当前位置:文档之家› An Empirical Comparison of Supervised Learning Algorithms

An Empirical Comparison of Supervised Learning Algorithms

An Empirical Comparison of Supervised Learning Algorithms
An Empirical Comparison of Supervised Learning Algorithms

An Empirical Comparison of Supervised Learning Algorithms Rich Caruana caruana@https://www.doczj.com/doc/d816085462.html,

Alexandru Niculescu-Mizil alexn@https://www.doczj.com/doc/d816085462.html, Department of Computer Science,Cornell University,Ithaca,NY14853USA

Abstract

A number of supervised learning methods

have been introduced in the last decade.Un-

fortunately,the last comprehensive empiri-

cal evaluation of supervised learning was the

Statlog Project in the early90’s.We present

a large-scale empirical comparison between

ten supervised learning methods:SVMs,

neural nets,logistic regression,naive bayes,

memory-based learning,random forests,de-

cision trees,bagged trees,boosted trees,and

boosted stumps.We also examine the e?ect

that calibrating the models via Platt Scaling

and Isotonic Regression has on their perfor-

mance.An important aspect of our study is

the use of a variety of performance criteria to

evaluate the learning methods.

1.Introduction

There are few comprehensive empirical studies com-paring learning algorithms.STATLOG is perhaps the best known study(King et al.,1995).STATLOG was very comprehensive when it was performed,but since then new learning algorithms have emerged(e.g.,bag-ging,boosting,SVMs,random forests)that have excel-lent performance.An extensive empirical evaluation of modern learning methods would be useful.

Learning algorithms are now used in many domains, and di?erent performance metrics are appropriate for each domain.For example Precision/Recall measures are used in information retrieval;medicine prefers ROC area;Lift is appropriate for some marketing tasks,etc.The di?erent performance metrics measure di?erent tradeo?s in the predictions made by a clas-si?er,and it is possible for learning methods to per-form well on one metric,but be suboptimal on other metrics.Because of this it is important to evaluate algorithms on a broad set of performance metrics. Appearing in Proceedings of the23rd International Con-ference on Machine Learning,Pittsburgh,PA,2006.Copy-right2006by the author(s)/owner(s).This paper presents results of a large-scale empirical comparison of ten supervised learning algorithms us-ing eight performance criteria.We evaluate the perfor-mance of SVMs,neural nets,logistic regression,naive bayes,memory-based learning,random forests,deci-sion trees,bagged trees,boosted trees,and boosted stumps on eleven binary classi?cation problems using a variety of performance metrics:accuracy,F-score, Lift,ROC Area,average precision,precision/recall break-even point,squared error,and cross-entropy. For each algorithm we examine common variations, and thoroughly explore the space of parameters.For example,we compare ten decision tree styles,neural nets of many sizes,SVMs with many kernels,etc. Because some of the performance metrics we examine interpret model predictions as probabilities and mod-els such as SVMs are not designed to predict probabil-ities,we compare the performance of each algorithm both before and after calibrating its predictions with Platt Scaling and Isotonic Regression.

The empirical results are surprising.To preview:prior to calibration,bagged trees,random forests,and neu-ral nets give the best average performance across all eight metrics and eleven test problems.Boosted trees, however,are best if we restrict attention to the six metrics that do not require probabilities.After cal-ibration with Platt’s Method,boosted trees predict better probabilities than all other methods and move into?rst place overall.Neural nets,on the other hand, are so well calibrated to begin with that they are hurt slightly by calibration.After calibration with Platt’s Method or Isotonic Regression,SVMs perform compa-rably to neural nets and nearly as well as boosted trees, random forests and bagged trees.Boosting full deci-sion trees dramatically outperforms boosting weaker stumps on most problems.On average,memory-based learning,boosted stumps,single decision trees,logistic regression,and naive bayes are not competitive with the best methods.These generalizations,however,do not always hold.For example,boosted stumps and logistic regression,which perform poorly on average, are the best models for some metrics on two of the test problems.

2.Methodology

2.1.Learning Algorithms

We attempt to explore the space of parameters and common variations for each learning algorithm as thor-oughly as is computationally feasible.This section summarizes the parameters used for each learning al-gorithm,and may safely be skipped by readers who are easily bored.

SVMs:we use the following kernels in SVMLight (Joachims,1999):linear,polynomial degree2&3,ra-dial with width{0.001,0.005,0.01,0.05,0.1,0.5,1,2}.We also vary the regularization parameter by factors of ten from10?7to103with each kernel.

ANN we train neural nets with gradient descent backprop and vary the number of hidden units {1,2,4,8,32,128}and the momentum{0,0.2,0.5,0.9}. We halt training the nets at many di?erent epochs and use validation sets to select the best nets. Logistic Regression(LOGREG):we train both unregularized and regularized models,varying the ridge(regularization)parameter by factors of10from 10?8to104.

Naive Bayes(NB):we use Weka(Witten&Frank, 2005)and try all three of the Weka options for han-dling continuous attributes:modeling them as a single normal,modeling them with kernel estimation,or dis-cretizing them using supervised discretization. KNN:we use26values of K ranging from K=1to K=|trainset|.We use KNN with Euclidean distance and Euclidean distance weighted by gain ratio.We also use distance weighted KNN,and locally weighted averaging.The kernel widths for locally weighted aver-aging vary from20to210times the minimum distance between any two points in the train set.

Random Forests(RF):we tried both the Breiman-Cutler and Weka implementations;Breiman-Cutler yielded better results so we report those here.The forests have1024trees.The size of the feature set considered at each split is1,2,4,6,8,12,16or20. Decision trees(DT):we vary the splitting crite-rion,pruning options,and smoothing(Laplacian or Bayesian smoothing).We use all of the tree models in Buntine’s IND package(Buntine&Caruana,1991): BAYES,ID3,CART,CART0,C4,MML,and SMML. We also generate trees of type C44LS(C4with no pruning and Laplacian smoothing),C44BS(C44with Bayesian smoothing),and MMLLS(MML with Lapla-cian smoothing).See(Provost&Domingos,2003)for a description of C44LS.Bagged trees(BAG-DT):we bag100trees of each type described above.With boosted trees (BST-DT)we boost each tree type as well.Boost-ing can over?t,so we consider boosted trees af-ter2,4,8,16,32,64,128,256,512,1024and2048steps of boosting.With boosted stumps(BST-STMP)we boost single level decision trees gener-ated with5di?erent splitting criteria,each boosted for 2,4,8,16,32,64,128,256,512,1024,2048,4096,8192steps. With LOGREG,ANN,SVM and KNN we scale at-tributes to0mean1std.With DT,RF,NB,BAG-DT,BST-DT and BST-STMP we don’t scale the data. In total,we train about2000di?erent models in each trial on each problem.

2.2.Performance Metrics

We divide the eight performance metrics into three groups:threshold metrics,ordering/rank metrics and probability metrics.

The threshold metrics are accuracy(ACC),F-score (FSC)and lift(LFT).See(Giudici,2003)for a de-scription of Lift https://www.doczj.com/doc/d816085462.html,ually ACC and FSC have a?xed threshold(we use0.5).For lift,often a?xed percent,p,of cases are predicted as positive and the rest as negative(we use p=25%).For thresholded metrics,it is not important how close a prediction is to a threshold,only if it is above or below threshold. The ordering/rank metrics depend only on the order-ing of the cases,not the actual predicted values.As long as ordering is preserved,it makes no di?erence if predicted values fall between0and1or0.89and0.90. These metrics measure how well the positive cases are ordered before negative cases and can be viewed as a summary of model performance across all possible thresholds.The rank metrics we use are area under the ROC curve(ROC),average precision(APR),and precision/recall break even point(BEP).See(Provost &Fawcett,1997)for a discussion of ROC from a ma-chine learning perspective.

The probability metrics,squared error(RMS)and cross-entropy(MXE),interpret the predicted value of each case as the conditional probability of that case being in the positive class.See(Caruana&Niculescu-Mizil,2004)for a more detailed description of the eight performance metrics.

https://www.doczj.com/doc/d816085462.html,paring Across Performance Metrics Performance metrics such as accuracy or squared er-ror have range[0,1],while others(lift,cross entropy) range from0to p where p depends on the data set. For some metrics lower values indicate better perfor-

mance.For others higher values are better.Metrics such as ROC area have baseline rates that are indepen-dent of the data,while others such as accuracy have baseline rates that depend on the data.If baseline ac-curacy is0.98,an accuracy of0.981probably is not good performance,but on another problem the Bayes optimal rate might be0.60and achieving an accuracy of0.59might be excellent performance.

To permit averaging across metrics and problems,per-formances must be placed on comparable scales.We do this by scaling performance for each problem and metric from0to1,where0is baseline performance and1is Bayes optimal.Since the Bayes optimal rate cannot be estimated on real problems,we use the best observed performance as a proxy.The following base-line model is used:predict p for every case,where p is the percent of positives in the data.1Performances are normalized to[0,1],where0is baseline and1rep-resents best performance.If a model performs worse than baseline,its normalized score will be negative. One disadvantage of normalized scores is that recov-ering raw performances requires knowing the perfor-mances at the top and bottom of the scale,and as new best models are found the top of the scale may change. The performances that de?ne the top and bottom of the scales for each problem and metric are available at https://www.doczj.com/doc/d816085462.html,\~caruana to allow comparison with our results.

