Types of classification problems

Three types of classification problems:

• binary classification
• multiclass
• multilabel

Multilabel classification - a sample can be assigned to more than one label from more than 2 label options
Multiclass classification - a sample can be assigned to one label but from more than 2 label options

Multiclass image classificaton: pizza, steak, sushi

Input_shape = [None, 224, 224, 3] - single image

Input shape = [32, 224, 224, 3] - common batch size of images

32 is a common batch size

...

Classification evaluation methods

---

Accuracy

tf.keras.metrics.Accuracy()

sklearn.metrics.accuracy_score()

Not the best for imbalanced classes

---

Precision

For imbalanced class problems. Higher precision leads to less false positives.

---

Recall

Higher recall leads to less false negatives.

Tradeoff between recall and precision.

---

F1-score

Combination of precision and recall, ususally a good overall metric for classification models.

---

keyboard_arrow_down

Confusion matrix

Can be hard to use whith large numbers of classes.

y-axis -> true label
x-axis -> predicted label

# Create confusion metrics

from sklearn.metrics import confusion_matrix

y_preds = model_8.predict(X_test)

confusion_matrix(y_test, y_preds)

important:

This time there is a problem with loss function.

• In case of categorical_crossentropy the labels have to be one-hot encoded

• In case of labels as integeres use SparseCategoricalCrossentropy

# Get the patterns of a layer in our network

weights, biases = model_35.layers[1].get_weights()

Getting dataset ready for tensorflow

1. Converting non-numerical columns

For example: Use pandas get_dummies() function

insurance_one_hot = pd.get_dummies(insurance,dtype="int32") #to avoid bool which generate problem with model fitting in TensorFlow
insurance_one_hot

How we can improve model (in the particular stage of the process)?

# 1. Creating model: add more layers, increase numbers of hidden neurons, change activation functions

# 2. Compiling: change optimizer or its parameters (eg. learning rate)

# 3. Fitting: more epochs, more data

### How?

# from smaller model to larger model

Typical workflow: build a model -> fit it -> evaulate -> tweak -> fit > evaluate -> ....

Deep Learning mantras: ;)

Building model: experiment
Evaluation model: visualize

What can visualize?

• the data
• model itself
• the training of a model
• predictions

## The 3 sets (or actually 2 sets: training and test set) - USING ONLY TensorFlow

tf.random.set_seed(999)

X_train, X_test = tf.split(tf.random.shuffle(X, seed=42), num_or_size_splits=[40, 10])

For regression problems:

• MAE
  • tf.keras.losses.MAE()
  • tf.metrics.mean_absolute_error()
  • great starter metrics for any regression problem
• MSE
  • tf.keras.losses.MSE()
  • tf.metrics.mean_square_error()
  • when larger errors are more significant that smaller errors
• Huber
  • tf.keras.losses.Huber()
  • combintion of MSE and MAE less sensitive to outliers than MSE

Take away: You should minimize the time between your experiments (that's way you should start with smaller models). The more experiments you do, the more things you figure out that don't work.

Some examples of sequence prediction problems include: Weather Forecasting . Given a sequence of observations about the weather over time, predict the expected weather tomorrow.

To achieve better explainability, in many e-commerce applications consumer behavior can be viewed on the level of sessions. A session is a well-defined visit of a consumer to a web- shop: a subsequence of events within the consumer’s history that lay no further apart than a predefined time difference.

We highlight our model’s flexibility and performance on two groups of valuable customers: those who keep making more and more transactions with the firm (denoted as ”opportunity” customers) and those who are at risk of defection. We demonstrate that the model also excels at automatically capturing seasonal trends in customer activity, such as the shopping period leading up to the December holidays.