Yet Another Blog in Statistical Computing

I can calculate the motion of heavenly bodies but not the madness of people. -Isaac Newton

Archive for the ‘Machine Learning’ Category

DART: Dropout Regularization in Boosting Ensembles

The dropout approach developed by Hinton has been widely employed in deep learnings to prevent the deep neural network from overfitting, as shown in https://statcompute.wordpress.com/2017/01/02/dropout-regularization-in-deep-neural-networks.

In the paper http://proceedings.mlr.press/v38/korlakaivinayak15.pdf, the dropout can also be used to address the overfitting in boosting tree ensembles, e.g. MART, caused by the so-called “over-specialization”. In particular, while first few trees added at the beginning of ensembles would dominate the model performance, the rest added later can only improve the prediction for a small subset, which increases the risk of overfitting. The idea of DART is to build an ensemble by randomly dropping boosting tree members. The percentage of dropouts can determine the degree of regularization for boosting tree ensembles.

Below is a demonstration showing the implementation of DART with the R xgboost package. First of all, after importing the data, we divided it into two pieces, one for training and the other for testing.

pkgs <- c('pROC', 'xgboost')
lapply(pkgs, require, character.only = T)
df1 <- read.csv("Downloads/credit_count.txt")
df2 <- df1[df1$CARDHLDR == 1, ]
set.seed(2017)
n <- nrow(df2)
sample <- sample(seq(n), size = n / 2, replace = FALSE)
train <- df2[sample, -1]
test <- df2[-sample, -1]

For the comparison purpose, we first developed a boosting tree ensemble without dropouts, as shown below. For the simplicity, all parameters were chosen heuristically. The max_depth is set to 3 due to the fact that the boosting tends to work well with so-called “weak” learners, e.g. simple trees. While ROC for the training set can be as high as 0.95, ROC for the testing set is only 0.60 in our case, implying the overfitting issue.

mart.parm <- list(booster = "gbtree", nthread = 4, eta = 0.1, max_depth = 3, subsample = 1, eval_metric = "auc")
mart <- xgboost(data = as.matrix(train[, -1]), label = train[, 1], params = mart.parm, nrounds = 500, verbose = 0, seed = 2017)
pred1 <- predict(mart, as.matrix(train[, -1]))
pred2 <- predict(mart, as.matrix(test[, -1]))
roc(as.factor(train$DEFAULT), pred1)
# Area under the curve: 0.9459
roc(as.factor(test$DEFAULT), pred2)
# Area under the curve: 0.6046

With the same set of parameters, we refitted the ensemble with dropouts, e.g. DART. As shown below, by dropping 10% tree members, ROC for the testing set can increase from 0.60 to 0.65. In addition, the performance disparity between training and testing sets with DART decreases significantly.

dart.parm <- list(booster = "dart", rate_drop = 0.1, nthread = 4, eta = 0.1, max_depth = 3, subsample = 1, eval_metric = "auc")
dart <- xgboost(data = as.matrix(train[, -1]), label = train[, 1], params = dart.parm, nrounds = 500, verbose = 0, seed = 2017)
pred1 <- predict(dart, as.matrix(train[, -1]))
pred2 <- predict(dart, as.matrix(test[, -1]))
roc(as.factor(train$DEFAULT), pred1)
# Area under the curve: 0.7734
roc(as.factor(test$DEFAULT), pred2)
# Area under the curve: 0.6517

Besides rate_drop = 0.1, a wide range of dropout rates have also been tested. In most cases, DART outperforms its counterpart without the dropout regularization.

Written by statcompute

August 20, 2017 at 5:50 pm

GLM with H2O in R

Below is an example showing how to fit a Generalized Linear Model with H2O in R. The output is much more comprehensive than the one generated by the generic R glm().

