learn_embed.py

def user_embedding_model(embedding_size = 50):
    
    #Embed items and users in vec space
    
    # Both inputs are 1-dimensional
    user = Input(name = 'user', shape = [1])
    item = Input(name = 'item', shape = [1])
    
    #  (None, 1, 50))
    user_embedding = Embedding(name = 'user_embedding',
                               input_dim = len(user_index),
                               output_dim = embedding_size)(user)
    
    #  (None, 1, 50))
    item_embedding = Embedding(name = 'item_embedding',
                               input_dim = len(item_index),
                               output_dim = embedding_size)(item)
    
    # Merge the layers with a dot product along the second axis (shape will be (None, 1, 1))
    merged = Dot(name = 'dot_product', normalize = True, axes = 2)([user_embedding, item_embedding])
    
    # Reshape to be a single number (shape will be (None, 1))
    merged = Reshape(target_shape = [1])(merged)
    
    merged = Dense(1, activation = 'sigmoid')(merged)
    model = Model(inputs = [user, item], outputs = merged)
    model.compile(optimizer = 'Adam', loss = 'binary_crossentropy', metrics = ['accuracy'])
    
    return model