Introduction to processpredictR:
workflow
library(processpredictR)
library(bupaR)
library(ggplot2)
library(dplyr)
library(keras)
library(purrr)
Introduction
The goal of processpredictR is to perform prediction tasks on
processes using event logs and Transformer models. The 5 process
monitoring tasks are defined as follows:
- outcome: predict the case outcome, which can be the last
activity, or a manually defined variable
- next activity: predict the next activity instance
- remaining trace: predict the sequence of all next activity
instances
- next time: predict the start time of the next activity
instance
- remaining time: predict the remaining time till the end of
the case
The overall approach using processpredictR is shown in
the Figure below. prepare_examples() transforms logs into a
dataset that can be used for training and prediction, which is
thereafter split into train and test set. Subsequently a model is made,
compiled and fit. Finally, the model can be used to predict and can be
evaluated
Different levels of customization are offered. Using
create_model(), a standard off-the-shelf model can be
created for each of the supported tasks, including standard
features.
A first customization is to include additional features, such as case
or event attributes. These can be configured in the
prepare_examples() step, and they will be processed
automatically (normalized for numerical features, or hot-encoded for
categorical features).
A further way to customize your model, is to only generate the input
layer of the model with create_model(), and define the
remainder of the model yourself by adding keras layers
using the provided stack_layers() function.
Going beyond that, you can also create the model entirely yourself
using keras, including the preprocessing of the data.
Auxiliary functions are provided to help you with, e.g., tokenizing
activity sequences.
In the remainder of this tutorial, each of the steps and possible
avenues for customization will be described in more detail.
Preprocessing
As a first step in the process prediction workflow we use
prepare_examples() to obtain a dataset, where:
- each row/observation is a unique activity instance id
- the prefix(_list) column stores the sequence of activities already
executed in the case
- necessary features and target variables are calculated and/or
added
The returned object is of class ppred_examples_df, which
inherits from tbl_df.
In this tutorial we will use the traffic_fines event log
from eventdataR. Note that both eventlog and
activitylog objects, as defined by bupaR are
supported.
df <- prepare_examples(traffic_fines, task = "outcome")
df
#> # A tibble: 34,724 × 11
#> ith_case case_id prefix prefix_list outcome k activity resource
#> <int> <chr> <chr> <list> <fct> <dbl> <chr> <fct>
#> 1 1 A2127 Create Fine <chr [1]> Payment 0 Create … 537
#> 2 1 A2127 Create Fine - P… <chr [2]> Payment 1 Payment <NA>
#> 3 2 A15 Create Fine <chr [1]> Send f… 0 Create … 561
#> 4 2 A15 Create Fine - S… <chr [2]> Send f… 1 Send Fi… <NA>
#> 5 2 A15 Create Fine - S… <chr [3]> Send f… 2 Insert … <NA>
#> 6 2 A15 Create Fine - S… <chr [4]> Send f… 3 Add pen… <NA>
#> 7 2 A15 Create Fine - S… <chr [5]> Send f… 4 Send fo… <NA>
#> 8 3 A1820 Create Fine <chr [1]> Payment 0 Create … 563
#> 9 3 A1820 Create Fine - P… <chr [2]> Payment 1 Payment <NA>
#> 10 4 A22 Create Fine <chr [1]> Payment 0 Create … 561
#> # ℹ 34,714 more rows
#> # ℹ 3 more variables: start_time <dttm>, end_time <dttm>,
#> # remaining_trace_list <list>
We split the transformed dataset df into train- and test
sets for later use in fit() and predict(),
respectively. The proportion of the train set is configured with the
split argument.
set.seed(123)
split <- df %>% split_train_test(split = 0.8)
split$train_df %>% head(5)
#> # A tibble: 5 × 11
#> ith_case case_id prefix prefix_list outcome k activity resource
#> <int> <chr> <chr> <list> <fct> <dbl> <chr> <fct>
#> 1 1 A2127 Create Fine <chr [1]> Payment 0 Create … 537
#> 2 1 A2127 Create Fine - Pa… <chr [2]> Payment 1 Payment <NA>
#> 3 2 A15 Create Fine <chr [1]> Send f… 0 Create … 561
#> 4 2 A15 Create Fine - Se… <chr [2]> Send f… 1 Send Fi… <NA>
#> 5 2 A15 Create Fine - Se… <chr [3]> Send f… 2 Insert … <NA>
#> # ℹ 3 more variables: start_time <dttm>, end_time <dttm>,
#> # remaining_trace_list <list>
split$test_df %>% head(5)
#> # A tibble: 5 × 11
#> ith_case case_id prefix prefix_list outcome k activity resource
#> <int> <chr> <chr> <list> <fct> <dbl> <chr> <fct>
#> 1 8001 A24869 Create Fine <chr [1]> Payment 0 Create … 559
#> 2 8001 A24869 Create Fine - Pa… <chr [2]> Payment 1 Payment <NA>
#> 3 8002 A24871 Create Fine <chr [1]> Payment 0 Create … 559
#> 4 8002 A24871 Create Fine - Pa… <chr [2]> Payment 1 Payment <NA>
#> 5 8003 A24872 Create Fine <chr [1]> Send f… 0 Create … 559
#> # ℹ 3 more variables: start_time <dttm>, end_time <dttm>,
#> # remaining_trace_list <list>
It’s important to note that the split is done at case level (a case
is fully part of either the train data or either the test data).
