Microsoft DS 2 Interview Experience | Selected waiting for offer.
Summary
I recently interviewed with Microsoft for a DS 2 position and was successfully selected, currently awaiting the official offer. The interview process spanned several rounds, including an online assessment with ML notebook challenges, followed by in-depth discussions on data science system design, machine learning concepts, project deep dives, and behavioral aspects.
Full Experience
I recently appeared for an Online Assessment (OA) and subsequent interviews with Microsoft for a DS 2 role. I'm sharing my experience to help others, especially since specific Data Scientist interview experiences for FAANG-type companies aren't always widely available. Each round was eliminatory.
Round 0: Online Assessment
This round consisted of three machine learning notebook questions, with a time limit of 165 minutes. The task was to code end-to-end, starting with Exploratory Data Analysis (EDA), handling missing values, scaling data, training a model, and finally submitting the results as a CSV. Importantly, there were no DSA questions in my OA, though I've heard some DS2 OAs might include DSA and SQL. My strategy was to use simpler but effective models to ensure at least two questions could be completed within the given time.
Round 1: Principal Engineer
This round was with a principal engineer. We started with two SQL questions, although I haven't detailed the specific problems. The discussion then moved into Data Science system design, focusing on how I would handle a new project from end to end. We also covered various LLM concepts, including how LLMs work, the time complexity of attention mechanisms, calculating the number of parameters in an LLM given vocabulary and other parameters, and decoding in LLMs. Finally, I was asked to explain cross-entropy and differentiate between Bagging and Boosting.
Round 2: DSA Round
This was a dedicated Data Structures & Algorithms round. There were two DSA questions with follow-ups, which I needed to code in my chosen language. The first question was based on BFS/DFS, and the second involved dynamic programming with space and time optimization follow-ups. The expectation was to provide working code, explain the complexities, and articulate my approach clearly. The interviewer tested my code with a few sample cases.
Round 3: Project Deep Dive
This round focused on a deep dive into my past projects. The discussion centered on ML system evaluation, data preparation strategies, the justification behind heuristic choices, and the rationale for model selection. The interviewer actively suggested alternate design choices, leading to a constructive discussion. We also talked about problems with RNNs, specifically why they face vanishing gradient issues, and I provided a detailed overview of how Transformers work, including the justification for skip connections and normalization. LLM evaluation was also a topic, along with various evaluation metrics like confusion matrix and AUC curves (I explained two types and their use cases).
Round 4: AA Round
My final round was with a group manager who had about 25 years of experience. This round involved discussions on ML system design, A/B testing, and methods to assess the quality of our tests. Standard behavioral questions were also posed, such as my strengths, weaknesses, and my plan for improvement.
Approximately an hour after this round, I received a call from the recruiter confirming positive feedback and requesting documents to initiate the offer process. I'm currently waiting for the official offer.
Interview Questions (9)
For the online assessment, I was given 3 machine learning notebook questions to complete in 165 minutes. The task involved coding end-to-end, starting from Exploratory Data Analysis (EDA), handling missing values, scaling features, training a machine learning model, and submitting results as a CSV file. The expectation was to go with simpler but good enough models to complete at least two questions within the time limit. No DSA questions were present in this OA, though some DS2 OAs apparently include DSA and SQL.
I had a discussion on general Data Science system design, specifically how I would handle a new project from end to end.
Questions were asked about how Large Language Models (LLMs) work, including the time complexity of the attention mechanism, calculating the number of parameters in an LLM given vocabulary and other parameters, and the decoding process in LLMs.
I was asked to explain cross-entropy loss and compare Bagging versus Boosting ensemble methods.
We delved into problems associated with Recurrent Neural Networks (RNNs), specifically why they suffer from vanishing gradient issues. I then had to explain how Transformers work, providing a detailed overview including the justification for skip connections and normalization layers.
I discussed methods and considerations for evaluating Large Language Models (LLMs).
I was asked to explain various machine learning evaluation metrics such as the confusion matrix and AUC curves (describing two types and their respective use cases).
This round involved a discussion on general ML system design, A/B testing methodologies, and how to determine if our tests are good.
I encountered standard behavioral questions, such as discussing my strengths and weaknesses and what I am doing to improve them.