Add RP3Beta

[deklanw]

Random walk method from Updatable, Accurate, Diverse, and Scalable Recommendations for Interactive Applications https://dl.acm.org/doi/10.1145/2955101

[linzihan-backforward]

It seems that the whole user_all_items matrix can be computed in advance to avoid multiplication in each batch.

[deklanw]

@linzihan-backforward Hello, sorry for the (very) late reply. I intentionally avoided computing the entire user_item matrix to save memory. I included the all-at-once implementation at the top for reference, and explained the former. See

RecBole/recbole/model/general_recommender/rp3beta.py

Lines 28 to 43 in e28e150

	# for reference, doing it in one computation
	# since the resultant matrix is dense, I'll refrain from doing this
	def calculate_rp3beta(B, beta):
	user_degree_inv = get_inv_degree_matrix(B)
	item_degree_inv = get_inv_degree_matrix(B.T)
	# multiplication on left for row-wise scaling
	user_transition = user_degree_inv @ B
	item_transition = item_degree_inv @ B.T
	P3 = user_transition @ item_transition @ user_transition
	# multiplication on right for column-wise scaling (i.e., we're reweighting by inverse item popularity to a power)
	RP3Beta = P3 @ item_degree_inv.power(beta)
	return RP3Beta

[linzihan-backforward]

It make sense. We will test the algorithm as soon as possible. It will be merged into the future version if everything is OK. Thanks again for your contribution !

[linzihan-backforward]

Same as above.