methods like neural networks and support vector machines.

Listen

This large-scale empirical study compares ten supervised learning algorithmssupport vector machines, neural networks, logistic regression, naive Bayes, memory-based learning, random forests, decision trees, bagged trees, boosted trees, and boosted stumpsacross nine performance metrics on eight binary classification problems drawn from standard repositories and real-world sources. The work addresses the scarcity of comprehensive benchmarks that include newer ensemble methods and metrics beyond simple accuracy, such as ranking measures used in information retrieval and probability calibration needed for risk assessment or decision thresholds. Such comparisons matter because algorithms optimize different internal criteria, and performance on one metric often fails to predict results on others.

The study trained roughly 2000 model variants per problem, selected the best parameter settings via cross-validation on held-out validation data, and evaluated final performance on large independent test sets. Results were normalized to a common scale to allow averaging across metrics and datasets, and predictions from models not designed to output probabilities were adjusted using Platt scaling and isotonic regression before re-evaluation.

Before calibration, neural networks, bagged trees, and random forests delivered the strongest average performance across all metrics. After calibration, boosted trees rose to the top overall, followed closely by random forests, while neural networks remained competitive without adjustment. Calibration produced large gains for boosted trees, support vector machines, and naive Bayes on squared error, cross-entropy, and calibration error, but offered little or no benefit to neural networks or bagged trees. Boosted stumps consistently underperformed boosted full trees, and logistic regression and naive Bayes ranked lowest on average.

These patterns indicate that modern ensemble methods can achieve excellent results on both ranking and probability metrics once their outputs are properly scaled, whereas neural networks provide robust performance with fewer post-processing steps. For applications that require well-calibrated probabilities, such as medical risk scoring or cost-sensitive decisions, calibration should be treated as a standard final step rather than an optional refinement. When ranking or threshold-based decisions dominate, uncalibrated boosted trees and random forests already perform strongly.

No algorithm dominated every problem or metric, confirming that selection must still be validated on the target data. Practitioners should therefore run modest calibration experiments on their own validation sets and compare a small number of top candidates rather than defaulting to any single method. The findings rest on eight datasets and five trials per problem; broader testing on additional domains, larger training sets, and newer algorithms would increase confidence in the rankings.

  • Paper: A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection, Ron Kohavi (1995). Reading this foundational study on cross-validation and bootstrap accuracy estimation improves understanding of the evaluation protocols used to compare supervised algorithms.
  • Paper: Statistical Comparisons of Classifiers over Multiple Data Sets, Janez Demšar (2006). Understanding these proper statistical tests for comparing classifiers across multiple datasets provides essential background for interpreting the performance differences reported in the source paper.
  • Paper: Random Forests, Leo Breiman (2001). Reviewing Breiman's introduction to random forests is essential for grasping the mechanics of one of the top-performing ensemble methods evaluated in the source comparison.
  • Paper: Greedy function approximation: A gradient boosting machine, Jerome H. Friedman (2001). Studying Friedman's gradient boosting machine provides the necessary algorithmic foundation for the boosted tree and stump methods benchmarked in the source study.
  • Paper: Experiments with a New Boosting Algorithm, Yoav Freund et al. (1996). Reading the original work on AdaBoost clarifies the boosting algorithms whose predictive performance and calibration are examined across multiple criteria in the source evaluation.
  • Paper: Bagging Predictors, L. Breiman (1996). Understanding bagging predictors provides critical context for the bagged tree methods whose accuracy and probability calibration are compared in the source paper.
  • Paper: Support-vector networks, Corinna Cortes et al. (1995). Examining the foundational support-vector machine paper helps clarify the optimization and probability calibration challenges that motivate the scaling techniques discussed in the source study.
  • Paper: Do we need hundreds of classifiers to solve real world classification problems?, Manuel Fernández Delgado et al. (2014). This later study naturally extends the source's empirical comparison by scaling up to 179 classifiers and 121 datasets to provide definitive guidance on top-performing algorithms.
  • Paper: On Calibration of Modern Neural Networks, Chuan Guo et al. (2017). Building directly on the source's findings regarding model calibration, this paper investigates the miscalibration of modern neural networks and evaluates post-hoc scaling techniques.
  • Paper: Extremely randomized trees, Pierre Geurts et al. (2006). This paper continues the exploration of tree-based ensembles evaluated in the source by introducing and benchmarking extremely randomized trees against standard bagging and random forests.
  • Paper: XGBoost: A Scalable Tree Boosting System, Tianqi Chen et al. (2016). Extending the gradient boosting methods compared in the source, this work introduces a highly scalable tree boosting system optimized for speed and large datasets.
  • Paper: CatBoost: unbiased boosting with categorical features, Liudmila Prokhorenkova et al. (2018). This paper advances gradient boosting technology beyond the algorithms evaluated in the source by introducing ordered boosting to eliminate target leakage with categorical features.
  • Paper: LightGBM: A Highly Efficient Gradient Boosting Decision Tree, Guolin Ke et al. (2017). Building on the gradient boosting algorithms examined in the source, this chapter presents LightGBM to dramatically accelerate training times on massive datasets.
Cover for An empirical comparison of supervised learning algorithms