> library(h2o)

> h2o.init(max_mem_size = "12g")

> df1 <- h2o.uploadFile("Documents/credit_count.txt", header = TRUE, sep = ",", parse_type = "CSV")

> df2 <- h2o.assign(df1[df1$CARDHLDR == 1, ], "glm_df")

> h2o.colnames(df2)
 [1] "CARDHLDR" "DEFAULT"  "AGE"      "ACADMOS"  "ADEPCNT"  "MAJORDRG"
 [7] "MINORDRG" "OWNRENT"  "INCOME"   "SELFEMPL" "INCPER"   "EXP_INC"
[13] "SPENDING" "LOGSPEND"

> Y <- "DEFAULT"

> X <- c("MAJORDRG", "MINORDRG", "INCOME", "OWNRENT")

> dist <- "binomial"

> link <- "logit"

> id <- "h2o_mdl01"

> mdl <- h2o.glm(X, Y, training_frame = h2o.getFrame("glm_df"), model_id = id, family = dist, link = link, lambda = 0, compute_p_values = TRUE, standardize = FALSE)

> show(h2o.getModel(id)@model$coefficients_table)
Coefficients: glm coefficients
      names coefficients std_error    z_value  p_value
1 Intercept    -1.204439  0.090811 -13.263121 0.000000
2  MAJORDRG     0.203135  0.069250   2.933370 0.003353
3  MINORDRG     0.202727  0.047971   4.226014 0.000024
4   OWNRENT    -0.201223  0.071619  -2.809636 0.004960
5    INCOME    -0.000442  0.000040 -10.942350 0.000000

> h2o.performance(h2o.getModel(id))
H2OBinomialMetrics: glm
** Reported on training data. **

MSE:  0.08414496
RMSE:  0.2900775
LogLoss:  0.3036585
Mean Per-Class Error:  0.410972
AUC:  0.6432189
Gini:  0.2864378
R^2:  0.02005004
Residual Deviance:  6376.221
AIC:  6386.221

Confusion Matrix (vertical: actual; across: predicted) for F1-optimal threshold:
          0    1    Error         Rate
0      7703 1800 0.189414   =1800/9503
1       630  366 0.632530     =630/996
Totals 8333 2166 0.231451  =2430/10499

Maximum Metrics: Maximum metrics at their respective thresholds
                        metric threshold    value idx
1                       max f1  0.126755 0.231499 142
2                       max f2  0.075073 0.376556 272
3                 max f0point5  0.138125 0.191828 115
4                 max accuracy  0.368431 0.905039   0
5                max precision  0.314224 0.250000   3
6                   max recall  0.006115 1.000000 399
7              max specificity  0.368431 0.999895   0
8             max absolute_mcc  0.126755 0.128940 142
9   max min_per_class_accuracy  0.106204 0.604546 196
10 max mean_per_class_accuracy  0.103730 0.605663 202

Written by statcompute

June 28, 2017 at 12:25 am

Finer Monotonic Binning Based on Isotonic Regression

In my early post (https://statcompute.wordpress.com/2017/01/22/monotonic-binning-with-smbinning-package/), I wrote a monobin() function based on the smbinning package by Herman Jopia to improve the monotonic binning algorithm. The function works well and provides robust binning outcomes. However, there are a couple potential drawbacks due to the coarse binning. First of all, the derived Information Value for each binned variable might tend to be low. Secondly, the binned variable might not be granular enough to reflect the data nature.

In light of the aforementioned, I drafted an improved function isobin() based on the isotonic regression (https://en.wikipedia.org/wiki/Isotonic_regression), as shown below.

isobin <- function(data, y, x) {
  d1 <- data[c(y, x)]
  d2 <- d1[!is.na(d1[x]), ]
  c <- cor(d2[, 2], d2[, 1], method = "spearman", use = "complete.obs")
  reg <- isoreg(d2[, 2], c / abs(c) * d2[, 1])
  k <- knots(as.stepfun(reg))
  sm1 <-smbinning.custom(d1, y, x, k)
  c1 <- subset(sm1$ivtable, subset = CntGood * CntBad > 0, select = Cutpoint)
  c2 <- suppressWarnings(as.numeric(unlist(strsplit(c1$Cutpoint, " "))))
  c3 <- c2[!is.na(c2)]
  return(smbinning.custom(d1, y, x, c3[-length(c3)]))
}

Compared with the legacy monobin(), the isobin() function is able to significantly increase the binning granularity as well as moderately improve the Information Value.

