model.md

Model

The model is based on the research paper Deep Neural Net with Attention for Multi-channel Multi-touch Attribution

Architecture

Input layer

name: input_layer shape: (None, journey_max_len, nb_campaigns)

Embedding layer

name: embedding_layer input: input_layer type: dense activation: linear output_shape: (None, journey_max_len, nb_units_embedding)

LSTM layer

name: lstm_layer input: embedding_layer type: lstm activation: tanh output_shape: (None, journey_max_len, nb_units_lstm)

Attention layer

name: attention_layer input: lstm_layer type: dense -> flatten -> activation activation: softmax output_shape: (None, journey_max_len)

Weighted activation layer

name: weighted_activation_layer input: lstm_layer, attention_layer type: repeat_vector, permute, multiply, lambda activation: none output_shape: (None, nb_lstm_units)

Output layer

name: output_layer input: weighted_activation_layer type: dense activation: softmax output_shape: (None, 1)

Optimizer

Adam optimizer with learning_rate as hyperparameters, beta_1 set to 0.9, beta_2 set to 0.999 and epsilon set to 1e-7

Attention model

The attention or attribution model is obtained by cutting the original model once trained such as the input is the input_layer and the output is the attention_layer.