It wasn't easy question, but I would approach it different way. 1. Train multiple regression models to predict different LTVs (Long term value), for example, 1 week of user's historical data to predict 1 year terms of value, 1 month of data to predict 1 year of value, 2 months of data to predict 2 years of value, etc. 2. Find top features in these models. 3. If we would use neural network then we can learn user embeddings from this process. 4. Now we have top features and user embeddings, so we can use them in multiple ways. 5.1 Use LTV model to predict their value right away and see if it can be good investor in year or so. On top of that, we can always use exploration/exploitation model to experiment with threshold for good investor as well. 5.2 One can still run clustering on those top features as well as user embeddings.
if we have historical data of investors, we can get good investors based on the ratio of profit received from them to investor's capital. After that we can try logreg + l1 to zeroing unnecessary features. Or pca+xgb if we have nonlinear relationship.
This was quite a poor performance that could be improved on. He didn't really seem confident about what he was saying especially when talking about clustering and talking about the technical details of these. It's quite humiliating that data science has come to this and fallen so far behind software engineering in professionalism, the whole interview is basically import sklearn, model.fit and then a few gushy words about automatic retraining without any technical details lol
![[UNCUT]“ฅนตื่นธรรม”ปะทะ“หมอปลา-ปู่มหามุนี-อั๋นโอกิ”สะเทือนกระทรวงเวทมนตร์ Iคนดังนั่งเคลียร์I10 ตค.67](http://i.ytimg.com/vi/peMD_SAkwNc/mqdefault.jpg)
It wasn't easy question, but I would approach it different way.
1. Train multiple regression models to predict different LTVs (Long term value), for example, 1 week of user's historical data to predict 1 year terms of value, 1 month of data to predict 1 year of value, 2 months of data to predict 2 years of value, etc.
2. Find top features in these models.
3. If we would use neural network then we can learn user embeddings from this process.
4. Now we have top features and user embeddings, so we can use them in multiple ways.
5.1 Use LTV model to predict their value right away and see if it can be good investor in year or so. On top of that, we can always use exploration/exploitation model to experiment with threshold for good investor as well.
5.2 One can still run clustering on those top features as well as user embeddings.
if we have historical data of investors, we can get good investors based on the ratio of profit received from them to investor's capital. After that we can try logreg + l1 to zeroing unnecessary features. Or pca+xgb if we have nonlinear relationship.
Hi Jay, could you please share the Google doc, would be helpful to read through the scenario later.. Thanks!
I like the style how Jay walks through his processing
This was quite a poor performance that could be improved on. He didn't really seem confident about what he was saying especially when talking about clustering and talking about the technical details of these.
It's quite humiliating that data science has come to this and fallen so far behind software engineering in professionalism, the whole interview is basically import sklearn, model.fit and then a few gushy words about automatic retraining without any technical details lol
Agree, he didn’t do it well
Thank you! It's very informative!
Thank for your insightful video.
Roles have been flipped!