Structured data classification with FeatureSpace

Introduction

This example demonstrates how to do structured data classification (also known as tabular data classification), starting from a raw CSV file. Our data includes numerical features, and integer categorical features, and string categorical features. We will use the utility layer_feature_space() to index, preprocess, and encode our features.

The code is adapted from the example Structured data classification from scratch. While the previous example managed its own low-level feature preprocessing and encoding with Keras preprocessing layers, in this example we delegate everything to layer_feature_space(), making the workflow extremely quick and easy.

The dataset

Our dataset is provided by the Cleveland Clinic Foundation for Heart Disease. It’s a CSV file with 303 rows. Each row contains information about a patient (a sample), and each column describes an attribute of the patient (a feature). We use the features to predict whether a patient has a heart disease (binary classification).

Here’s the description of each feature:

Column	Description	Feature Type
Age	Age in years	Numerical
Sex	(1 = male; 0 = female)	Categorical
CP	Chest pain type (0, 1, 2, 3, 4)	Categorical
Trestbpd	Resting blood pressure (in mm Hg on admission)	Numerical
Chol	Serum cholesterol in mg/dl	Numerical
FBS	fasting blood sugar in 120 mg/dl (1 = true; 0 = false)	Categorical
RestECG	Resting electrocardiogram results (0, 1, 2)	Categorical
Thalach	Maximum heart rate achieved	Numerical
Exang	Exercise induced angina (1 = yes; 0 = no)	Categorical
Oldpeak	ST depression induced by exercise relative to rest	Numerical
Slope	Slope of the peak exercise ST segment	Numerical
CA	Number of major vessels (0-3) colored by fluoroscopy	Both numerical & categorical
Thal	3 = normal; 6 = fixed defect; 7 = reversible defect	Categorical
Target	Diagnosis of heart disease (1 = true; 0 = false)	Target

Setup

library(readr)
library(dplyr, warn.conflicts = FALSE)
library(keras3)
library(tensorflow, exclude = c("shape", "set_random_seed"))
library(tfdatasets, exclude = "shape")

conflicted::conflicts_prefer(
  keras3::shape(),
  keras3::set_random_seed(),
  dplyr::filter(),
  .quiet = TRUE
)

use_backend("tensorflow")

Preparing the data

Let’s download the data and load it into a Pandas dataframe:

file_url <-
  "http://storage.googleapis.com/download.tensorflow.org/data/heart.csv"
df <- read_csv(file_url, col_types = cols(
  oldpeak = col_double(),
  thal = col_character(),
  .default = col_integer()
))

# the dataset has two malformed rows, filter them out
df <- df |> filter(!thal %in% c("1", "2"))

The dataset includes 303 samples with 14 columns per sample (13 features, plus the target label)

glimpse(df)

## Rows: 301
## Columns: 14
## $ age      <int> 63, 67, 67, 37, 41, 56, 62, 57, 63, 53, 57, 56, 56, 44, 5…
## $ sex      <int> 1, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 0, …
## $ cp       <int> 1, 4, 4, 3, 2, 2, 4, 4, 4, 4, 4, 2, 3, 2, 3, 3, 2, 4, 3, …
## $ trestbps <int> 145, 160, 120, 130, 130, 120, 140, 120, 130, 140, 140, 14…
## $ chol     <int> 233, 286, 229, 250, 204, 236, 268, 354, 254, 203, 192, 29…
## $ fbs      <int> 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0, 0, …
## $ restecg  <int> 2, 2, 2, 0, 2, 0, 2, 0, 2, 2, 0, 2, 2, 0, 0, 0, 0, 0, 0, …
## $ thalach  <int> 150, 108, 129, 187, 172, 178, 160, 163, 147, 155, 148, 15…
## $ exang    <int> 0, 1, 1, 0, 0, 0, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, …
## $ oldpeak  <dbl> 2.3, 1.5, 2.6, 3.5, 1.4, 0.8, 3.6, 0.6, 1.4, 3.1, 0.4, 1.…
## $ slope    <int> 3, 2, 2, 3, 1, 1, 3, 1, 2, 3, 2, 2, 2, 1, 1, 1, 3, 1, 1, …
## $ ca       <int> 0, 3, 2, 0, 0, 0, 2, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, …
## $ thal     <chr> "fixed", "normal", "reversible", "normal", "normal", "nor…
## $ target   <int> 0, 1, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, …