Furthermore, the split is done chronologically, meaning that the train
set contains the split% first cases, and the test set contains the
(1-split)% last cases.
Note that because the split is done at case level, the percentage of
all examples in the train set can be slightly different, as cases differ
with respect their length.
nrow(split$train_df) / nrow(df)
#> [1] 0.8016934
n_distinct(split$train_df$case_id) / n_distinct(df$case_id)
#> [1] 0.8
Transformer model
The next step in the workflow is to build a model.
processpredictR provides a default set of functions that
are wrappers of generics provided by keras. For ease of
use, the preprocessing steps, such as tokenizing of sequences,
normalizing numerical features, etc. happen within the
create_model() function and are abstracted from the
user.
Define model
Based on the train set we define the default transformer model, using
create_model().
model <- split$train_df %>% create_model(name = "my_model")
# pass arguments as ... that are applicable to keras::keras_model()
model # is a list
#> Model: "my_model"
#> ________________________________________________________________________________
#> Layer (type) Output Shape Param #
#> ================================================================================
#> input_1 (InputLayer) [(None, 9)] 0
#> token_and_position_embedding (Toke (None, 9, 36) 792
#> nAndPositionEmbedding)
#> transformer_block (TransformerBloc (None, 9, 36) 26056
#> k)
#> global_average_pooling1d (GlobalAv (None, 36) 0
#> eragePooling1D)
#> dropout_3 (Dropout) (None, 36) 0
#> dense_3 (Dense) (None, 64) 2368
#> dropout_2 (Dropout) (None, 64) 0
#> dense_2 (Dense) (None, 6) 390
#> ================================================================================
#> Total params: 29,606
#> Trainable params: 29,606
#> Non-trainable params: 0
#> ________________________________________________________________________________
Some useful information and metrics are stored for a tracebility and
an easy extraction if needed.
model %>% names() # objects from a returned list
#> $names
#> [1] "model" "max_case_length" "number_features" "task"
#> [5] "num_outputs" "vocabulary"
Note that create_model() returns a list, in which the
actual keras model is stored under the element name model.
Thus, we can use functions from the keras-package as follows:
model$model$name # get the name of a model
#> [1] "my_model"
model$model$non_trainable_variables # list of non-trainable parameters of a model
#> list()
The result of create_model() is assigned it’s own class
(ppred_model) for which the processpredictR
provides the methods compile(), fit(),
predict() and evaluate().
Compilation
The following step is to compile the model. By default, the loss
function is the log-cosh or the categorical cross entropy, for
regression tasks (next time and remaining time) and classification
tasks, respectively. It is of course possible to override the
defaults.
model %>% compile() # model compilation
#> Compilation complete!
Training
Training of the model is done with the fit() function.
During training, a visualization window will open in the Viewer-pane to
show the progress in terms of loss. Optionally, the result of
fit() can be assigned to an object to access the training
metrics specified in compile().
hist <- fit(object = model, train_data = split$train_df, epochs = 5)
#> $verbose
#> [1] 1
#>
#> $epochs
#> [1] 5
#>
#> $steps
#> [1] 2227
#> $loss
#> [1] 0.7875332 0.7410239 0.7388409 0.7385073 0.7363014
#>
#> $sparse_categorical_accuracy
#> [1] 0.6539739 0.6713067 0.6730579 0.6735967 0.6747193
#>
#> $val_loss
#> [1] 0.7307042 0.7261314 0.7407018 0.7326428 0.7317348
#>
#> $val_sparse_categorical_accuracy
#> [1] 0.6725934 0.6727730 0.6725934 0.6725934 0.6722342
Make predictions
The method predict() can return 3 types of output, by
setting the argument output to “append”, “y_pred” or
“raw”.
