# Chapter 10 Convolutional neural networks

The first neural networks we built in Chapter 8 did not learn much about structure, sequences, or long-range dependencies in our text data. The LSTM networks we trained in Chapter 9 were especially suited to learning long-range dependencies. In this final chapter, we will focus on convolutional neural network (CNN) architecture, which can learn local, spatial structure within a dataset.

CNNs can be well-suited for modeling text data text often contains quite a lot of local structure. A CNN does not learn long-range structure within a sequence like an LSTM, but instead detects local patterns. A CNN takes data (like text) as input and then hopefully produces output that represents specific structures in the data.

Let’s take more time with CNNs in this chapter to explore their construction, different features, and the hyperparameters we can tune.

## 10.1 What are CNNs?

CNNs can work with data of different dimensions (like two-dimensional images or three-dimensional video), but for text modeling, we typically work in one dimension. The illustrations and explanations in this chapter use only one dimension to match the text use case. Figure 10.1 illustrates a typical CNN architecture. The input sequence in this example uses character tokens, but it could also use word tokens. A convolutional filter slides along the sequence to produce a new, smaller sequence. This is repeated multiple times, typically with different parameters for each layer, until we are left with a small tensor which we can transform into our required output shape, a value between 0 and 1 in the case of binary classification.

This figure isn’t entirely accurate because we technically don’t feed characters into a CNN, but instead use sequence one-hot encoding (Section 8.2.2) with a possible word embedding. Let’s talk about two of the most important CNN concepts, filters and kernel size.

### 10.1.1 Filters

The kernel is a small tensor of the same dimensionality as the input tensor that slides along the input tensor. When it is sliding, it performs element-wise multiplication of the values in the input tensor and its weights and then sums up the values to get a single value. Sometimes an activation function is applied as well. It is these weights that are trained with gradient descent to find the best fit. In Keras, the filters represent how many different kernels are trained in each layer. You typically start with fewer filters at the beginning of your network and then increase them as you go along.

### 10.1.2 Kernel size

The most prominent hyperparameter is the kernel size. The kernel size is the size of the tensor (one-dimensional is this case) that contains the weights. A kernel of size 5 will have 5 weights. These kernels will similarly capture local information about how n-grams capture location patterns. Increasing the size of the kernel decreases the size of the output tensor, as we see in Figure 10.2.

Larger kernels will detect larger and less frequent patterns where smaller kernels will find fine-grained features. Notice how the choice of token will affect how we think about kernel size. For character tokenization, a kernel size of 5 will (in early layers) find patterns in subwords more often than patterns across words, since five characters will not span multiple words. By contrast, a kernel size of 5 for word tokenization will learn patterns within sentences instead.

Kernels must have an odd length.

## 10.2 A first CNN model

We will be using the same data which we examine in Section 8.1 and use in Chapters 8 and 9. This dataset contains short text blurbs for prospective crowdfunding campaigns on Kickstarter, along with if they were successful. Our goal of this modeling is to predict successful campaigns from the text contained in the blurb. We will also use the same preprocessing and feature engineering recipe that we created and described in Sections 8.2.1 and 9.1.

Our first start CNN will look a lot like what is shown in Figure 10.1. We start with an embedding layer, followed by a single one-dimensional convolution layer layer_conv_1d(), then a global max pooling layer layer_global_max_pooling_1d(), a densely connected layer, and ending with a dense layer with a sigmoid activation function to give us one value between 0 and 1 to use in our binary classification task.

library(keras)

simple_cnn_model <- keras_model_sequential() %>%
layer_embedding(
input_dim = max_words + 1, output_dim = 16,
input_length = max_length
) %>%
layer_conv_1d(filter = 32, kernel_size = 5, activation = "relu") %>%
layer_global_max_pooling_1d() %>%
layer_dense(units = 64, activation = "relu") %>%
layer_dense(units = 1, activation = "sigmoid")

simple_cnn_model
## Model
## Model: "sequential"
## ________________________________________________________________________________
## Layer (type)                        Output Shape                    Param #
## ================================================================================
## embedding (Embedding)               (None, 30, 16)                  320016
## ________________________________________________________________________________
## conv1d (Conv1D)                     (None, 26, 32)                  2592
## ________________________________________________________________________________
## global_max_pooling1d (GlobalMaxPool (None, 32)                      0
## ________________________________________________________________________________
## dense_1 (Dense)                     (None, 64)                      2112
## ________________________________________________________________________________
## dense (Dense)                       (None, 1)                       65
## ================================================================================
## Total params: 324,785
## Trainable params: 324,785
## Non-trainable params: 0
## ________________________________________________________________________________

We are using the same embedding layer with the same max_length as in the previous networks so there is nothing new there. The layer_global_max_pooling_1d() layer collapses the remaining CNN output into one dimension so we can finish it off with a densely connected layer and the sigmoid activation function.

