Questions about the time of day embedding

[mb-Ma]

I noticed that only the last time step of each sample is kept by the following code:

STID/stid/stid_arch/stid_arch.py

Lines 76 to 88 in f9801a5

	if self.if_time_in_day:
	t_i_d_data = history_data[..., 1]
	# In the datasets used in STID, the time_of_day feature is normalized to [0, 1]. We multiply it by 288 to get the index.
	# If you use other datasets, you may need to change this line.
	time_in_day_emb = self.time_in_day_emb[(t_i_d_data[:, -1, :] * self.time_of_day_size).type(torch.LongTensor)]
	else:
	time_in_day_emb = None
	if self.if_day_in_week:
	d_i_w_data = history_data[..., 2]
	day_in_week_emb = self.day_in_week_emb[(
	d_i_w_data[:, -1, :]).type(torch.LongTensor)]
	else:
	day_in_week_emb = None

More specifically, the operation "t_i_d_data[:, -1, :]" is used to get the last time step from input data.

Could you please provide more interpretations?

[zezhishao]

Thanks for your interests.
For more details, you can refer to the discussion below the blog of STID in Zhihu:

[mb-Ma]

Thank you for your fast response. I did not see these comments. :-)

Yeah. It seems an efficient solution. The purpose is just to identify different times when the current sample is, only one time step is okay which meets Ockham's Razor. From another perspective, (B, [all time step concat], N) compared to (B, [single time step embedding], N) is more redundant since all nodes are same time embedding in one sample.

[zezhishao]

Thank you very much for your acknowledgment! Any further discussions are welcome~