LTV Binning with isobin() Function

   Cutpoint CntRec CntGood CntBad CntCumRec CntCumGood CntCumBad PctRec GoodRate BadRate    Odds LnOdds     WoE     IV
1     <= 46     81      78      3        81         78         3 0.0139   0.9630  0.0370 26.0000 3.2581  1.9021 0.0272
2     <= 71    312     284     28       393        362        31 0.0535   0.9103  0.0897 10.1429 2.3168  0.9608 0.0363
3     <= 72     22      20      2       415        382        33 0.0038   0.9091  0.0909 10.0000 2.3026  0.9466 0.0025
4     <= 73     27      24      3       442        406        36 0.0046   0.8889  0.1111  8.0000 2.0794  0.7235 0.0019
5     <= 81    303     268     35       745        674        71 0.0519   0.8845  0.1155  7.6571 2.0356  0.6797 0.0194
6     <= 83    139     122     17       884        796        88 0.0238   0.8777  0.1223  7.1765 1.9708  0.6149 0.0074
7     <= 90    631     546     85      1515       1342       173 0.1081   0.8653  0.1347  6.4235 1.8600  0.5040 0.0235
8     <= 94    529     440     89      2044       1782       262 0.0906   0.8318  0.1682  4.9438 1.5981  0.2422 0.0049
9     <= 95    145     119     26      2189       1901       288 0.0248   0.8207  0.1793  4.5769 1.5210  0.1651 0.0006
10   <= 100    907     709    198      3096       2610       486 0.1554   0.7817  0.2183  3.5808 1.2756 -0.0804 0.0010
11   <= 101    195     151     44      3291       2761       530 0.0334   0.7744  0.2256  3.4318 1.2331 -0.1229 0.0005
12   <= 110   1217     934    283      4508       3695       813 0.2085   0.7675  0.2325  3.3004 1.1940 -0.1619 0.0057
13   <= 112    208     158     50      4716       3853       863 0.0356   0.7596  0.2404  3.1600 1.1506 -0.2054 0.0016
14   <= 115    253     183     70      4969       4036       933 0.0433   0.7233  0.2767  2.6143 0.9610 -0.3950 0.0075
15   <= 136    774     548    226      5743       4584      1159 0.1326   0.7080  0.2920  2.4248 0.8857 -0.4702 0.0333
16   <= 138     27      18      9      5770       4602      1168 0.0046   0.6667  0.3333  2.0000 0.6931 -0.6628 0.0024
17    > 138     66      39     27      5836       4641      1195 0.0113   0.5909  0.4091  1.4444 0.3677 -0.9882 0.0140
18  Missing      1       0      1      5837       4641      1196 0.0002   0.0000  1.0000  0.0000   -Inf    -Inf    Inf
19    Total   5837    4641   1196        NA         NA        NA 1.0000   0.7951  0.2049  3.8804 1.3559  0.0000 0.1897

LTV Binning with monobin() Function

  Cutpoint CntRec CntGood CntBad CntCumRec CntCumGood CntCumBad PctRec GoodRate BadRate   Odds LnOdds     WoE     IV
1    <= 85   1025     916    109      1025        916       109 0.1756   0.8937  0.1063 8.4037 2.1287  0.7727 0.0821
2    <= 94   1019     866    153      2044       1782       262 0.1746   0.8499  0.1501 5.6601 1.7334  0.3775 0.0221
3   <= 100   1052     828    224      3096       2610       486 0.1802   0.7871  0.2129 3.6964 1.3074 -0.0486 0.0004
4   <= 105    808     618    190      3904       3228       676 0.1384   0.7649  0.2351 3.2526 1.1795 -0.1765 0.0045
5   <= 114    985     748    237      4889       3976       913 0.1688   0.7594  0.2406 3.1561 1.1493 -0.2066 0.0076
6    > 114    947     665    282      5836       4641      1195 0.1622   0.7022  0.2978 2.3582 0.8579 -0.4981 0.0461
7  Missing      1       0      1      5837       4641      1196 0.0002   0.0000  1.0000 0.0000   -Inf    -Inf    Inf
8    Total   5837    4641   1196        NA         NA        NA 1.0000   0.7951  0.2049 3.8804 1.3559  0.0000 0.1628