This might not end up being the best CNN configuration, but it is a good starting point. One of the challenges when working with CNNs is to ensure that we manage the dimensionality correctly. The length of the sequence decreases by (kernel_size - 1) for each layer. For this input, we have a sequence of length max_length = 30, which is decreased by (5 - 1) = 4 resulting in a sequence of 26, as show in the printed output of simple_cnn_model. We could create seven layers with kernel_size = 5, since we would end with 30 - 4 - 4 - 4 - 4 - 4 - 4 - 4 = 2 elements in the resulting sequence. However, we would not be able to do a network with 3 layers of kernel_size = 7 followed by 3 layers of kernel_size = 5 since the resulting sequence would be 30 - 6 - 6 - 6 - 4 - 4 - 4 = 0 and we must have a positive length for our sequence. Remember that kernel_size is not the only argument that will change the length of the resulting sequence.

Constructing a sequence layer by layer and using Keras’ print method to check the configuration is a great way to make sure your architecture is valid.

The compilation and fitting are the same as we have seen before, using a validation split created with tidymodels as shown in Sections 8.2.4 and 9.1.2.

simple_cnn_model %>% compile(
loss = "binary_crossentropy",
metrics = c("accuracy")
)

cnn_history <- simple_cnn_model %>% fit(
x = kick_analysis,
y = state_analysis,
batch_size = 512,
epochs = 10,
validation_data = list(kick_assess, state_assess)
)

We are using the “adam” optimizer since it performs well for many kinds of models. You may have to experiment to find the optimizer that works best for your model and data.

Now that the model is done fitting, we can evaluate it on the validation data set using the same keras_predict() function we created in Section 8.2.4 and used throughout Chapters 8 and 9.

val_res <- keras_predict(simple_cnn_model, kick_assess, state_assess)
val_res
## # A tibble: 50,522 x 3
##         .pred_1 .pred_class state
##           <dbl> <fct>       <fct>
##  1 0.0000000596 0           0
##  2 0.0130       0           0
##  3 0.000214     0           0
##  4 0.00112      0           0
##  5 0.000127     0           0
##  6 0.992        1           0
##  7 0.000208     0           0
##  8 0.00109      0           0
##  9 0.00170      0           0
## 10 0.000959     0           0
## # … with 50,512 more rows

We can calculate some standard metrics with metrics().

metrics(val_res, state, .pred_class, .pred_1)
## # A tibble: 4 x 3
##   .metric     .estimator .estimate
##   <chr>       <chr>          <dbl>
## 1 accuracy    binary         0.813
## 2 kap         binary         0.624
## 3 mn_log_loss binary         0.971
## 4 roc_auc     binary         0.863

We see some improvement over the densely connected network from Chapter 8, our best performing model on the Kickstarter data so far.

The heatmap in Figure 10.3 shows that the model performs about the same for the two classes, success and failure for the crowdfunding campaigns; we are getting fairly good results from a baseline CNN model.

val_res %>%
conf_mat(state, .pred_class) %>%
autoplot(type = "heatmap")

The ROC curve in Figure 10.4 shows how the model performs at different thresholds.

val_res %>%
roc_curve(truth = state, .pred_1) %>%
autoplot() +
labs(
title = "Receiver operator curve for Kickstarter blurbs"
)

## 10.3 Case study: adding more layers

Now that we know how our basic CNN performs, we can see what happens when we apply some common modifications to it. This case study will examine:

• how we can add additional convolutional layers to our base model and

Let’s start by adding another fully connected layer. We take the architecture we used in simple_cnn_model and add another layer_dense() after the first layer_dense() in the model. Increasing the depth of the model via the fully connected layers allows the model to find more complex patterns. There is, however, a trade-off. Adding more layers adds more weights to the model, making it more complex and harder to train. If you don’t have enough data or the patterns you are trying to classify aren’t that complex, then model performance will suffer since the model will start overfitting as it starts picking up on patterns in the data that aren’t there.

When working with CNNs, the different layers perform different tasks. A convolutional layer extracts local patterns as it slides along the sequences, while a fully connected layer finds global patterns.

We can think of the convolutional layers as doing preprocessing on the text, which is then fed into the dense neural network that tries to fit the best curve. Adding more fully connected layers allows the network to create more intricate curves, and adding more convolutional layers gives richer features that are used when fitting the curves. Your job when constructing a CNN is to make the architecture just complex enough to match the data without overfitting. Yoshua Bengio has a simple rule for this (Bengio 2012):

Just keep adding layers until the test error does not improve anymore.

cnn_double_dense <- keras_model_sequential() %>%
layer_embedding(
input_dim = max_words + 1, output_dim = 16,
input_length = max_length
) %>%
layer_conv_1d(filter = 32, kernel_size = 5, activation = "relu") %>%
layer_global_max_pooling_1d() %>%
layer_dense(units = 64, activation = "relu") %>%
layer_dense(units = 64, activation = "relu") %>%
layer_dense(units = 1, activation = "sigmoid")

cnn_double_dense
## Model
## Model: "sequential_1"
## ________________________________________________________________________________
## Layer (type)                        Output Shape                    Param #
## ================================================================================
## embedding_1 (Embedding)             (None, 30, 16)                  320016
## ________________________________________________________________________________
## conv1d_1 (Conv1D)                   (None, 26, 32)                  2592
## ________________________________________________________________________________
## global_max_pooling1d_1 (GlobalMaxPo (None, 32)                      0
## ________________________________________________________________________________
## dense_4 (Dense)                     (None, 64)                      2112
## ________________________________________________________________________________
## dense_3 (Dense)                     (None, 64)                      4160
## ________________________________________________________________________________
## dense_2 (Dense)                     (None, 1)                       65
## ================================================================================
## Total params: 328,945
## Trainable params: 328,945
## Non-trainable params: 0
## ________________________________________________________________________________