2.4.Calibration Methods

Some of the learning algorithms we examine are not designed to predict probabilities.For example the out-puts of an SVM are normalized distances to the deci-sion boundary.And naive bayes models are known to predict poorly calibrated probabilities because of the unrealistic independence assumption.

A number of methods have been proposed for map-ping predictions to posterior probabilities.Platt (1999)proposed transforming SVM predictions to pos-terior probabilities by passing them through a sigmoid. Platt’s method also works well for boosted trees and boosted stumps(Niculescu-Mizil&Caruana,2005).A sigmoid,however,might not be the correct transfor-mation for all learning algorithms.

Zadrozny and Elkan(2002;2001)used a more gen-eral calibration method based on Isotonic Regression (Robertson et al.,1988)to calibrate predictions from SVMs,naive bayes,boosted naive bayes,and decision trees.Isotonic Regression is more general in that the 1For F-Score,baseline is calculated by predicting1for all cases since predicting p can yield0F-Score.

Table1.Description of problems

problem#attr train size test size%poz adult14/10450003522225% bact11/17050003426269% cod15/6050001400050% calhous950001464052% cov type5450002500036% hs2005000436624% letter.p1165000140003% letter.p21650001400053% medis635000819911% mg12450001280717% slac5950002500050% only restriction it makes is that the mapping function be isotonic(monotonically increasing).A standard al-gorithm for Isotonic Regression that?nds a piecewise constant solution in linear time,is the pair-adjacent violators(PAV)algorithm(Ayer et al.,1955).

To calibrate models,we use the same1000points val-idation set that will be used for model selection.

2.5.Data Sets

We compare the algorithms on11binary classi?ca-tion problems.ADULT,COV TYPE and LETTER are from the UCI Repository(Blake&Merz,1998). COV TYPE has been converted to a binary problem by treating the largest class as the positive and the rest as negative.We converted LETTER to boolean in two ways.LETTER.p1treats”O”as positive and the remaining25letters as negative,yielding a very unbal-anced problem.LETTER.p2uses letters A-M as posi-tives and the rest as negatives,yielding a well balanced problem.HS is the IndianPine92data set(Gualtieri et al.,1999)where the di?cult class Soybean-mintill is the positive class.SLAC is a problem from the Stan-ford Linear Accelerator.MEDIS and MG are medical data sets.COD,BACT,and CALHOUS are three of the datasets used in(Perlich et al.,2003).ADULT, COD,and BACT contain nominal attributes.For ANNs,SVMs,KNNs,and LOGREG we transform nominal attributes to boolean(one boolean per value). Each DT,BAG-DT,BST-DT,BST-STMP,RF,and NB model is trained twice,once with transformed at-tributes and once with the original ones.See Table1 for characteristics of these problems.

3.Performances by Metric

For each test problem we randomly select5000cases for training and use the rest of the cases as a large ?nal test set.We use5-fold cross validation on the

5000cases to obtain?ve trials.For each trial we use 4000cases to train the di?erent models,1000cases to calibrate the models and select the best parameters, and then report performance on the large?nal test set.We would like to run more trials,but this is a very expensive set of experiments.Fortunately,even with only?ve trials we are able to discern interesting di?erences between methods.

Table2shows the normalized score for each algorithm on each of the eight metrics.For each problem and metric we?nd the best parameter settings for each al-gorithm using the1k validation sets set aside by cross-validation,then report that model’s normalized score on the?nal test set.Each entry in the table averages these scores across the?ve trials and eight test prob-lems.The second column tells if model predictions were calibrated.A“–”means the model predictions were not calibrated–they are the raw model predic-tions.(The one exception is SVMs,where distances to the separating hyperplane are linearly scaled to[0,1] before computing the three probability metrics.)A “PLT”or“ISO”in the second column indicates that the model predictions were scaled after the model was trained using Platt Scaling or Isotonic Regression,re-spectively.These scaling methods were discussed in Section2.4.In the table,higher scores always indicate better performance.

The second to last column,MEAN,is the mean nor-malized score over the eight metrics,eleven problems, and?ve trials.The models in the table are sorted by the mean normalized score in this column.For now, ignore the last column,OPT-SEL.This column will be discussed later in this section.

In the table,the algorithm with the best performance on each metric is boldfaced.Other algorithm’s whose performance is not statistically distinguishable from the best algorithm at p=0.05using paired t-tests on the5trials are*’ed.2Entries in the table that are neither bold nor starred indicate performance that is signi?cantly lower than the best models at p=0.05.3 2Performing this many independent t-tests,each at p=0.05,is problematic.Some di?erences that are labeled signi?cant in the table probably are not truly signi?cant at p=0.05.We considered applying a more stringent experiment-wise p-value that takes into account the num-ber of tests performed,but the strong correlations between performances on di?erent metrics,and on calibrated and uncalibrated models,makes this problematic as well,so we decided to keep it simple.Most of the di?erences in the ta-ble are signi?cant well beyond p=0.05.Doing the t-tests at p=0.01adds few additional stars to the table.

3Note that it is possible for the di?erence between the scores0.90and0.89to be statistically signi?cant,and yet for the same0.90score to be statistically indistinguishable Averaging across all eight metrics,the strongest mod-els are calibrated boosted trees,calibrated random forests,bagged trees,PLT-calibrated SVMs and neu-ral nets.If not calibrated,the best models overall are bagged trees,random forests,and neural nets.With or without calibration,the poorest performing models are naive bayes,logistic regression,and decision trees. Memory-based methods(e.g.KNN)are remarkably una?ected by calibration,but exhibit mediocre overall performance.Boosted stumps,even after calibration, also have mediocre performance,and do not perform nearly as well as boosted full trees.

Looking at individual metrics,boosted trees,which have poor squared error and cross-entropy prior to cal-ibration,dominate the other algorithms on these met-rics after calibration.Bagged trees,random forests and neural nets also predict good probabilities.Inter-estingly,calibrating neural nets with PLT or ISO hurts their calibration.If neural nets are trained well to be-gin with it is better not to adjust their predictions. Boosted trees also have excellent performance on or-dering metrics:area under the ROC,average precision, and the precision/recall break-even point.4Random forests and bagged trees have similar performance as boosted trees on these metrics.The neural nets and SVMs also order cases extremely well.

On metrics that compare predictions to a thresh-old:accuracy,F-score,and Lift,the best models are calibrated or uncalibrated random forests,calibrated boosted trees,and bagged trees with or without cali-bration.We are not sure why ISO-calibration signi?-cantly improves the F-Score of all models except NB, including well-calibrated models such as neural nets and bagged trees.

The last column,OPT-SEL,is the mean normalized score for the eight metrics when model selection is done by cheating and looking at the?nal test sets.The means in this column represent the best performance that could be achieved with each learning method if model selection were done optimally.We present these numbers because parameter optimization is more crit-ical(and more di?cult)with some algorithms than with others.For example,bagging works well with most decision tree types and requires little tuning,but neural nets and SVMs require careful parameter selec-tion.As expected,the mean normalized scores in the from a poorer score of0.88if the variance of the0.88score is higher than the variance of the0.89score.

4PLT calibration does not change the ordering predicted by models,so it does not a?ect these metrics.ISO calibra-tion,however,can introduce ties in the predicted values that may a?ect performance on the ordering metrics.

Table2.Normalized scores for each learning algorithm by metric(average over eleven problems)

model cal acc fsc lft roc apr bep rms mxe mean opt-sel bst-dt plt.843*.779.939.963.938.929*.880.896.896.917 rf plt.872*.805.934*.957.931.930.851.858.892.898 bag-dt–.846.781.938*.962*.937*.918.845.872.887*.899 bst-dt iso.826*.860*.929*.952.921.925*.854.815.885.917* rf–.872.790.934*.957.931.930.829.830.884.890 bag-dt plt.841.774.938*.962*.937*.918.836.852.882.895

rf iso.861*.861.923.946.910.925.836.776.880.895 bag-dt iso.826.843*.933*.954.921.915.832.791.877.894 svm plt.824.760.895.938.898.913.831.836.862.880 ann–.803.762.910.936.892.899.811.821.854.885 svm iso.813.836*.892.925.882.911.814.744.852.882 ann plt.815.748.910.936.892.899.783.785.846.875 ann iso.803.836.908.924.876.891.777.718.842.884 bst-dt–.834*.816.939.963.938.929*.598.605.828.851 knn plt.757.707.889.918.872.872.742.764.815.837 knn–.756.728.889.918.872.872.729.718.810.830 knn iso.755.758.882.907.854.869.738.706.809.844 bst-stmp plt.724.651.876.908.853.845.716.754.791.808 svm–.817.804.895.938.899.913.514.467.781.810 bst-stmp iso.709.744.873.899.835.840.695.646.780.810 bst-stmp–.741.684.876.908.853.845.394.382.710.726 dt iso.648.654.818.838.756.778.590.589.709.774 dt–.647.639.824.843.762.777.562.607.708.763 dt plt.651.618.824.843.762.777.575.594.706.761