Abstract

We present results from a large-scale empirical comparison between ten learning methods: SVMs, neural nets, logistic regression, naive bayes, memory-based learning, random forests, decision trees, bagged trees, boosted trees, and boosted stumps. We evaluate the methods on binary classification problems using nine performance criteria: accuracy, squared error, cross-entropy, ROC Area, F-score, precision/recall break-even point, average precision, lift, and calibration. Because some models (e.g. SVMs and boosted trees) do not predict well-calibrated probabilities, we compare the performance of the algorithms both before and after calibrating their predictions with Platt Scaling and Isotonic Regression. Before scaling, the models with the best overall performance are neural nets, bagged trees, and random forests. After scaling, the best models are boosted trees, random forests, and unscaled neural nets.

Table of Contents

  • 1 Introduction
  • 2 Methodology
  • 2.1 Learning Algorithms
  • 2.2 Performance Metrics
  • 2.3 Comparing Across Performance Metrics
  • 2.4 Calibration Methods
  • 2.5 Data Sets
  • 3 Performances by Metric
  • 4 Performances by Problem
  • 5 Relative Computational Cost
  • 6 Related Work
  • 7 Conclusions
  • References

Knowls

  1. Knowl 1 — Normalized Empirical Performance of Supervised Learning Algorithms Across Nine Evaluation Metrics

    data/table

    Across eight binary classification tasks evaluated on nine performance criteria, supervised learning algorithms exhibit substantial performance differences depending on their family and whether post-training probability calibration (Platt Scaling [P] or Isotonic Regression [I]) is applied. The performance scores below are normalized such that 0.00.0 represents baseline performance (predicting the positive class prevalence pp for every case) and 1.01.0 represents the best observed performance across all models on that problem and metric (serving as a proxy for Bayes optimal performance). The calibration metric (CAL) is centered at 0.00.0 (mean observed calibration error) and scaled to 1.01.0 for the best observed calibration.

    Model Cal ACC FSC LFT ROC APR BEP RMS MXE CAL MEAN OPT
    BST-DT P .860* .854 .956* .977 .958 .952 .929 .932 .808* .914 .941
    BST-DT I .812* .921* .948 .965 .937 .942 .896 .859 .798* .897 .939*
    RF P .866 .871 .958 .977* .957* .948* .892 .898 .702 .897 .906
    RF I .847* .915 .948 .966 .937 .940* .881 .828 .791 .895 .910
    ANN - .817* .875 .947* .963 .926 .929 .872 .878 .826 .892 .932
    SVM P .823 .851 .928 .961 .931 .929 .882 .880 .769 .884 .909
    BAG-DT I .820 .886 .947 .961 .930 .925 .859 .822 .771 .880 .905
    BAG-DT - .836 .849 .953 .972 .950* .928 .875 .901 .637 .878 .899
    RF - .844* .845 .958 .977* .957* .948* .882 .899 .567 .875 .882
    BAG-DT P .822 .843 .953 .972 .950* .929 .863 .874 .666 .875 .893
    ANN I .816* .934 .943 .949 .906 .920 .835 .782 .767 .873 .927
    SVM I .806 .914 .924 .946 .910 .928 .859 .799 .759* .872 .912
    ANN P .833 .863 .947* .963 .926 .929 .842 .839 .651 .866 .903
    KNN P .759 .820 .914 .937 .893 .898 .786 .805 .706 .835 .869
    KNN I .753 .866 .905 .926 .873 .893 .782 .756 .746 .834 .883
    KNN - .759 .839 .914 .937 .893 .898 .783 .769 .684 .831 .858
    BST-STMP P .698 .760 .898 .926 .871 .854 .740 .783 .678 .801 .834
    BST-STMP I .677 .821 .892 .916 .850 .852 .708 .674 .679 .785 .832
    BST-DT - .861* .885 .956* .977 .958 .952 .596 .598 .045 .758 .795
    DT I .631 .787 .848 .864 .778 .807 .617 .627 .645 .734 .814
    DT P .611 .771 .856 .871 .789 .808 .586 .625 .688 .734 .794
    DT - .612 .789 .856 .871 .789 .808 .583 .638 .512 .717 .782
    LOGREG - .602 .623 .829 .849 .732 .714 .614 .620 .678 .696 .704
    LOGREG I .590 .640 .827 .848 .721 .726 .607 .594 .636 .688 .699
    NB I .537 .616 .786 .830 .721 .731 .582 .576 .635 .668 .682
    BST-STMP - .701 .782 .898 .926 .871 .854 .355 .339 .123 .650 .676
    SVM - .810 .891 .928 .961 .931 .929 .484 .447 -.546 .648 .712
    LOGREG P .597 .606 .832 .855 .733 .717 .584 .591 .302 .646 .653
    NB P .536 .615 .786 .833 .733 .730 .539 .565 .161 .611 .625
    NB - .414 .637 .746 .767 .698 .689 .271 -.980 -.918 .258 .293

