There were four interviews in total, scheduled over two days, with one coding round and one ML knowledge Q&A each day.
Round 1, coding: You are given two matrices A and B, both of which are compressed representations of an original sparse matrix (only the positions of the non-zero entries and their values are shown). For example, if the original matrix is [[0,3,0],[4,7,0],[0,0,0]], its compressed representation is [(0,1,3),(1,0,4),(1,1,7)]. You have to implement A+B and A*B and output the result in the same kind of compressed representation. Use a dictionary.
Round 2, ML: You are given a problem and then you discuss it together. The problem: we want to build a classification system for images (pins) that identifies bad pins such as toxic and porn content. Each pin has the image itself, a text description and other information, plus user interactions such as reports. How do you use all of this to build a good detection system? It is very open-ended, but the questions go into quite fine detail, so you need a solid ML foundation.
Round 3, coding: The pin record problem that many other people have already talked about. I pasted my full answer here.
Round 4, ML: First I briefly introduced a paper I had written before, and they asked me a few questions about it. Then another open-ended problem: for the current pin, there are several existing boards (collections) the user could save it to, and you need to recommend which board the user is most likely to put it in. It leans toward recommender systems, and again they test how you collect and organize multiple features and build a model to do the task.
Overall the feedback was fairly positive, and the interviewers give you hints and discuss the problems with you. In the end I got rejected.
Discussion
Loading comments…