lr–.636.545.823.852.743.734.620.645.700.710 lr iso.627.567.818.847.735.742.608.589.692.703

lr plt.630.500.823.852.743.734.593.604.685.695 nb iso.579.468.779.820.727.733.572.555.654.661 nb plt.576.448.780.824.738.735.537.559.650.654 nb–.496.562.781.825.738.735.347-.633.481.489

cheating column(OPT-SEL)tend to be higher than the mean normalized scores when selection is done us-ing1k validation sets because model selection using the validation sets does not always select the model with the best performance on the?nal test set. Comparing the MEAN and OPT-SEL columns,selec-tion using1k validation sets yields on average about 0.023decrease in normalized score compared to op-timal selection.As expected,high variance models have the biggest drop in performance.Neural nets, for example,which have relatively high variance,lose substantial performance when selecting models with 1k validation sets,while other models such as random forests,which have small variance,lose very little per-formance when selection is done using1k validation sets.SVM’s have variance between RF and ANNs,and thus lose more than RF,but less than ANN.Boosted trees also have relatively high variance,but their over-all performance after PLT or ISO calibration is so strong that they remain the best model overall even when selection is done using1k validation sets.4.Performances by Problem

Table3shows the normalized score for each algorithm on each of the11test problems.Each entry is an average over the eight performance metrics and?ve trials when selection is done using1k validation sets. As the No Free Lunch Theorem suggests,there is no universally best learning algorithm.Even the best models(calibrated boosted trees,random forests and bagged trees)perform poorly on some problems,and models that have poor average performance perform well on a few problems or metrics.For example,the best models on ADULT are calibrated boosted stumps, random forests and bagged trees.Boosted trees per-form much worse.Bagged trees and random forests also perform very well on MG and SLAC.On MEDIS, the best models are random forests,neural nets and lo-gistic regression.The only models that never exhibit excellent performance on any problem are naive bayes and memory-based learning.

Boosting full decision trees yields better performance

Table3.Normalized scores of each learning algorithm by problem(averaged over eight metrics)

model cal covt adult ltr.p1ltr.p2medis slac hs mg calhous cod bact mean bst-dt plt.938.857.959.976.700.869.933.855.974.915.878*.896* rf plt.876.930.897.941.810.907*.884.883.937.903*.847.892 bag-dt–.878.944*.883.911.762.898*.856.898.948.856.926.887* bst-dt iso.922*.865.901*.969.692*.878.927.845.965.912*.861.885* rf–.876.946*.883.922.785.912*.871.891*.941.874.824.884 bag-dt plt.873.931.877.920.752.885.863.884.944.865.912*.882 rf iso.865.934.851.935.767*.920.877.876.933.897*.821.880 bag-dt iso.867.933.840.915.749.897.856.884.940.859.907*.877 svm plt.765.886.936.962.733.866.913*.816.897.900*.807.862 ann–.764.884.913.901.791*.881.932*.859.923.667.882.854 svm iso.758.882.899.954.693*.878.907.827.897.900*.778.852 ann plt.766.872.898.894.775.871.929*.846.919.665.871.846 ann iso.767.882.821.891.785*.895.926*.841.915.672.862.842 bst-dt–.874.842.875.913.523.807.860.785.933.835.858.828 knn plt.819.785.920.937.626.777.803.844.827.774.855.815 knn–.807.780.912.936.598.800.801.853.827.748.852.810 knn iso.814.784.879.935.633.791.794.832.824.777.833.809 bst-stmp plt.644.949.767.688.723.806.800.862.923.622.915*.791 svm–.696.819.731.860.600.859.788.776.833.864.763.781 bst-stmp iso.639.941.700.681.711.807.793.862.912.632.902*.780 bst-stmp–.605.865.540.615.624.779.683.799.817.581.906*.710 dt iso.671.869.729.760.424.777.622.815.832.415.884.709 dt–.652.872.723.763.449.769.609.829.831.389.899*.708 dt plt.661.863.734.756.416.779.607.822.826.407.890*.706 lr–.625.886.195.448.777*.852.675.849.838.647.905*.700 lr iso.616.881.229.440.763*.834.659.827.833.636.889*.692 lr plt.610.870.185.446.738.835.667.823.832.633.895.685 nb iso.574.904.674.557.709.724.205.687.758.633.770.654 nb plt.572.892.648.561.694.732.213.690.755.632.756.650 nb–.552.843.534.556.011.714-.654.655.759.636.688.481

than boosting stumps only on seven problems.Occa-sionally boosted stumps perform very well,but some-times they perform very poorly so their average per-formance is low.On ADULT,when boosting trees,the ?rst iteration of boosting hurts the performance of all tree types,and never recovers in subsequent rounds. When this happens even single decision trees outper-form their boosted counterparts.Bagged trees and random forests,however,consistently outperform sin-gle trees on all problems.Bagging and random forests are“safer”than boosting,even on the metrics for which boosting yields the best overall performance.

5.Bootstrap Analysis

The results depend on the choice of problems and met-rics.What impact might selecting other problems,or evaluating performance on other metrics,have on the results?For example,neural nets perform well on all metrics on10of11problems,but perform poorly on COD.If we hadn’t included the COD problem,neural nets would move up1-2places in the rankings.To help evaluate the impact of the choice of problems and metrics we performed a bootstrap analysis.We randomly select a bootstrap sample(sampling with replacement)from the original11problems.For this sample of problems we then randomly select a boot-strap sample of8metrics from the original8metrics (again sampling with replacement).For this bootstrap sample of problems and metrics we rank the ten algo-rithms by mean performance across the sampled prob-lems and metrics(and the5folds).This bootstrap sampling is repeated1000times,yielding1000poten-tially di?erent rankings of the learning methods. Table4shows the frequency that each method ranks 1st,2nd,3rd,etc.The0.228entry for boosted trees in the column for2nd place,tells us that there is a 22.8%chance that boosted decision trees would have placed2nd in the table of results(instead of1st)had we selected other problems and/or metrics.

The bootstrap analysis complements the t-tests in Ta-bles2and 3.The results suggest that if we had se-lected other problems/metrics,there is a58%chance that boosted decision trees would still have ranked1st

Table4.Bootstrap Analysis of Overall Rank by Mean Performance Across Problems and Metrics

model1st2nd3rd4th5th6th7th8th9th10th bst-dt0.5800.2280.1600.0230.0090.0000.0000.0000.0000.000 rf0.3900.5250.0840.0010.0000.0000.0000.0000.0000.000 bag-dt0.0300.2320.5710.1500.0170.0000.0000.0000.0000.000 svm0.0000.0080.1480.5740.2400.0290.0010.0000.0000.000 ann0.0000.0070.0350.2300.6060.1220.0000.0000.0000.000 knn0.0000.0000.0000.0090.1140.5920.2450.0380.0020.000 bst-stmp0.0000.0000.0020.0130.0140.2570.7100.0040.0000.000 dt0.0000.0000.0000.0000.0000.0000.0040.6160.2910.089 logreg0.0000.0000.0000.0000.0000.0000.0400.3120.4230.225 nb0.0000.0000.0000.0000.0000.0000.0000.0300.2840.686

overall,and only a4.2%chance of seeing them rank lower than3rd place.Random forests would come in 1st place39%of the time,2nd place53%of the time, with little chance(0.1%)of ranking below third place. There is less than a20%chance that a method other than boosted trees,random forests,and bagged trees would rank in the top three,and no chance(0.0%) that another method would rank1st—it appears to be a clean sweep for ensembles of trees.SVMs probably would rank4th,and neural nets probably would rank 5th,but there is a1in3chance that SVMs would rank after neural nets.The bootstrap analysis clearly shows that MBL,boosted1-level stumps,plain decision trees, logistic regression,and naive bayes are not competitive on average with the top?ve models on these problems and metrics when trained on5k samples.

6.Related Work

STATLOG is perhaps the best known study(King et al.,1995).STATLOG was a very comprehensive study when it was performed,but since then important new learning algorithms have been introduced such as bagging,boosting,SVMs,and random forests.LeCun et al.(1995)presents a study that compares several learning algorithms(including SVMs)on a handwrit-ing recognition problem using three performance crite-ria:accuracy,rejection rate,and computational cost. Cooper et al.(1997)present results from a study that evaluates nearly a dozen learning methods on a real medical data set using both accuracy and an ROC-like metric.Lim et al.(2000)perform an empirical com-parison of decision trees and other classi?cation meth-ods using accuracy as the main criterion.Bauer and Kohavi(1999)present an impressive empirical analy-sis of ensemble methods such as bagging and boosting. Perlich et al.(2003)conducts an empirical comparison between decision trees and logistic regression.Provost and Domingos(2003)examine the issue of predicting probabilities with decision trees,including smoothed and bagged trees.Provost and Fawcett(1997)discuss the importance of evaluating learning algorithms on metrics other than accuracy such as ROC.