We can compile and fit this new model. We will try to keep as much as we can constant as we compare the different models.

cnn_double_dense %>% compile(
loss = "binary_crossentropy",
metrics = c("accuracy")
)

history <- cnn_double_dense %>% fit(
x = kick_analysis,
y = state_analysis,
batch_size = 512,
epochs = 10,
validation_data = list(kick_assess, state_assess)
)
val_res_double_dense <- keras_predict(
cnn_double_dense,
kick_assess,
state_assess
)

metrics(val_res_double_dense, state, .pred_class, .pred_1)
## # A tibble: 4 x 3
##   .metric     .estimator .estimate
##   <chr>       <chr>          <dbl>
## 1 accuracy    binary         0.809
## 2 kap         binary         0.618
## 3 mn_log_loss binary         1.04
## 4 roc_auc     binary         0.859

This model performs well, but it is not entirely clear that it is working much better than the first CNN model we tried. This could be an indication that the original model had enough fully connected layers for the amount of training data we have available.

If we have two models with nearly identical performance, we should choose the less complex of the two, since it will have faster performance.

We can also change the number of convolutional layers, by adding more such layers.

cnn_double_conv <- keras_model_sequential() %>%
layer_embedding(
input_dim = max_words + 1, output_dim = 16,
input_length = max_length
) %>%
layer_conv_1d(filter = 32, kernel_size = 5, activation = "relu") %>%
layer_max_pooling_1d(pool_size = 2) %>%
layer_conv_1d(filter = 64, kernel_size = 3, activation = "relu") %>%
layer_global_max_pooling_1d() %>%
layer_dense(units = 64, activation = "relu") %>%
layer_dense(units = 1, activation = "sigmoid")

cnn_double_conv
## Model
## Model: "sequential_2"
## ________________________________________________________________________________
## Layer (type)                        Output Shape                    Param #
## ================================================================================
## embedding_2 (Embedding)             (None, 30, 16)                  320016
## ________________________________________________________________________________
## conv1d_3 (Conv1D)                   (None, 26, 32)                  2592
## ________________________________________________________________________________
## max_pooling1d (MaxPooling1D)        (None, 13, 32)                  0
## ________________________________________________________________________________
## conv1d_2 (Conv1D)                   (None, 11, 64)                  6208
## ________________________________________________________________________________
## global_max_pooling1d_2 (GlobalMaxPo (None, 64)                      0
## ________________________________________________________________________________
## dense_6 (Dense)                     (None, 64)                      4160
## ________________________________________________________________________________
## dense_5 (Dense)                     (None, 1)                       65
## ================================================================================
## Total params: 333,041
## Trainable params: 333,041
## Non-trainable params: 0
## ________________________________________________________________________________

There are a lot of different ways we can extend the network by adding convolutional layers with layer_conv_1d(). We must consider the individual characteristics of each layer, with respect to kernel size, as well as other CNN parameters we have not discussed in detail yet like stride, padding, and dilation rate. We also have to consider the progression of these layers within the network itself. The model is using an increasing number of filters in each layer, doubling the number of filters for each layer. This is to ensure that there are more filters later on to capture enough of the global information.

This model is using kernel size 5 twice. There aren’t any hard rules about how you structure kernel sizes, but the sizes you choose will change what features the model can detect.

The early layers extract general or low-level features while the later layers learn finer detail or high-level features in the data. The choice of kernel size determines the size of these features.

Having a small kernel size in the first layer will let the model detect low-level features locally.

We are also including a max-pooling layer with layer_max_pooling_1d() between the convolutional layers. This layer performs a pooling operation that calculates the maximum values in its pooling window; in this model, that is set to 2. This is done in the hope that the pooled features will be able to perform better by weeding out the small weights. This is another parameter you can tinker with when you are designing the network.

We compile this model like the others, again trying to keep as much as we can constant. The only thing that changed in this model compared to the first is the addition of a layer_max_pooling_1d() and a layer_conv_1d().

cnn_double_conv %>% compile(
loss = "binary_crossentropy",
metrics = c("accuracy")
)

history <- cnn_double_conv %>% fit(
x = kick_analysis,
y = state_analysis,
batch_size = 512,
epochs = 10,
validation_data = list(kick_assess, state_assess)
)
val_res_double_conv <- keras_predict(
cnn_double_conv,
kick_assess,
state_assess
)

metrics(val_res_double_conv, state, .pred_class, .pred_1)
## # A tibble: 4 x 3
##   .metric     .estimator .estimate
##   <chr>       <chr>          <dbl>
## 1 accuracy    binary         0.807
## 2 kap         binary         0.614
## 3 mn_log_loss binary         1.09
## 4 roc_auc     binary         0.856

This model also performs well compared to earlier results. Let us extract the the prediction using keras_predict() we defined in 8.2.4.