    In the table, learning algorithms are denoted as: Boosted Decision Trees (BST-DT), Random Forests (RF), Artificial Neural Networks (ANN), Support Vector Machines (SVM), Bagged Decision Trees (BAG-DT), KK-Nearest Neighbors (KNN), Boosted Decision Stumps (BST-STMP), Decision Trees (DT), Logistic Regression (LOGREG), and Naive Bayes (NB). Calibration schemes in column Cal are uncalibrated (-), Platt scaling (P), and Isotonic Regression (I). Metrics include Accuracy (ACC), F-score (FSC), Lift (LFT), Area under the ROC Curve (ROC), Average Precision (APR), Precision/Recall Break-Even Point (BEP), Root Mean Squared Error (RMS), Mean Cross-Entropy (MXE), and Calibration Error (CAL). MEAN denotes the cross-problem, cross-metric average under 1k-sample validation selection, and OPT denotes theoretical optimal model selection on the test set. Values marked with * are statistically indistinguishable from the top-performing model on that metric at p=0.05p = 0.05 (paired tt-test).

    Key takeaways:

    1. Calibrated boosted decision trees (BST-DT + Platt) achieve the highest overall score (0.914 MEAN), dominating across ordering, threshold, and probability metrics.
    2. Without calibration, neural networks achieve the highest overall score (0.892 MEAN), closely followed by uncalibrated bagged trees (0.878) and uncalibrated random forests (0.875).
    3. Logistic regression, decision trees, naive bayes, and boosted stumps consistently occupy the lowest overall tiers regardless of calibration.
  2. Knowl 2 — Impact of Probability Calibration on Classifier Families Across Performance Metric Types

    empirical result

    The impact of post-hoc probability calibration (Platt Scaling and Isotonic Regression) depends heavily on whether a model's raw output represents posterior probabilities or uncalibrated decision margins:

    1. Algorithms with Substantial Gains:

      • Boosted Decision Trees (BST-DT): Uncalibrated BST-DT achieves strong ranking metrics (ROC = 0.977, APR = 0.958) but fails on probability metrics (RMS = 0.596, MXE = 0.598, CAL = 0.045). Platt scaling increases RMS to 0.929, MXE to 0.932, and CAL to 0.808, elevating overall MEAN score from 0.758 to 0.914.
      • Support Vector Machines (SVM): Uncalibrated SVM predictions (scaled linearly to [0,1][0, 1]) have poor probability performance (RMS = 0.484, MXE = 0.447, CAL = -0.546). Platt scaling elevates RMS to 0.882, MXE to 0.880, and CAL to 0.769, raising overall MEAN from 0.648 to 0.884.
      • Naive Bayes (NB) and Boosted Stumps (BST-STMP): Both show massive gains on probability metrics under calibration (e.g., NB CAL improves from -0.918 uncalibrated to 0.635 with Isotonic Regression).
      • Random Forests (RF): Calibration produces modest but consistent improvements on probability metrics (CAL increases from 0.567 uncalibrated to 0.702 with Platt and 0.791 with Isotonic Regression).
    2. Algorithms with Negligible or Negative Gains:

      • Artificial Neural Networks (ANN): Trained with gradient descent backpropagation, ANNs produce well-calibrated uncalibrated probabilities (CAL = 0.826, RMS = 0.872, MXE = 0.878). Applying Platt scaling degrades CAL to 0.651 and Isotonic Regression degrades CAL to 0.767.
      • Bagged Decision Trees (BAG-DT), KK-Nearest Neighbors (KNN), and Logistic Regression (LOGREG): Calibration provides virtually no improvement to overall metric scores, and Platt scaling significantly hurts calibration on logistic regression (CAL drops from 0.678 to 0.302).
  3. Knowl 3 — Benchmark Dataset Performance Profiles and Algorithm Robustness

    data/table

    Across individual binary classification problems, no single learning algorithm dominates universally, consistent with the No Free Lunch theorem. The table below presents the normalized score for each algorithm variant averaged across all nine performance metrics on each of the eight test datasets:

    Model Cal COVT ADULT LTR.P1 LTR.P2 MEDIS SLAC HS MG MEAN
    BST-DT P .971 .875 .965 .982 .765 .882 .965 .908 .914
    BST-DT I .954* .871 .910* .975* .727 .903 .949 .892 .897
    RF P .919 .840 .914 .948 .816* .915* .924 .897 .897
    RF I .912 .882 .878 .944 .787 .947 .911 .899 .895
    ANN - .808 .908 .919 .901 .826* .910 .947 .921 .893
    SVM P .805 .898 .941 .957 .774 .890 .937* .869 .884
    BAG-DT I .900 .936* .847 .900 .725 .920 .877 .938* .880
    BAG-DT - .918 .948 .877 .804 .765 .934* .829 .949 .878
    RF - .895 .918 .902 .827 .794 .926* .831 .909 .875
    BAG-DT P .904 .902 .863 .899 .736 .897 .879 .917 .875
    ANN I .806 .890 .837 .890 .805* .921 .941* .890 .873
    SVM I .793 .888 .911* .957 .725* .906 .925 .868 .872
    ANN P .799 .868 .895 .853 .802* .893 .928 .890 .866
    KNN P .847 .809 .918 .924 .679 .808 .815 .882 .835
    KNN I .847 .802 .889 .931 .681 .819 .820 .879 .834
    KNN - .844 .799 .911 .925 .612 .838 .820 .896 .831
    BST-STMP P .711 .943* .775 .677 .759* .813 .822 .909 .801
    BST-STMP I .690 .945* .705 .658 .726 .825 .825 .909 .785
    BST-DT - .857 .768 .850 .850 .401 .830 .825 .688 .758
    DT I .725 .882 .731 .748 .474 .802 .657 .851 .734
    DT P .729 .880 .724 .755 .458 .804 .646 .874 .734
    DT - .682 .893 .708 .729 .467 .792 .609 .858 .717
    LOGREG - .701 .855 .203 .448 .837 .885 .720 .916 .696
    LOGREG I .679 .891 .254 .433 .800 .857 .695 .892 .688
    NB I .646 .922 .678 .534 .744 .757 .328 .736 .668
    BST-STMP - .669 .838 .334 .569 .580 .766 .642 .803 .650
    SVM - .615 .684 .495 .769 .552 .849 .555 .668 .648
    LOGREG P .667 .849 .133 .425 .725 .854 .678 .840 .646
    NB P .605 .801 .632 .457 .671 .734 .309 .679 .611
    NB - .518 .617 .462 .398 -.180 .662 -1.03 .618 .258

    Problem-specific observations:

    • On ADULT, bagged trees (0.948 uncalibrated) and boosted stumps (0.945* with Isotonic Regression) perform best. Boosted trees degrade immediately on iteration 1 and fail to recover (scoring only 0.768 uncalibrated), performing worse than single decision trees (0.893).
    • On MEDIS, logistic regression is the top-performing individual algorithm (0.837 uncalibrated).
    • Boosting full trees outperforms boosted stumps on 5 of the 8 problems, but stumps suffer catastrophic failure on problems like LTR.P1 (0.334) and LTR.P2 (0.569).
    • Bagged trees consistently outperform unbagged decision trees across all 8 datasets.
    • Neural networks exhibit high robustness: while rarely ranking #1 on any single task, they never suffer catastrophic failure on any dataset, maintaining an uncalibrated mean score of 0.893.
  4. Knowl 4 — Performance Degradation Under Validation-Based Versus Optimal Model Selection

    empirical result

    Evaluating models using a 1,000-example validation set (MEAN) incurs an average performance drop of approximately 0.030.03 in normalized score compared to optimal parameter selection evaluated directly on the final test set (OPT). The magnitude of this degradation depends on the variance and hyperparameter sensitivity of the model family:

    1. High-Variance / Highly Sensitive Models:

      • Artificial Neural Networks (ANN): Experience substantial drops when tuned via validation sets due to the high sensitivity of architecture and training epochs. For example, uncalibrated ANN achieves an OPT score of 0.9320.932 versus a MEAN score of 0.8920.892 (a 0.0400.040 drop), causing ANNs to fall multiple ranks in empirical standing.
      • Boosted Decision Trees (BST-DT): Suffer noticeable drops (e.g., uncalibrated BST-DT drops from 0.7950.795 OPT to 0.7580.758 MEAN), but when calibrated with Platt scaling, their overall performance (0.9410.941 OPT, 0.9140.914 MEAN) remains superior to all other models.
    2. Moderate-Variance Models:

      • Support Vector Machines (SVM): Uncalibrated SVM drops from 0.7120.712 OPT to 0.6480.648 MEAN (0.0640.064 drop), and Platt-calibrated SVM drops from 0.9090.909 OPT to 0.8840.884 MEAN (0.0250.025 drop).
    3. Low-Variance Models:

      • Random Forests (RF): Exhibit low variance across hyperparameter choices, resulting in minimal validation-tuning loss (Platt-calibrated RF has OPT of 0.9060.906 vs. MEAN of 0.8970.897, a drop of only 0.0090.009; uncalibrated RF has OPT of 0.8820.882 vs. MEAN of 0.8750.875).
      • Bagged Decision Trees (BAG-DT): Require minimal parameter tuning across tree variants, losing little performance under validation selection (uncalibrated BAG-DT OPT is 0.8990.899 vs. MEAN of 0.8780.878).
  5. Knowl 5 — Cross-Metric and Cross-Dataset Performance Normalization Framework

    model/method

    To allow meaningful averaging of classifier performance across disparate evaluation metrics (which differ in scale, range, and direction of optimality) and diverse datasets (which differ in baseline difficulty and Bayes optimal rates), raw scores are linearly normalized onto a common scale:

    1. Standard Normalization (Eight Metrics): For Accuracy (ACC), F-score (FSC), Lift (LFT), ROC Area (ROC), Average Precision (APR), Precision/Recall Break-Even Point (BEP), Root Mean Squared Error (RMS), and Mean Cross-Entropy (MXE), the normalized score SnormS_{\text{norm}} is defined by: Snorm=SrawSbaselineSbestSbaselineS_{\text{norm}} = \frac{S_{\text{raw}} - S_{\text{baseline}}}{S_{\text{best}} - S_{\text{baseline}}} where:

      • SrawS_{\text{raw}} is the raw metric score achieved by the model on the test set.
      • SbaselineS_{\text{baseline}} is the score achieved by a naive baseline model that predicts the test set positive prevalence p=N+/Np = N_+ / N for every instance. A model performing at baseline receives Snorm=0S_{\text{norm}} = 0, and models worse than baseline receive Snorm<0S_{\text{norm}} < 0.
      • SbestS_{\text{best}} is the best observed raw score achieved by any tested model on that specific problem and metric, serving as an empirical proxy for the Bayes optimal rate (Snorm=1.0S_{\text{norm}} = 1.0). If lower raw scores indicate better performance (such as RMS or MXE), the signs are inverted so that Snorm=1.0S_{\text{norm}} = 1.0 remains the best performance.
    2. Calibration Error Normalization (CAL): Because the baseline model predicting a constant prevalence pp is well calibrated, SbaselineS_{\text{baseline}} and SbestS_{\text{best}} are nearly identical under the standard formula. Therefore, CAL is normalized separately by setting the mean observed raw calibration error across all evaluated models to 0.00.0 and the best observed raw calibration error to 1.01.0.

  6. Knowl 6 — Sliding-Window Probability Calibration Error Metric

    algorithm

    The probability calibration metric (CAL) quantifies whether a model's predicted probability y^[0,1]\hat{y} \in [0, 1] matches the true empirical positive fraction y{0,1}y \in \{0, 1\} across continuous probability ranges using an overlapping sliding window of sorted predictions.

    Input: Test instances with true binary labels yj{0,1}y_j \in \{0, 1\} and model probability predictions y^j[0,1]\hat{y}_j \in [0, 1] for j=1,,Nj = 1, \dots, N, sliding window size W=100W = 100.
    Output: Calibration error CAL[0,1]\text{CAL} \in [0, 1].
    Sort the NN instances in ascending order of predicted probabilities such that y^(1)y^(2)y^(N)\hat{y}_{(1)} \le \hat{y}_{(2)} \le \dots \le \hat{y}_{(N)} with corresponding labels y(1),y(2),,y(N)y_{(1)}, y_{(2)}, \dots, y_{(N)}.
    Initialize total_error=0.0\text{total\_error} = 0.0
    Set K=NW+1K = N - W + 1
    for i=1i = 1 to KK do
        Define window bin Bi={(y(j),y^(j))}j=ii+W1B_i = \{ (y_{(j)}, \hat{y}_{(j)}) \}_{j=i}^{i+W-1}
        Compute observed positive frequency yˉi=1Wj=ii+W1y(j)\bar{y}_i = \frac{1}{W} \sum_{j=i}^{i+W-1} y_{(j)}
        Compute mean predicted probability pˉi=1Wj=ii+W1y^(j)\bar{p}_i = \frac{1}{W} \sum_{j=i}^{i+W-1} \hat{y}_{(j)}
        Compute window calibration error ei=yˉipˉie_i = |\bar{y}_i - \bar{p}_i|
        total_error=total_error+ei\text{total\_error} = \text{total\_error} + e_i
    end for
    return CAL=total_errorK\text{CAL} = \frac{\text{total\_error}}{K}

    CAL represents the expected absolute deviation between the model's confidence and the empirical posterior probability along the prediction distribution.