Bureau_Score Binning with isobin() Function

   Cutpoint CntRec CntGood CntBad CntCumRec CntCumGood CntCumBad PctRec GoodRate BadRate    Odds  LnOdds     WoE     IV
1    <= 491      4       1      3         4          1         3 0.0007   0.2500  0.7500  0.3333 -1.0986 -2.4546 0.0056
2    <= 532     24       9     15        28         10        18 0.0041   0.3750  0.6250  0.6000 -0.5108 -1.8668 0.0198
3    <= 559     51      24     27        79         34        45 0.0087   0.4706  0.5294  0.8889 -0.1178 -1.4737 0.0256
4    <= 560      2       1      1        81         35        46 0.0003   0.5000  0.5000  1.0000  0.0000 -1.3559 0.0008
5    <= 572     34      17     17       115         52        63 0.0058   0.5000  0.5000  1.0000  0.0000 -1.3559 0.0143
6    <= 602    153      84     69       268        136       132 0.0262   0.5490  0.4510  1.2174  0.1967 -1.1592 0.0459
7    <= 605     56      31     25       324        167       157 0.0096   0.5536  0.4464  1.2400  0.2151 -1.1408 0.0162
8    <= 606     14       8      6       338        175       163 0.0024   0.5714  0.4286  1.3333  0.2877 -1.0683 0.0035
9    <= 607     17      10      7       355        185       170 0.0029   0.5882  0.4118  1.4286  0.3567 -0.9993 0.0037
10   <= 632    437     261    176       792        446       346 0.0749   0.5973  0.4027  1.4830  0.3940 -0.9619 0.0875
11   <= 639    150      95     55       942        541       401 0.0257   0.6333  0.3667  1.7273  0.5465 -0.8094 0.0207
12   <= 653    451     300    151      1393        841       552 0.0773   0.6652  0.3348  1.9868  0.6865 -0.6694 0.0412
13   <= 662    295     213     82      1688       1054       634 0.0505   0.7220  0.2780  2.5976  0.9546 -0.4014 0.0091
14   <= 665    100      77     23      1788       1131       657 0.0171   0.7700  0.2300  3.3478  1.2083 -0.1476 0.0004
15   <= 667     57      44     13      1845       1175       670 0.0098   0.7719  0.2281  3.3846  1.2192 -0.1367 0.0002
16   <= 677    381     300     81      2226       1475       751 0.0653   0.7874  0.2126  3.7037  1.3093 -0.0466 0.0001
17   <= 679     66      53     13      2292       1528       764 0.0113   0.8030  0.1970  4.0769  1.4053  0.0494 0.0000
18   <= 683    160     129     31      2452       1657       795 0.0274   0.8062  0.1938  4.1613  1.4258  0.0699 0.0001
19   <= 689    203     164     39      2655       1821       834 0.0348   0.8079  0.1921  4.2051  1.4363  0.0804 0.0002
20   <= 699    304     249     55      2959       2070       889 0.0521   0.8191  0.1809  4.5273  1.5101  0.1542 0.0012
21   <= 707    312     268     44      3271       2338       933 0.0535   0.8590  0.1410  6.0909  1.8068  0.4509 0.0094
22   <= 717    368     318     50      3639       2656       983 0.0630   0.8641  0.1359  6.3600  1.8500  0.4941 0.0132
23   <= 721    134     119     15      3773       2775       998 0.0230   0.8881  0.1119  7.9333  2.0711  0.7151 0.0094
24   <= 723     49      44      5      3822       2819      1003 0.0084   0.8980  0.1020  8.8000  2.1748  0.8188 0.0043
25   <= 739    425     394     31      4247       3213      1034 0.0728   0.9271  0.0729 12.7097  2.5424  1.1864 0.0700
26   <= 746    166     154     12      4413       3367      1046 0.0284   0.9277  0.0723 12.8333  2.5520  1.1961 0.0277
27   <= 756    234     218     16      4647       3585      1062 0.0401   0.9316  0.0684 13.6250  2.6119  1.2560 0.0422
28   <= 761    110     104      6      4757       3689      1068 0.0188   0.9455  0.0545 17.3333  2.8526  1.4967 0.0260
29   <= 763     46      44      2      4803       3733      1070 0.0079   0.9565  0.0435 22.0000  3.0910  1.7351 0.0135
30   <= 767     96      92      4      4899       3825      1074 0.0164   0.9583  0.0417 23.0000  3.1355  1.7795 0.0293
31   <= 772     77      74      3      4976       3899      1077 0.0132   0.9610  0.0390 24.6667  3.2055  1.8495 0.0249
32   <= 787    269     260      9      5245       4159      1086 0.0461   0.9665  0.0335 28.8889  3.3635  2.0075 0.0974
33   <= 794     95      93      2      5340       4252      1088 0.0163   0.9789  0.0211 46.5000  3.8395  2.4835 0.0456
34    > 794    182     179      3      5522       4431      1091 0.0312   0.9835  0.0165 59.6667  4.0888  2.7328 0.0985
35  Missing    315     210    105      5837       4641      1196 0.0540   0.6667  0.3333  2.0000  0.6931 -0.6628 0.0282
36    Total   5837    4641   1196        NA         NA        NA 1.0000   0.7951  0.2049  3.8804  1.3559  0.0000 0.8357