Here’s a preview of a few samples:

df

## # A tibble: 301 × 14
##      age   sex    cp trestbps  chol   fbs restecg thalach exang oldpeak
##    <int> <int> <int>    <int> <int> <int>   <int>   <int> <int>   <dbl>
##  1    63     1     1      145   233     1       2     150     0     2.3
##  2    67     1     4      160   286     0       2     108     1     1.5
##  3    67     1     4      120   229     0       2     129     1     2.6
##  4    37     1     3      130   250     0       0     187     0     3.5
##  5    41     0     2      130   204     0       2     172     0     1.4
##  6    56     1     2      120   236     0       0     178     0     0.8
##  7    62     0     4      140   268     0       2     160     0     3.6
##  8    57     0     4      120   354     0       0     163     1     0.6
##  9    63     1     4      130   254     0       2     147     0     1.4
## 10    53     1     4      140   203     1       2     155     1     3.1
## # ℹ 291 more rows
## # ℹ 4 more variables: slope <int>, ca <int>, thal <chr>, target <int>

The last column, “target”, indicates whether the patient has a heart disease (1) or not (0).

Let’s split the data into a training and validation set:

val_idx <- nrow(df) %>% sample.int(., . * 0.2)
val_df <- df[val_idx, ]
train_df <- df[-val_idx, ]

cat(sprintf(
  "Using %d samples for training and %d for validation",
  nrow(train_df), nrow(val_df)
))

## Using 241 samples for training and 60 for validation

Let’s generate tf_dataset objects for each dataframe:

dataframe_to_dataset <- function(df) {
  labels <- df |> pull(target) |> as.integer()
  inputs <- df |> select(-target) |> as.list()

  ds <- tensor_slices_dataset(list(inputs, labels)) |>
    dataset_shuffle(nrow(df))

  ds
}

train_ds <- dataframe_to_dataset(train_df)
val_ds <- dataframe_to_dataset(val_df)

Each tf_dataset yields a tuple (input, target) where input is a dictionary (a named list) of features and target is the value 0 or 1:

c(x, y) %<-% iter_next(as_iterator(train_ds))
cat("Input: "); str(x)
cat("Target: "); str(y)

## Input: List of 13
##  $ age     :<tf.Tensor: shape=(), dtype=int32, numpy=45>
##  $ sex     :<tf.Tensor: shape=(), dtype=int32, numpy=1>
##  $ cp      :<tf.Tensor: shape=(), dtype=int32, numpy=1>
##  $ trestbps:<tf.Tensor: shape=(), dtype=int32, numpy=110>
##  $ chol    :<tf.Tensor: shape=(), dtype=int32, numpy=264>
##  $ fbs     :<tf.Tensor: shape=(), dtype=int32, numpy=0>
##  $ restecg :<tf.Tensor: shape=(), dtype=int32, numpy=0>
##  $ thalach :<tf.Tensor: shape=(), dtype=int32, numpy=132>
##  $ exang   :<tf.Tensor: shape=(), dtype=int32, numpy=0>
##  $ oldpeak :<tf.Tensor: shape=(), dtype=float32, numpy=1.2>
##  $ slope   :<tf.Tensor: shape=(), dtype=int32, numpy=2>
##  $ ca      :<tf.Tensor: shape=(), dtype=int32, numpy=0>
##  $ thal    :<tf.Tensor: shape=(), dtype=string, numpy=b'reversible'>
## Target: <tf.Tensor: shape=(), dtype=int32, numpy=0>

Let’s batch the datasets:

train_ds <- train_ds |> dataset_batch(32)
val_ds <- val_ds |> dataset_batch(32)

Configuring a FeatureSpace

To configure how each feature should be preprocessed, we instantiate a layer_feature_space(), and we pass to it a dictionary (named list with unique names) that maps the name of our features to a string that describes the feature type.

