In rstudio/keras: R Interface to 'Keras'
library(keras3)
use_backend("jax")
Introduction
This example looks at the
Kaggle Credit Card Fraud Detection
dataset to demonstrate how
to train a classification model on data with highly imbalanced classes.
You can download the data by clicking "Download" at
the link, or if you're setup with a kaggle API key at
"~/.kaggle/kagle.json", you can run the following:
reticulate::py_install("kaggle", pip = TRUE)
reticulate::py_available(TRUE) # ensure 'kaggle' is on the PATH
system("kaggle datasets download -d mlg-ulb/creditcardfraud")
zip::unzip("creditcardfraud.zip", files = "creditcard.csv")
First, load the data
library(readr)
df <- read_csv("creditcard.csv", col_types = cols(
Class = col_integer(),
.default = col_double()
))
tibble::glimpse(df)
## Rows: 284,807
## Columns: 31
## $ Time [3m[38;5;246m<dbl>[39m[23m 0, 0, 1, 1, 2, 2, 4, 7, 7, 9, 10, 10, 10, 11, 12, 12, 12, 1…
## $ V1 [3m[38;5;246m<dbl>[39m[23m -1.3598071, 1.1918571, -1.3583541, -0.9662717, -1.1582331, …
## $ V2 [3m[38;5;246m<dbl>[39m[23m -0.07278117, 0.26615071, -1.34016307, -0.18522601, 0.877736…
## $ V3 [3m[38;5;246m<dbl>[39m[23m 2.53634674, 0.16648011, 1.77320934, 1.79299334, 1.54871785,…
## $ V4 [3m[38;5;246m<dbl>[39m[23m 1.37815522, 0.44815408, 0.37977959, -0.86329128, 0.40303393…
## $ V5 [3m[38;5;246m<dbl>[39m[23m -0.33832077, 0.06001765, -0.50319813, -0.01030888, -0.40719…
## $ V6 [3m[38;5;246m<dbl>[39m[23m 0.46238778, -0.08236081, 1.80049938, 1.24720317, 0.09592146…
## $ V7 [3m[38;5;246m<dbl>[39m[23m 0.239598554, -0.078802983, 0.791460956, 0.237608940, 0.5929…
## $ V8 [3m[38;5;246m<dbl>[39m[23m 0.098697901, 0.085101655, 0.247675787, 0.377435875, -0.2705…
## $ V9 [3m[38;5;246m<dbl>[39m[23m 0.3637870, -0.2554251, -1.5146543, -1.3870241, 0.8177393, -…
## $ V10 [3m[38;5;246m<dbl>[39m[23m 0.09079417, -0.16697441, 0.20764287, -0.05495192, 0.7530744…
## $ V11 [3m[38;5;246m<dbl>[39m[23m -0.55159953, 1.61272666, 0.62450146, -0.22648726, -0.822842…
## $ V12 [3m[38;5;246m<dbl>[39m[23m -0.61780086, 1.06523531, 0.06608369, 0.17822823, 0.53819555…
## $ V13 [3m[38;5;246m<dbl>[39m[23m -0.99138985, 0.48909502, 0.71729273, 0.50775687, 1.34585159…
## $ V14 [3m[38;5;246m<dbl>[39m[23m -0.31116935, -0.14377230, -0.16594592, -0.28792375, -1.1196…
## $ V15 [3m[38;5;246m<dbl>[39m[23m 1.468176972, 0.635558093, 2.345864949, -0.631418118, 0.1751…
## $ V16 [3m[38;5;246m<dbl>[39m[23m -0.47040053, 0.46391704, -2.89008319, -1.05964725, -0.45144…
## $ V17 [3m[38;5;246m<dbl>[39m[23m 0.207971242, -0.114804663, 1.109969379, -0.684092786, -0.23…
## $ V18 [3m[38;5;246m<dbl>[39m[23m 0.02579058, -0.18336127, -0.12135931, 1.96577500, -0.038194…
## $ V19 [3m[38;5;246m<dbl>[39m[23m 0.40399296, -0.14578304, -2.26185710, -1.23262197, 0.803486…
## $ V20 [3m[38;5;246m<dbl>[39m[23m 0.25141210, -0.06908314, 0.52497973, -0.20803778, 0.4085423…
## $ V21 [3m[38;5;246m<dbl>[39m[23m -0.018306778, -0.225775248, 0.247998153, -0.108300452, -0.0…
## $ V22 [3m[38;5;246m<dbl>[39m[23m 0.277837576, -0.638671953, 0.771679402, 0.005273597, 0.7982…
## $ V23 [3m[38;5;246m<dbl>[39m[23m -0.110473910, 0.101288021, 0.909412262, -0.190320519, -0.13…
## $ V24 [3m[38;5;246m<dbl>[39m[23m 0.06692807, -0.33984648, -0.68928096, -1.17557533, 0.141266…
## $ V25 [3m[38;5;246m<dbl>[39m[23m 0.12853936, 0.16717040, -0.32764183, 0.64737603, -0.2060095…
## $ V26 [3m[38;5;246m<dbl>[39m[23m -0.18911484, 0.12589453, -0.13909657, -0.22192884, 0.502292…
## $ V27 [3m[38;5;246m<dbl>[39m[23m 0.133558377, -0.008983099, -0.055352794, 0.062722849, 0.219…
## $ V28 [3m[38;5;246m<dbl>[39m[23m -0.021053053, 0.014724169, -0.059751841, 0.061457629, 0.215…
## $ Amount [3m[38;5;246m<dbl>[39m[23m 149.62, 2.69, 378.66, 123.50, 69.99, 3.67, 4.99, 40.80, 93.…
## $ Class [3m[38;5;246m<int>[39m[23m 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,…
Prepare a validation set
val_idx <- nrow(df) %>% sample.int(., round( . * 0.2))
val_df <- df[val_idx, ]
train_df <- df[-val_idx, ]
cat("Number of training samples:", nrow(train_df), "\n")
## Number of training samples: 227846
cat("Number of validation samples:", nrow(val_df), "\n")
## Number of validation samples: 56961
Analyze class imbalance in the targets
counts <- table(train_df$Class)
counts
##
## 0 1
## 227450 396
cat(sprintf("Number of positive samples in training data: %i (%.2f%% of total)",
counts["1"], 100 * counts["1"] / sum(counts)))