Bureau_Score Binning with monobin() Function

   Cutpoint CntRec CntGood CntBad CntCumRec CntCumGood CntCumBad PctRec GoodRate BadRate    Odds LnOdds     WoE     IV
1    <= 617    513     284    229       513        284       229 0.0879   0.5536  0.4464  1.2402 0.2153 -1.1407 0.1486
2    <= 642    515     317    198      1028        601       427 0.0882   0.6155  0.3845  1.6010 0.4706 -0.8853 0.0861
3    <= 657    512     349    163      1540        950       590 0.0877   0.6816  0.3184  2.1411 0.7613 -0.5946 0.0363
4    <= 672    487     371    116      2027       1321       706 0.0834   0.7618  0.2382  3.1983 1.1626 -0.1933 0.0033
5    <= 685    494     396     98      2521       1717       804 0.0846   0.8016  0.1984  4.0408 1.3964  0.0405 0.0001
6    <= 701    521     428     93      3042       2145       897 0.0893   0.8215  0.1785  4.6022 1.5265  0.1706 0.0025
7    <= 714    487     418     69      3529       2563       966 0.0834   0.8583  0.1417  6.0580 1.8014  0.4454 0.0144
8    <= 730    489     441     48      4018       3004      1014 0.0838   0.9018  0.0982  9.1875 2.2178  0.8619 0.0473
9    <= 751    513     476     37      4531       3480      1051 0.0879   0.9279  0.0721 12.8649 2.5545  1.1986 0.0859
10   <= 775    492     465     27      5023       3945      1078 0.0843   0.9451  0.0549 17.2222 2.8462  1.4903 0.1157
11    > 775    499     486     13      5522       4431      1091 0.0855   0.9739  0.0261 37.3846 3.6213  2.2653 0.2126
12  Missing    315     210    105      5837       4641      1196 0.0540   0.6667  0.3333  2.0000 0.6931 -0.6628 0.0282
13    Total   5837    4641   1196        NA         NA        NA 1.0000   0.7951  0.2049  3.8804 1.3559  0.0000 0.7810

Written by statcompute

June 15, 2017 at 5:24 pm

Random Search for Optimal Parameters

Practices of manual search, grid search, or the combination of both have been successfully employed in the machine learning to optimize hyper-parameters. However, in the arena of deep learning, both approaches might become impractical. For instance, the computing cost of grid search for hyper-parameters in a multi-layer deep neural network (DNN) could be prohibitively high.

In light of aforementioned hurdles, Bergstra and Bengio proposed a novel idea of random search in the paper http://www.jmlr.org/papers/volume13/bergstra12a/bergstra12a.pdf. In their study, it was found that random search is more efficient than grid search for the hyper-parameter optimization in terms of computing costs.