We have a few “integer categorical” features such as "FBS", one “string categorical” feature ("thal"), and a few numerical features, which we’d like to normalize – except "age", which we’d like to discretize into a number of bins.

We also use the crosses argument to capture feature interactions for some categorical features, that is to say, create additional features that represent value co-occurrences for these categorical features. You can compute feature crosses like this for arbitrary sets of categorical features – not just tuples of two features. Because the resulting co-occurences are hashed into a fixed-sized vector, you don’t need to worry about whether the co-occurence space is too large.

feature_space <- layer_feature_space(
  features = list(
    # Categorical features encoded as integers
    sex = "integer_categorical",
    cp = "integer_categorical",
    fbs = "integer_categorical",
    restecg = "integer_categorical",
    exang = "integer_categorical",
    ca = "integer_categorical",
    # Categorical feature encoded as string
    thal = "string_categorical",
    # Numerical features to discretize
    age = "float_discretized",
    # Numerical features to normalize
    trestbps = "float_normalized",
    chol = "float_normalized",
    thalach = "float_normalized",
    oldpeak = "float_normalized",
    slope = "float_normalized"
  ),
  # We create additional features by hashing
  # value co-occurrences for the
  # following groups of categorical features.
  crosses = list(c("sex", "age"), c("thal", "ca")),
  # The hashing space for these co-occurrences
  # wil be 32-dimensional.
  crossing_dim = 32,
  # Our utility will one-hot encode all categorical
  # features and concat all features into a single
  # vector (one vector per sample).
  output_mode = "concat"
)

Further customizing a FeatureSpace

Specifying the feature type via a string name is quick and easy, but sometimes you may want to further configure the preprocessing of each feature. For instance, in our case, our categorical features don’t have a large set of possible values – it’s only a handful of values per feature (e.g. 1 and 0 for the feature "FBS"), and all possible values are represented in the training set. As a result, we don’t need to reserve an index to represent “out of vocabulary” values for these features – which would have been the default behavior. Below, we just specify num_oov_indices=0 in each of these features to tell the feature preprocessor to skip “out of vocabulary” indexing.

Other customizations you have access to include specifying the number of bins for discretizing features of type "float_discretized", or the dimensionality of the hashing space for feature crossing.

feature_space <- layer_feature_space(
  features = list(
    # Categorical features encoded as integers
    sex       = feature_integer_categorical(num_oov_indices = 0),
    cp        = feature_integer_categorical(num_oov_indices = 0),
    fbs       = feature_integer_categorical(num_oov_indices = 0),
    restecg   = feature_integer_categorical(num_oov_indices = 0),
    exang     = feature_integer_categorical(num_oov_indices = 0),
    ca        = feature_integer_categorical(num_oov_indices = 0),
    # Categorical feature encoded as string
    thal      = feature_string_categorical(num_oov_indices = 0),
    # Numerical features to discretize
    age       = feature_float_discretized(num_bins = 30),
    # Numerical features to normalize
    trestbps  = feature_float_normalized(),
    chol      = feature_float_normalized(),
    thalach   = feature_float_normalized(),
    oldpeak   = feature_float_normalized(),
    slope     = feature_float_normalized()
  ),
  # Specify feature cross with a custom crossing dim.
  crosses = list(
    feature_cross(
      feature_names = c("sex", "age"),
      crossing_dim = 64
    ),
    feature_cross(
      feature_names = c("thal", "ca"),
      crossing_dim = 16
    )
  ),
  output_mode = "concat"
)

Adapt the FeatureSpace to the training data

Before we start using the FeatureSpace to build a model, we have to adapt it to the training data. During adapt(), the FeatureSpace will:

• Index the set of possible values for categorical features.
• Compute the mean and variance for numerical features to normalize.
• Compute the value boundaries for the different bins for numerical features to discretize.