7.Conclusions

The?eld has made substantial progress in the last decade.Learning methods such as boosting,random forests,bagging,and SVMs achieve excellent perfor-mance that would have been di?cult to obtain just15 years ago.Of the earlier learning methods,feedfor-ward neural nets have the best performance and are competitive with some of the newer methods,particu-larly if models will not be calibrated after training. Calibration with either Platt’s method or Isotonic Re-gression is remarkably e?ective at obtaining excellent performance on the probability metrics from learning algorithms that performed well on the ordering met-rics.Calibration dramatically improves the perfor-mance of boosted trees,SVMs,boosted stumps,and Naive Bayes,and provides a small,but noticeable im-provement for random forests.Neural nets,bagged trees,memory based methods,and logistic regression are not signi?cantly improved by calibration.

With excellent performance on all eight metrics,cali-brated boosted trees were the best learning algorithm overall.Random forests are close second,followed by uncalibrated bagged trees,calibrated SVMs,and un-calibrated neural nets.The models that performed poorest were naive bayes,logistic regression,decision trees,and boosted stumps.Although some methods clearly perform better or worse than other methods on average,there is signi?cant variability across the problems and metrics.Even the best models some-times perform poorly,and models with poor average

performance occasionally perform exceptionally well. Acknowledgments

We thank F.Provost and C.Perlich for the BACT, COD,and CALHOUS data,C.Young for the SLAC data,A.Gualtieri for the HS data,and B.Zadrozny and C.Elkan for the Isotonic Regression code.This work was supported by NSF Award0412930. References

Ayer,M.,Brunk,H.,Ewing,G.,Reid,W.,&Silver-man,E.(1955).An empirical distribution function for sampling with incomplete information.Annals of Mathematical Statistics,5,641–647.

Bauer,E.,&Kohavi,R.(1999).An empirical com-parison of voting classi?cation algorithms:Bagging, boosting,and variants.Machine Learning,36.

Blake,C.,&Merz,C.(1998).UCI repository of ma-chine learning databases.

Breiman,L.(1996).Bagging predictors.Machine Learning,24,123–140.

Breiman,L.(2001).Random forests.Machine Learn-ing,45,5–32.

Buntine,W.,&Caruana,R.(1991).Introduction to ind and recursive partitioning(Technical Report FIA-91-28).NASA Ames Research Center. Caruana,R.,&Niculescu-Mizil,A.(2004).Data min-ing in metric space:An empirical analysis of sup-pervised learning performance criteria.Knowledge Discovery and Data Mining(KDD’04).

Cooper,G.F.,Aliferis,C.F.,Ambrosino,R.,Aronis, J.,Buchanan,B.G.,Caruana,R.,Fine,M.J.,Gly-mour,C.,Gordon,G.,Hanusa,B.H.,Janosky,J.E., Meek,C.,Mitchell,T.,Richardson,T.,&Spirtes, P.(1997).An evaluation of machine learning meth-ods for predicting pneumonia mortality.Arti?cial Intelligence in Medicine,9.

Giudici,P.(2003).Applied data mining.New York: John Wiley and Sons.

Gualtieri,A.,Chettri,S.R.,Cromp,R.,&Johnson, L.(1999).Support vector machine classi?ers as ap-plied to aviris data.Proc.Eighth JPL Airborne Geo-science Workshop.

Joachims,T.(1999).Making large-scale svm learning practical.Advances in Kernel Methods.King,R.,Feng,C.,&Shutherland,A.(1995).Statlog: comparison of classi?cation algorithms on large real-world problems.Applied Arti?cial Intelligence,9. 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.(1995).Comparison of learning algorithms for handwritten digit recognition.International Con-ference on Arti?cial Neural Networks(pp.53–60). Paris:EC2&Cie.

Lim,T.-S.,Loh,W.-Y.,&Shih,Y.-S.(2000).A comparison of prediction accuracy,complexity,and training time of thirty-three old and new classi?ca-tion algorithms.Machine Learning,40,203–228. Niculescu-Mizil,A.,&Caruana,R.(2005).Predicting good probabilities with supervised learning.Proc. 22nd International Conference on Machine Learn-ing(ICML’05).

Perlich,C.,Provost,F.,&Simono?,J.S.(2003).Tree induction vs.logistic regression:a learning-curve analysis.J.Mach.Learn.Res.,4,211–255. Platt,J.(1999).Probabilistic outputs for support vec-tor machines and comparison to regularized likeli-hood methods.Adv.in Large Margin Classi?ers. Provost,F.,&Domingos,P.(2003).Tree induction for probability-based rankings.Machine Learning. Provost,F.J.,&Fawcett,T.(1997).Analysis and visualization of classi?er performance:Comparison under imprecise class and cost distributions.Knowl-edge Discovery and Data Mining(pp.43–48). Robertson,T.,Wright,F.,&Dykstra,R.(1988).Or-der restricted statistical inference.New York:John Wiley and Sons.

Schapire,R.(2001).The boosting approach to ma-chine learning:An overview.In MSRI Workshop on Nonlinear Estimation and Classi?cation. Vapnik,V.(1998).Statistical learning theory.New York:John Wiley and Sons.

Witten,I.H.,&Frank,E.(2005).Data mining:Prac-tical machine learning tools and techniques.San Francisco:Morgan Kaufmann.Second edition. Zadrozny, B.,&Elkan,C.(2001).Obtaining cali-brated probability estimates from decision trees and naive bayesian classi?ers.ICML.

Zadrozny,B.,&Elkan,C.(2002).Transforming clas-si?er scores into accurate multiclass probability es-timates.KDD.

美国历史与文化

《美国历史与文化》 结课论文 专业:化学工程与工艺 学号:041114116 姓名:杨乐

Columbo's influence on the American continent Columbo is a famous Spanish navigator, is a pioneer of the great geographical discovery. Columbus young is garden said believers, he so haunting had in Genoa made prison of Marco Polo, determined to be a navigator.1502 he crossed the Atlantic four times in 1492, discovered the American continent, he also became a famous navigator. On August 3, 1492, Columbus by the king of Spain dispatch, with credentials of Indian monarchs and emperors of China, led the three tons of Baishi of sailing, from Spain Palos Yang Fan of the Atlantic, straight towards to the West. After seventy days and nights of hard sailing, in the early morning of October 12, 1492 finally found the land. Columbo thought he had arrived in India. Later know that Columbus landed on a piece of land belonging to the now Balak America than the sea in the Bahamas, he was it named San Salvador. Columbo's discovery has changed the course of world history. It opens up a new era of development and colonization in the new world. At that time, the European population was expanding, with this discovery, the Europeans have settled in two new continents, there will be able to make a difference in the European economy and the resources of the mineral resources and raw materials. This discovery led to the destruction of the American Indian civilization. From a long-term point of view, there have been a number of new countries in the Western

美国历史与文化

浅谈美国历史 ——见证从蚂蚁到大象历程 引导语:中华文化源远流长,五千年的华夏文明留给我们太多的回忆。分久必合、合久必分;从繁荣昌盛到民族衰落;从压迫受辱到当家作主、从璀璨奇葩到复兴中华……可是,跨过大洋的彼岸,初出茅庐的美国却在近两百多年的历史跨度下完成了从蚂蚁到大象的历程,创造出美国独特的发展文化。今天的世界,“汤姆大叔”在全球“维护着世界和平”;好莱坞大片充斥着各大荧屏;NBA回荡在茶前饭后的娱乐中……两百多年来,美国历史一直都是民主制度的试验。早年被提出的问题如今持续被提出并获得解决;强大政府对抗弱小政府、个人权利对抗群体权利、自有资本主义对抗受到管理的商业与劳工以及参与世界对抗孤独主义。美国对于民主制度有很高的期待,而现实又是不如人意。然而国家经过妥协,已见成长与繁荣。在今天的发展过程中笔者认为有必要借鉴美国蚂蚁变大象的历程。 自从哥伦布发现新大陆之后,这片土地上开始了她不平凡的发展。十七世纪初,英国开始向北美殖民。最初的北美移民主要是一些失去土地的农民,生活艰难的工人以及受宗教迫害的清教徒。在殖民地时代,伴随着与北美洲原住民印第安人的长期战争,对当地印第安人的肆意屠杀,严重的劳力缺乏产生了像奴隶和契约奴隶这类的非自由劳力。万恶的黑奴贸易盛行起来。从1607年到1733年,英国殖民者先后在北美洲东岸建立了十三个殖民地。由于英国移民北美是为了追求自由和财富,如被迫害的清教徒和贫农。地方政府享受自治权。殖民地居民有比英人更广泛参与政治的机会和权利,培养了自治的意识和能力,所以他们相信社会契约中,政府是人民需要保护而得人民支持才组成的。在十八世纪中期,殖民地的经济,文化,政治相对成熟,殖民地议会仍信奉英王乔治三世,不过他们追求与英国国会同等的地位,并不想成为英国的次等公民,但是此时英法的七年战争结束,急于巩固领土,使向北美殖民地人民征租重税及英王乔治三世一改放任政策,主张高压手段。因此引发殖民地人民反抗,如“没有代表就不纳税”宣言、“波士顿惨案”、“不可容忍的法案”等。1775年4月在来克星顿和康科特打响“来克星顿的枪声”揭开美国独立战争的前奏。后来,这些殖民地便成为美国北美独立十三州最初的十三个州。 1774年, 来自12州的代表,聚集在费城, 召开所谓第一次大陆会议,希望能寻出一条合理的途径, 与英国和平解决问题,然而英王却坚持殖民地必须无条件臣服于英国国王, 并接受处分。 1775年,在麻州点燃战火, 5月,召开第二次大陆会议, 坚定了战争与独立的决心,并起草有名的独立宣言, 提出充分的理由来打这场仗,这也是最后致胜的要素. 1776年7月4日,宣告了美国的独立,1776年7月4日大陆会议在费城乔治·华盛顿发表了《独立宣言》。《独立宣言》开宗明义地阐明,一切人生而平等,具有追求幸福与自由的天赋权利;淋漓尽致地历数了英国殖民主义者在美洲大陆犯下的罪行;最后庄严宣告美利坚合众国脱离英国而独立。《独立宣言》是具有世界历史意义的伟大文献。完全脱离英国,目的是为‘图生存、求自由、谋幸福’,实现启蒙运动的理想。 1781年, 美军赢得了决定性的胜利, 1783年, 美英签订巴黎条约,结束了独立战争。这也充分展现出