## Number of positive samples in training data: 396 (0.17% of total)
weight_for_0 = 1 / counts["0"]
weight_for_1 = 1 / counts["1"]
Normalize the data using training set statistics
feature_names <- colnames(train_df) %>% setdiff("Class")
train_features <- as.matrix(train_df[feature_names])
train_targets <- as.matrix(train_df$Class)
val_features <- as.matrix(val_df[feature_names])
val_targets <- as.matrix(val_df$Class)
train_features %<>% scale()
val_features %<>% scale(center = attr(train_features, "scaled:center"),
scale = attr(train_features, "scaled:scale"))
Build a binary classification model
model <-
keras_model_sequential(input_shape = ncol(train_features)) |>
layer_dense(256, activation = "relu") |>
layer_dense(256, activation = "relu") |>
layer_dropout(0.3) |>
layer_dense(256, activation = "relu") |>
layer_dropout(0.3) |>
layer_dense(1, activation = "sigmoid")
model
## [1mModel: "sequential"[0m
## ┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
## ┃[1m [0m[1mLayer (type) [0m[1m [0m┃[1m [0m[1mOutput Shape [0m[1m [0m┃[1m [0m[1m Param #[0m[1m [0m┃
## ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
## │ dense ([38;5;33mDense[0m) │ ([38;5;45mNone[0m, [38;5;34m256[0m) │ [38;5;34m7,936[0m │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dense_1 ([38;5;33mDense[0m) │ ([38;5;45mNone[0m, [38;5;34m256[0m) │ [38;5;34m65,792[0m │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dropout ([38;5;33mDropout[0m) │ ([38;5;45mNone[0m, [38;5;34m256[0m) │ [38;5;34m0[0m │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dense_2 ([38;5;33mDense[0m) │ ([38;5;45mNone[0m, [38;5;34m256[0m) │ [38;5;34m65,792[0m │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dropout_1 ([38;5;33mDropout[0m) │ ([38;5;45mNone[0m, [38;5;34m256[0m) │ [38;5;34m0[0m │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dense_3 ([38;5;33mDense[0m) │ ([38;5;45mNone[0m, [38;5;34m1[0m) │ [38;5;34m257[0m │
## └─────────────────────────────────┴────────────────────────┴───────────────┘
## [1m Total params: [0m[38;5;34m139,777[0m (546.00 KB)
## [1m Trainable params: [0m[38;5;34m139,777[0m (546.00 KB)
## [1m Non-trainable params: [0m[38;5;34m0[0m (0.00 B)
Train the model with class_weight argument
metrics <- list(
metric_false_negatives(name = "fn"),
metric_false_positives(name = "fp"),
metric_true_negatives(name = "tn"),
metric_true_positives(name = "tp"),
metric_precision(name = "precision"),
metric_recall(name = "recall")
)
model |> compile(
optimizer = optimizer_adam(1e-2),
loss = "binary_crossentropy",
metrics = metrics
)
callbacks <- list(
callback_model_checkpoint("fraud_model_at_epoch_{epoch}.keras")
)
class_weight <- list("0" = weight_for_0,
"1" = weight_for_1)
model |> fit(
train_features, train_targets,
validation_data = list(val_features, val_targets),
class_weight = class_weight,
batch_size = 2048,
epochs = 30,
callbacks = callbacks,
verbose = 2
)
## Epoch 1/30
## 112/112 - 3s - 27ms/step - fn: 40.0000 - fp: 23705.0000 - loss: 2.2728e-06 - precision: 0.0148 - recall: 0.8990 - tn: 203745.0000 - tp: 356.0000 - val_fn: 8.0000 - val_fp: 1502.0000 - val_loss: 0.1089 - val_precision: 0.0553 - val_recall: 0.9167 - val_tn: 55363.0000 - val_tp: 88.0000
## Epoch 2/30
## 112/112 - 1s - 6ms/step - fn: 33.0000 - fp: 8009.0000 - loss: 1.4949e-06 - precision: 0.0434 - recall: 0.9167 - tn: 219441.0000 - tp: 363.0000 - val_fn: 10.0000 - val_fp: 1176.0000 - val_loss: 0.0912 - val_precision: 0.0681 - val_recall: 0.8958 - val_tn: 55689.0000 - val_tp: 86.0000
## Epoch 3/30
## 112/112 - 0s - 2ms/step - fn: 32.0000 - fp: 7704.0000 - loss: 1.3369e-06 - precision: 0.0451 - recall: 0.9192 - tn: 219746.0000 - tp: 364.0000 - val_fn: 9.0000 - val_fp: 1202.0000 - val_loss: 0.0870 - val_precision: 0.0675 - val_recall: 0.9062 - val_tn: 55663.0000 - val_tp: 87.0000
## Epoch 4/30
## 112/112 - 0s - 2ms/step - fn: 28.0000 - fp: 8366.0000 - loss: 1.2269e-06 - precision: 0.0421 - recall: 0.9293 - tn: 219084.0000 - tp: 368.0000 - val_fn: 10.0000 - val_fp: 1560.0000 - val_loss: 0.0967 - val_precision: 0.0522 - val_recall: 0.8958 - val_tn: 55305.0000 - val_tp: 86.0000
## Epoch 5/30
## 112/112 - 0s - 2ms/step - fn: 19.0000 - fp: 6217.0000 - loss: 8.5492e-07 - precision: 0.0572 - recall: 0.9520 - tn: 221233.0000 - tp: 377.0000 - val_fn: 9.0000 - val_fp: 2626.0000 - val_loss: 0.1236 - val_precision: 0.0321 - val_recall: 0.9062 - val_tn: 54239.0000 - val_tp: 87.0000