Note that adapt() should be called on a tf_dataset which yields dicts (named lists) of feature values – no labels.

train_ds_with_no_labels <- train_ds |> dataset_map(\(x, y) x)
feature_space |> adapt(train_ds_with_no_labels)

At this point, the FeatureSpace can be called on a dict of raw feature values, and will return a single concatenate vector for each sample, combining encoded features and feature crosses.

c(x, y) %<-% iter_next(as_iterator(train_ds))
preprocessed_x <- feature_space(x)
preprocessed_x

## tf.Tensor(
## [[0. 0. 0. ... 0. 0. 0.]
##  [0. 0. 0. ... 0. 0. 0.]
##  [0. 0. 0. ... 0. 0. 0.]
##  ...
##  [0. 0. 0. ... 0. 0. 0.]
##  [0. 0. 0. ... 0. 0. 0.]
##  [0. 0. 0. ... 0. 0. 0.]], shape=(32, 136), dtype=float32)

Two ways to manage preprocessing: as part of the tf.data pipeline, or in the model itself

There are two ways in which you can leverage your FeatureSpace:

Asynchronous preprocessing in tf.data

You can make it part of your data pipeline, before the model. This enables asynchronous parallel preprocessing of the data on CPU before it hits the model. Do this if you’re training on GPU or TPU, or if you want to speed up preprocessing. Usually, this is always the right thing to do during training.

Synchronous preprocessing in the model

You can make it part of your model. This means that the model will expect dicts of raw feature values, and the preprocessing batch will be done synchronously (in a blocking manner) before the rest of the forward pass. Do this if you want to have an end-to-end model that can process raw feature values – but keep in mind that your model will only be able to run on CPU, since most types of feature preprocessing (e.g. string preprocessing) are not GPU or TPU compatible.

Do not do this on GPU / TPU or in performance-sensitive settings. In general, you want to do in-model preprocessing when you do inference on CPU.

In our case, we will apply the FeatureSpace in the tf.data pipeline during training, but we will do inference with an end-to-end model that includes the FeatureSpace.

Let’s create a training and validation dataset of preprocessed batches:

preprocessed_train_ds <- train_ds |>
  dataset_map(\(x, y) list(feature_space(x), y),
              num_parallel_calls = tf$data$AUTOTUNE) |>
  dataset_prefetch(tf$data$AUTOTUNE)

preprocessed_val_ds <- val_ds |>
  dataset_map(\(x, y) list(feature_space(x), y),
              num_parallel_calls = tf$data$AUTOTUNE) |>
  dataset_prefetch(tf$data$AUTOTUNE)

Build a model

Time to build a model – or rather two models:

• A training model that expects preprocessed features (one sample = one vector)
• An inference model that expects raw features (one sample = dict of raw feature values)

dict_inputs <- feature_space$get_inputs()
encoded_features <- feature_space$get_encoded_features()

predictions <- encoded_features |>
  layer_dense(32, activation="relu") |>
  layer_dropout(0.5) |>
  layer_dense(1, activation="sigmoid")

training_model <- keras_model(inputs = encoded_features,
                              outputs = predictions)
training_model |> compile(optimizer = "adam",
                          loss = "binary_crossentropy",
                          metrics = "accuracy")

inference_model <- keras_model(inputs = dict_inputs,
                               outputs = predictions)

Train the model

Let’s train our model for 20 epochs. Note that feature preprocessing is happening as part of the tf.data pipeline, not as part of the model.

training_model |> fit(
  preprocessed_train_ds,
  epochs = 20,
  validation_data = preprocessed_val_ds,
  verbose = 2
)

