Imbalanced classification: credit card fraud detection
Source:vignettes/examples/structured_data/imbalanced_classification.Rmd
imbalanced_classification.Rmd
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 <dbl> 0, 0, 1, 1, 2, 2, 4, 7, 7, 9, 10, 10, 10, 11, 12, 12, 12, 1…
## $ V1 <dbl> -1.3598071, 1.1918571, -1.3583541, -0.9662717, -1.1582331, …
## $ V2 <dbl> -0.07278117, 0.26615071, -1.34016307, -0.18522601, 0.877736…
## $ V3 <dbl> 2.53634674, 0.16648011, 1.77320934, 1.79299334, 1.54871785,…
## $ V4 <dbl> 1.37815522, 0.44815408, 0.37977959, -0.86329128, 0.40303393…
## $ V5 <dbl> -0.33832077, 0.06001765, -0.50319813, -0.01030888, -0.40719…
## $ V6 <dbl> 0.46238778, -0.08236081, 1.80049938, 1.24720317, 0.09592146…
## $ V7 <dbl> 0.239598554, -0.078802983, 0.791460956, 0.237608940, 0.5929…
## $ V8 <dbl> 0.098697901, 0.085101655, 0.247675787, 0.377435875, -0.2705…
## $ V9 <dbl> 0.3637870, -0.2554251, -1.5146543, -1.3870241, 0.8177393, -…
## $ V10 <dbl> 0.09079417, -0.16697441, 0.20764287, -0.05495192, 0.7530744…
## $ V11 <dbl> -0.55159953, 1.61272666, 0.62450146, -0.22648726, -0.822842…
## $ V12 <dbl> -0.61780086, 1.06523531, 0.06608369, 0.17822823, 0.53819555…
## $ V13 <dbl> -0.99138985, 0.48909502, 0.71729273, 0.50775687, 1.34585159…
## $ V14 <dbl> -0.31116935, -0.14377230, -0.16594592, -0.28792375, -1.1196…
## $ V15 <dbl> 1.468176972, 0.635558093, 2.345864949, -0.631418118, 0.1751…
## $ V16 <dbl> -0.47040053, 0.46391704, -2.89008319, -1.05964725, -0.45144…
## $ V17 <dbl> 0.207971242, -0.114804663, 1.109969379, -0.684092786, -0.23…
## $ V18 <dbl> 0.02579058, -0.18336127, -0.12135931, 1.96577500, -0.038194…
## $ V19 <dbl> 0.40399296, -0.14578304, -2.26185710, -1.23262197, 0.803486…
## $ V20 <dbl> 0.25141210, -0.06908314, 0.52497973, -0.20803778, 0.4085423…
## $ V21 <dbl> -0.018306778, -0.225775248, 0.247998153, -0.108300452, -0.0…
## $ V22 <dbl> 0.277837576, -0.638671953, 0.771679402, 0.005273597, 0.7982…
## $ V23 <dbl> -0.110473910, 0.101288021, 0.909412262, -0.190320519, -0.13…
## $ V24 <dbl> 0.06692807, -0.33984648, -0.68928096, -1.17557533, 0.141266…
## $ V25 <dbl> 0.12853936, 0.16717040, -0.32764183, 0.64737603, -0.2060095…
## $ V26 <dbl> -0.18911484, 0.12589453, -0.13909657, -0.22192884, 0.502292…
## $ V27 <dbl> 0.133558377, -0.008983099, -0.055352794, 0.062722849, 0.219…
## $ V28 <dbl> -0.021053053, 0.014724169, -0.059751841, 0.061457629, 0.215…
## $ Amount <dbl> 149.62, 2.69, 378.66, 123.50, 69.99, 3.67, 4.99, 40.80, 93.…
## $ Class <int> 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## Number of validation samples: 56961Analyze class imbalance in the targets
counts <- table(train_df$Class)
counts##
## 0 1
## 227455 391
cat(sprintf("Number of positive samples in training data: %i (%.2f%% of total)",