## Epoch 6/30
## 112/112 - 0s - 2ms/step - fn: 18.0000 - fp: 6566.0000 - loss: 7.4300e-07 - precision: 0.0544 - recall: 0.9545 - tn: 220884.0000 - tp: 378.0000 - val_fn: 8.0000 - val_fp: 2302.0000 - val_loss: 0.1081 - val_precision: 0.0368 - val_recall: 0.9167 - val_tn: 54563.0000 - val_tp: 88.0000
## Epoch 7/30
## 112/112 - 0s - 2ms/step - fn: 14.0000 - fp: 6819.0000 - loss: 7.1838e-07 - precision: 0.0530 - recall: 0.9646 - tn: 220631.0000 - tp: 382.0000 - val_fn: 7.0000 - val_fp: 1376.0000 - val_loss: 0.0800 - val_precision: 0.0608 - val_recall: 0.9271 - val_tn: 55489.0000 - val_tp: 89.0000
## Epoch 8/30
## 112/112 - 0s - 2ms/step - fn: 19.0000 - fp: 7345.0000 - loss: 8.3921e-07 - precision: 0.0488 - recall: 0.9520 - tn: 220105.0000 - tp: 377.0000 - val_fn: 10.0000 - val_fp: 416.0000 - val_loss: 0.0406 - val_precision: 0.1713 - val_recall: 0.8958 - val_tn: 56449.0000 - val_tp: 86.0000
## Epoch 9/30
## 112/112 - 0s - 2ms/step - fn: 13.0000 - fp: 6282.0000 - loss: 8.3488e-07 - precision: 0.0575 - recall: 0.9672 - tn: 221168.0000 - tp: 383.0000 - val_fn: 10.0000 - val_fp: 933.0000 - val_loss: 0.0450 - val_precision: 0.0844 - val_recall: 0.8958 - val_tn: 55932.0000 - val_tp: 86.0000
## Epoch 10/30
## 112/112 - 0s - 2ms/step - fn: 15.0000 - fp: 7597.0000 - loss: 1.1098e-06 - precision: 0.0478 - recall: 0.9621 - tn: 219853.0000 - tp: 381.0000 - val_fn: 10.0000 - val_fp: 2417.0000 - val_loss: 0.4261 - val_precision: 0.0344 - val_recall: 0.8958 - val_tn: 54448.0000 - val_tp: 86.0000
## Epoch 11/30
## 112/112 - 0s - 2ms/step - fn: 24.0000 - fp: 10269.0000 - loss: 1.9062e-06 - precision: 0.0350 - recall: 0.9394 - tn: 217181.0000 - tp: 372.0000 - val_fn: 9.0000 - val_fp: 1434.0000 - val_loss: 0.1261 - val_precision: 0.0572 - val_recall: 0.9062 - val_tn: 55431.0000 - val_tp: 87.0000
## Epoch 12/30
## 112/112 - 0s - 2ms/step - fn: 17.0000 - fp: 6634.0000 - loss: 1.0711e-06 - precision: 0.0540 - recall: 0.9571 - tn: 220816.0000 - tp: 379.0000 - val_fn: 8.0000 - val_fp: 2263.0000 - val_loss: 0.1654 - val_precision: 0.0374 - val_recall: 0.9167 - val_tn: 54602.0000 - val_tp: 88.0000
## Epoch 13/30
## 112/112 - 0s - 2ms/step - fn: 10.0000 - fp: 5620.0000 - loss: 5.3600e-07 - precision: 0.0643 - recall: 0.9747 - tn: 221830.0000 - tp: 386.0000 - val_fn: 12.0000 - val_fp: 914.0000 - val_loss: 0.0414 - val_precision: 0.0842 - val_recall: 0.8750 - val_tn: 55951.0000 - val_tp: 84.0000
## Epoch 14/30
## 112/112 - 0s - 2ms/step - fn: 8.0000 - fp: 5103.0000 - loss: 5.1168e-07 - precision: 0.0707 - recall: 0.9798 - tn: 222347.0000 - tp: 388.0000 - val_fn: 11.0000 - val_fp: 1382.0000 - val_loss: 0.0660 - val_precision: 0.0579 - val_recall: 0.8854 - val_tn: 55483.0000 - val_tp: 85.0000
## Epoch 15/30
## 112/112 - 0s - 2ms/step - fn: 7.0000 - fp: 4497.0000 - loss: 4.4632e-07 - precision: 0.0796 - recall: 0.9823 - tn: 222953.0000 - tp: 389.0000 - val_fn: 10.0000 - val_fp: 1805.0000 - val_loss: 0.0786 - val_precision: 0.0455 - val_recall: 0.8958 - val_tn: 55060.0000 - val_tp: 86.0000
## Epoch 16/30
## 112/112 - 0s - 2ms/step - fn: 5.0000 - fp: 5056.0000 - loss: 6.1038e-07 - precision: 0.0718 - recall: 0.9874 - tn: 222394.0000 - tp: 391.0000 - val_fn: 13.0000 - val_fp: 889.0000 - val_loss: 0.0676 - val_precision: 0.0854 - val_recall: 0.8646 - val_tn: 55976.0000 - val_tp: 83.0000
## Epoch 17/30
## 112/112 - 0s - 2ms/step - fn: 3.0000 - fp: 4639.0000 - loss: 4.5763e-07 - precision: 0.0781 - recall: 0.9924 - tn: 222811.0000 - tp: 393.0000 - val_fn: 11.0000 - val_fp: 1185.0000 - val_loss: 0.0669 - val_precision: 0.0669 - val_recall: 0.8854 - val_tn: 55680.0000 - val_tp: 85.0000
## Epoch 18/30
## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 4355.0000 - loss: 4.5299e-07 - precision: 0.0826 - recall: 0.9899 - tn: 223095.0000 - tp: 392.0000 - val_fn: 9.0000 - val_fp: 1405.0000 - val_loss: 0.0772 - val_precision: 0.0583 - val_recall: 0.9062 - val_tn: 55460.0000 - val_tp: 87.0000