all_cnn_model_predictions <- bind_rows(
mutate(val_res, model = "Basic CNN"),
mutate(val_res_double_dense, model = "Double Dense"),
mutate(val_res_double_conv, model = "Double Conv")
)

all_cnn_model_predictions
## # A tibble: 151,566 x 4
##         .pred_1 .pred_class state model
##           <dbl> <fct>       <fct> <chr>
##  1 0.0000000596 0           0     Basic CNN
##  2 0.0130       0           0     Basic CNN
##  3 0.000214     0           0     Basic CNN
##  4 0.00112      0           0     Basic CNN
##  5 0.000127     0           0     Basic CNN
##  6 0.992        1           0     Basic CNN
##  7 0.000208     0           0     Basic CNN
##  8 0.00109      0           0     Basic CNN
##  9 0.00170      0           0     Basic CNN
## 10 0.000959     0           0     Basic CNN
## # … with 151,556 more rows

Now that the results are combined in all_cnn_model_predictions we can calculate group-wise evaluation statistics by grouping them by the model variable.

all_cnn_model_predictions %>%
group_by(model) %>%
metrics(state, .pred_class, .pred_1)
## # A tibble: 12 x 4
##    model        .metric     .estimator .estimate
##    <chr>        <chr>       <chr>          <dbl>
##  1 Basic CNN    accuracy    binary         0.813
##  2 Double Conv  accuracy    binary         0.807
##  3 Double Dense accuracy    binary         0.809
##  4 Basic CNN    kap         binary         0.624
##  5 Double Conv  kap         binary         0.614
##  6 Double Dense kap         binary         0.618
##  7 Basic CNN    mn_log_loss binary         0.971
##  8 Double Conv  mn_log_loss binary         1.09
##  9 Double Dense mn_log_loss binary         1.04
## 10 Basic CNN    roc_auc     binary         0.863
## 11 Double Conv  roc_auc     binary         0.856
## 12 Double Dense roc_auc     binary         0.859

We can also compute ROC curves for all our models so far. Figure 10.5 shows the three different ROC curves together in one chart.

all_cnn_model_predictions %>%
group_by(model) %>%
roc_curve(truth = state, .pred_1) %>%
autoplot() +
labs(
title = "Receiver operator curve for Kickstarter blurbs"
)

The curves are very close in this chart, indicating that we don’t have much to gain by adding more layers and that they don’t improve performance substantively. This doesn’t mean that we are done with CNNs! There are still many things we can explore, like different tokenization approahces and hyperparameters that can be trained.

## 10.4 Case study: byte pair encoding

In our models in this chapter so far we have used words as the token of interest. We saw in Chapters 6 and 7 how n-grams could be used in modeling as well. One of the reasons why the Kickstarter dataset is hard to work with is because the text is quite short so we don’t get that many individual tokens to work with. Another choice of token is subwords, where we split the text into smaller units than words and especially some longer words will be broken into multiple subword units. One way to tokenize text into subword units is byte pair encoding proposed by Gage (1994). This algorithm has been repurposed to work on text by iteratively merging frequently occurring subword pairs. Using subwords in text is used in methods such as BERT and GPT-2 with great success. The byte pair encoding algorithm has a hyperparameter controlling the size of the vocabulary. Setting it to higher values allows the models to find more rarely used character sequences in the text.

Byte pair encoding offers a good trade-off between character level and word level information, and can also encode unknown words. For example, suppose that the model is aware of the word “woman”. A simple tokenizer such as those we used before would have to put a word such as “womanhood” into an unknown bucket or ignore it completely, whereas the byte pair encoding should be able to pick up on the subwords “woman” and “hood” (or “woman”, “h”, and “ood”, depending on if the model found “hood” as a common enough subword). Using a subword tokenizer such as byte pair encoding should let us see the text with more granularity since we will have more and smaller tokens for each observation.

Character level CNNs have also proven successful in some contexts. They have been explored by Zhang, Zhao, and LeCun (2015) and work quite well on some shorter texts such as headlines and tweets (Vosoughi, Vijayaraghavan, and Roy 2016).

We need to remind ourselves that these models don’t contain any linguistic knowledge at all, they only “learn” the morphological patterns of sequences of characters (Section 1.2) in the training set. This does not make the models useless, but it should set our expectations about what any given model is capable of.

Since we are using a completely different preprocessing approach, we need to specify a new feature engineering recipe. The textrecipes package has a tokenization engine to perform byte pair encoding, but we need to determine the vocabulary size and the appropriate sequence length. This function takes a character vector and a vocabulary size and returns a dataframe with the number of tokens in each observation.

library(textrecipes)

get_bpe_token_dist <- function(vocab_size, x) {
recipe(~text, data = tibble(text = x)) %>%
step_mutate(text = tolower(text)) %>%
step_tokenize(text,
engine = "tokenizers.bpe",
training_options = list(vocab_size = vocab_size)
) %>%
prep() %>%
bake(new_data = NULL) %>%
transmute(
n_tokens = lengths(textrecipes:::get_tokens(text)),
vocab_size = vocab_size
)
}

