the relevance probability seems to no effect on final result

[HuangLK]

Hi, I have used RippleNet on my dataset but can't get the explainability by tracking the paths to the item with high relevance probability in the KG.
So I try to modify the source code:

RippleNet/src/model.py

Line 98 in ba7ce72

o = tf.reduce_sum(self.t_emb_list[hop] * probs_expanded, axis=1)

After replacing the reduce_sum with reduce_mean, the line be:

    o = tf.reduce_mean(self.t_emb_list[hop], axis=1)

and I run these two version code on the movie dataset, and they both get the best result(AUC 0.92 ACC 0.84).
BEFORE(ORIGINAL):

AFTER:

Could these results tell the truth that the relevance probabilities of the heads have no effect on the final result?
If I misunderstand something, please let me know. Thx :)

[HuangLK]

@hwwang55

[hwwang55]

Relevance probabilities are more to provide explainability in MovieLens, since the performance on MovieLens is already high. Relevance probabilities are more effective in news dataset according to our results.

[Liane-Wang]

Hi, can you print the values of 'probs_expanded'? I have tired to do that, but met an issue ' You must feed a value for placeholder tensor 'memories_r_0' '. Can I get some help? thanks a lot

[Liane-Wang]

@HuangLK