60 minutes, coding + ML knowledge.
The coding had to be done with numpy. The input was a user embedding vector and N item embedding vectors. Taking the dot product of the two gives a similarity score. The task was to sort the N items by score from high to low, and the output is the list of item indices after sorting.
The follow-up: if the user has already seen some of those items, filter the seen items out of the final result.
I forgot the numpy syntax. The interviewer said that was fine, but he wouldn't let me google it myself. I could explain my approach and then ask him. In total I wrote under 10 lines of code... It's the easiest coding interview I've ever seen.
The ML knowledge questions were all the common ones: overfitting, underfitting, bias-variance tradeoff, class imbalance, how to handle missing data.
Overall a pretty easy interview, and the interviewer was really nice.
During the Q&A at the end, the interviewer said Yahoo used to be very chill, but after the recent reorg things have gotten busy, and now they're doing the "move fast" thing too.
Discussion
Loading comments…