  7. Knowl 7 — Parameter Space Exploration and Training Configurations for Evaluated Classifier Families

    experimental setup

    To evaluate each supervised learning family at its best achievable performance, approximately 2,000 model configurations are trained per trial per problem across the following parameter spaces:

    1. Support Vector Machines (SVM): Implemented using SVMLight. Kernels: Linear, Polynomial (degree 2 and 3), Radial Basis Function (RBF) with kernel width γ{0.001,0.005,0.01,0.05,0.1,0.5,1,2}\gamma \in \{0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1, 2\}. Regularization parameter CC varied by factors of 10 from 10710^{-7} to 10310^{3} for each kernel.
    2. Artificial Neural Networks (ANN): Multi-layer perceptrons trained with gradient descent backpropagation. Hidden units {1,2,4,8,32,128}\in \{1, 2, 4, 8, 32, 128\}; momentum {0,0.2,0.5,0.9}\in \{0, 0.2, 0.5, 0.9\}. Nets are checkpointed and evaluated across multiple epochs without early stopping or weight decay.
    3. Logistic Regression (LOGREG): Unregularized and L2L_2-regularized models with regularization weight λ\lambda varied by factors of 10 from 10810^{-8} to 10410^{4}.
    4. Naive Bayes (NB): Weka implementation with continuous features handled via: (a) single Gaussian, (b) kernel density estimation, or (c) supervised discretization.
    5. Memory-Based Learning (KNN): 26 values of K[1,Ntrain]K \in [1, N_{\text{train}}]. Distance metrics: standard Euclidean, gain-ratio weighted Euclidean, distance-weighted KNN, and locally weighted averaging with kernel widths from 202^0 to 2102^{10} times the minimum inter-point training distance.
    6. Random Forests (RF): Weka implementation with 1,024 trees and split feature subset sizes {1,2,4,6,8}\in \{1, 2, 4, 6, 8\}.
    7. Decision Trees (DT): Splitting criteria and pruning variants from Buntine's IND package (BAYES, ID3, CART, CART0, C4, MML, SMML) and smoothed variants (C44LS, C44BS, MMLLS) using Laplacian or Bayesian smoothing.
    8. Bagged Decision Trees (BAG-DT): Bootstrap aggregation of 100 decision trees of each aforementioned DT type.
    9. Boosted Decision Trees (BST-DT): AdaBoost applied to each DT type, evaluated at M{2,4,8,16,32,64,128,256,512,1024,2048}M \in \{2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048\} boosting steps.
    10. Boosted Decision Stumps (BST-STMP): Single-level decision trees trained with 5 splitting criteria and boosted for M{2,4,8,,8192}M \in \{2, 4, 8, \dots, 8192\} steps.

    Data Preprocessing: Attributes are normalized to zero mean and unit standard deviation for LOGREG, ANN, SVM, and KNN. Attributes remain unscaled for DT, RF, NB, BAG-DT, BST-DT, and BST-STMP.

  8. Knowl 8 — Benchmark Binary Classification Datasets and Preprocessing

    data/table

    The empirical comparison is conducted across eight diverse binary classification problems spanning demographic, vision, medical, physical, and agricultural domains:

    Problem # Attributes Train Size Test Size % Positive
    ADULT 14 / 104 4000 35222 25%
    COVT 54 4000 25000 36%
    LTR.P1 16 4000 14000 3%
    LTR.P2 16 4000 14000 53%
    MEDIS 63 4000 8199 11%
    MG 124 4000 12807 17%
    SLAC 59 4000 25000 50%
    HS 200 4000 4366 24%

    Problem specifications and binarization procedures:

    • ADULT: UCI census income dataset; the only dataset containing nominal attributes (14 original nominal attributes expanded to 104 binary attributes for ANN, SVM, and KNN; tree-based and NB models are evaluated on both representations).
    • COVT: UCI Forest CoverType; converted to binary by designating the largest class as positive and all remaining classes as negative.
    • LTR.P1: UCI Letter Recognition; treats the letter "O" as positive and all 25 other letters as negative (highly unbalanced: 3% positive).
    • LTR.P2: UCI Letter Recognition; treats letters "A"–"M" as positive and "N"–"Z" as negative (balanced: 53% positive).
    • MEDIS and MG: Medical informatics classification datasets.
    • SLAC: Particle physics dataset from the Stanford Linear Accelerator.
    • HS: Hyperspectral remote sensing dataset (IndianPine92) with "Soybean-mintill" as the positive class.