宗教对美国社会的影响

宗教对美国社会的影响 梁子毓英语学院英语137班 摘要:美国是当今世界上最发达的资本主义国家,同时又是发达国家中最具宗教色彩的国家,可以说美国的发展与宗教有着密不可分的关系。宗教从美国建国伊始至今对人们思想的影响是根本性的,这种影响绝不仅仅只是停留在道德方面,在政体的确立、民主制度的促进等方面都有着同样深远的作用。如今,宗教在美国的影响力有增无减,信仰上帝的人越来越多了,宗教在美国国家社会生活中的作用也越来越大。因此,研究宗教对美国发展的影响更有助于我们了解美国的社会现实和文化特征,可以使我们对宗教在美国社会的影响有更深刻的认识。 关键词:宗教美国社会影响 引言 美国文化的一大特征就是其宗教性。来自世界各地不同国家不同种族的移民带来了他们自己的宗教,使美国成为多宗教的国家。各种不同的宗教必然会对美国的社会产生影响,同样也融入了美国的文化。在众多宗教中,影响最大的是新教众多教派中的清教,清教在发展过程中,其影响超越了宗教领域,渐渐渗透到了美国宗教以外的政治、经济、文化领域,使美国政治、经济、文化都带有明显的清教主义特征。清教主义因而成为美国文化和社会形成过程中的重要因素。美国的文明都刻有明显的清教主义印记,清教主义文化也造就了独特的美利坚精神与文化。 一、基督教新教对美国历史的影响 美国虽然只有四百多年的历史,但却从一个英属殖民地逐渐成为世界上的头号强国。在这片土地上诞生了世界上第一部成文宪法,产生了世界上第一个民选总统,建立了三权分立的民主政体,这些都对世界历史产生了深远的影响。然而这些都与基督新教及其伦理道德有着密不可分的关系,这种关系甚至是决定性的。 早期殖民北美的新教1徒的新教信仰,构成了北美早期社会思想风潮的主调,也构成了以后美利坚的民族精神以及国家意识形态的基础。美国独立战争的发生,是因为早期移民北美的多数人都是新教教徒的缘故。由逃亡的清教徒们建立的美洲殖民地,在宗教上,一直与英国本土的宗教处于对立状态。美洲大陆的宗教主流为清教徒,英国的国教则为天主教与新教的混合体圣公会安力甘宗,安力甘宗作为英国国教一直是清教徒改革的对象。在美国独立战争及18世纪二十年代,英国本土和美洲殖民地发生了一场轰轰烈烈的宗教“伟大复兴”运动,这场运动在美国是新教教义的普及和强化运动,是一场彻头彻尾的新教教义在新大陆被强化的运动,这场运动最后导致了新教的进一步振兴,从而在思想上与英国国教彻底脱离了关系。宗教运动进一步促进了北美殖民地人群的主体意识,进一步加强了殖民地与英国本土的在宗教上和政治上的离心力,为独立战争做了思想和意识形态的准备。北美殖民地与英国本土上的宗教对立,对美国独立战争的爆发有着深远的影响。 对美国近代发展奠定基础的南北战争,实际上也有着深刻的宗教原因——美国清教对南方奴隶制的憎恶而引起的南北方对立。废奴主义一直是基督教教义的传统。在新教中这被理解为上帝爱世上的每一个人。这种思想成为西方人权思想的源头。既然上帝爱每一个人,个 1新教:新教(Protestantism)是由16世纪宗教改革运动中脱离罗马天主教的一系列新教派的统称,主流教派有路德宗、加尔文宗和圣公会。

4. 美国的宗教历史-清教徒

Puritans and Harvard 清教徒和哈佛 Puritans played a very important role in Harvard history as well as in American history. 清教徒不仅在美国历史中扮演过重要角色,同样在哈佛大学的历史中也起到了重要作用。 In the 16th and 17th century, a series of religious reforms took place in England Finally, the Church of England was established as the Established Church. People did not have religious freedom. Those people who did not agree with the Church of England were regarded as Separatists and were persecuted. To escape religious persecution, many people fled from England to other countries. Many Puritans chose to go to the New World. 在16和17世纪的英国发生了一系列宗教改革,最终圣公会被定为英国国教。人们没有宗教自由。那些不同意英国国教观点的人被视为宗教分裂者,并遭到迫害。为了逃脱宗教迫害,许多人逃离英国,前往其他国家。很多清教徒选择移民到新大陆。 In 1620, a ship named Mayflower left England. It transported the English Separatists, better known as Pilgrims, from Plymouth, England to Plymouth, Massachusetts, the United States. There were 102 passengers, more than one third of whom were Puritans. During the following decade, the number of Puritans grew. Then in 1628 these Puritans established the Massachusetts Bay Colony. 1620年,一艘名为“五月花号”的船离开了英国,它载着英国宗教分离派——更熟悉的称号是“清教徒前辈移民”,离开了英国的普利茅斯前往美国马萨诸塞州的普利茅斯。船上有102名乘客,其中1/3多的人是清教徒。在接下来的10年中,清教徒队伍逐渐壮大起来,并于1628年建立了马萨诸塞湾殖民地。 Many Puritans had received classic style of higher education in Oxford University and Cambridge University in England. They wanted to pass on to their descendants this kind of education. Besides, these colonists saw colleges as an effective way to disperse religious belief. Therefore, in 1636 the first and oldest institution of higher learning in the United States was established by vote of the Great and General Court of Massachusetts Bay Colony, 140 years earlier than the foundation of the United States. 许多清教徒都曾在英国的牛津和剑桥大学接受过古典式的高等教育,他们想把这种教育传给子孙后代。此外,这些殖民者把大学看做是传播宗教信仰的有效途径。所以,1636年马萨诸塞湾殖民地法院投票通过并建立了美国第一所也是最古老的一所高等教育机构,比美国建国早了140年。 The institution was initially named “New College” or “the College at New Towne"(or “Cambridge College”). The town where the college located was named Cambridge, becau se many Puritans had studied in Cambridge University. The name of the town demonstrated that many people still cherished their memory of life in England. Although they had been persecuted in England, many people still saw England as their motherland. The college followed the style of English universities as well. Then in 1639 the college was renamed Harvard College after a benefactor called John Harvard who was a clergyman. Therefore, in a sense, Harvard University was the product of religious activities.

