Recently, I have been getting a lot of requests to share my interview experience with Microsoft for the role of data and applied scientist, so I have compiled this article to help the community and other data science enthusiasts. Hopefully, my experience will add value to your preparations.

I had applied off campus in the Bing Ads team as I was looking for an NLP based role.

Interview timeline : Jan-Feb 2022

The overall process took around 2.5 months to get completed and comprised of a screening around, 4 Technical rounds and one final AA (managerial round). All the rounds were of one hour duration.

Screening round: After the initial introduction, I was asked about what is overfitting, explain bias variance tradeoff, what are the methods for overcoming overfitting, explain dropout and was asked to do a sample implementation of dropout on editor. After this I asked a couple of questions related to team’s work and my role.

Round 1: After initial introduction, we discussed a bit about my projects. Then I was given a design question the details of which I don’t remember, but the solution which I proposed involved creating clusters and a lookup table for the associated values. Then creating a random forest classifier to finally answer the design question. Questions about how decision tree is constructed, why not SVM in place of Decision trees, details on SVM’s formulation was asked. In the end, I asked a couple of questions to the interviewer and learnt from his insights.

Round 2: In this round, the discussion was related to word2vec models, negative sampling, n-gram language modelling, smoothing techniques, Later we discussed about LSTM’s vs RNN’s, vanishing gradients and why is it caused, how and why transformer is better compared to LSTM, drawback of Bert type models. Later, I asked some questions to the interviewer related to my role, which I don’t remember exactly.

Round 3: This round initially started with detailed discussion around my masters thesis project and involved questions related to my design choices. Later, I was asked to implement self attention using sample values in editor. The idea is to clearly know how it is calculated and the steps involved during it. After this questions on comparing different optimizers were asked and which one should be used when and why. In the end, I asked some of my queries which were patiently answered by the interviewer.

Round 4: In this round, there were initial questions on Knowledge distillation, subword tokenization methods, a question on showing the relationship between euclidean distance and cosine similarity. After this a design question was asked where I had to design an autosuggest system with low latency, and how it can be personalized for each user. Detailed discussion on design choices followed. I was not sure how to proceed, but the interviewer helped at crucial junctions by giving inputs and clarifying the requirements.

AA round: In this round I was initially asked about my educational background, then a question on what is AUC-ROC curve, how do you plot it and compute the AUC. With some random sample points was asked to plot the curve and compute the AUC-ROC score on editor, what is the difference between precision and roc curve. After this I asked questions regarding team’s work and other doubts that I had regarding my role.

Thanks a lot, I hope it helps the community in their preparation.