Statistics & Experimentation Interview Questions

Dhruv S. - "Before proceeding, I just wanted to clarify we wanted to check for the impact of showing content from non-friends in users’ feeds, and here non-friends I would assume could be anyone, but mainly like content creators, and I am not including ads here. But I wanted to ask if there is any current logic as to what posts to show based on users' affinity to those posts, maybe basis the user engagement to Insta feed. now objective of this would be to improve the engagement of the platform, as if users"See full answer

Himanshu G. - "To model ROI for a product launch, the first step is to define the timeline you're targeting Example 6 months post-launch, 1 year, or even 5 years. Tip: Start with a 1-year ROI projection to estimate near-term returns, and build a 3-year projection to evaluate growth and scalability. ROI is essentially the net return over that period: Profit=Revenue (within timeline)−Total Cost (from project start) Total Cost includes both fixed and variable costs incurred since t"See full answer

Saurabh K. - "Missing Item - User ordered multiple items, few items are missing Wrong Item - Entire order is wrong / there are items in the order that were never ordered How is this measured ? CSAT Missing Items Wrong Items Step 1 : Collect data on orders that reported missing / wrong items. Dive deep to understand if the problem is isolated to a specific metro/zip code/restaurant type (say fast food vs fine dine), time of day (lunch vs dinner), tenure of the courier on th"See full answer

Aarav G. - "Is there a reason a confidence interval was used to solve this problem over just using the mean/expected value directly?"See full answer

Ranj A. - "In the Transformer architecture, the decoder differs from the encoder primarily in its additional mechanisms designed to handle autoregressive sequence generation. Here's a breakdown of the key differences: Self-Attention Mechanism: Encoder: The encoder has a standard self-attention mechanism that allows each token to attend to all other tokens in the input sequence. Decoder: The decoder has two types of self-attention. The first is the same as in the encoder, but the second is mas"See full answer

Saurabh K. - "P(A) = 0.6 P(B) = 0.4 P(D|A) = 0.05 P(D|B) = 0.03 Question asks to solve for P(A|D) P(A|D) = (P(D|A) x P(A))/P(D) = (0.05 x 0.6)/(P(D|A) x P(A) + P(D|B) x P(B)) = (0.05 x 0.6)/(0.05 x 0.6+0.03 x 0.4) = 30/42 = 5/7 = 0.714 Notice above that P(D) = P(D|A) x P(A) + P(D|B) x P (B)"See full answer

Jiin S. - "I would use A/B testing to see if the new feature would be incrementally beneficial. To begin the testing, we should define what's the goal of this testing. Let's say the new feature would increase the average number of trade by X. Then randomly assign the clients to two groups, control and test group. Control group doesn't see the new feature and the test group see the new feature. We could also stratified sampling if we want to make sure cover different customer segmentation. During this desig"See full answer

Cesar F. - "Outliers are data points that significantly deviate from the majority of the data distribution. They can arise due to various reasons, such as measurement errors, natural variability, or rare events. Outliers can distort statistical analyses and machine learning models, making it crucial to detect and handle them properly."See full answer

Will I. - "Is it bad to get the answer a different way? Will they mark that as not knowing Bayes Theorem or just correct as it is an easier way to get the answer? The way I went is to look at what happens when the factory makes 100 light bulbs. Machine A makes 60 of which 3 are faulty, Machine B makes 40 of which 1.2 are faulty. Therefore the pool of faulty lightbulbs is 3/4.2 = 5/7 from machine A and 1.2/4.2 = 3/7 from Machine B."See full answer

Abhinav J. - "Use Normalization when: When using pixel values (0-255) into a Neural Network. ➔ Normalize the data between [0,1] to avoid huge input values that could slow down training. When using k-Nearest Neighbors (kNN) or K-Means Clustering. ➔ Because distance metrics like Euclidean distance are highly sensitive to magnitude differences. You are building a Recommender System using Cosine Similarity.➔ Cosine similarity needs data to be unit norm. Use **Sta"See full answer

Connor W. - "In the expected value of a coupon problem, you calculated variance of a binomial distribution, and used the satandard deviation, square root of variance, to calculate the confidence interval. Will that approach work the same here? For fair coin: (Heads = 0, tails = 1) Var = 10 * (.5)(1-.5) = 2.5 Stdev = Sqrt(2.5) = 1.581 Mean = 5 Z-score = (Observed Val - Mean) / Stdev = (10 - 5) / 1.581 = 3.164 P val = 0.0008% (Slightly different from the video's solution of 0.00097) Pros of this approach: It"See full answer