We can use map() to try a handful of different vocabulary sizes

bpe_token_dist <- map_dfr(
c(2500, 5000, 10000, 20000),
get_bpe_token_dist,
kickstarter_train$blurb ) bpe_token_dist ## # A tibble: 808,372 x 2 ## n_tokens vocab_size ## <int> <dbl> ## 1 9 2500 ## 2 35 2500 ## 3 27 2500 ## 4 32 2500 ## 5 22 2500 ## 6 45 2500 ## 7 35 2500 ## 8 29 2500 ## 9 39 2500 ## 10 33 2500 ## # … with 808,362 more rows If we compare with the word count distribution we saw in Figure 8.3, then we see in Figure 10.6 that any of these choices for vocabulary size will result in more tokens overall. bpe_token_dist %>% ggplot(aes(n_tokens)) + geom_bar() + facet_wrap(~vocab_size) + labs( x = "Number of subwords per campaign blurb", y = "Number of campaign blurbs" ) Let’s pick a vocabulary size of 10,000 and a corresponding sequence length of 40. To use byte pair encoding as a tokenizer in textrecipes set engine = "tokenizers.bpe"; the vocabulary size can be denoted using the training_options argument. Everything else in the recipe stays the same. max_subwords <- 10000 bpe_max_length <- 40 bpe_rec <- recipe(~blurb, data = kickstarter_train) %>% step_mutate(blurb = tolower(blurb)) %>% step_tokenize(blurb, engine = "tokenizers.bpe", training_options = list(vocab_size = max_subwords) ) %>% step_sequence_onehot(blurb, sequence_length = bpe_max_length) bpe_prep <- prep(bpe_rec) bpe_analysis <- bake(bpe_prep, new_data = analysis(kick_val$splits[[1]]),
composition = "matrix"
)
bpe_assess <- bake(bpe_prep,
new_data = assessment(kick_val$splits[[1]]), composition = "matrix" ) Our model will be very similar to the baseline CNN model from Section 10.2; we’ll use a larger kernel size of 7 to account for the finer detail in the tokens. cnn_bpe <- keras_model_sequential() %>% layer_embedding( input_dim = max_words + 1, output_dim = 16, input_length = bpe_max_length ) %>% layer_conv_1d(filter = 32, kernel_size = 7, activation = "relu") %>% layer_global_max_pooling_1d() %>% layer_dense(units = 64, activation = "relu") %>% layer_dense(units = 1, activation = "sigmoid") cnn_bpe ## Model ## Model: "sequential_3" ## ________________________________________________________________________________ ## Layer (type) Output Shape Param # ## ================================================================================ ## embedding_3 (Embedding) (None, 40, 16) 320016 ## ________________________________________________________________________________ ## conv1d_4 (Conv1D) (None, 34, 32) 3616 ## ________________________________________________________________________________ ## global_max_pooling1d_3 (GlobalMaxPo (None, 32) 0 ## ________________________________________________________________________________ ## dense_8 (Dense) (None, 64) 2112 ## ________________________________________________________________________________ ## dense_7 (Dense) (None, 1) 65 ## ================================================================================ ## Total params: 325,809 ## Trainable params: 325,809 ## Non-trainable params: 0 ## ________________________________________________________________________________ We can compile and train like we have done so many times now. cnn_bpe %>% compile( optimizer = "adam", loss = "binary_crossentropy", metrics = c("accuracy") ) bpe_history <- cnn_bpe %>% fit( bpe_analysis, state_analysis, epochs = 10, validation_data = list(bpe_assess, state_assess), batch_size = 512 ) bpe_history ## ## Final epoch (plot to see history): ## loss: 0.03634 ## accuracy: 0.9933 ## val_loss: 0.9621 ## val_accuracy: 0.8092 The performance is doing quite well, which is a pleasant surprise! This is what we hoped would happen if we switched to a higher detail tokenizer. The confusion matrix in 10.7 also gives us clear information that there isn’t much bias between the two classes with this new tokenizer. val_res_bpe <- keras_predict(cnn_bpe, bpe_assess, state_assess) val_res_bpe %>% conf_mat(state, .pred_class) %>% autoplot(type = "heatmap") What are the subwords being used in this model? We can extract them from step_sequence_onehot() using tidy() on the prepped recipe. All the tokens that start with an "h" are seen here. bpe_rec %>% prep() %>% tidy(3) %>% filter(str_detect(token, "^h")) %>% pull(token) ## [1] "h" "ha" "hab" "ham" "hand" "he" "head" "heart" "heast" ## [10] "hed" "heim" "hel" "help" "hem" "hen" "her" "here" "hern" ## [19] "hero" "hes" "hes," "hes." "hest" "het" "hetic" "hett" "hib" ## [28] "hic" "hing" "hing." "hip" "hist" "hn" "ho" "hol" "hold" ## [37] "hood" "hop" "hor" "hous" "house" "how" "hr" "hs" "hu" Notice how some of these subword tokens are full words and some are part of words. This is what allows the model to be able to “read” long unknown words by combining many smaller sub words. We can also look at common long words. bpe_rec %>% prep() %>% tidy(3) %>% arrange(desc(nchar(token))) %>% slice_head(n = 25) %>% pull(token) ## [1] "▁singer-songwriter" "▁singer/songwriter" "▁post-apocalyptic" ## [4] "▁environmentally" "▁interchangeable" "▁post-production" ## [7] "▁singer/songwrit" "▁entertainment." "▁feature-length" ## [10] "▁groundbreaking" "▁illustrations." "▁professionally" ## [13] "▁relationships." "▁self-published" "▁sustainability" ## [16] "▁transformation" "▁unconventional" "▁architectural" ## [19] "▁automatically" "▁award-winning" "▁collaborating" ## [22] "▁collaboration" "▁collaborative" "▁coming-of-age" ## [25] "▁communication" These twenty-five words were common enough to get their own subword token, and helps us understand the nature of these Kickstarter crowdfunding campaigns. Examining the longest subword tokens gives you a good sense of the data you are working with! ## 10.5 Case study: explainability with LIME We noted in Section 8.6 that one of the significant limitations of deep learning models is that they are hard to reason about. One of the ways to understand a predictive model, even a “black box” one, is using the Local Interpretable Model-Agnostic Explanations (Ribeiro, Singh, and Guestrin 2016) algorithm, or LIME for short. The lime