## Epoch 19/30
## 112/112 - 0s - 2ms/step - fn: 6.0000 - fp: 5779.0000 - loss: 4.8324e-07 - precision: 0.0632 - recall: 0.9848 - tn: 221671.0000 - tp: 390.0000 - val_fn: 10.0000 - val_fp: 1135.0000 - val_loss: 0.0546 - val_precision: 0.0704 - val_recall: 0.8958 - val_tn: 55730.0000 - val_tp: 86.0000
## Epoch 20/30
## 112/112 - 0s - 2ms/step - fn: 3.0000 - fp: 4748.0000 - loss: 3.7424e-07 - precision: 0.0764 - recall: 0.9924 - tn: 222702.0000 - tp: 393.0000 - val_fn: 9.0000 - val_fp: 895.0000 - val_loss: 0.0396 - val_precision: 0.0886 - val_recall: 0.9062 - val_tn: 55970.0000 - val_tp: 87.0000
## Epoch 21/30
## 112/112 - 0s - 2ms/step - fn: 3.0000 - fp: 3798.0000 - loss: 3.1417e-07 - precision: 0.0938 - recall: 0.9924 - tn: 223652.0000 - tp: 393.0000 - val_fn: 10.0000 - val_fp: 652.0000 - val_loss: 0.0309 - val_precision: 0.1165 - val_recall: 0.8958 - val_tn: 56213.0000 - val_tp: 86.0000
## Epoch 22/30
## 112/112 - 0s - 2ms/step - fn: 0.0000e+00 - fp: 2188.0000 - loss: 1.8112e-07 - precision: 0.1533 - recall: 1.0000 - tn: 225262.0000 - tp: 396.0000 - val_fn: 10.0000 - val_fp: 427.0000 - val_loss: 0.0250 - val_precision: 0.1676 - val_recall: 0.8958 - val_tn: 56438.0000 - val_tp: 86.0000
## Epoch 23/30
## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 3066.0000 - loss: 3.4932e-07 - precision: 0.1134 - recall: 0.9899 - tn: 224384.0000 - tp: 392.0000 - val_fn: 11.0000 - val_fp: 1053.0000 - val_loss: 0.0494 - val_precision: 0.0747 - val_recall: 0.8854 - val_tn: 55812.0000 - val_tp: 85.0000
## Epoch 24/30
## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 3175.0000 - loss: 5.1994e-07 - precision: 0.1099 - recall: 0.9899 - tn: 224275.0000 - tp: 392.0000 - val_fn: 11.0000 - val_fp: 1012.0000 - val_loss: 0.0997 - val_precision: 0.0775 - val_recall: 0.8854 - val_tn: 55853.0000 - val_tp: 85.0000
## Epoch 25/30
## 112/112 - 0s - 2ms/step - fn: 7.0000 - fp: 5874.0000 - loss: 6.5862e-07 - precision: 0.0621 - recall: 0.9823 - tn: 221576.0000 - tp: 389.0000 - val_fn: 9.0000 - val_fp: 3090.0000 - val_loss: 0.1296 - val_precision: 0.0274 - val_recall: 0.9062 - val_tn: 53775.0000 - val_tp: 87.0000
## Epoch 26/30
## 112/112 - 0s - 2ms/step - fn: 6.0000 - fp: 5395.0000 - loss: 5.9253e-07 - precision: 0.0674 - recall: 0.9848 - tn: 222055.0000 - tp: 390.0000 - val_fn: 9.0000 - val_fp: 1320.0000 - val_loss: 0.0928 - val_precision: 0.0618 - val_recall: 0.9062 - val_tn: 55545.0000 - val_tp: 87.0000
## Epoch 27/30
## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 3887.0000 - loss: 7.1278e-07 - precision: 0.0916 - recall: 0.9899 - tn: 223563.0000 - tp: 392.0000 - val_fn: 12.0000 - val_fp: 572.0000 - val_loss: 0.0352 - val_precision: 0.1280 - val_recall: 0.8750 - val_tn: 56293.0000 - val_tp: 84.0000
## Epoch 28/30
## 112/112 - 0s - 2ms/step - fn: 3.0000 - fp: 2209.0000 - loss: 5.8153e-07 - precision: 0.1510 - recall: 0.9924 - tn: 225241.0000 - tp: 393.0000 - val_fn: 11.0000 - val_fp: 1373.0000 - val_loss: 0.0808 - val_precision: 0.0583 - val_recall: 0.8854 - val_tn: 55492.0000 - val_tp: 85.0000
## Epoch 29/30
## 112/112 - 0s - 2ms/step - fn: 1.0000 - fp: 2506.0000 - loss: 2.9031e-07 - precision: 0.1362 - recall: 0.9975 - tn: 224944.0000 - tp: 395.0000 - val_fn: 11.0000 - val_fp: 690.0000 - val_loss: 0.0401 - val_precision: 0.1097 - val_recall: 0.8854 - val_tn: 56175.0000 - val_tp: 85.0000
## Epoch 30/30
## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 2719.0000 - loss: 6.1097e-07 - precision: 0.1260 - recall: 0.9899 - tn: 224731.0000 - tp: 392.0000 - val_fn: 13.0000 - val_fp: 861.0000 - val_loss: 0.0362 - val_precision: 0.0879 - val_recall: 0.8646 - val_tn: 56004.0000 - val_tp: 83.0000
val_pred <- model %>%
predict(val_features) %>%
{ as.integer(. > 0.5) }
## 1781/1781 - 1s - 311us/step
pred_correct <- val_df$Class == val_pred
cat(sprintf("Validation accuracy: %.2f", mean(pred_correct)))
## Validation accuracy: 0.98
fraudulent <- val_df$Class == 1
n_fraudulent_detected <- sum(fraudulent & pred_correct)
n_fraudulent_missed <- sum(fraudulent & !pred_correct)
n_legitimate_flagged <- sum(!fraudulent & !pred_correct)
Conclusions
At the end of training, out of
56,961 validation transactions, we
are:
- Correctly identifying
83 of them as
fraudulent
- Missing 13
fraudulent transactions
- At the cost of incorrectly flagging
861 legitimate
transactions
In the real world, one would put an even higher weight on class 1,
so as to reflect that False Negatives are more costly than False Positives.
Next time your credit card gets declined in an online purchase -- this is why.
rstudio/keras documentation built on July 8, 2024, 3:07 p.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
library(keras3) use_backend("jax")
Introduction
This example looks at the
Kaggle Credit Card Fraud Detection
dataset to demonstrate how
to train a classification model on data with highly imbalanced classes.