Test dataset with appended predicted values (output = “append”)
predictions <- model %>% predict(test_data = split$test_df,
output = "append") # default
predictions %>% head(5)
#> # A tibble: 5 × 13
#> ith_case case_id prefix prefix_…¹ outcome k activ…² resou…³
#> <int> <chr> <chr> <list> <fct> <dbl> <chr> <fct>
#> 1 8001 A24869 Create Fine <chr [1]> Payment 0 Create… 559
#> 2 8001 A24869 Create Fine - Payment <chr [2]> Payment 1 Payment <NA>
#> 3 8002 A24871 Create Fine <chr [1]> Payment 0 Create… 559
#> 4 8002 A24871 Create Fine - Payment <chr [2]> Payment 1 Payment <NA>
#> 5 8003 A24872 Create Fine <chr [1]> Send f… 0 Create… 559
#> # … with 5 more variables: start_time <dttm>, end_time <dttm>,
#> # remaining_trace_list <list>, y_pred <dbl>, pred_outcome <chr>, and
#> # abbreviated variable names ¹prefix_list, ²activity, ³resource
raw predicted values (output = “raw”)
#> Payment Send for Credit Collection Send Fine
#> [1,] 4.966056e-01 0.344094276 1.423686e-01
#> [2,] 9.984029e-01 0.001501600 8.890528e-05
#> [3,] 4.966056e-01 0.344094276 1.423686e-01
#> [4,] 9.984029e-01 0.001501600 8.890528e-05
#> [5,] 4.966056e-01 0.344094276 1.423686e-01
#> [6,] 1.556145e-01 0.518976271 2.884890e-01
#> [7,] 2.345311e-01 0.715000629 5.147375e-06
#> [8,] 2.627363e-01 0.726804197 5.480492e-06
#> [9,] 3.347774e-05 0.999961376 2.501280e-08
#> [10,] 4.966056e-01 0.344094276 1.423686e-01
predicted values with postprocessing (output = “y_pred”)
#> [1] "Payment" "Payment"
#> [3] "Payment" "Payment"
#> [5] "Payment" "Send for Credit Collection"
#> [7] "Send for Credit Collection" "Send for Credit Collection"
#> [9] "Send for Credit Collection" "Payment"
#> [11] "Send for Credit Collection" "Payment"
#> [13] "Send for Credit Collection" "Payment"
#> [15] "Send for Credit Collection" "Send for Credit Collection"
#> [17] "Send for Credit Collection" "Send for Credit Collection"
#> [19] "Payment" "Send for Credit Collection"
Visualize predictions
For the classification tasks outcome and next activity a
confusion_matrix() function is provided.
#> [1] "ppred_predictions" "ppred_examples_df" "ppred_examples_df"
#> [4] "ppred_examples_df" "tbl_df" "tbl"
#> [7] "data.frame"
confusion_matrix(predictions)
#>
#> Payment Send Appeal to Prefecture
#> Appeal to Judge 2 6
#> Notify Result Appeal to Offender 0 0
#> Payment 1903 7
#> Send Appeal to Prefecture 34 90
#> Send Fine 387 0
#> Send for Credit Collection 688 22
#>
#> Send for Credit Collection
#> Appeal to Judge 10
#> Notify Result Appeal to Offender 0
#> Payment 617
#> Send Appeal to Prefecture 89
#> Send Fine 387
#> Send for Credit Collection 2644
Plot method for the confusion matrix (classification) or a scatter
plot (regression).
plot(predictions) +
theme(axis.text.x = element_text(angle = 90))
knitr::include_graphics("confusion_matrix.PNG")
Evaluate model
Returns loss and metrics specified in compile().
model %>% evaluate(split$test_df)
#> loss sparse_categorical_accuracy
#> 0.7779053 0.6716526
Customize your transformer model
Instead of using the standard off the shelf transformer
model that comes with processpredictR, you can customize
the model. One way to do this, is by using the custom
argument of the create_model() function. The resulting
model will then only contain the input layers of the model, as shown
below.
df <- prepare_examples(traffic_fines, task = "next_activity") %>% split_train_test()
custom_model <- df$train_df %>% create_model(custom = TRUE, name = "my_custom_model")
custom_model
#> Model: "my_custom_model"
#> ________________________________________________________________________________
#> Layer (type) Output Shape Param #
#> ================================================================================
#> input_2 (InputLayer) [(None, 9)] 0
#> token_and_position_embedding_1 (To (None, 9, 36) 828
#> kenAndPositionEmbedding)
#> transformer_block_1 (TransformerBl (None, 9, 36) 26056
#> ock)
#> global_average_pooling1d_1 (Global (None, 36) 0
#> AveragePooling1D)
#> ================================================================================
#> Total params: 26,884
#> Trainable params: 26,884
#> Non-trainable params: 0
#> ________________________________________________________________________________
You can than stack layers on top of your custom model as you prefer,
using the stack_layers() function. This function provides
an abstraction from a little bit more code work if keras is
used (see later).
