the result of recommendation using two domain rating matrix is very bad,would you help me?

[david-cortes]

the result of recommendation using two domain rating matrix is very bad,would you help me?

[david-cortes]

I'd need a more detailed piece of code to check what's wrong, including what kind of objects are _testNegatives, _testRatings, and how you generated them from triplets (user, item, rating), but for now:

• Why are you appending gtItems? If these are lists of lists, wouldn't that leave you with a list having another list as last element?
• Shouldn't the NDCG be evaluated only with the ratings of a user in a hold-out sample? Are you trying to use this for implicit-feedback data?
• I assume HR is hit ratio - shouldn't this take a topN ranking among the non-training items?
• Are you using some existing package for calculating these metrics?

lstA = [[1, 2, 3], [4, 5], [6]]
lstB = [[7, 8], [9, 10, 11], [12]]

def modify_list_el(ix):
    entryA = lstA[ix]
    entryB = lstB[ix]
    entryB.append(entryA)
    return entryB

modify_list_el(0)
>>> [7, 8, [1, 2, 3]] ## Last el. is a list
np.array(modify_list_el(0))
>>> array([7, 8, list([1, 2, 3])], dtype=object)
### This is not the expected input for _model.predict

[david-cortes]

The model object will actually accept both numpy arrays and lists - e.g.

CMF_obj.predict([0, 0, 0], [1, 2, 3])
CMF_obj.predict(np.array([0, 0, 0]), np.array([1, 2, 3]))

But would not work correctly with this:

CMF_obj.predict([0, 0, 0], [1, 2, [3, 4, 5]])

Although if you are not getting any error, I guess you're passing flat arrays already. Still, NDCG for explicit-feedback data (e.g. ratings) I think should be computed only with the test items, not the negative sample, unless you are trying to evaluate implicit-feedback metrics. If you are trying to fit a model to implicit-feedback data (e.g. clicks) and want to evaluate on the negative entries, you might want to try methods meant for that kind of data instead (e.g. https://www.github.com/david-cortes/ctpfrec).

[sunshinelium]

the predict be used to rank, then eval model by hr,ngcf

[sunshinelium]

thanks for your reply. the item is negative sample of the user. acturely, the input of cmf.predict is the form of [array[0,0,0], array[1,2,3]] . so it should be input as list not array ?i will try to change to list . i am so sorry for disturbe. and thanks for your time

[sunshinelium]

thanks for your reply. acturely. I recommend for implicit ratings. Each user is sorted from 100 samples by the leave-one-out evalution , among which 99 negative samples ,denoted by item in code , dtype:list, and 1 positive sample ,denoted by gtItem in code, dtype:int. Therefore, the input is[ user(dtype: array), item(dtype:array)]. There were no errors during the run, but the results were much worse than the BPRMF. So this model is not suitable for implicit rating recommendations? thanks for your time and your resources. I will study it carefully

[david-cortes]

That's right, this model is not suitable for implicit-feedback data. Unlike BPR or implicit-ALS, it will not fit the model to the negative entries in the training data, so if you're looking to optimize those kinds of metrics, this is not the right choice.

[sunshinelium]

thanks for your help. ^_^ I will try another method. I have been studying cross-domain recommendation recently and hope to have more opportunities to communicate later

[david-cortes]

Closing as the latest version includes an implicit-feedback version.