In the example below, it is shown that both grid search and random search have reached similar results in the SVM parameter optimization based on cross-validations.

import pandas as pd
import numpy as np
from sklearn import preprocessing
from sklearn.model_selection import GridSearchCV, RandomizedSearchCV 
from sklearn.svm import SVC as svc 
from sklearn.metrics import make_scorer, roc_auc_score
from scipy import stats

# DATA PREPARATION
df = pd.read_csv("credit_count.txt")
y = df[df.CARDHLDR == 1].DEFAULT.values 
x = preprocessing.scale(df[df.CARDHLDR == 1].ix[:, 2:12], axis = 0) 

# DEFINE MODEL AND PERFORMANCE MEASURE
mdl = svc(probability = True, random_state = 1)
auc = make_scorer(roc_auc_score)

# GRID SEARCH FOR 20 COMBINATIONS OF PARAMETERS
grid_list = {"C": np.arange(2, 10, 2),
             "gamma": np.arange(0.1, 1, 0.2)}

grid_search = GridSearchCV(mdl, param_grid = grid_list, n_jobs = 4, cv = 3, scoring = auc) 
grid_search.fit(x, y) 
grid_search.cv_results_

# RANDOM SEARCH FOR 20 COMBINATIONS OF PARAMETERS
rand_list = {"C": stats.uniform(2, 10),
             "gamma": stats.uniform(0.1, 1)}
             
rand_search = RandomizedSearchCV(mdl, param_distributions = rand_list, n_iter = 20, n_jobs = 4, cv = 3, random_state = 2017, scoring = auc) 
rand_search.fit(x, y) 
rand_search.cv_results_

Written by statcompute

April 10, 2017 at 12:07 am

A Simple Convolutional Neural Network for The Binary Outcome

Since CNN(Convolutional Neural Networks) have achieved a tremendous success in various challenging applications, e.g. image or digit recognitions, one might wonder how to employ CNNs in classification problems with binary outcomes.

Below is an example showing how to use a simple 1D convolutional neural network to predict credit card defaults.

### LOAD PACKAGES 
from numpy.random import seed
from pandas import read_csv, DataFrame
from sklearn.preprocessing import minmax_scale
from keras.layers.convolutional import Conv1D, MaxPooling1D
from keras.optimizers import SGD
from keras.models import Sequential
from keras.layers import Dense, Flatten

### PREPARE THE DATA 
df = read_csv("credit_count.txt")
Y = df[df.CARDHLDR == 1].DEFAULT
X = minmax_scale(df[df.CARDHLDR == 1].ix[:, 2:12], axis = 0)
y_train = Y.values
x_train = X.reshape(X.shape[0], X.shape[1], 1)

### FIT A 1D CONVOLUTIONAL NEURAL NETWORK
seed(2017)
conv = Sequential()
conv.add(Conv1D(20, 4, input_shape = x_train.shape[1:3], activation = 'relu'))
conv.add(MaxPooling1D(2))
conv.add(Flatten())
conv.add(Dense(1, activation = 'sigmoid'))
sgd = SGD(lr = 0.1, momentum = 0.9, decay = 0, nesterov = False)
conv.compile(loss = 'binary_crossentropy', optimizer = sgd, metrics = ['accuracy'])
conv.fit(x_train, y_train, batch_size = 500, epochs = 100, verbose = 0)

Considering that 1D is the special case of 2D, we can also solve the same problem with a 2D convolutional neural network by changing the input shape, as shown below.

from numpy.random import seed
from pandas import read_csv, DataFrame
from sklearn.preprocessing import minmax_scale
from keras_diagram import ascii
from keras.layers.convolutional import Conv2D, MaxPooling2D
from keras.optimizers import SGD
from keras.models import Sequential
from keras.layers import Dense, Flatten

df = read_csv("credit_count.txt")
Y = df[df.CARDHLDR == 1].DEFAULT
X = minmax_scale(df[df.CARDHLDR == 1].ix[:, 2:12], axis = 0)
y_train = Y.values
x_train = X.reshape(X.shape[0], 1, X.shape[1], 1)

seed(2017)
conv = Sequential()
conv.add(Conv2D(20, (1, 4), input_shape = x_train.shape[1:4], activation = 'relu'))
conv.add(MaxPooling2D((1, 2)))
conv.add(Flatten())
conv.add(Dense(1, activation = 'sigmoid'))
sgd = SGD(lr = 0.1, momentum = 0.9, decay = 0, nesterov = False)
conv.compile(loss = 'binary_crossentropy', optimizer = sgd, metrics = ['accuracy'])
conv.fit(x_train, y_train, batch_size = 500, epochs = 100, verbose = 0)