counts["1"], 100 * counts["1"] / sum(counts)))## Number of positive samples in training data: 391 (0.17% of total)
weight_for_0 = 1 / counts["0"]
weight_for_1 = 1 / counts["1"]Normalize the data using training set statistics
feature_names <- colnames(train_df) %>% setdiff("Class")
train_features <- as.matrix(train_df[feature_names])
train_targets <- as.matrix(train_df$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## Model: "sequential"
## ┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
## ┃ Layer (type) ┃ Output Shape ┃ Param # ┃
## ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
## │ dense (Dense) │ (None, 256) │ 7,936 │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dense_1 (Dense) │ (None, 256) │ 65,792 │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dropout (Dropout) │ (None, 256) │ 0 │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dense_2 (Dense) │ (None, 256) │ 65,792 │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dropout_1 (Dropout) │ (None, 256) │ 0 │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dense_3 (Dense) │ (None, 1) │ 257 │
## └─────────────────────────────────┴────────────────────────┴───────────────┘
## Total params: 139,777 (546.00 KB)
## Trainable params: 139,777 (546.00 KB)
## Non-trainable params: 0 (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 - 28ms/step - fn: 45.0000 - fp: 23707.0000 - loss: 2.2081e-06 - precision: 0.0144 - recall: 0.8849 - tn: 203748.0000 - tp: 346.0000 - val_fn: 9.0000 - val_fp: 2721.0000 - val_loss: 0.1728 - val_precision: 0.0327 - val_recall: 0.9109 - val_tn: 54139.0000 - val_tp: 92.0000
## Epoch 2/30
## 112/112 - 1s - 6ms/step - fn: 35.0000 - fp: 9264.0000 - loss: 1.6928e-06 - precision: 0.0370 - recall: 0.9105 - tn: 218191.0000 - tp: 356.0000 - val_fn: 9.0000 - val_fp: 1195.0000 - val_loss: 0.1100 - val_precision: 0.0715 - val_recall: 0.9109 - val_tn: 55665.0000 - val_tp: 92.0000
## Epoch 3/30
## 112/112 - 0s - 2ms/step - fn: 28.0000 - fp: 8603.0000 - loss: 1.2127e-06 - precision: 0.0405 - recall: 0.9284 - tn: 218852.0000 - tp: 363.0000 - val_fn: 6.0000 - val_fp: 4727.0000 - val_loss: 0.2568 - val_precision: 0.0197 - val_recall: 0.9406 - val_tn: 52133.0000 - val_tp: 95.0000
## Epoch 4/30
## 112/112 - 0s - 2ms/step - fn: 25.0000 - fp: 6697.0000 - loss: 1.0623e-06 - precision: 0.0518 - recall: 0.9361 - tn: 220758.0000 - tp: 366.0000 - val_fn: 12.0000 - val_fp: 1079.0000 - val_loss: 0.0645 - val_precision: 0.0762 - val_recall: 0.8812 - val_tn: 55781.0000 - val_tp: 89.0000
## Epoch 5/30
## 112/112 - 0s - 2ms/step - fn: 25.0000 - fp: 6461.0000 - loss: 9.4882e-07 - precision: 0.0536 - recall: 0.9361 - tn: 220994.0000 - tp: 366.0000 - val_fn: 8.0000 - val_fp: 2180.0000 - val_loss: 0.1322 - val_precision: 0.0409 - val_recall: 0.9208 - val_tn: 54680.0000 - val_tp: 93.0000
## Epoch 6/30
## 112/112 - 0s - 2ms/step - fn: 14.0000 - fp: 6436.0000 - loss: 7.0306e-07 - precision: 0.0553 - recall: 0.9642 - tn: 221019.0000 - tp: 377.0000 - val_fn: 12.0000 - val_fp: 927.0000 - val_loss: 0.0489 - val_precision: 0.0876 - val_recall: 0.8812 - val_tn: 55933.0000 - val_tp: 89.0000
## Epoch 7/30
## 112/112 - 0s - 2ms/step - fn: 20.0000 - fp: 8652.0000 - loss: 9.5322e-07 - precision: 0.0411 - recall: 0.9488 - tn: 218803.0000 - tp: 371.0000 - val_fn: 9.0000 - val_fp: 1213.0000 - val_loss: 0.0587 - val_precision: 0.0705 - val_recall: 0.9109 - val_tn: 55647.0000 - val_tp: 92.0000
## Epoch 8/30
## 112/112 - 0s - 2ms/step - fn: 16.0000 - fp: 8275.0000 - loss: 8.6911e-07 - precision: 0.0434 - recall: 0.9591 - tn: 219180.0000 - tp: 375.0000 - val_fn: 13.0000 - val_fp: 771.0000 - val_loss: 0.0680 - val_precision: 0.1024 - val_recall: 0.8713 - val_tn: 56089.0000 - val_tp: 88.0000
## Epoch 9/30
## 112/112 - 0s - 2ms/step - fn: 11.0000 - fp: 5889.0000 - loss: 6.6022e-07 - precision: 0.0606 - recall: 0.9719 - tn: 221566.0000 - tp: 380.0000 - val_fn: 9.0000 - val_fp: 564.0000 - val_loss: 0.0321 - val_precision: 0.1402 - val_recall: 0.9109 - val_tn: 56296.0000 - val_tp: 92.0000
## Epoch 10/30
## 112/112 - 0s - 2ms/step - fn: 9.0000 - fp: 5360.0000 - loss: 4.7577e-07 - precision: 0.0665 - recall: 0.9770 - tn: 222095.0000 - tp: 382.0000 - val_fn: 8.0000 - val_fp: 1314.0000 - val_loss: 0.0550 - val_precision: 0.0661 - val_recall: 0.9208 - val_tn: 55546.0000 - val_tp: 93.0000
## Epoch 11/30
## 112/112 - 0s - 2ms/step - fn: 20.0000 - fp: 9677.0000 - loss: 1.0933e-06 - precision: 0.0369 - recall: 0.9488 - tn: 217778.0000 - tp: 371.0000 - val_fn: 10.0000 - val_fp: 1706.0000 - val_loss: 0.0974 - val_precision: 0.0506 - val_recall: 0.9010 - val_tn: 55154.0000 - val_tp: 91.0000
## Epoch 12/30
## 112/112 - 0s - 2ms/step - fn: 9.0000 - fp: 5843.0000 - loss: 5.5728e-07 - precision: 0.0614 - recall: 0.9770 - tn: 221612.0000 - tp: 382.0000 - val_fn: 7.0000 - val_fp: 2226.0000 - val_loss: 0.1175 - val_precision: 0.0405 - val_recall: 0.9307 - val_tn: 54634.0000 - val_tp: 94.0000
## Epoch 13/30
## 112/112 - 0s - 2ms/step - fn: 12.0000 - fp: 8078.0000 - loss: 8.7332e-07 - precision: 0.0448 - recall: 0.9693 - tn: 219377.0000 - tp: 379.0000 - val_fn: 8.0000 - val_fp: 2571.0000 - val_loss: 0.0926 - val_precision: 0.0349 - val_recall: 0.9208 - val_tn: 54289.0000 - val_tp: 93.0000
## Epoch 14/30
## 112/112 - 0s - 2ms/step - fn: 9.0000 - fp: 6035.0000 - loss: 5.1023e-07 - precision: 0.0595 - recall: 0.9770 - tn: 221420.0000 - tp: 382.0000 - val_fn: 9.0000 - val_fp: 998.0000 - val_loss: 0.0420 - val_precision: 0.0844 - val_recall: 0.9109 - val_tn: 55862.0000 - val_tp: 92.0000
## Epoch 15/30
## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 4814.0000 - loss: 4.2882e-07 - precision: 0.0744 - recall: 0.9898 - tn: 222641.0000 - tp: 387.0000 - val_fn: 7.0000 - val_fp: 2388.0000 - val_loss: 0.2131 - val_precision: 0.0379 - val_recall: 0.9307 - val_tn: 54472.0000 - val_tp: 94.0000
## Epoch 16/30
## 112/112 - 0s - 2ms/step - fn: 14.0000 - fp: 6866.0000 - loss: 6.4459e-07 - precision: 0.0521 - recall: 0.9642 - tn: 220589.0000 - tp: 377.0000 - val_fn: 8.0000 - val_fp: 1106.0000 - val_loss: 0.0472 - val_precision: 0.0776 - val_recall: 0.9208 - val_tn: 55754.0000 - val_tp: 93.0000
## Epoch 17/30
## 112/112 - 0s - 2ms/step - fn: 8.0000 - fp: 7243.0000 - loss: 5.1227e-07 - precision: 0.0502 - recall: 0.9795 - tn: 220212.0000 - tp: 383.0000 - val_fn: 8.0000 - val_fp: 1185.0000 - val_loss: 0.0483 - val_precision: 0.0728 - val_recall: 0.9208 - val_tn: 55675.0000 - val_tp: 93.0000
## Epoch 18/30
## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 3529.0000 - loss: 3.2284e-07 - precision: 0.0988 - recall: 0.9898 - tn: 223926.0000 - tp: 387.0000 - val_fn: 8.0000 - val_fp: 1655.0000 - val_loss: 0.0717 - val_precision: 0.0532 - val_recall: 0.9208 - val_tn: 55205.0000 - val_tp: 93.0000
## Epoch 19/30
## 112/112 - 0s - 2ms/step - fn: 2.0000 - fp: 4189.0000 - loss: 3.3201e-07 - precision: 0.0850 - recall: 0.9949 - tn: 223266.0000 - tp: 389.0000 - val_fn: 10.0000 - val_fp: 775.0000 - val_loss: 0.0351 - val_precision: 0.1051 - val_recall: 0.9010 - val_tn: 56085.0000 - val_tp: 91.0000
## Epoch 20/30
## 112/112 - 0s - 2ms/step - fn: 3.0000 - fp: 3593.0000 - loss: 4.7423e-07 - precision: 0.0975 - recall: 0.9923 - tn: 223862.0000 - tp: 388.0000 - val_fn: 9.0000 - val_fp: 982.0000 - val_loss: 0.0459 - val_precision: 0.0857 - val_recall: 0.9109 - val_tn: 55878.0000 - val_tp: 92.0000
## Epoch 21/30
## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 4057.0000 - loss: 5.4189e-07 - precision: 0.0871 - recall: 0.9898 - tn: 223398.0000 - tp: 387.0000 - val_fn: 12.0000 - val_fp: 603.0000 - val_loss: 0.0544 - val_precision: 0.1286 - val_recall: 0.8812 - val_tn: 56257.0000 - val_tp: 89.0000
## Epoch 22/30
## 112/112 - 0s - 2ms/step - fn: 4.0000 - fp: 4709.0000 - loss: 5.5967e-07 - precision: 0.0759 - recall: 0.9898 - tn: 222746.0000 - tp: 387.0000 - val_fn: 8.0000 - val_fp: 1645.0000 - val_loss: 0.0862 - val_precision: 0.0535 - val_recall: 0.9208 - val_tn: 55215.0000 - val_tp: 93.0000
## Epoch 23/30
## 112/112 - 0s - 2ms/step - fn: 13.0000 - fp: 7337.0000 - loss: 9.8862e-07 - precision: 0.0490 - recall: 0.9668 - tn: 220118.0000 - tp: 378.0000 - val_fn: 11.0000 - val_fp: 683.0000 - val_loss: 0.0500 - val_precision: 0.1164 - val_recall: 0.8911 - val_tn: 56177.0000 - val_tp: 90.0000
## Epoch 24/30
## 112/112 - 0s - 2ms/step - fn: 8.0000 - fp: 4145.0000 - loss: 5.6664e-07 - precision: 0.0846 - recall: 0.9795 - tn: 223310.0000 - tp: 383.0000 - val_fn: 9.0000 - val_fp: 3412.0000 - val_loss: 0.1944 - val_precision: 0.0263 - val_recall: 0.9109 - val_tn: 53448.0000 - val_tp: 92.0000
## Epoch 25/30
## 112/112 - 0s - 2ms/step - fn: 2.0000 - fp: 4869.0000 - loss: 4.9343e-07 - precision: 0.0740 - recall: 0.9949 - tn: 222586.0000 - tp: 389.0000 - val_fn: 10.0000 - val_fp: 879.0000 - val_loss: 0.0453 - val_precision: 0.0938 - val_recall: 0.9010 - val_tn: 55981.0000 - val_tp: 91.0000
## Epoch 26/30
## 112/112 - 0s - 2ms/step - fn: 5.0000 - fp: 3203.0000 - loss: 3.2283e-07 - precision: 0.1076 - recall: 0.9872 - tn: 224252.0000 - tp: 386.0000 - val_fn: 9.0000 - val_fp: 1163.0000 - val_loss: 0.0681 - val_precision: 0.0733 - val_recall: 0.9109 - val_tn: 55697.0000 - val_tp: 92.0000
## Epoch 27/30
## 112/112 - 0s - 2ms/step - fn: 3.0000 - fp: 3169.0000 - loss: 3.0544e-07 - precision: 0.1091 - recall: 0.9923 - tn: 224286.0000 - tp: 388.0000 - val_fn: 9.0000 - val_fp: 1845.0000 - val_loss: 0.0807 - val_precision: 0.0475 - val_recall: 0.9109 - val_tn: 55015.0000 - val_tp: 92.0000
## Epoch 28/30
## 112/112 - 0s - 2ms/step - fn: 2.0000 - fp: 3077.0000 - loss: 2.6160e-07 - precision: 0.1122 - recall: 0.9949 - tn: 224378.0000 - tp: 389.0000 - val_fn: 12.0000 - val_fp: 275.0000 - val_loss: 0.0220 - val_precision: 0.2445 - val_recall: 0.8812 - val_tn: 56585.0000 - val_tp: 89.0000
## Epoch 29/30
## 112/112 - 0s - 2ms/step - fn: 2.0000 - fp: 2628.0000 - loss: 2.5775e-07 - precision: 0.1289 - recall: 0.9949 - tn: 224827.0000 - tp: 389.0000 - val_fn: 13.0000 - val_fp: 1013.0000 - val_loss: 0.0520 - val_precision: 0.0799 - val_recall: 0.8713 - val_tn: 55847.0000 - val_tp: 88.0000
## Epoch 30/30
## 112/112 - 0s - 2ms/step - fn: 2.0000 - fp: 4318.0000 - loss: 4.4087e-07 - precision: 0.0826 - recall: 0.9949 - tn: 223137.0000 - tp: 389.0000 - val_fn: 11.0000 - val_fp: 903.0000 - val_loss: 0.0561 - val_precision: 0.0906 - val_recall: 0.8911 - val_tn: 55957.0000 - val_tp: 90.0000
val_pred <- model %>%
predict(val_features) %>%
{ as.integer(. > 0.5) }## 1781/1781 - 1s - 424us/step
pred_correct <- val_df$Class == val_pred
cat(sprintf("Validation accuracy: %.2f", mean(pred_correct)))## Validation accuracy: 0.98Conclusions
At the end of training, out of 56,961 validation transactions, we are:
- Correctly identifying 90 of them as fraudulent
- Missing 11 fraudulent transactions
- At the cost of incorrectly flagging 903 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.