keras_FM

This is a tensorflow.keras implementation of Factorized Machine(FM)

Dataset

• The dataset is a Click-Through Rate Prediction dataset provided by Avazu in a Kaggle competition.
• You can find the data here

Data preprocessing

• 'hour' contains year, month, day, and hour information. Only hour information is kept for training. Thus, this feature can be treated as a catagorical feature(hour = 0, 1, 2, ..., 23).
• All other features are catagorical(discrete), one hot encoding is used on the datset to make it suitable for FM training.

FMCrossLayer class

• Set the weights(the hidden vectors in FM) in build().
• Define the forward propagatation in call().
• Here is how we simply crossed features:

Generate FM model

• Generate input_layer with tf.keras.Input()
• Generate linear_layer with tf.keras.Dense()(input_layer)
• Generate cross_layer with FMCrossLayer()(input_layer)
• Add the output of linear_layer and cross_layer to get the final FM model

Some settings

• Sigmoid is used as the activation function
• Loss is set to "binary_crossentropy"
• Optimizer is set to tf.optimizers.Adam(learning_rate=0.001)
• Metrics is set to "binary_accuracy"
• The length of hidden vactor in FM is set to 5 by default