Written by statcompute

April 2, 2017 at 11:45 pm

Autoencoder for Dimensionality Reduction

We often use ICA or PCA to extract features from the high-dimensional data. The autoencoder is another interesting algorithm to achieve the same purpose in the context of Deep Learning.

with the purpose of learning a function to approximate the input data itself such that F(X) = X, an autoencoder consists of two parts, namely encoder and decoder. While the encoder aims to compress the original input data into a low-dimensional representation, the decoder tries to reconstruct the original input data based on the low-dimension representation generated by the encoder. As a result, the autoencoder has been widely used to remove the data noise as well to reduce the data dimension.

First of all, we will show the basic structure of an autoencoder with 1-layer encoder and 1-layer decoder, as below. In the example, we will compress the input data with 10 columns into a compressed on with 3 columns.

from pandas import read_csv, DataFrame
from numpy.random import seed
from sklearn.preprocessing import minmax_scale
from sklearn.model_selection import train_test_split
from keras.layers import Input, Dense
from keras.models import Model

df = read_csv("credit_count.txt")
Y = df[df.CARDHLDR == 1].DEFAULTS
X = df[df.CARDHLDR == 1].ix[:, 2:12]
# SCALE EACH FEATURE INTO [0, 1] RANGE
sX = minmax_scale(X, axis = 0)
ncol = sX.shape[1]
X_train, X_test, Y_train, Y_test = train_test_split(sX, Y, train_size = 0.5, random_state = seed(2017))

### AN EXAMPLE OF SIMPLE AUTOENCODER ###
# InputLayer (None, 10)
#      Dense (None, 5)
#      Dense (None, 10)

input_dim = Input(shape = (ncol, ))
# DEFINE THE DIMENSION OF ENCODER ASSUMED 3
encoding_dim = 3
# DEFINE THE ENCODER LAYER
encoded = Dense(encoding_dim, activation = 'relu')(input_dim)
# DEFINE THE DECODER LAYER
decoded = Dense(ncol, activation = 'sigmoid')(encoded)
# COMBINE ENCODER AND DECODER INTO AN AUTOENCODER MODEL
autoencoder = Model(input = input_dim, output = decoded)
# CONFIGURE AND TRAIN THE AUTOENCODER
autoencoder.compile(optimizer = 'adadelta', loss = 'binary_crossentropy')
autoencoder.fit(X_train, X_train, nb_epoch = 50, batch_size = 100, shuffle = True, validation_data = (X_test, X_test))
# THE ENCODER TO EXTRACT THE REDUCED DIMENSION FROM THE ABOVE AUTOENCODER
encoder = Model(input = input_dim, output = encoded)
encoded_input = Input(shape = (encoding_dim, ))
encoded_out = encoder.predict(X_test)
encoded_out[0:2]
#array([[ 0.        ,  1.26510417,  1.62803197],
#       [ 2.32508397,  0.99735016,  2.06461048]], dtype=float32)

In the next example, we will relax the constraint of layers and employ a stack of layers to achievement the same purpose as above.

### AN EXAMPLE OF DEEP AUTOENCODER WITH MULTIPLE LAYERS
# InputLayer (None, 10)
#      Dense (None, 20)
#      Dense (None, 10)
#      Dense (None, 5)
#      Dense (None, 3)
#      Dense (None, 5)
#      Dense (None, 10)
#      Dense (None, 20)
#      Dense (None, 10)