package in R implements the LIME algorithm; it can take a prediction from a model and determine a small set of features in the original data that has driven the outcome of the prediction. As indicated by its name, LIME is an approach to compute local feature importance, or explainability at the individual observation level. It does not offer global feature importance, or explainability for the model as a whole. To use this package we need to write a helper function to get the data in the format we want. The lime() function takes two mandatory arguments, x and model. The model argument is the trained model we are trying to explain. The lime() function works out of the box with Keras models so we should be good to go. The x argument is the training data used for training the model. This is where we need to to create a helper function; the lime package is expecting x to be a character vector so we’ll need a function that takes a character vector as input and returns the matrix the Keras model is expecting. kick_prepped_rec <- prep(kick_rec) text_to_matrix <- function(x) { bake( kick_prepped_rec, new_data = tibble(blurb = x), composition = "matrix" ) } Since the function needs to be able to work with just the x parameter alone, we need to put prepped_recipe inside the function rather than passing it in as an argument. This will work with R’s scoping rules but does require you to create a new function for each recipe. Let’s select a couple of training observations to explain. sentence_to_explain <- kickstarter_train %>% slice(c(1, 2)) %>% pull(blurb) sentence_to_explain ## [1] "Exploring paint and its place in a digital world." ## [2] "Mike Fassio wants a side-by-side photo of me and Hazel eating cake with our bare hands. Let's make this a reality!" We now load the lime package and pass observations into lime() along with the model we are trying to explain and the preprocess function. Be sure that the preprocess function matches the preprocessing that was used to train the model. library(lime) explainer <- lime( x = sentence_to_explain, model = simple_cnn_model, preprocess = text_to_matrix ) This explainer object can now be used with explain() to generate explanations for the sentences. We set n_labels = 1 to only get explanations for the first label, since we are working with a binary classification model15. We set n_features = 12 so we can look at the 12 most important features. If we were dealing with longer text you might want to change n_features to capture the effect of as many features you want. explanation <- explain( x = sentence_to_explain, explainer = explainer, n_labels = 1, n_features = 12 ) explanation ## # A tibble: 21 x 13 ## model_type case label label_prob model_r2 model_intercept model_prediction ## * <chr> <int> <chr> <dbl> <dbl> <dbl> <dbl> ## 1 classific… 1 1 0.997 0.288 0.751 1.01 ## 2 classific… 1 1 0.997 0.288 0.751 1.01 ## 3 classific… 1 1 0.997 0.288 0.751 1.01 ## 4 classific… 1 1 0.997 0.288 0.751 1.01 ## 5 classific… 1 1 0.997 0.288 0.751 1.01 ## 6 classific… 1 1 0.997 0.288 0.751 1.01 ## 7 classific… 1 1 0.997 0.288 0.751 1.01 ## 8 classific… 1 1 0.997 0.288 0.751 1.01 ## 9 classific… 1 1 0.997 0.288 0.751 1.01 ## 10 classific… 2 1 1.00 0.524 0.379 0.913 ## # … with 11 more rows, and 6 more variables: feature <chr>, ## # feature_value <chr>, feature_weight <dbl>, feature_desc <chr>, data <chr>, ## # prediction <list> The output comes in a tibble format where feature and feature_weight are included, but fortunately lime contains some functions to visualize these weights. Figure 10.8 shows the result of using plot_features(), with each facet containing an observation-label pair and the bars showing the weight of the different tokens. Blue bars indicate that the weights support the prediction in the direction and red bars indicate contradictions. This chart is great for finding the most prominent features in an observation. plot_features(explanation) Figure 10.9 shows the weights by highlighting the words directly in the text. This gives us a way to see if any local patterns contain a lot of weight. plot_text_explanations(explanation) FIGURE 10.9: Feature highlighting of words for two examples explained by a CNN model. The interactive_text_explanations() function can be used to launch an interactive Shiny app where you can explore the model weights. One of the ways a deep learning model is hard to explain is that changes to a part of the input can affect how the input is being used as a whole. Remember that in bag-of-words models adding another token when predicting would just add another unit in the weight; this is not always the case when using deep learning models. The following example shows this effect. We have created two very similar sentences in fake_sentences. fake_sentences <- c( "Fun and exciting dice game for the whole family", "Fun and exciting dice game for the family" ) explainer <- lime( x = fake_sentences, model = simple_cnn_model, preprocess = text_to_matrix ) explanation <- explain( x = fake_sentences, explainer = explainer, n_labels = 1, n_features = 12 ) Explanations based on these two sentences are fairly similar as we can see in Figure 10.10. However, notice how the removal of the word “whole” affects the weights of the other words in the examples, in some cases switching the sign from supporting to contradicting. plot_text_explanations(explanation) FIGURE 10.10: Feature highlighting of words in two examples explained by a CNN model. It is these kinds of correlated patterns that can make deep learning models hard to reason about and can deliver surprising results. ## 10.7 Cross-validation for evaluation In Section 8.4, we saw how we can use resampling to create cross-validation folds for evaluation. The Kickstarter dataset we are using is big enough that we have enough data to use a single training set, validation set, and testing set that all contain enough observations in them to give reliable performance metrics. However, it is important to understand how to implement other resampling