Something that wasn't touched on that I think might be important is where inference will happen, on the edge vs on the cloud? I'd imagine something like this would be better run on the edge, in that case a logistic regression or SVM model might be a better choice when taking into consideration computation power. If NN still outperform significantly enough, quantization and pruning may be helpful (although I doubt with a model of that size that matters).
@tankieslayer6927
4 ай бұрын
I doubt this model outperforms XGB for tabular data like this.
@jacobsimon4699
6 ай бұрын
Super detailed and helpful, great work! 🤯
@sophiophile
6 ай бұрын
I would have asked whether we might have access to (or be able to scrape) data related add some sentiment data from places like Twitter/FB/news articles. This can help make a model which is more robust against (somewhat) external events. This can used to create a model that is able to predict better across different demographics (younger people who are looking at social media more are more likely to delete if there is negative sentiment online). Also, regarding the end discussion- you can still get a probability of deletion out of a gradient boosted tree model using (for example) XGBRegressor. The idea that trees can only give a binary classification is something that a lot of people seem to say for some reason. They can even output long horizoned time-series predictions.
@tryexponent
6 ай бұрын
Thanks for sharing your thoughts, sophiophile!
@DamjanDimitrioski
6 ай бұрын
Is it useful to check if for instance Android OS version is 4 times lower than current and we know it's in end of life, then we know we won't support that version in our app any more, will it predict then that will lose those users?
@MateusSantos97
6 ай бұрын
Anyone here is a paid user at the exponent platform? Is it worth it or just the KZitem videos here already do the job?
Пікірлер: 7