    For every dataset, 5,000 instances are randomly sampled: 5-fold cross-validation is used where each fold assigns 4,000 instances to training and 1,000 instances to validation/calibration, while the remaining instances form the final independent test set.

  9. Knowl 9 — Multi-Criteria Taxonomy of Binary Classification Evaluation Metrics

    definition

    The nine evaluation metrics used to assess classifier performance are categorized into three mathematical families:

    1. Threshold Metrics: Depend solely on whether predictions fall above or below a fixed decision threshold θ\theta, invariant to the magnitude of the margin:

      • Accuracy (ACC): Proportion of correct classifications at threshold θ=0.5\theta = 0.5.
      • F-score (FSC): Harmonic mean of precision and recall at threshold θ=0.5\theta = 0.5, defined as F1=2PrecisionRecallPrecision+RecallF_1 = \frac{2 \cdot \text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}}.
      • Lift (LFT): Ratio of positive instances among the top p=25%p = 25\% highest-scoring cases relative to the dataset average prevalence.
    2. Ordering and Ranking Metrics: Depend exclusively on the rank order of predictions, measuring how well positive cases are sorted ahead of negative cases across all possible decision thresholds:

      • Area Under the ROC Curve (ROC): Probability that a randomly chosen positive instance receives a higher score than a randomly chosen negative instance.
      • Average Precision (APR): Average of precision values computed at the rank position of each positive instance.
      • Precision/Recall Break-Even Point (BEP): Precision at the threshold where Precision equals Recall.
    3. Probability Metrics: Minimized in expectation only when the model's predicted probability y^i\hat{y}_i matches the true conditional posterior probability P(Y=1xi)P(Y=1 \mid x_i):

      • Root Mean Squared Error (RMS): 1Ni=1N(yiy^i)2\sqrt{\frac{1}{N} \sum_{i=1}^N (y_i - \hat{y}_i)^2}.
      • Mean Cross-Entropy (MXE): 1Ni=1N(yilogy^i+(1yi)log(1y^i))-\frac{1}{N} \sum_{i=1}^N \left( y_i \log \hat{y}_i + (1 - y_i) \log(1 - \hat{y}_i) \right).
      • Calibration Error (CAL): Mean absolute difference between observed positive frequency and mean predicted probability across overlapping bins of size 100 sorted predictions.
  10. Knowl 10 — Relative Computational Cost and Parallelization Trade-Offs Among Supervised Learners

    empirical result

    The computational training cost and parallelizability vary substantially across the ten supervised learning algorithms:

    1. Most Expensive Algorithms:

      • Artificial Neural Networks (ANN): Require training hundreds of distinct networks per dataset to explore hidden unit counts, momentum terms, learning rates, and training epoch checkpoints, making ANNs one of the most computationally demanding methods.
      • Support Vector Machines (SVM): While many SVM hyperparameter configurations train quickly, specific combinations of kernel bandwidth γ\gamma and large regularization parameter CC lead to slow quadratic/cubic optimization times, dominating total run time.
    2. Ensemble Scalability and Parallelism:

      • Bagged Trees (BAG-DT): Highly computationally efficient. Because individual trees are independent, bagging is embarrassingly parallel, and performance typically converges within 100 bootstrap iterations.
      • Random Forests (RF): Less expensive than boosting but more expensive than bagging because performance requires more trees (e.g., 1,024 trees) to converge, alongside tuning the subset feature size.
      • Boosted Trees and Stumps (BST-DT, BST-STMP): Computationally expensive in practice because boosting is strictly sequential and difficult to parallelize, requiring thousands of boosting iterations (up to 2,048 for full trees and 8,192 for stumps).
    3. Fastest Algorithms:

      • Logistic Regression (LOGREG) and Naive Bayes (NB): Exhibit the lowest computational training and tuning overhead.
      • Memory-Based Learning (KNN): Inexpensive during training on 4,000-instance training sets, but inference cost scales linearly with training set size.

Coverage note — No substantial contributed material was omitted; the knowls cover all empirical comparisons, algorithm configurations, calibration methods, metric normalization procedures, and dataset evaluations presented in the study.

