Shall I use cosine or dot product to calculate the similarity between user & item?

[jackyhawk]

Shall I user consine or dot product to calculate the similarity between user & item?

(since there would be some negative latent fatore for user & item, is it still suitable to cosine?

thanks very much

[chtran]

Hi, if you care only about similarity, not popularity, you can use cosine similarity. If you care about popularity of items, you can use dot product

[jackyhawk]

Thanks very much,chtran.

And what's more ,shall I use the item bias when I calculate the score?

would the result be like this:

score = user_factor dot_product item_factor + item_bias

or just:
score = user_factor dot_product item_factor

?

[jackyhawk]

And btw, it seemed that qmf does not support user bias?

[jackyhawk]

Any suggestions for this? Thanks very much

[albietz]

Hi @jackyhawk,

Yes, if you train your model with item biases, you should also use them when making predictions.
qmf does not have user biases because the models try to predict preferences / rankings for each user, as opposed to absolute scores (like ratings in the netflix dataset), so adding a user offset does not change anything.

-Alberto

[jackyhawk]

Thanks very much albietz.

and it would be like this?
score = user_factor dot_product item_factor + item_bias
or
score = cosine(user_factor , item_factor) + item_bias

And i've just paste the test metrics of my real training output( about 5 million,0.5 million items, and about 0.5 billion click,which are our users behavior within latest 30 days of (after filtering some outliers) )
it seems that the auc is very high ,but the precision and recall (@10) is very low,
is this the normal scenario?

And what about the train loss? it seems that value of [train loss] betweem 0.05~0.08 works for me.
but it is not good for [test loss = 0.223691]

====================================================
test metrics:
train loss = 0.0759575, test loss = 0.223691

18:29:02.830051 26919 MetricsEngine.cpp:41] begin metrics: epoch 9: recorded metric test_avg_auc = 0.91341,log_:1
18:29:02.830128 26919 MetricsEngine.cpp:45] epoch 9: recorded metric test_avg_auc = 0.91341
18:29:11.645699 26919 MetricsEngine.cpp:41] begin metrics: epoch 9: recorded metric test_avg_ap = 0.00194532,log_:1
18:29:11.645777 26919 MetricsEngine.cpp:45] epoch 9: recorded metric test_avg_ap = 0.00194532
18:29:14.434798 26919 MetricsEngine.cpp:41] begin metrics: epoch 9: recorded metric test_avg_p@10 = 0.0016,log_:1
18:29:14.434895 26919 MetricsEngine.cpp:45] epoch 9: recorded metric test_avg_p@10 = 0.0016
18:29:17.357926 26919 MetricsEngine.cpp:41] begin metrics: epoch 9: recorded metric test_avg_r@10 = 0.00135481,log_:1
18:29:17.358002 26919 MetricsEngine.cpp:45] epoch 9: recorded metric test_avg_r@10 = 0.00135481

[jackyhawk]

And what's more, as for the input data file, should the old data in the beginning or the new data in the beginning? Thanks very much

[jackyhawk]

And what's more, Shall I use more (longer-period user's behaviro ) data as the training dataset to improve the precison & recall @10?

and currently, i just use 100 as the nfactor dimension, shall I also increase that value?

Thanks very much

[jackyhawk]

Any suggestions for that? Thanks very much

[jackyhawk]

I tried to increase the dimension of latent factor (from 100 to 200 and then to 300), it seemed that the results got better

[jackyhawk]

Any suggestions for this? Thanks very much