Question on Temporal Filtering

[ThisIsForReview]

Thanks to the authors to release the code. Could anyone ask a question on the code?

[ThisIsForReview]

In the original paper, see Section 5.1, Filtering along the Temporal Domain: The Temporal Filtering (ie. Fourier transform) should be applied to the temporal sequence for each node. However this is implemented in EFT_spectral_attn where Fourier transform is applied on the message sequence of a node (i.e., the features of neighbors of a node according to their edge time). See for example the definition of def item_fourier_reduce_func(self, nodes).

Thanks

[ansonb]

Hi,

Thanks for your interest in our work and for the question.

Regarding the frequency transform along temporal domain we apply it to the sequence of signals for a node. In the sequential recommendation setting, we consider the immediate neighbors as the time-varying signal for a node. Specifically, we have the interacted items over time as a sequence of signals for a single user. For the sequential recommendation task, the formulation comes naturally as a user would be known by the items interacted with, and the changing preferences would be reflected in the time-varying signals. Hence, we apply the frequency transform dimension-wise to these signals. Note this implementation is only applicable to the sequential recommendation setting. Apologies for not being able to present the implementation detail well enough in the paper due to space constraints.

Regards