美国历史与文化第五讲 二战中的美国

第五讲二战中的美国 一、摆脱孤立主义 孤立主义也有不同的意见:自由主义者强烈反对美国卷入任何欧洲战争,包括亨利福特和飞行英雄查尔斯林白。 罗斯福作为政治家的最艰难的功绩,就是使全国确信必须抱起孤立主义,把国家的力量投入反侵略的斗争。 这些对美国意味着什么?美国人只想躲避风雨,20世纪30年代,美国再次孤立主义盛行。 各个方面都主张,美国参加第一次世界大战时错误的,是英国的宣传和银行家与军火商的阴谋欺骗了我们,我们坚持中立国的权利是有勇有谋的。 1,1935年8月31日,国会在这种普遍的情绪中,通过了《中立法》,禁止在未来的冲突中给与交战双方贷款,禁止出售武器给任何一方。由于对侵略者和被侵略者同等对待,新的中立法鼓励了独裁者,使其相信他们可以继续他们征服而不用担心美国的干涉。 2,1937年10月,罗斯福在芝加哥演讲——“隔离演说”:“当某种传染性疾病开始蔓延的时候,为了保护居民的健康,防止疾病的流行,社会许可并且对患者实行隔离。”“战争都会蔓延。战争可以席卷远离原来战场的国家和人民。我们决心置于战争之外,然而我们并不能保证我们不受战争灾难的影响和避免卷入战争的危机。” 但隔离开演讲遭到了猛烈的抨击,但他向美国公众指出战争恐怖的存在。 3,1939年9月,德波战争爆发,第二次世界大战全面爆发,战争的胜负已直接关系到美国的安危,美国舆论和人民转向同情英法,罗斯福政府抓住时机为废除中立法进行宣传,他对国会说,援助英国就是帮助自己,他强调指出废除武器禁运能更好地保护国家不卷入战争,还可以为成千上万的人提供就业机会,这种就业又会自动地帮助建设美国的国防。罗斯福的宣传获得了成功。1939年11月4日,美国国会通过了对交战国解除军火禁运的新中立法,但仍须"现购自运"。尽管从原则上讲,"'现购自运'原则对欧洲所有交战国,也包括德国在内,都是有效的,但由于只有英国和法国才拥有制海权,因而只要他们有美元现金,就能自由自在地运输。"因此,这无疑是对英法作战的巨大支持。同时,也为美国借助英法军事订货,摆脱1937年以来新的经济萧条,加速扩军备战提供了有利条件。 4,1940年5月,1940年5-6月,德国闪击西北欧和法国,打破了欧洲力量的平衡,法兰西败局已定,不列颠前途难卜,被美国视为根本利益的安全线--莱茵河边界已被德国越过。罗斯福要求国会追加国防拨款加强战备。国会批准了陆海军的扩充计划。 5,1940年9月2日,英美两国达成协议:美国用50艘旧驱逐舰交换英国在加勒比海8个岛屿的军事基地,美国租用99年,这是第二次世界大战爆发以后,美国首次向英国进行租借。这项协定意味着中立的结束,标志着美国有限参战的开始。 6,1940年12月9日,丘吉尔致信罗斯福,声称英国国库已经空虚,而军用物资极为短缺,希望美国能给与帮助。12月16日,罗斯福在记者招待会上说了一条美国历史上那个最不平凡的新闻:假如我的邻居失火,…… "保卫美国的最好的直接办法就是英国能够保卫其本身。""历史上还没有一

American Religion 美国宗教发展历程

Religion in America In a Christian world, many countries in the West have experienced declines in religious observance and increases in secularization in the twentieth century. This is often attributed to the influences of industrialization, consumerism, materialism, hedonism, mass culture, and universal education. The United States, however, seems to be an exception. Despite its materialistic image and intense worship of “mighty dollars”, the U. S. still remains the most religious country in the Western countries. In comparison with European countries, America not only has a greater number of religious believers, but also enjoys a much higher church survey, The Economist reported that about 95 percent of Americans believed in God; four out of five believed in miracles, life after death and the Virgin Mary birth; 6.5 percent believed in the devil; 75 percent believed in angels; and nine out of ten owned a bible. Similarly, surveys by the Gallup Organization in the early 1990s indicated that among Americans under 30 years old, about 36 percent attended church on regular basis, while close to 47 percent of the people at or over 50 went to church once in a week. Is America a religious culture, shaped by men who sought freedom of worship, with God constantly present in their minds even when the Church has become formalized? Or is it a secular culture with religion playing only a marginal role in men’s daily lives since the Untied States long time ago separated Church and

美国宗教与文化

美国的宗教与法律 当我们提及美国时,很多人经常想到是什么呢?想到的是美国很富有,美元很重要,汇率降一分牵动多少万人的心,美国的次贷危机间接引发了全球性金融危机;其次是美国科技很发达。阿波罗登月计划的成功鉴证了一个人的一小步可以成为全人类的一大步。微软巨头把电脑这个原本由技术人员才能操作的复杂高科技工具,变为普遍老百姓的日常生活工具,开创了电脑平民化时代。另外还有美国的军事。有一条的手机短信说,“我是美国,我想打谁我打谁”。从扔到日本的广岛长崎的两颗原子弹到伊拉克战争中空地压制武器“哈姆”反辐射导弹,其实力之强毫无疑问。 美国经济、科技和军事实力确实是有目共睹的,但很少有人想到美国的宗教。让我们讲讲宗教在美国的地位。 一方面,根据美国一些比较有名的科研机构,比方说盖洛普这样的权威机构在几十年间的跟踪统计,美国大概每十个人里头,有九个人自称是相信上帝,也就是92%到93%的人是相信上帝的;有八个人认为,宗教们生活中非常重要的、非常关键的内容;有七个人属于某种宗教组织,但不一定是基督教,可能是各种其他宗教:有六个人每天祈祷;有四个人每周去教堂,也就是说有40%的人每周上教堂。这个

数字一百多年来都没有改变,只是在9.11以后,突然上升到70%多,后来有一个回落,但是基本稳定在40%多。美国是所有西方国家里,进教堂人数最高的国家。美国的教堂有30多万个,如果你去美国,会看到一个十字街角会有好几个教堂矗立在那儿,不同的风格,没有一个教堂和另一个教堂是相同的。30多万个教堂分布在美国城乡的各个角落里。另一方面,美国人每年观看橄榄球、棒球、篮球这些重要比赛的人数,大约是4亿人次,但是参加宗教活动的人数,是52亿人次。美国人在20世纪90年代,每年给宗教事业的捐款达500亿美元;但是花在棒球、篮球、橄榄球三大球上的钱是50亿美元。由此可见宗教对美国人的公共生活有多么重要。 除此以外宗教对美国人的政治也有十分深远的影响。“我们信赖上帝”一语不仅铭铸在美国政府发行的硬币上,而且也赫然悬刻在国会大厦的墙壁。联邦国会开会的开场白是牧师的祷告,最高法院也要先诵读“上帝拯救合众国和这个可尊敬的法院”后才可以开庭,联邦军队中也设有随军牧师,即使你去法庭作证,也要像总统宣誓就职一样,必须手按圣经,发誓你所说的一切句句属实,否则你是必输无疑,凡此种种,不一而足。 当然对于美国教育和文化,宗教也有其重要的影响。美国的私立中小学基本上都是教会办的。美国有一些很著名的

美国历史与文化感想完整版

美 国 历 史 与 文 化 之 感 想 姓名:xxx 班级:xxx 学号:xxxxxx

美国历史与文化之感想 当初报选选修课的时候,自己一直在那里纠结该报什么课程。我不想像好多同学那样抱着“既然是选修课那就让自己轻松点”的想法,我是真的想通过选修课来充实自己的文化知识。在浏览了选修课名之后我被‘美国历史与文化’这一课名所吸引,说起自己是完全被课题所吸引而报的这个课是假的,在被课题吸引的同时,我更加希望能在这个课里边学到一些关于美国的一些文化习俗,或者是能对我的英语有所提高!!! 记得初次来到课堂,看到讲课的老师竟然是我在英语角常碰到的老师,当时我就想这门课肯定会给自己增加很多的知识。在英语角的时候就听老师谈过关于美国总统林肯的一些东西,当时我就想老师课堂上讲的应该也有与林肯相关的知识。果不其然,老师一开始讲的就是美国的优秀的领袖——林肯的事迹。对于林肯,老师曾经还举办过一个讲座,专门来讲诉林肯。曾有人这样评价他--不会被困难所吓倒,不会为成功所迷惑的人,他不屈不挠地迈向自己的伟大目标,而从不轻举妄动,他稳步向前,而从不倒退;……总之,他是一位达到了伟大境界而仍然保持自己优良品质的罕有的人物。 美国奴隶制猖獗,1854年南部奴隶主竟派遣一批暴徒拥入堪萨斯州、用武力强制推行奴隶制度,引起了堪萨斯内战。这一事件激起了林肯的斗争热情,他明确地宣布了他要“为争取自由和废除奴隶制而斗争”的政治主张。1860年他当选为总统。南方奴隶主对林肯的政治主张是清楚的,他们当然不愿坐以待毙。1861年,南部7个州的代表脱离联邦,宣布独立,自组“南部联盟”,并于4月12日开始向联邦军队发起攻击,内战爆发初期,联邦军队一再失利。1862年9月22日,林肯宣布了亲自起草的具有伟大历史意义的文献——《解放黑奴宣言》草案(即后来的《解放宣言》),从此战争形势才开始发生了明显的变化,北部军队很快地由防御转入了进攻1865年终于获得了彻底胜利!!!此时,林肯在美国人民中的声望已愈来愈高了,1864年,林肯再度当选为总统。但在内战结束之后,林肯在剧院不幸被人暗杀致死! 虽然未曾接受高深的教育,林肯具有卓越的口才和文采,他是美国历史上最伟大的一位总统。人们怀念他的正直、仁慈和坚强的个性,他一直是美国历史上最受人景仰的总统之一。他拥有敏锐的洞察力和卓越的口才,这也是他在竞选总统和去的内战的胜利的原因之一。林肯一生致力于美国的独立事业,不惜为解放奴隶而牺牲自己,在他去世后,他的遗体在14个城市供群众凭吊了两个多星期,后被安葬在斯普林菲尔德(Springfield)。 在上完有关林肯的课后,让我懂得了林肯一生是多么的有价值,尽管有许多的牺牲,有许多的流血,但是他推翻了美国的奴隶制,解放了无数的生活在黑暗中的奴隶。 上了美国历史与文化这一课程,因为课程的紧张没有讲解许多的美国文化,但偶尔老师也会放一些美国经典的歌曲和电影片段给我们,让我们从哪些方面来了解美国的文化。因为老师上课都是放听力让我们来了解课程的,在上课的同时,我们还可以联系自己的听力和单词的拼写能力。每次上课老师都会点名回答问题,这样也真正的让一些想通过选修课来增长自己的英语知识的人能够学到一些很有用的东西,比如说听力和拼写能力。能够选到这门课程让我自己感到很高兴,在了解美国名人事件的同时又让我的英语听力的到了练习,谢谢老师的教学!