You can download the data by clicking "Download" at
the link, or if you're setup with a kaggle API key at
"~/.kaggle/kagle.json", you can run the following:
reticulate::py_install("kaggle", pip = TRUE) reticulate::py_available(TRUE) # ensure 'kaggle' is on the PATH system("kaggle datasets download -d mlg-ulb/creditcardfraud") zip::unzip("creditcardfraud.zip", files = "creditcard.csv")
First, load the data
library(readr) df <- read_csv("creditcard.csv", col_types = cols( Class = col_integer(), .default = col_double() )) tibble::glimpse(df)
## Rows: 284,807 ## Columns: 31 ## $ Time [3m[38;5;246m<dbl>[39m[23m 0, 0, 1, 1, 2, 2, 4, 7, 7, 9, 10, 10, 10, 11, 12, 12, 12, 1… ## $ V1 [3m[38;5;246m<dbl>[39m[23m -1.3598071, 1.1918571, -1.3583541, -0.9662717, -1.1582331, … ## $ V2 [3m[38;5;246m<dbl>[39m[23m -0.07278117, 0.26615071, -1.34016307, -0.18522601, 0.877736… ## $ V3 [3m[38;5;246m<dbl>[39m[23m 2.53634674, 0.16648011, 1.77320934, 1.79299334, 1.54871785,… ## $ V4 [3m[38;5;246m<dbl>[39m[23m 1.37815522, 0.44815408, 0.37977959, -0.86329128, 0.40303393… ## $ V5 [3m[38;5;246m<dbl>[39m[23m -0.33832077, 0.06001765, -0.50319813, -0.01030888, -0.40719… ## $ V6 [3m[38;5;246m<dbl>[39m[23m 0.46238778, -0.08236081, 1.80049938, 1.24720317, 0.09592146… ## $ V7 [3m[38;5;246m<dbl>[39m[23m 0.239598554, -0.078802983, 0.791460956, 0.237608940, 0.5929… ## $ V8 [3m[38;5;246m<dbl>[39m[23m 0.098697901, 0.085101655, 0.247675787, 0.377435875, -0.2705… ## $ V9 [3m[38;5;246m<dbl>[39m[23m 0.3637870, -0.2554251, -1.5146543, -1.3870241, 0.8177393, -… ## $ V10 [3m[38;5;246m<dbl>[39m[23m 0.09079417, -0.16697441, 0.20764287, -0.05495192, 0.7530744… ## $ V11 [3m[38;5;246m<dbl>[39m[23m -0.55159953, 1.61272666, 0.62450146, -0.22648726, -0.822842… ## $ V12 [3m[38;5;246m<dbl>[39m[23m -0.61780086, 1.06523531, 0.06608369, 0.17822823, 0.53819555… ## $ V13 [3m[38;5;246m<dbl>[39m[23m -0.99138985, 0.48909502, 0.71729273, 0.50775687, 1.34585159… ## $ V14 [3m[38;5;246m<dbl>[39m[23m -0.31116935, -0.14377230, -0.16594592, -0.28792375, -1.1196… ## $ V15 [3m[38;5;246m<dbl>[39m[23m 1.468176972, 0.635558093, 2.345864949, -0.631418118, 0.1751… ## $ V16 [3m[38;5;246m<dbl>[39m[23m -0.47040053, 0.46391704, -2.89008319, -1.05964725, -0.45144… ## $ V17 [3m[38;5;246m<dbl>[39m[23m 0.207971242, -0.114804663, 1.109969379, -0.684092786, -0.23… ## $ V18 [3m[38;5;246m<dbl>[39m[23m 0.02579058, -0.18336127, -0.12135931, 1.96577500, -0.038194… ## $ V19 [3m[38;5;246m<dbl>[39m[23m 0.40399296, -0.14578304, -2.26185710, -1.23262197, 0.803486… ## $ V20 [3m[38;5;246m<dbl>[39m[23m 0.25141210, -0.06908314, 0.52497973, -0.20803778, 0.4085423… ## $ V21 [3m[38;5;246m<dbl>[39m[23m -0.018306778, -0.225775248, 0.247998153, -0.108300452, -0.0… ## $ V22 [3m[38;5;246m<dbl>[39m[23m 0.277837576, -0.638671953, 0.771679402, 0.005273597, 0.7982… ## $ V23 [3m[38;5;246m<dbl>[39m[23m -0.110473910, 0.101288021, 0.909412262, -0.190320519, -0.13… ## $ V24 [3m[38;5;246m<dbl>[39m[23m 0.06692807, -0.33984648, -0.68928096, -1.17557533, 0.141266… ## $ V25 [3m[38;5;246m<dbl>[39m[23m 0.12853936, 0.16717040, -0.32764183, 0.64737603, -0.2060095… ## $ V26 [3m[38;5;246m<dbl>[39m[23m -0.18911484, 0.12589453, -0.13909657, -0.22192884, 0.502292… ## $ V27 [3m[38;5;246m<dbl>[39m[23m 0.133558377, -0.008983099, -0.055352794, 0.062722849, 0.219… ## $ V28 [3m[38;5;246m<dbl>[39m[23m -0.021053053, 0.014724169, -0.059751841, 0.061457629, 0.215… ## $ Amount [3m[38;5;246m<dbl>[39m[23m 149.62, 2.69, 378.66, 123.50, 69.99, 3.67, 4.99, 40.80, 93.… ## $ Class [3m[38;5;246m<int>[39m[23m 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,…
Prepare a validation set
val_idx <- nrow(df) %>% sample.int(., round( . * 0.2)) val_df <- df[val_idx, ] train_df <- df[-val_idx, ] cat("Number of training samples:", nrow(train_df), "\n")
## Number of training samples: 227846
cat("Number of validation samples:", nrow(val_df), "\n")
## Number of validation samples: 56961
Analyze class imbalance in the targets
counts <- table(train_df$Class) counts
## ## 0 1 ## 227450 396
cat(sprintf("Number of positive samples in training data: %i (%.2f%% of total)", counts["1"], 100 * counts["1"] / sum(counts)))
## Number of positive samples in training data: 396 (0.17% of total)
weight_for_0 = 1 / counts["0"] weight_for_1 = 1 / counts["1"]
Normalize the data using training set statistics
