zgw21cn / deepfm Goto Github PK

second_order = np.zeros([sample_size,1])

second_order = second_order+tf.reshape(tf.reduce_sum(tf.multiply(former, later), 1, keep_dims=True ),[sample_size,1])

It seems all the second order product are added together before sending to the output layer.

However, in the paper each dot product node are connected to the output. How should we understand this piece of code?

sparse_index = tf.placeholder(tf.int64, [None, 2])
sparse_ids = tf.placeholder(tf.int64, [None])
sparse_values = tf.placeholder(tf.float32, [None])
sparse_shape = tf.placeholder(tf.int64, [2])
ids = tf.SparseTensor(sparse_index, sparse_ids, sparse_shape)
values = tf.SparseTensor(sparse_index, sparse_values, sparse_shape)

sparse input like this above
how do you handle it?

Thanks for sharing!

According to the original paper, DeepFM should concatenate linear, 2-order and DNN layer first, then apply a final DNN layer to generate the final CTR. In your code base, these 3 layers are added before feeding to the final DNN layer. Am I wrong?

In the build_model method:

V =tf.reshape(tf.nn.embedding_lookup(inter_vector, X_code),[sample_size,feature_code_size*embed_size])

Then will different code variables share the same embedding vector? e.g. Variable A code 1, and Variable B code 1 will use the same piece of embedding vector after the embedding_lookup.

Please, are you implementing deep FMDeepFM: A Factorization-Machine based Neural Network for CTR Prediction.