References

  1. 1.King, R., Feng, C., Shutherland, A.: Statlog: comparison of classification algorithms on large real-world problems. Applied Artificial Intelligence 9 (1995)
  2. 2.Joachims, T.: Making large-scale svm learning practical. In: Advances in Kernel Methods. (1999)
  3. 3.Provost, F., Domingos, P.: Tree induction for probability-based rankings. Machine Learning (2003)
  4. 4.Giudici, P.: Applied Data Mining. John Wiley and Sons, New York (2003)
  5. 5.Provost, F.J., Fawcett, T.: Analysis and visualization of classifier performance: Comparison under imprecise class and cost distributions. In: Knowledge Discovery and Data Mining. (1997) 43–48
  6. 6.Platt, J.: Probabilistic outputs for support vector machines and comparison to regularized likelihood methods. In: Adv. in Large Margin Classifiers. (1999)
  7. 7.Zadrozny, B., Elkan, C.: Transforming classifier scores into accurate multiclass probability estimates. In: KDD. (2002)
  8. 8.Zadrozny, B., Elkan, C.: Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. In: ICML. (2001)
  9. 9.Robertson, T., Wright, F., Dykstra, R.: Order Restricted Statistical Inference. John Wiley and Sons, New York (1988)
  10. 10.Ayer, M., Brunk, H., Ewing, G., Reid, W., Silverman, E.: An empirical distribution function for sampling with incomplete information. Annals of Mathematical Statistics 5 (1955) 641–647
  11. 11.Blake, C., Merz, C.: UCI repository of machine learning databases (1998)
  12. 12.Gualtieri, A., Chettri, S.R., Cromp, R., Johnson, L.: Support vector machine classifiers as applied to aviris data. In: Proc. Eighth JPL Airborne Geoscience Workshop. (1999)
  13. 13.LeCun, Y., Jackel, L.D., Bottou, L., Brunot, A., Cortes, C., Denker, J.S., Drucker, H., Guyon, I., Muller, U.A., Sackinger, E., Simard, P., Vapnik, V.: Comparison of learning algorithms for handwritten digit recognition. In Fogelman, F., Gallinari, P., eds.: International Conference on Artificial Neural Networks, Paris, EC2 & Cie (1995) 53–60
  14. 14.Cooper, G.F., Aliferis, C.F., Ambrosino, R., Aronis, J., Buchanan, B.G., Caruana, R., Fine, M.J., Glymour, C., Gordon, G., Hanusa, B.H., Janosky, J.E., Meek, C., Mitchell, T., Richardson, T., Spirtes, P.: An evaluation of machine learning methods for predicting pneumonia mortality. Artificial Intelligence in Medicine 9 (1997)
  15. 15.Lim, T.S., Loh, W.Y., Shih, Y.S.: An empirical comparison of decision trees and other classification methods. Technical Report 979, Madison, WI (1997)
  16. 16.Bauer, E., Kohavi, R.: An empirical comparison of voting classification algorithms: Bagging, boosting, and variants. Machine Learning 36 (1999)
  17. 17.Perlich, C., Provost, F., Simonoff, J.S.: Tree induction vs. logistic regression: a learning-curve analysis. J. Mach. Learn. Res. 4 (2003) 211–255
  18. 18.Vapnik, V.: Statistical Learning Theory. John Wiley and Sons, New York (1998)
  19. 19.Schapire, R.: The boosting approach to machine learning: An overview. In: In MSRI Workshop on Nonlinear Estimation and Classification. (2001)
  20. 20.Breiman, L.: Bagging predictors. Machine Learning 24 (1996) 123–140

Citation

MLA
Caruana, R., and A. Niculescu-Mizil. “An Empirical Comparison of Supervised Learning Algorithms”. Proceedings of the 23rd International Conference on Machine Learning - ICML '06, 2006, pp. 161–68, https://doi.org/10.1145/1143844.1143865.
APA
Caruana, R., & Niculescu-Mizil, A. (2006). An empirical comparison of supervised learning algorithms. Proceedings of the 23rd International Conference on Machine Learning - ICML '06, 161–168. https://doi.org/10.1145/1143844.1143865
Chicago
Caruana, R., and A. Niculescu-Mizil. 2006. “An Empirical Comparison of Supervised Learning Algorithms”. Proceedings of the 23rd International Conference on Machine Learning - ICML '06, 161–68. https://doi.org/10.1145/1143844.1143865.
Harvard
Caruana, R. and Niculescu-Mizil, A. (2006) “An empirical comparison of supervised learning algorithms”, Proceedings of the 23rd international conference on Machine learning - ICML '06. ACM Press, pp. 161–168. Available at: https://doi.org/10.1145/1143844.1143865.
Vancouver
1. Caruana R, Niculescu-Mizil A (2006) An empirical comparison of supervised learning algorithms. In: Proceedings of the 23rd international conference on Machine learning - ICML '06. ACM Press, pp 161–168

BibTeX

@inproceedings{Caruana_2006, series={ICML ’06}, title={An empirical comparison of supervised learning algorithms}, url={http://dx.doi.org/10.1145/1143844.1143865}, DOI={10.1145/1143844.1143865}, booktitle={Proceedings of the 23rd international conference on Machine learning  - ICML ’06}, publisher={ACM Press}, author={Caruana, Rich and Niculescu-Mizil, Alexandru}, year={2006}, pages={161–168}, collection={ICML ’06} }
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: Authors