In DeepFm.py line 102, why the second order term's size is (None, K) rather than (None, )?

[Wang-Yu-Qing]

From the equation (2) in the paper, the second order term should be a scalar. However, in the code, it is computed as a vector. I think the line 102 should be something like this: self.y_second_order = 0.5 * tf.reduce_sum(tf.subtract(self.summed_features_emb_square, self.squared_sum_features_emb))

[nkqiaolin]

IMO, the original paper said each of three components should all be one scalar, then sum together to fed to sigmod() . But you can also treat the sum of scalar as sum(w1x1, w2x2...), and set w1=w2=..=1。In this git repo, the author treat all W as trainable variables, which is an extension to original paper. Although effect of this modification is unknown, this implementation is still following the core ideas of the paper.

[alpha008]

我是健翔，您的邮件我收到了。