Nail Your Next Machine Learning Interview

Sorting

• Introduction to Sorting
• Basics of Asymptotic Analysis and Worst Case & Average Case Analysis
• Different Sorting Algorithms and their comparison
• Algorithm paradigms like Divide & Conquer, Decrease & Conquer, Transform & Conquer
• Presorting
• Extensions of Merge Sort, Quick Sort, Heap Sort
• Common sorting-related coding interview problems

Recursion

• Recursion as a Lazy Manager's Strategy
• Recursive Mathematical Functions
• Combinatorial Enumeration
• Backtracking
• Exhaustive Enumeration & General Template
• Common recursion- and backtracking-related coding interview problems

Trees

• Dictionaries & Sets, Hash Tables 
• Modeling data as Binary Trees and Binary Search Tree and performing different operations over them
• Tree Traversals and Constructions 
• BFS Coding Patterns
• DFS Coding Patterns
• Tree Construction from its traversals 
• Common trees-related coding interview problems

Graphs

• Overview of Graphs
• Problem definition of the 7 Bridges of Konigsberg and its connection with Graph theory
• What is a graph, and when do you model a problem as a Graph?
• How to store a Graph in memory (Adjacency Lists, Adjacency Matrices, Adjacency Maps)
• Graphs traversal: BFS and DFS, BFS Tree, DFS stack-based implementation
• A general template to solve any problems modeled as Graphs
• Graphs in Interviews
• Common graphs-related coding interview problems

Dynamic Programming

• Dynamic Programming Introduction
• Modeling problems as recursive mathematical functions
• Detecting overlapping subproblems
• Top-down Memorization
• Bottom-up Tabulation
• Optimizing Bottom-up Tabulation
• Common DP-related coding interview problems

Online Processing Systems

• The client-server model of Online processing
• Top-down steps for system design interview
• Depth and breadth analysis
• Cryptographic hash function
• Network Protocols, Web Server, Hash Index
• Scaling
• Performance Metrics of a Scalable System
• SLOs and SLAs
• Proxy: Reverse and Forward
• Load balancing
• CAP Theorem
• Content Distribution Networks
• Cache
• Sharding
• Consistent Hashing
• Storage
• Case Studies: URL Shortener, Instagram, Uber, Twitter, Messaging/Chat Services

Batch Processing Systems

• Inverted Index
• External Sort Merge
• K-way External Sort-Merge
• Distributed File System
• Map-reduce Framework
• Distributed Sorting
• Case Studies: Search Engine, Graph Processor, Typeahead Suggestions, Recommendation Systems

Supervised Learning I - Rank Relevant Search Results

• Deep dive into the design of a search relevance system like Google Search (a popular FAANG interview question).
• Comprehensive coverage of document indexing, retrieval, similarity scoring, filtering, and ranking.
• Model training: Build a ranking model using Linear or Logistic Regression. Which method performs better for ranking a search result?
• Online testing scenarios: How to run an A/B test to evaluate if new features improve the model? Which metrics to use?
• Interesting follow-up questions: How do you offset the lack of negative samples for training without adding new data? Does a “no-click” impression correspond to a negative label?
• Non-trivial questions on fundamental topics: Does L1/L2 regularization always increase model performance? Which metric is best for evaluating models on imbalanced datasets? When is the k-nearest neighbors algorithm better than logistic regression for classification analysis?

Supervised Learning II - Design a YouTube Video Recommendation System

• Build a video recommendation system for YouTube users. Answer questions like: How to maximize user engagement? How to recommend new content to users?
• Best practices of feature engineering, data collection, feature encoding, and video embeddings.
• In-depth coverage of content-based and collaborative filtering, matrix factorization, and maximizing the optimum objective function.
• Triage critical concerns when building a recommender system: How do you account for positional bias when ranking a video? Strategies to ensure the freshness, diversity, and fairness of the recommendations.
• Non-trivial questions on fundamental topics: How do you train a State Vector Machine (SVM) on non-linear data? Why only a subset of features for each tree in a random forest? What is the bias-variance tradeoff, and how can ensemble learning techniques like bagging and boosting address it?

Unsupervised Learning - Detect Fraud Transactions for Airbnb

• Design an anomaly detection system for Airbnb transactions.
• Comprehensive coverage of fraud detection techniques: Reputation lists, rules-based detection, classification vs. clustering.
• A top-down approach to building a high-level architecture: User and agent data, feature aggregation, model dashboard, data embedding
• Difficult follow-up questions: How to speed up computation time for unsupervised anomaly detection? Strategies to combine clustering with supervised learning techniques.
• Non-trivial questions on fundamental topics: What could be the possible reason(s) for producing two different dendrograms using an agglomerative clustering algorithm for the same dataset? Dimensionality reduction with computational power constraints - t-SNE vs. PCA?

Deep Learning I - Detect and Process Objects in a Scene

• Design an Image Processing system for Object Detection (frequently asked in FAANG ML interviews).
• Deep-dive into object detection workflow: Preprocessing, Candidate Generation and Selection, Unprocessing, and Postprocessing.
• Multiple strategies to build the object detector: Convolutional Networks, Region-based CNNs, You Only Look Once (YOLO), Transfer Learning, etc.
• Interview-relevant strategies: What questions to ask an interviewer? How to structure your solution?
• Address possible follow-up questions: How do you detect and replace multiple instances of the same object from an image? What if direct ground truth labels are absent? 
• Non-trivial questions on fundamental topics: Using dropout layers in a small neural network. How to combat vanishing and exploding gradients in CNNs? What is the best learning rate optimizer for improving model performance on large datasets?

Deep Learning II - Build a Tech Support Chatbot

• Design an intelligent Discord bot to provide Technical Support for a software Bootcamp.
• In-depth coverage of functional and non-functional requirements: Scale and latency estimation, throughput, passive feedback mechanisms.
• Knowledge base creation: Embedding, Sharding, Caching, etc. Strategies to expand the knowledge base.
• Challenges in bot design: How to deal with a cold-start with no knowledge base? How do you generate answers to previously unasked questions?
• Logical follow-up questions: How to handle increasing complexity and scale? How can we introduce a continuous learning mechanism in the chatbot design?
• Non-trivial questions on fundamental topics: Which type of word embedding method is more suitable for measuring context similarity? Why is “Exploding Gradients'' a problem in the context of RNNs? When should we not use a bi-directional LSTM?

Additional Topics:

• How is ML System Design different from General System Design?
• What does an interviewer expect from this round?
• How do you breakdown and answer open-ended questions like: 

• Why do we learn a distribution instead of a deterministic model during encoding? How do we introduce variability in a variable autoencoder?
• What is the difference between Discriminative and Generative models? 

• How do you evaluate the state and responses of an agent?
• How is value iteration different from policy iteration? What problem does it address?