第一章 美国历史和政治传统

第一章美国历史和政治传统 很多人认为美国是欧洲的延续,是欧洲人种、语言、文化、历史、传统的延续,是欧洲政治思想和制度的延续与实践。大多数美国人确实来自欧洲,美国的人种、思想、文化、语言、行为方式,政治思想、理论、制度、历史经验和教训基本源于欧洲。但美国不是欧洲的翻版,美国有自己的地理、历史、文化、思想和制度,美国创造了与当时欧洲国家根本不同的国家和制度,政治思想、理论、价值、观念、制度中具有很多欧洲当时和后来所不具有的东西。美国有自己的政治思想、理论、制度和实践,具有同欧洲相似又根本不同的历史和政治传统。第一节美利坚民族的形成及历史和文化传统 美国诞生不到两个世纪,就成为世界领先的经济、军事、科技和文化强国,这其中有地理条件、继承英国等欧洲国家近代文明等原因,更重要的是美国人民的努力创造了今天的辉煌。“美国人民的光荣在于他们的理想主义、他们对自己的资源的利用,以及他们在骄傲和贪婪的诱惑面前所显示的品格。" Harold Evans. The American Century. U.S. News & World Report, 1998-10- 05. 美国的形成有其独特性,构成美国较为鲜明的特征。研究美国的著名法国学者托克维尔写道:“在美国,任何一种见解,任何一种习惯,任何一项法律,而且我敢说任何一个事件,都不难从这个国家的起源当中找到解释。" \ 托克维尔:《论美国的民主》,董果良译,32页,北京,商务印书馆,1991. 他认为,有助于维护美国民主制度的原因有三:自然环境、法制和民情。“自然环境不如法制,而法制又不如民情。”他认为美国民主的民情扎根于历史上形成的新英格兰乡镇自治制度。这个早在17世纪开始形成,后经基督教新教的地方教会自治思想培养壮大起来的制度,促进了美国独立运动的发展,提高了人民积极参加公共事务的觉悟,并为后来被联邦宪法肯定下来的中央和地方分权的制度奠定了基础。托克维尔把乡镇自治的传统看成是人民主权和美国人在实践中确立的公民自由原则的根源。\ 托克维尔:《论美国的民主》,董果良译,VI页,北京,商务印书馆,1991. 托克维尔写道: "17世纪初在美洲定居下来的移民,从他们在欧洲旧社会所反对的一切原则中析出民主原则,独自把它移植到新大陆的海岸上。在这里,民主原则得到自由成长,并在同民情的一并前进中和平地发展成为法律。" \ 托克维尔:《论美国的民主》,董果良译,15页,北京,商务印书馆,1991. 一位英国史学家眼中的美国较有代表性:美国人的行为方式和思想方式经常令其他国家的人感到困惑。《美国人民的历史》 (A History of the American People)一书作者、英国历史学家保罗·约翰逊(Paul Johnson)就这些方式及其产生的根源做了介绍,他提出关于认识美国的“十诫”,供大家了解美国时遵循。第一诫:美国是一个地域广袤、地貌多样的大国。在内燃机尤其是飞机问世之前,美国不得不与空间距离进行艰苦卓绝的斗争。首批开拓者不知道内陆有什么、居住着什么人,之后的200年间,仍然没有人知道。因此第一诫是记住美国是那么辽阔、那么多姿多彩的国家。学习美国历史时,身边就必须有一本翻开的大比例尺美国地图册。 第二诫:记住,美国目前是,而且始终是一个宗教国家。它的诞生主要应归因于宗教。直到17世纪末,宗教信仰和宗教冲突毫无疑问一直占据主导地位。甚至在这之后,“大觉醒”(great awakenings) 是美国变化的决定因素。起始于18世纪20年代的第一次大觉醒是导致美利坚合众国诞生的美国革命的精神动力和情感动力。19世纪初的第二次大觉醒成为不可避免的南北战争的导火索。 《宪法》(Constitution)和《权利法案》(Bill of Rights)将美国从强烈的宗教信仰带来的政治影响中解脱出来。从这个意义上说,美国又是一个非宗教的国家。但是,在其他任何方面,美国绝对是一个敬畏上帝的国家。美国是世界上唯一一个绝大多数国民仍旧自愿地过着积极的宗教生活的大国。这也是美国与众不同的主要原因。

美国历史与文化参考资料

第一讲从殖民地到“超级大国”:美国的发展历程 美国的奠基时代(1607-1775年) 1.北美13个殖民地 1607到1733年,英国在北美陆续建立了13个殖民地 殖民地的社会结构: 上层是以总督为首的大商人、大土地所有者或种植园奴隶主; 中间是小土地所有者、小工厂主、技师和自耕农等; 底层是白人契约奴和黑人奴隶 美利坚民族的形成 殖民地间经济的差异性促进了彼此商品的流通,逐渐形成了统一的市场。 经济的交往也促进了文化的交流,如哈佛学院。 最早具有民族意识的知识分子,如托马斯?杰斐逊、詹姆斯?麦迪逊、亚历山大?汉密尔顿等;英语逐渐成为他们的共同语言; 共同的心理素质。 3.英国与北美殖民地矛盾的激化 ¨1773年茶叶税法,波士顿倾茶事件 ¨ 1774年3月始,英国政府接连颁布了5项高压法令(“不可容忍的法令Intolerable Acts”)。 二、美国的建立与初步繁荣(1776-1860年) 1.美国独立革命(1775-1783年)The War of American Independence “列克星敦枪声Lexington”和独立战争的爆发(1775年4月19日) Someone fire d the “shot heard round the world” 1775年5月10日召开的第二届大陆会议开始招募军队,华盛顿被任命为统一的美利坚军队的总司令 托马斯?潘恩于1776年1月10日发表《常识》Common Sense 《独立宣言》与美国的建立1776年7月4日 《独立宣言》Declaration of Independence July 4, 1776 天赋人权和社会契约论social contract,认为人人生而平等,享有不可剥夺的“生命权、自由权和追求幸福的权利” “We hold these truths to be self-evident, that all men are created equal, that they are endowed by their Creator with certain unalienable rights, that among these are life, liberty, and the pursuit of happiness.” 美国革命通过《独立宣言》以及由此导致的一系列立法,创造了一个全新的国家 从邦联到联邦 邦联制下的美国《邦联条例》Article of Confederation (1777) 谢斯起义Rebel of Shays

美国宗教及影响

美国宗教及其影响 当我们提到美国时,许多人往往想到的是美国经济上的富有,科技、教育上的发达,军事上的强大,也有人可能会想到美国的政治、法律制度以及困扰美国的各种社会问题。的确,这些都是反映美国社会的重要方面,但这还不足以勾画一个完整的美国。要说明美国,还有一个不可缺少的方面,这就是美国的宗教。 可以说,美国人生活的各个方面都和宗教有着密切的关系,正如美国著名神学家尼布尔说的那样,美国“是世界上最世俗的国家,也是宗教性最强的国家。” 美国不是一个以神权为中心的宗教国家,但离开了宗教,很难想象会有今天的美国。宗教对于美国,不是可有可无,而是须臾不可或缺。 一、美国宗教的影响 事实上,如果稍微对美国社会和美国人的生活做一点观察,就会发现宗教深深地植根于美国社会之中并对美国社会生活的各个方面有着极为广泛的影响。 美国的钞票上,赫然印着“我们信仰上帝”。美国的国歌里,有“上帝保佑美国”的歌词。美国总统就职,要手按圣经进行宣誓。国会参众两院的每一届会议都是以国会牧师主持的祈祷开始。美国的军队里有牧师、神甫等各种不同宗教的随军神职人员,身穿军官制服,在军中提供宗教服务。美国的大学校园里,活动着大量的学生宗教团体。美国的医院、监狱、机场及其他许多公共与民间机构中也都有专职或兼职的宗教职业人员提供宗教服务。 美国85%以上私立中小学校的学生就读于教会学校。哈佛大学、耶鲁大学、普林斯顿大学等许多著名的美国大学最初都是由教会创办的。今天,美国有1200多家宗教广播电台播放宗教节目,每12家电视台中就有一家是宗教电视台,在20世纪的最后10年里,美国的宗教节目增加了75%。美国的宗教报刊杂志有5000多种,《新约圣经》在美国的印数超过了1亿册,宗教音乐的音像制品销售量远远超过了爵士乐、古典音乐及其他各种流行音乐。一半以上的美国成年人参加过宗教组织的慈善服务活动或做过志愿者,在纽约、芝加哥、洛杉矶、费城等大城市中提供社区服务的主要力量是宗教团体。大多数美国人的婚礼是在教堂举行的,而他们的丧礼要由牧师、神甫主持。 宗教在美国有着如此广泛的影响。 二、美国的主要宗教类型 美国是个多民族、多宗教的国家。居民主要信奉基督教和天主教,犹太教、佛教、伊斯兰教等宗教亦有一定信众,信仰宗教的公民在总人口中约占91%。