input_dim = Input(shape = (ncol, ))
# DEFINE THE DIMENSION OF ENCODER ASSUMED 3
encoding_dim = 3
# DEFINE THE ENCODER LAYERS
encoded1 = Dense(20, activation = 'relu')(input_dim)
encoded2 = Dense(10, activation = 'relu')(encoded1)
encoded3 = Dense(5, activation = 'relu')(encoded2)
encoded4 = Dense(encoding_dim, activation = 'relu')(encoded3)
# DEFINE THE DECODER LAYERS
decoded1 = Dense(5, activation = 'relu')(encoded4)
decoded2 = Dense(10, activation = 'relu')(decoded1)
decoded3 = Dense(20, activation = 'relu')(decoded2)
decoded4 = Dense(ncol, activation = 'sigmoid')(decoded3)
# COMBINE ENCODER AND DECODER INTO AN AUTOENCODER MODEL
autoencoder = Model(input = input_dim, output = decoded4)
# CONFIGURE AND TRAIN THE AUTOENCODER
autoencoder.compile(optimizer = 'adadelta', loss = 'binary_crossentropy')
autoencoder.fit(X_train, X_train, nb_epoch = 100, batch_size = 100, shuffle = True, validation_data = (X_test, X_test))
# THE ENCODER TO EXTRACT THE REDUCED DIMENSION FROM THE ABOVE AUTOENCODER
encoder = Model(input = input_dim, output = encoded4)
encoded_input = Input(shape = (encoding_dim, ))
encoded_out = encoder.predict(X_test)
encoded_out[0:2]
#array([[ 3.74947715,  0.        ,  3.22947764],
#       [ 3.93903661,  0.17448257,  1.86618853]], dtype=float32)

Written by statcompute

January 15, 2017 at 6:19 pm

An Example of Merge Layer in Keras

The power of a DNN does not only come from its depth but also come from its flexibility of accommodating complex network structures. For instance, the DNN shown below consists of two branches, the left with 4 inputs and the right with 6 inputs. In addition, the right branch shows a more complicated structure than the left.

                                                InputLayer (None, 6)
                                                     Dense (None, 6)
                                        BatchNormalization (None, 6)
                                                     Dense (None, 6)
         InputLayer (None, 4)           BatchNormalization (None, 6)
              Dense (None, 4)                        Dense (None, 6)
 BatchNormalization (None, 4)           BatchNormalization (None, 6)
                    \____________________________________/
                                      |
                                 Merge (None, 10)
                                 Dense (None, 1)

To create a DNN as the above, both left and right branches are defined separately with corresponding inputs and layers. In the line 29, both branches would be combined with a MERGE layer. There are multiple benefits of such merged DNNs. For instance, the DNN has the flexibility to handle various inputs differently. In addition, new features can be added conveniently without messing around with the existing network structure.

from pandas import read_csv, DataFrame
from numpy.random import seed
from sklearn.preprocessing import scale
from keras.models import Sequential
from keras.constraints import maxnorm
from keras.optimizers import SGD
from keras.layers import Dense, Merge
from keras.layers.normalization import BatchNormalization
from keras_diagram import ascii

df = read_csv("credit_count.txt")
Y = df[df.CARDHLDR == 1].DEFAULTS
X1 = scale(df[df.CARDHLDR == 1][["MAJORDRG", "MINORDRG", "OWNRENT", "SELFEMPL"]])
X2 = scale(df[df.CARDHLDR == 1][["AGE", "ACADMOS", "ADEPCNT", "INCPER", "EXP_INC", "INCOME"]])

branch1 = Sequential()
branch1.add(Dense(X1.shape[1], input_shape = (X1.shape[1],), init = 'normal', activation = 'relu'))
branch1.add(BatchNormalization())

branch2 = Sequential()
branch2.add(Dense(X2.shape[1], input_shape =  (X2.shape[1],), init = 'normal', activation = 'relu'))
branch2.add(BatchNormalization())
branch2.add(Dense(X2.shape[1], init = 'normal', activation = 'relu', W_constraint = maxnorm(5)))
branch2.add(BatchNormalization())
branch2.add(Dense(X2.shape[1], init = 'normal', activation = 'relu', W_constraint = maxnorm(5)))
branch2.add(BatchNormalization())

model = Sequential()
model.add(Merge([branch1, branch2], mode = 'concat'))
model.add(Dense(1, init = 'normal', activation = 'sigmoid'))
sgd = SGD(lr = 0.1, momentum = 0.9, decay = 0, nesterov = False)
model.compile(loss = 'binary_crossentropy', optimizer = sgd, metrics = ['accuracy'])
seed(2017)
model.fit([X1, X2], Y.values, batch_size = 2000, nb_epoch = 100, verbose = 1)

Written by statcompute

January 8, 2017 at 4:42 pm