feature_names <- colnames(train_df) %>% setdiff("Class") train_features <- as.matrix(train_df[feature_names]) train_targets <- as.matrix(train_df$Class) val_features <- as.matrix(val_df[feature_names]) val_targets <- as.matrix(val_df$Class) train_features %<>% scale() val_features %<>% scale(center = attr(train_features, "scaled:center"), scale = attr(train_features, "scaled:scale"))
Build a binary classification model
model <- keras_model_sequential(input_shape = ncol(train_features)) |> layer_dense(256, activation = "relu") |> layer_dense(256, activation = "relu") |> layer_dropout(0.3) |> layer_dense(256, activation = "relu") |> layer_dropout(0.3) |> layer_dense(1, activation = "sigmoid") model
## [1mModel: "sequential"[0m ## ┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓ ## ┃[1m [0m[1mLayer (type) [0m[1m [0m┃[1m [0m[1mOutput Shape [0m[1m [0m┃[1m [0m[1m Param #[0m[1m [0m┃ ## ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩ ## │ dense ([38;5;33mDense[0m) │ ([38;5;45mNone[0m, [38;5;34m256[0m) │ [38;5;34m7,936[0m │ ## ├─────────────────────────────────┼────────────────────────┼───────────────┤ ## │ dense_1 ([38;5;33mDense[0m) │ ([38;5;45mNone[0m, [38;5;34m256[0m) │ [38;5;34m65,792[0m │ ## ├─────────────────────────────────┼────────────────────────┼───────────────┤ ## │ dropout ([38;5;33mDropout[0m) │ ([38;5;45mNone[0m, [38;5;34m256[0m) │ [38;5;34m0[0m │ ## ├─────────────────────────────────┼────────────────────────┼───────────────┤ ## │ dense_2 ([38;5;33mDense[0m) │ ([38;5;45mNone[0m, [38;5;34m256[0m) │ [38;5;34m65,792[0m │ ## ├─────────────────────────────────┼────────────────────────┼───────────────┤ ## │ dropout_1 ([38;5;33mDropout[0m) │ ([38;5;45mNone[0m, [38;5;34m256[0m) │ [38;5;34m0[0m │ ## ├─────────────────────────────────┼────────────────────────┼───────────────┤ ## │ dense_3 ([38;5;33mDense[0m) │ ([38;5;45mNone[0m, [38;5;34m1[0m) │ [38;5;34m257[0m │ ## └─────────────────────────────────┴────────────────────────┴───────────────┘ ## [1m Total params: [0m[38;5;34m139,777[0m (546.00 KB) ## [1m Trainable params: [0m[38;5;34m139,777[0m (546.00 KB) ## [1m Non-trainable params: [0m[38;5;34m0[0m (0.00 B)
Train the model with class_weight argument
metrics <- list( metric_false_negatives(name = "fn"), metric_false_positives(name = "fp"), metric_true_negatives(name = "tn"), metric_true_positives(name = "tp"), metric_precision(name = "precision"), metric_recall(name = "recall") ) model |> compile( optimizer = optimizer_adam(1e-2), loss = "binary_crossentropy", metrics = metrics ) callbacks <- list( callback_model_checkpoint("fraud_model_at_epoch_{epoch}.keras") ) class_weight <- list("0" = weight_for_0, "1" = weight_for_1) model |> fit( train_features, train_targets, validation_data = list(val_features, val_targets), class_weight = class_weight, batch_size = 2048, epochs = 30, callbacks = callbacks, verbose = 2 )
## Epoch 1/30 ## 112/112 - 3s - 27ms/step - fn: 40.0000 - fp: 23705.0000 - loss: 2.2728e-06 - precision: 0.0148 - recall: 0.8990 - tn: 203745.0000 - tp: 356.0000 - val_fn: 8.0000 - val_fp: 1502.0000 - val_loss: 0.1089 - val_precision: 0.0553 - val_recall: 0.9167 - val_tn: 55363.0000 - val_tp: 88.0000 ## Epoch 2/30 ## 112/112 - 1s - 6ms/step - fn: 33.0000 - fp: 8009.0000 - loss: 1.4949e-06 - precision: 0.0434 - recall: 0.9167 - tn: 219441.0000 - tp: 363.0000 - val_fn: 10.0000 - val_fp: 1176.0000 - val_loss: 0.0912 - val_precision: 0.0681 - val_recall: 0.8958 - val_tn: 55689.0000 - val_tp: 86.0000 ## Epoch 3/30 ## 112/112 - 0s - 2ms/step - fn: 32.0000 - fp: 7704.0000 - loss: 1.3369e-06 - precision: 0.0451 - recall: 0.9192 - tn: 219746.0000 - tp: 364.0000 - val_fn: 9.0000 - val_fp: 1202.0000 - val_loss: 0.0870 - val_precision: 0.0675 - val_recall: 0.9062 - val_tn: 55663.0000 - val_tp: 87.0000 ## Epoch 4/30 ## 112/112 - 0s - 2ms/step - fn: 28.0000 - fp: 8366.0000 - loss: 1.2269e-06 - precision: 0.0421 - recall: 0.9293 - tn: 219084.0000 - tp: 368.0000 - val_fn: 10.0000 - val_fp: 1560.0000 - val_loss: 0.0967 - val_precision: 0.0522 - val_recall: 0.8958 - val_tn: 55305.0000 - val_tp: 86.0000 ## Epoch 5/30 ## 112/112 - 0s - 2ms/step - fn: 19.0000 - fp: 6217.0000 - loss: 8.5492e-07 - precision: 0.0572 - recall: 0.9520 - tn: 221233.0000 - tp: 377.0000 - val_fn: 9.0000 - val_fp: 2626.0000 - val_loss: 