美国文化历史演变

美国文化历史演变 早期的移民把欧洲文化带到美国。很快地,这些文化遍及美国各地。时至今日,美国已经成为世界文化的主流之一。许多的美国艺术家们对于发展新的风格,新的自我表现方式,甚至新的文化型式都作出巨大的贡献。 文学 美国的文学被翻译为世界各种语言。早在美国历史初期,美国就产生了许多杰出的作家。最早的古柏JamesFenimoreCooper,霍桑NathanielHawthorne,以及欧文WashingtonLrving描写出一个年轻并且不断成长的美国。后来,梅尔维尔HermanMelville写出有关海洋的小说以探讨道德问题。马克。叶法国则表现了密西西比河上生活的情趣与幽默。7个美国人曾经获得诺贝尔文学奖:剧作家尤金。欧尼尔与小说家索尔拜娄SaulBellow、赛珍珠、福克纳、海明威、辛克莱。路易士SinclairLewis以及史坦贝克。 绘画 1913年,一个现代艺术的展览会在纽约市的兵工大厦举行。这个兵工大厦画展大大改革了美国艺术,一群写实主义的画家们开了这个画展,以抗议那些拒绝陈列他们作品的保守画廊。这个展览同时也展出欧洲艺术家的作品,这些抽象作品极具吸引力,群众们觉得现代的欧洲艺术是可以被了解的、或者说是够吓人的;而美国艺术家们则发现它们很富刺激。于是许多美国画家开始采取欧洲的风格,三十年代的一些画家,如葛兰特伍德GrantWood等,都是只描绘他们自己家乡

的地域主义者。然而到了四十年代,许多画家哈佩EdwardHopper之类,都属写实主义派画家。抽象艺术则是在五十年代才确立它的地位。今天,美国的画家们仍然不断地尝试新风格。举例而言,普普艺术(意即风行普遍,或者是因为看来像海报,或是利用喜剧式的线条),首次出现在五十年代。而在六十年代,引人注意的则是那些视觉幻像的图画——欧普艺术。 建筑 在美国的早期历史中,建筑一直是呈现美国风貌的。19世纪末,苏利文LouisSullivan设计出摩天大楼。而怀特FrankLloydWright 的想像力也影响世界各地的建筑师。今天,更多的美国建筑师利用钢筋、玻璃以及混凝土设计出宏伟、出色的建筑物。 雕塑 美国的雕塑在20世纪前,一直受到欧洲风格的影响。圣高登AugustusSaint-Gaudens是19世纪能在其作品表现出真正想像力的美国雕塑家。今日,美国雕塑完全是一种个人的表现,它以各种不同的形式出现,而且材料包罗万象。举例来说,德维森JoDavidson是以他为名人雕塑栩栩如生的青铜像而出名。史坦其卫兹RichardStankiewicz则是以金属原料焊接成的雕塑创造出惊人的抽象作品。 舞蹈 舞蹈在美国反映出美国人不断求变的欲望,以及他们对新奇事物的喜好。一些如萧恩和丹尼斯等著名的舞者,并且不断地在传统舞蹈

美国的宗教文化与政治

2008年6月号 制还是激励机制的发挥,归根到底还要体现在教师身上。所以,教师本身的自觉自律才是师德建设的关键。文明难以移植,信仰不能强制,道德需要自律。高校应该创造多方面条件,努力加强教师的自我修养,最终达到教师的慎独境界。比如高校应该努力发掘和形成独具特色的校园德治环境,并形成良好的校风、教风,使教师在其中受到陶冶,从而实现自我教育、自觉自律。如果广大教师对先进道德文化和环境普遍认同,那么自身思想道德素质就能普遍提高,就能够自觉地用普遍认同的人格追求和道德标准来衡量和规范自己的行为举止。 教师应该有针对性地学习,掌握师德的基础理论,坚定我们的立场和行为方向;在善恶的观念冲突中,要自己跟自 己“打官司”,在自己的头脑中进行正确 的选择。教师加强自我修养,还应该树 立强烈的动机,积极主动、持之以恒的进 行师德修养,不断的反思和评价自己,发 扬成绩,克服不良的行为习惯,从而不断 的提高自我修养的境界。 总之,高校教师的思想道德素质,直 接影响到大学生的成才,进而影响到国 家的兴旺发达,高校必须进一步加强和 改进师德建设,以邓小平理论、“三个代 表”重要思想和科学发展观为指导,紧紧 围绕全面实施素质教育、全面加强青少 年思想道德建设和思想政治教育的目标 要求,以热爱学生、教书育人为核心,以 “学为人师、行为世范”为准则,以提高教 师思想政治素质、职业理想和职业道德 水平为重点,弘扬高尚师德,力行师德规 范,强化师德教育,优化制度环境,不断 提高师德水平,造就忠诚于人民教育事 业、为人民服务、让人民满意的教师队 伍,为培养德智体美全面发展的社会主 义建设者和接班人做出新贡献。 参考文献 [1]教育部.关于进一步加强和改进 师德建设的意见[N].人民日报, 2005-10-25. [2]罗国杰.伦理学教程[M].北京:中 国人民大学出版社,2001. [3]教育部人事司组.高等学校教师 职业道德修养[M].北京:北京师范大学出 版社,2000. [4]师远.浅谈新时期高校师德建设 [J].思想教育研究,2007(10). 美国的宗教文化与政治 云南民族大学哲学学院王滨 [摘要]宗教是美国民族的精神源泉。宗教是美国文化的中枢,美国文化可以称之为一种宗教文化。宗教作为一种社会文化形式在美国历史和现实中都具有不可忽略的作用和价值。 [关键词]宗教文化美国政治 一、宗教是美国民族的精神源泉 世俗化了的宗教是美国国家政治的基本依托。美国宗教与政治的关系过去是、现在是、将来一定还是密不可分的;研究和探索美国宗教与政治及其相互关系是我们了解美国文化和美国文明的基础,理解宗教因素在美国社会政治生活中的作用是我们认识美国社会的关键所在。所以,宗教与政治之间的关系作为对美国方方面面多具有深刻的影响,下面让我们深入探讨美国宗教、政治等一系列问题。同时,有助于对我们深入、全面地了解美国及其社会具有重大意义。 首先,我们先从文化价值角度入手。以基督新教、天主教、犹太教三位一体的基督宗教为主流的文化体制,没有宗教信仰的多元化就失去了美国文化赖以存在的根基。在美国人判断人格的价值尺度中,宗教因素具有重要地位。 宗教是美国文化的中枢,美国文化可以称之为一种宗教文化。这主要表现在两个方面:一是宗教在美国政治文化形成过程中具有重要价值;二是宗教持续不断地服务于美国自由主义及其政治和社会秩序的合法化。我们既可以用教会宗教在各个时期的“奋兴”作用来归纳美国文化发展的基本脉络,也可以用公民宗教在美国民族统一道德观和价值观上的作用来表述美国文化存在的基本特征。再次,宗教信仰多元化是美国社会存在和发展的基本依托。 其次,从文化层面过度到政治层面上,这样使我们就不难深入了解宗教对美国政治的影响了。美国是一个从宗教历史走来的世俗国家,这个过程本身就是一个宗教世俗化的过程。在这个过程中,宗教在社会政治生活中的作用非但没有减少,反而更加重要。从总统、公民、国会、利益集团、法院和外交等多方面和领域都有鲜明的体现;在美国社会政治生活中,宗教的作用非同小可;从道德、种族、到利益、权利、竞选等诸领域;信仰的价值不可小视,宗教与政治相结合、相配合、相融合、相吻合,是美国政治的突出特征,宗教作为国家意识形态发挥过和正在发挥着不可或缺的作用。在国内政治生活中,不论是公民、总统、国会,还是法院,都与宗教息息相关。 从根本上讲,不论是宗教的多样性决定了政治的多元化还是政治的多样性决定了宗教的多元化,都证明是美国宗教与政治相辅相成的关系塑铸了美国民族和国家。美国作为一个历史上形成的多种族、多宗教的移民国家,没有宗教信仰的多元化就难以体现宗教信仰自由的原则;美国历史上的政教关系有一个变化的过程,在这个变化过程中,美国走过了一条从政教合一到政教分离的路径。因此,它体现了美国移民始祖神权政体的兴衰过程,阐述了领导美国独立的政治精英们如何走上民主政治的的道路,介绍了立国之父政教分离制度的抉择,阐发了政治神学家在内战后南方重建中注重“道德重建”的成功。 学术论坛 探索 ?147?

相关主题
相关文档 最新文档