strategies for situations when your data budget may not be as plentiful or when your needs to computing performance metrics are more precise. set.seed(345) kick_folds <- vfold_cv(kickstarter_train, v = 5) kick_folds ## # 5-fold cross-validation ## # A tibble: 5 x 2 ## splits id ## <list> <chr> ## 1 <split [161.7K/40.4K]> Fold1 ## 2 <split [161.7K/40.4K]> Fold2 ## 3 <split [161.7K/40.4K]> Fold3 ## 4 <split [161.7K/40.4K]> Fold4 ## 5 <split [161.7K/40.4K]> Fold5 Each of these folds has an analysis/training set and an assessment/validation set. Instead of training our model one time and getting one measure of performance, we can train our model v times and get v measures, for more reliability. Last time we saw how to create a custom function to handle preprocessing, fitting, and evaluation. We will use the same approach of creating the function, but this time use the model specification from Section 10.2. fit_split <- function(split, prepped_rec) { ## preprocessing x_train <- bake(prepped_rec, new_data = analysis(split), composition = "matrix" ) x_val <- bake(prepped_rec, new_data = assessment(split), composition = "matrix" ) ## create model y_train <- analysis(split) %>% pull(state) y_val <- assessment(split) %>% pull(state) mod <- keras_model_sequential() %>% layer_embedding( input_dim = max_words + 1, output_dim = 16, input_length = max_length ) %>% layer_conv_1d(filter = 32, kernel_size = 5, activation = "relu") %>% layer_global_max_pooling_1d() %>% layer_dense(units = 64, activation = "relu") %>% layer_dense(units = 1, activation = "sigmoid") %>% compile( optimizer = "adam", loss = "binary_crossentropy", metrics = c("accuracy") ) ## fit model mod %>% fit( x_train, y_train, epochs = 10, validation_data = list(x_val, y_val), batch_size = 512, verbose = FALSE ) ## evaluate model keras_predict(mod, x_val, y_val) %>% metrics(state, .pred_class, .pred_1) } We can map() this function across all our cross-validation folds. This takes longer than our previous models to train, since we are training for 10 epochs each on five folds. cv_fitted <- kick_folds %>% mutate(validation = map(splits, fit_split, kick_prep)) cv_fitted ## # 5-fold cross-validation ## # A tibble: 5 x 3 ## splits id validation ## <list> <chr> <list> ## 1 <split [161.7K/40.4K]> Fold1 <tibble [4 × 3]> ## 2 <split [161.7K/40.4K]> Fold2 <tibble [4 × 3]> ## 3 <split [161.7K/40.4K]> Fold3 <tibble [4 × 3]> ## 4 <split [161.7K/40.4K]> Fold4 <tibble [4 × 3]> ## 5 <split [161.7K/40.4K]> Fold5 <tibble [4 × 3]> Now we can use unnest() to find the metrics we computed. cv_fitted %>% unnest(validation) ## # A tibble: 20 x 5 ## splits id .metric .estimator .estimate ## <list> <chr> <chr> <chr> <dbl> ## 1 <split [161.7K/40.4K]> Fold1 accuracy binary 0.823 ## 2 <split [161.7K/40.4K]> Fold1 kap binary 0.646 ## 3 <split [161.7K/40.4K]> Fold1 mn_log_loss binary 0.909 ## 4 <split [161.7K/40.4K]> Fold1 roc_auc binary 0.872 ## 5 <split [161.7K/40.4K]> Fold2 accuracy binary 0.827 ## 6 <split [161.7K/40.4K]> Fold2 kap binary 0.654 ## 7 <split [161.7K/40.4K]> Fold2 mn_log_loss binary 0.878 ## 8 <split [161.7K/40.4K]> Fold2 roc_auc binary 0.873 ## 9 <split [161.7K/40.4K]> Fold3 accuracy binary 0.826 ## 10 <split [161.7K/40.4K]> Fold3 kap binary 0.651 ## 11 <split [161.7K/40.4K]> Fold3 mn_log_loss binary 0.908 ## 12 <split [161.7K/40.4K]> Fold3 roc_auc binary 0.874 ## 13 <split [161.7K/40.4K]> Fold4 accuracy binary 0.824 ## 14 <split [161.7K/40.4K]> Fold4 kap binary 0.647 ## 15 <split [161.7K/40.4K]> Fold4 mn_log_loss binary 0.888 ## 16 <split [161.7K/40.4K]> Fold4 roc_auc binary 0.871 ## 17 <split [161.7K/40.4K]> Fold5 accuracy binary 0.825 ## 18 <split [161.7K/40.4K]> Fold5 kap binary 0.649 ## 19 <split [161.7K/40.4K]> Fold5 mn_log_loss binary 0.895 ## 20 <split [161.7K/40.4K]> Fold5 roc_auc binary 0.871 We can even create a visualization in Figure 10.11 to see the distribution of these metrics for these five folds. cv_fitted %>% unnest(validation) %>% mutate(.metric = fct_reorder(.metric, .estimate)) %>% ggplot(aes(.metric, .estimate, color = .metric)) + geom_boxplot() + geom_point(alpha = 0.5) + theme(legend.position = "none") The metrics have little variance just like they did last time, which is reassuring that our model is robust with respect to the evaluation metrics. ## 10.8 The full game: CNN We’ve come a long way in this chapter, and looked at the many different modifications to the simple CNN model we started with. Most of the alterations didn’t add much so this final model is not going to be much different than what we have seen so far. ### 10.8.1 Preprocess the data For this final model, we are not going to use our validation split again, so we only need to preprocess the training data. max_words <- 2e4 max_length <- 30 kick_rec <- recipe(~blurb, data = kickstarter_train) %>% step_tokenize(blurb) %>% step_tokenfilter(blurb, max_tokens = max_words) %>% step_sequence_onehot(blurb, sequence_length = max_length) kick_prep <- prep(kick_rec) kick_matrix <- bake(kick_prep, new_data = NULL, composition = "matrix") dim(kick_matrix) ## [1] 202093 30 ### 10.8.2 Specify the model Instead of using specific validation data that we can then compute performance metrics for, let’s go back to specifying validation_split = 0.1 and let the Keras model choose the validation set. final_mod <- keras_model_sequential() %>% layer_embedding( input_dim = max_words + 1, output_dim = 16, input_length = max_length ) %>% layer_conv_1d( filter = 32, kernel_size = 7, strides = 1, activation = "relu" ) %>% layer_global_max_pooling_1d() %>% layer_dense(units = 64, activation = "relu") %>% layer_dense(units = 1, activation = "sigmoid") final_mod %>% compile( optimizer = "adam", loss = "binary_crossentropy", metrics = c("accuracy") ) final_history <- final_mod %>% fit( kick_matrix, kickstarter_train$state,
epochs = 10,
validation_split = 0.1,
batch_size = 512,
verbose = FALSE
)
final_history
##
## Final epoch (plot to see history):
##         loss: 0.03343
##     accuracy: 0.9927
##     val_loss: 0.7211
## val_accuracy: 0.8653

