Add C-EASE and ADD-EASE

[deklanw]

Both models from "Closed-Form Models for Collaborative Filtering with Side-Information" by Olivier Jeunen, et al. https://dl.acm.org/doi/abs/10.1145/3383313.3418480

These are variations of EASE which incorporate item token and token_seq features. Small accuracy improvements over EASE (indeed, EASE is a special case of ADD-EASE with item_feat_proportion = 0.0). The paper claims large improvements for cold-start items, but unfortunately there is no way to check that in RecBole #671

The model in the paper allows feature-specific importance weighting, but didn't test that variant (probably because there is no obvious way in general to assign importance). So, in the case of C-EASE I just implemented a single parameter controlling the importance of item features (as is tested in the paper).

[yarncraft]

@deklanw, I ran into the following issue:

File "/home/yarncraft/.local/lib/python3.8/site-packages/recbole/model/general_recommender/addease.py", line 104, in __init__
    self.item_similarity = (1-item_feat_proportion) * inter_S + item_feat_proportion * item_S
ValueError: operands could not be broadcast together with shapes (2149,2149) (2786,2786) 

I manually checked this and it seems that your code does not account for the possibility of having more items in the .item file than (actually used) in the .inter file. This can be fixed by first filtering the item feature matrix with those items that are actually being used in the interactions I think?

Anyways, thanks for the nice work in the other PR's! 😄

[linzihan-backforward]

To sovle the above issue, you can clip the tag_item_matrix by num_items.
tag_item_matrix = tag_item_matrix[:, :self.num_items]

Or just filter the items by config parameter.
tem_inter_num_interval: [1,inf)