## Epoch 1/20
## 8/8 - 2s - 280ms/step - accuracy: 0.4689 - loss: 0.7471 - val_accuracy: 0.5167 - val_loss: 0.7019
## Epoch 2/20
## 8/8 - 1s - 140ms/step - accuracy: 0.5602 - loss: 0.6785 - val_accuracy: 0.6333 - val_loss: 0.6491
## Epoch 3/20
## 8/8 - 0s - 46ms/step - accuracy: 0.6307 - loss: 0.6478 - val_accuracy: 0.7000 - val_loss: 0.6053
## Epoch 4/20
## 8/8 - 0s - 12ms/step - accuracy: 0.6432 - loss: 0.6246 - val_accuracy: 0.7667 - val_loss: 0.5692
## Epoch 5/20
## 8/8 - 0s - 13ms/step - accuracy: 0.7178 - loss: 0.5813 - val_accuracy: 0.7667 - val_loss: 0.5359
## Epoch 6/20
## 8/8 - 0s - 13ms/step - accuracy: 0.7344 - loss: 0.5371 - val_accuracy: 0.7833 - val_loss: 0.5067
## Epoch 7/20
## 8/8 - 0s - 13ms/step - accuracy: 0.7884 - loss: 0.5158 - val_accuracy: 0.8333 - val_loss: 0.4810
## Epoch 8/20
## 8/8 - 0s - 13ms/step - accuracy: 0.7759 - loss: 0.5011 - val_accuracy: 0.8500 - val_loss: 0.4569
## Epoch 9/20
## 8/8 - 0s - 13ms/step - accuracy: 0.7676 - loss: 0.4865 - val_accuracy: 0.8500 - val_loss: 0.4354
## Epoch 10/20
## 8/8 - 0s - 13ms/step - accuracy: 0.7925 - loss: 0.4601 - val_accuracy: 0.8333 - val_loss: 0.4161
## Epoch 11/20
## 8/8 - 0s - 13ms/step - accuracy: 0.7967 - loss: 0.4617 - val_accuracy: 0.8667 - val_loss: 0.3976
## Epoch 12/20
## 8/8 - 0s - 13ms/step - accuracy: 0.7967 - loss: 0.4316 - val_accuracy: 0.8667 - val_loss: 0.3796
## Epoch 13/20
## 8/8 - 0s - 13ms/step - accuracy: 0.8506 - loss: 0.4058 - val_accuracy: 0.8833 - val_loss: 0.3643
## Epoch 14/20
## 8/8 - 0s - 13ms/step - accuracy: 0.8174 - loss: 0.4197 - val_accuracy: 0.8833 - val_loss: 0.3510
## Epoch 15/20
## 8/8 - 0s - 14ms/step - accuracy: 0.8299 - loss: 0.3888 - val_accuracy: 0.8833 - val_loss: 0.3405
## Epoch 16/20
## 8/8 - 0s - 13ms/step - accuracy: 0.8257 - loss: 0.3820 - val_accuracy: 0.8833 - val_loss: 0.3294
## Epoch 17/20
## 8/8 - 0s - 13ms/step - accuracy: 0.8299 - loss: 0.3746 - val_accuracy: 0.8833 - val_loss: 0.3223
## Epoch 18/20
## 8/8 - 0s - 13ms/step - accuracy: 0.8506 - loss: 0.3487 - val_accuracy: 0.8833 - val_loss: 0.3153
## Epoch 19/20
## 8/8 - 0s - 14ms/step - accuracy: 0.8465 - loss: 0.3558 - val_accuracy: 0.8667 - val_loss: 0.3093
## Epoch 20/20
## 8/8 - 0s - 14ms/step - accuracy: 0.8672 - loss: 0.3570 - val_accuracy: 0.8667 - val_loss: 0.3036

We quickly get to 80% validation accuracy.

Inference on new data with the end-to-end model

Now, we can use our inference model (which includes the FeatureSpace) to make predictions based on dicts of raw features values, as follows:

sample <- list(
  age = 60,
  sex = 1,
  cp = 1,
  trestbps = 145,
  chol = 233,
  fbs = 1,
  restecg = 2,
  thalach = 150,
  exang = 0,
  oldpeak = 2.3,
  slope = 3,
  ca = 0,
  thal = "fixed"
)

input_dict <- lapply(sample, \(x) op_convert_to_tensor(array(x)))
predictions <- inference_model |> predict(input_dict)

## 1/1 - 0s - 394ms/step

glue::glue(r"---(
  This particular patient had a {(100 * predictions) |> signif(3)}% probability
  of having a heart disease, as evaluated by our model.
)---")

## This particular patient had a 44.8% probability
## of having a heart disease, as evaluated by our model.