This looks promising! Let’s finally turn to the testing set, for the first time during this chapter, to evaluate this last model on data that has never been touched as part of the fitting process.

kick_matrix_test <- bake(kick_prep,
new_data = kickstarter_test,
composition = "matrix"
)
final_res <- keras_predict(final_mod, kick_matrix_test, kickstarter_test\$state)
final_res %>% metrics(state, .pred_class, .pred_1)
## # A tibble: 4 x 3
##   .metric     .estimator .estimate
##   <chr>       <chr>          <dbl>
## 1 accuracy    binary         0.852
## 2 kap         binary         0.704
## 3 mn_log_loss binary         0.778
## 4 roc_auc     binary         0.894

This is our best performing model in this chapter on CNN models, although not by much. We can again create an ROC curve, this time using the test data in Figure 10.12.

final_res %>%
roc_curve(state, .pred_1) %>%
autoplot()

We have been able to incrementally improve our model by adding to the structure and making good choices about preprocessing. We can visualize this final CNN model’s performance using a confusion matrix as well, in Figure 10.13.

final_res %>%
conf_mat(state, .pred_class) %>%
autoplot(type = "heatmap")

Notice that this final model performs better then any of the models we have tried so far in this chapter, Chapter 8, and Chapter 9.

For this particular dataset of short text blurbs, a CNN model able to learn local features performed the best, better than either a densely connected neural network or an LSTM.

## 10.9 Summary

CNNs are a type of neural network that can learn local spatial patterns. They essentially perform feature extraction, which can then be used efficiently in later layers of a network. Their simplicity and fast running time, compared to models like LSTMs, makes them excellent candidates for supervised models for text.

### 10.9.1 In this chapter, you learned:

• how to preprocess text data for CNN models
• how CNN layers can be stacked to extract patterns of varying detail
• how byte pair encoding can be used to tokenize for finer detail
• how to do hyperparameter search in Keras with tfruns
• how to evaluate CNN models for text

### References

Bengio, Yoshua. 2012. “Practical Recommendations for Gradient-Based Training of Deep Architectures.” http://arxiv.org/abs/1206.5533.

Gage, P. 1994. “A New Algorithm for Data Compression.” The C Users Journal Archive 12: 23–38.

Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. 2016. “"Why Should I Trust You?": Explaining the Predictions of Any Classifier.” http://arxiv.org/abs/1602.04938.

Vosoughi, Soroush, Prashanth Vijayaraghavan, and Deb Roy. 2016. “Tweet2Vec: Learning Tweet Embeddings Using Character-Level Cnn-Lstm Encoder-Decoder.” In, 1041–4. https://doi.org/10.1145/2911451.2914762.

Zhang, Xiang, Junbo Zhao, and Yann LeCun. 2015. “Character-Level Convolutional Networks for Text Classification.” In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, 649–57. NIPS’15. Cambridge, MA, USA: MIT Press.

1. The explanations of the second label would just be the inverse of the first label. If you have more than two labels, it makes sense to explore some or all of them.↩︎