I am working on an app recommendation problem. I don't have any app features, but I have user features. I've tried different similarity based models and also using a multiclass classification model's predicted probabilities to generate candidates. The later works simply because most users only use 1 app.

When I generate top 10 candidates, I get a hit rate of about 65% but the NDCG and other ranking metrics aren't great.

Once I generate candidates, is there a way to re-rank them?

I considered the XGBRanker but re-ranking on a massive dataset is hard. I have over 130 apps and over 70K users data right now. Pairwise ranking just blows up the dataset.

After generating candidates, do you re-train a model with the candidates for each user? That seems exhaustive again. This is largely a function of popularity bias in my dataset. 30 apps cover 80% of users.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.