0.1236 - val_precision: 0.0321 - val_recall: 0.9062 - val_tn: 54239.0000 - val_tp: 87.0000 ## Epoch 6/30 ## 112/112 - 0s - 2ms/step - fn: 18.0000 - fp: 6566.0000 - loss: 7.4300e-07 - precision: 0.0544 - recall: 0.9545 - tn: 220884.0000 - tp: 378.0000 - val_fn: 8.0000 - val_fp: 2302.0000 - val_loss: 0.1081 - val_precision: 0.0368 - val_recall: 0.9167 - val_tn: 54563.0000 - val_tp: 88.0000 ## Epoch 7/30 ## 112/112 - 0s - 2ms/step - fn: 14.0000 - fp: 6819.0000 - loss: 7.1838e-07 - precision: 0.0530 - recall: 0.9646 - tn: 220631.0000 - tp: 382.0000 - val_fn: 7.0000 - val_fp: 1376.0000 - val_loss: 0.0800 - val_precision: 0.0608 - val_recall: 0.9271 - val_tn: 55489.0000 - val_tp: 89.0000 ## Epoch 8/30 ## 112/112 - 0s - 2ms/step - fn: 19.0000 - fp: 7345.0000 - loss: 8.3921e-07 - precision: 0.0488 - recall: 0.9520 - tn: 220105.0000 - tp: 377.0000 - val_fn: 10.0000 - val_fp: 416.0000 - val_loss: 0.0406 - val_precision: 0.1713 - val_recall: 0.8958 - val_tn: 56449.0000 - val_tp: 86.0000 ## Epoch 9/30 ## 112/112 - 0s - 2ms/step - fn: 13.0000 - fp: 6282.0000 - loss: 8.3488e-07 - precision: 0.0575 - recall: 0.9672 - tn: 221168.0000 - tp: 383.0000 - val_fn: 10.0000 - val_fp: 933.0000 - val_loss: 0.0450 - val_precision: 0.0844 - val_recall: 0.8958 - val_tn: 55932.0000 - val_tp: 86.0000 ## Epoch 10/30 ## 112/112 - 0s - 2ms/step - fn: 15.0000 - fp: 7597.0000 - loss: 1.1098e-06 - precision: 0.0478 - recall: 0.9621 - tn: 219853.0000 - tp: 381.0000 - val_fn: 10.0000 - val_fp: 2417.0000 - val_loss: 0.4261 - val_precision: 0.0344 - val_recall: 0.8958 - val_tn: 54448.0000 - val_tp: 86.0000 ## Epoch 11/30 ## 112/112 - 0s - 2ms/step - fn: 24.0000 - fp: 10269.0000 - loss: 1.9062e-06 - precision: 0.0350 - recall: 0.9394 - tn: 217181.0000 - tp: 372.0000 - val_fn: 9.0000 - val_fp: 1434.0000 - val_loss: 0.1261 - val_precision: 0.0572 - val_recall: 0.9062 - val_tn: 55431.0000 - val_tp: 87.0000 ## Epoch 12/30 ## 112/112 - 0s - 2ms/step - fn: 17.0000 - fp: 6634.0000 - loss: 1.0711e-06 - precision: 0.0540 - recall: 0.9571 - tn: 220816.0000 - tp: 379.0000 - val_fn: 8.0000 - val_fp: 2263.0000 - val_loss: 0.1654 - val_precision: 0.0374 - val_recall: 0.9167 - val_tn: 54602.0000 - val_tp: 88.0000 ## Epoch 13/30 ## 112/112 - 0s - 2ms/step - fn: 10.0000 - fp: 5620.0000 - loss: 5.3600e-07 - precision: 0.0643 - recall: 0.9747 - tn: 221830.0000 - tp: 386.0000 - val_fn: 12.0000 - val_fp: 914.0000 - val_loss: 0.0414 - val_precision: 0.0842 - val_recall: 0.8750 - val_tn: 55951.0000 - val_tp: 84.0000 ## Epoch 14/30 ## 112/112 - 0s - 2ms/step - fn: 8.0000 - fp: 5103.0000 - loss: 5.1168e-07 - precision: 0.0707 - recall: 0.9798 - tn: 222347.0000 - tp: 388.0000 - val_fn: 11.0000 - val_fp: 1382.0000 - val_loss: 0.0660 - val_precision: 0.0579 - val_recall: 0.8854 - val_tn: 55483.0000 - val_tp: 85.0000 ## Epoch 15/30 ## 112/112 - 0s - 2ms/step - fn: 7.0000 - fp: 4497.0000 - loss: 4.4632e-07 - precision: 0.0796 - recall: 0.9823 - tn: 222953.0000 - tp: 389.0000 - val_fn: 10.0000 - val_fp: 1805.0000 - val_loss: 0.0786 - val_precision: 0.0455 - val_recall: 0.8958 - val_tn: 55060.0000 - val_tp: 86.0000 ## Epoch 16/30 ## 112/112 - 0s - 2ms/step - fn: 5.0000 - fp: 5056.0000 - loss: 6.1038e-07 - precision: 0.0718 - recall: 0.9874 - tn: 222394.0000 - tp: 391.0000 - val_fn: 13.0000 - val_fp: 889.0000 - val_loss: 0.0676 - val_precision: 0.0854 - val_recall: 0.8646 - val_tn: 55976.0000 - val_tp: 83.0000 ## Epoch 17/30 ## 112/112 - 0s - 2ms/step - fn: 3.0000 - fp: 4639.0000 - loss: 4.5763e-07 - precision: 0.0781 - recall: 0.9924 - tn: 222811.0000 - tp: 393.0000 - val_fn: 11.0000 - val_fp: 1185.0000 - val_loss: 0.0669 - val_precision: 0.0669 - val_recall: 0.8854 - val_tn: 55680.0000 - val_tp: 85.0000 ## Epoch 18/30 ## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 4355.0000 - loss: 4.5299e-07 - precision: 0.0826 - recall: 0.9899 - tn: 223095.0000 - tp: 392.0000 - val_fn: 9.0000 - val_fp: 1405.0000 - val_loss: 0.0772 - val_precision: 0.0583 - val_recall: 0.9062 - val_tn: 55460.0000 - val_tp: 87.0000 ## Epoch 19/30 ## 112/112 - 0s - 2ms/step - fn: 6.0000 - fp: 5779.0000 - loss: 4.8324e-07 - precision: 0.0632 - recall: 0.9848 - tn: 221671.0000 - tp: 390.0000 - val_fn: 10.0000 - val_fp: 1135.0000 - val_loss: 0.0546 - val_precision: 0.0704 - val_recall: 0.8958 - val_tn: 55730.0000 - val_tp: 