predict on new unseen users or test set users

[AlexVaith]

Hey, the package, in general, is very nice. However, I have not found a way to use a model on new data or even on the test set. From my personal perspective, it is quite unnecessary to test the predictions on the training data. Maybe I have misunderstood the predict function from the model, but from what I can tell, the user id is mapped to the training set right?

Could you provide a solution to my problem?

[massquantity]

Which algorithm are you using? If you are using a pure algorithm such as SVD or itemCF, you can pass users and items as list to the predict function. To deal with the new user problem, set the cold_start parameter.

>>> model.predict(user=[1,2,3], item=[4,5,6], cold_start="average")

If you are using a feat algorithm such as wide_deep or deepFM, you can use the predict_data_with_feats function, which can predict on new data and new users. You can find some brief explanations in User Guide. The example script changing_feature_example.py also show some usages of it in line 65 - 68.

I didn't put much focus on this problem, so I guess that's why people are hard to find them:) Actually, I don't consider the prediction problem important because, it is a recommender system. In real-world scenarios, the main job of it should be recommending instead of predicting. I've seen a lot of open-source recommender systems on Github, but few of them actually provides a recommend function. That really makes me confused, so I built this library in order to create a "real" recommender system.

[AlexVaith]

I get your idea, but for evaluation purposes it is nice to get an overview about how the prediction error is distributed. Therefore I would like to use the evaluation set as I would do with other machine learning tasks.

I will have a look at the predict data with feats function.

[AlexVaith]

In general, do you plan to transform the current tf 1.x code to tf 2.0?

[massquantity]

Yes you can use the evaluate function to evaluate on the new data directly, see line 76 in changing_feature_example.py.

I do have a tf2.x plan, but certainly it's not my top priority. It's really painful to transfer all the codes. Besides, I implemented some of the algorithms in tf2.x, but found that the training speed is much slower than tf1.x, so.... But this is not a formal benchmark, since I implemented it very casually.