Incremental update with old users but new items

[levrone1987]

I want to use model.partial_fit_items for incremental model update. The users pool stays the same, but new items may arrive. This is the code I have:

mat = csr_matrix((test_df['amount_in_eur'], (test_df['player_index'], test_df['item_index'])))
model.partial_fit_items(test_df['item_index'], mat)

and I get the following error:

    314 """Incrementally updates item factors
    315 
    316 This method updates factors for items specified by itemids, given a
   (...)
    327     Sparse matrix containing the liked users for each item in itemids
    328 """
    329 if len(itemids) != item_users.shape[0]:
--> 330     raise ValueError("item_users must contain 1 row for every user in itemids")
    332 # recalculate factors for each item in the input
    333 item_factors = self.recalculate_item(itemids, item_users)

ValueError: item_users must contain 1 row for every user in itemids

If my assumption that what I want to achieve is possible, please help me write correct code.

[benfred]

it looks like you're passing a (user, items) matrix to the partial_fit_items - but the method expects a (items, users) matrix.

Does it work if you create it like

mat = csr_matrix((test_df['amount_in_eur'], (test_df['item_index'], test_df['player_index'])))

?

[levrone1987]

Same error @benfred

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[38], line 1
----> 1 model.partial_fit_items(test_df['item_index'], mat)

File ~/miniconda3/envs/whizdom-env/lib/python3.9/site-packages/implicit/cpu/als.py:330, in AlternatingLeastSquares.partial_fit_items(self, itemids, item_users)
    314 """Incrementally updates item factors
    315 
    316 This method updates factors for items specified by itemids, given a
   (...)
    327     Sparse matrix containing the liked users for each item in itemids
    328 """
    329 if len(itemids) != item_users.shape[0]:
--> 330     raise ValueError("item_users must contain 1 row for every user in itemids")
    332 # recalculate factors for each item in the input
    333 item_factors = self.recalculate_item(itemids, item_users)

ValueError: item_users must contain 1 row for every user in itemids

Is my assumption correct: is it possible to retrain the model with some new items (so, same users as in the pretrained model, but new items)? If yes, could you please provide a small working example on how this could be done?

[levrone1987]

I tried all variations, but it still does not work @benfred

[benfred]

The best example I have is probably in the unittest here

implicit/tests/als_test.py

Lines 272 to 301 in cb2a66d

	@pytest.mark.parametrize("use_gpu", [True, False] if HAS_CUDA else [False])
	def test_incremental_retrain(use_gpu):
	likes = get_checker_board(50)
	model = AlternatingLeastSquares(factors=2, regularization=0, use_gpu=use_gpu, random_state=23)
	model.fit(likes, show_progress=False)
	ids, _ = model.recommend(0, likes[0])
	assert ids[0] == 0
	# refit the model for user 0, make them like the same thing as user 1
	model.partial_fit_users([0], likes[1])
	ids, _ = model.recommend(0, likes[1])
	assert ids[0] == 1
	# add a new user at position 100, make sure we can also use that for recommendations
	model.partial_fit_users([100], likes[1])
	ids, _ = model.recommend(100, likes[1])
	assert ids[0] == 1
	# add a new item at position 100, make sure it gets recommended for user right away
	model.partial_fit_items([100], likes[1])
	ids, _ = model.recommend(1, likes[1], N=2)
	assert set(ids) == {1, 100}
	# check to make sure we can index only a single extra item/user
	model.partial_fit_users([101], likes[1])
	model.partial_fit_items([101], likes[1])
	ids, _ = model.recommend(101, likes[1], N=3)
	assert set(ids) == {1, 100, 101}