custom_model <- custom_model %>%
stack_layers(layer_dropout(rate = 0.1)) %>%
stack_layers(layer_dense(units = 64, activation = 'relu'))
custom_model
#> Model: "my_custom_model"
#> ________________________________________________________________________________
#> Layer (type) Output Shape Param #
#> ================================================================================
#> input_2 (InputLayer) [(None, 9)] 0
#> token_and_position_embedding_1 (To (None, 9, 36) 828
#> kenAndPositionEmbedding)
#> transformer_block_1 (TransformerBl (None, 9, 36) 26056
#> ock)
#> global_average_pooling1d_1 (Global (None, 36) 0
#> AveragePooling1D)
#> dropout_6 (Dropout) (None, 36) 0
#> dense_6 (Dense) (None, 64) 2368
#> ================================================================================
#> Total params: 29,252
#> Trainable params: 29,252
#> Non-trainable params: 0
#> ________________________________________________________________________________
# this works too
custom_model %>%
stack_layers(layer_dropout(rate = 0.1), layer_dense(units = 64, activation = 'relu'))
Once you have finalized your model, with an appropriate output-layer
(which should have the correct amount of outputs, as recorded in
customer_model$num_outputs and an appropriate activation
function), you can use the compile(), fit(),
predict() and evaluate() functions as
before.
Custom training and prediction
We can also opt for setting up and training our model manually,
instead of using the provided methods. Note that after defining a model
with keras::keras_model() the model no longer is of class
ppred_model.
new_outputs <- custom_model$model$output %>% # custom_model$model to access a model and $output to access the outputs of that model
keras::layer_dropout(rate = 0.1) %>%
keras::layer_dense(units = custom_model$num_outputs, activation = 'softmax')
custom_model <- keras::keras_model(inputs = custom_model$model$input, outputs = new_outputs, name = "new_custom_model")
custom_model
#> Model: "new_custom_model"
#> ________________________________________________________________________________
#> Layer (type) Output Shape Param #
#> ================================================================================
#> input_2 (InputLayer) [(None, 9)] 0
#> token_and_position_embedding_1 (To (None, 9, 36) 828
#> kenAndPositionEmbedding)
#> transformer_block_1 (TransformerBl (None, 9, 36) 26056
#> ock)
#> global_average_pooling1d_1 (Global (None, 36) 0
#> AveragePooling1D)
#> dropout_6 (Dropout) (None, 36) 0
#> dense_6 (Dense) (None, 64) 2368
#> dropout_8 (Dropout) (None, 64) 0
#> dense_8 (Dense) (None, 11) 715
#> ================================================================================
#> Total params: 29,967
#> Trainable params: 29,967
#> Non-trainable params: 0
#> ________________________________________________________________________________
# class of the model
custom_model %>% class
#> [1] "keras.engine.functional.Functional"
#> [2] "keras.engine.training.Model"
#> [3] "keras.engine.base_layer.Layer"
#> [4] "tensorflow.python.module.module.Module"
#> [5] "tensorflow.python.trackable.autotrackable.AutoTrackable"
#> [6] "tensorflow.python.trackable.base.Trackable"
#> [7] "keras.utils.version_utils.LayerVersionSelector"
#> [8] "keras.utils.version_utils.ModelVersionSelector"
#> [9] "python.builtin.object"
# compile
compile(object=custom_model, optimizer = "adam",
loss = loss_sparse_categorical_crossentropy(),
metrics = metric_sparse_categorical_crossentropy())
Before training the model we first must prepare the data, using the
tokenize() function.
# the trace of activities must be tokenized
tokens_train <- df$train_df %>% tokenize()
map(tokens_train, head) # the output of tokens is a list
#> $token_x
#> $token_x[[1]]
#> [1] 2
#>
#> $token_x[[2]]
#> [1] 2 3
#>
#> $token_x[[3]]
#> [1] 2
#>
#> $token_x[[4]]
#> [1] 2 4
#>
#> $token_x[[5]]
#> [1] 2 4 5
#>
#> $token_x[[6]]
#> [1] 2 4 5 6
#>
#>
#> $numeric_features
#> NULL
#>
#> $categorical_features
#> NULL
#>
#> $token_y
#> [1] 0 1 2 3 4 5
# make sequences of equal length
x <- tokens_train$token_x %>% pad_sequences(maxlen = max_case_length(df$train_df), value = 0)
y <- tokens_train$token_y
We are now ready to train our custom model (the code below is not
being evaluated).
# train
fit(object = custom_model, x, y, epochs = 10, batch_size = 10) # see also ?keras::fit.keras.engine.training.Model
# predict
tokens_test <- df$test_df %>% tokenize()
x <- tokens_test$token_x %>% pad_sequences(maxlen = max_case_length(df$train_df), value = 0)
predict(custom_model, x)
# evaluate
tokens_test <- df$test_df %>% tokenize()
x <- tokens_test$token_x
# normalize by dividing y_test over the standard deviation of y_train
y <- tokens_test$token_y / sd(tokens_train$token_y)
evaluate(custom_model, x, y)