86.0000 ## Epoch 20/30 ## 112/112 - 0s - 2ms/step - fn: 3.0000 - fp: 4748.0000 - loss: 3.7424e-07 - precision: 0.0764 - recall: 0.9924 - tn: 222702.0000 - tp: 393.0000 - val_fn: 9.0000 - val_fp: 895.0000 - val_loss: 0.0396 - val_precision: 0.0886 - val_recall: 0.9062 - val_tn: 55970.0000 - val_tp: 87.0000 ## Epoch 21/30 ## 112/112 - 0s - 2ms/step - fn: 3.0000 - fp: 3798.0000 - loss: 3.1417e-07 - precision: 0.0938 - recall: 0.9924 - tn: 223652.0000 - tp: 393.0000 - val_fn: 10.0000 - val_fp: 652.0000 - val_loss: 0.0309 - val_precision: 0.1165 - val_recall: 0.8958 - val_tn: 56213.0000 - val_tp: 86.0000 ## Epoch 22/30 ## 112/112 - 0s - 2ms/step - fn: 0.0000e+00 - fp: 2188.0000 - loss: 1.8112e-07 - precision: 0.1533 - recall: 1.0000 - tn: 225262.0000 - tp: 396.0000 - val_fn: 10.0000 - val_fp: 427.0000 - val_loss: 0.0250 - val_precision: 0.1676 - val_recall: 0.8958 - val_tn: 56438.0000 - val_tp: 86.0000 ## Epoch 23/30 ## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 3066.0000 - loss: 3.4932e-07 - precision: 0.1134 - recall: 0.9899 - tn: 224384.0000 - tp: 392.0000 - val_fn: 11.0000 - val_fp: 1053.0000 - val_loss: 0.0494 - val_precision: 0.0747 - val_recall: 0.8854 - val_tn: 55812.0000 - val_tp: 85.0000 ## Epoch 24/30 ## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 3175.0000 - loss: 5.1994e-07 - precision: 0.1099 - recall: 0.9899 - tn: 224275.0000 - tp: 392.0000 - val_fn: 11.0000 - val_fp: 1012.0000 - val_loss: 0.0997 - val_precision: 0.0775 - val_recall: 0.8854 - val_tn: 55853.0000 - val_tp: 85.0000 ## Epoch 25/30 ## 112/112 - 0s - 2ms/step - fn: 7.0000 - fp: 5874.0000 - loss: 6.5862e-07 - precision: 0.0621 - recall: 0.9823 - tn: 221576.0000 - tp: 389.0000 - val_fn: 9.0000 - val_fp: 3090.0000 - val_loss: 0.1296 - val_precision: 0.0274 - val_recall: 0.9062 - val_tn: 53775.0000 - val_tp: 87.0000 ## Epoch 26/30 ## 112/112 - 0s - 2ms/step - fn: 6.0000 - fp: 5395.0000 - loss: 5.9253e-07 - precision: 0.0674 - recall: 0.9848 - tn: 222055.0000 - tp: 390.0000 - val_fn: 9.0000 - val_fp: 1320.0000 - val_loss: 0.0928 - val_precision: 0.0618 - val_recall: 0.9062 - val_tn: 55545.0000 - val_tp: 87.0000 ## Epoch 27/30 ## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 3887.0000 - loss: 7.1278e-07 - precision: 0.0916 - recall: 0.9899 - tn: 223563.0000 - tp: 392.0000 - val_fn: 12.0000 - val_fp: 572.0000 - val_loss: 0.0352 - val_precision: 0.1280 - val_recall: 0.8750 - val_tn: 56293.0000 - val_tp: 84.0000 ## Epoch 28/30 ## 112/112 - 0s - 2ms/step - fn: 3.0000 - fp: 2209.0000 - loss: 5.8153e-07 - precision: 0.1510 - recall: 0.9924 - tn: 225241.0000 - tp: 393.0000 - val_fn: 11.0000 - val_fp: 1373.0000 - val_loss: 0.0808 - val_precision: 0.0583 - val_recall: 0.8854 - val_tn: 55492.0000 - val_tp: 85.0000 ## Epoch 29/30 ## 112/112 - 0s - 2ms/step - fn: 1.0000 - fp: 2506.0000 - loss: 2.9031e-07 - precision: 0.1362 - recall: 0.9975 - tn: 224944.0000 - tp: 395.0000 - val_fn: 11.0000 - val_fp: 690.0000 - val_loss: 0.0401 - val_precision: 0.1097 - val_recall: 0.8854 - val_tn: 56175.0000 - val_tp: 85.0000 ## Epoch 30/30 ## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 2719.0000 - loss: 6.1097e-07 - precision: 0.1260 - recall: 0.9899 - tn: 224731.0000 - tp: 392.0000 - val_fn: 13.0000 - val_fp: 861.0000 - val_loss: 0.0362 - val_precision: 0.0879 - val_recall: 0.8646 - val_tn: 56004.0000 - val_tp: 83.0000
val_pred <- model %>% predict(val_features) %>% { as.integer(. > 0.5) }
## 1781/1781 - 1s - 311us/step
pred_correct <- val_df$Class == val_pred cat(sprintf("Validation accuracy: %.2f", mean(pred_correct)))
## Validation accuracy: 0.98
fraudulent <- val_df$Class == 1 n_fraudulent_detected <- sum(fraudulent & pred_correct) n_fraudulent_missed <- sum(fraudulent & !pred_correct) n_legitimate_flagged <- sum(!fraudulent & !pred_correct)
Conclusions
At the end of training, out of 56,961 validation transactions, we are:
- Correctly identifying 83 of them as fraudulent
- Missing 13 fraudulent transactions
- At the cost of incorrectly flagging 861 legitimate transactions
In the real world, one would put an even higher weight on class 1, so as to reflect that False Negatives are more costly than False Positives.
Next time your credit card gets declined in an online purchase -- this is why.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.