← LinkedIn Interview Insights

LinkedIn·Data Scientist·Technical Phone Screen·Intermediate

Intermediate
Jun 2026

Summary

LinkedIn data science technical screen, pretty much all stats fundamentals. Six questions back to back covering queuing theory, height distributions, social network degree distributions, and regularization. Nothing crazy hard but a couple of them required more careful thinking than I expected.

Questions Asked (6)

Q1

A bank has 5 tellers. System A has all 5 sharing one common queue. System B gives each teller their own separate queue. As a customer, which would you prefer and why? Compare the two on expected wait time, variance of wait time, and fairness.

Technical Trade-offsProduct Analytics & Metrics
Author's notes

This was the most interesting question of the screen.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying that as a customer, you prefer System A (single queue) because it minimizes expected wait time and variance, and is fairer. Then explain the underlying queueing theory and practical implications, using metrics to compare both systems.

Pro tip: Mention that while single queue is generally better for customers, multi-queue can be preferred by tellers due to perceived autonomy, but that's a trade-off between customer experience and employee satisfaction. Also, note that the analysis assumes homogeneous tellers and customers; in reality, variability in service times and customer types can affect the comparison.

1. Clarify the scenario and assumptions

State that you assume customers arrive randomly, service times are variable, and tellers are equally skilled. Clarify that you are evaluating from a customer perspective.

2. Compare expected wait time

Explain that in System A, the single queue feeds all tellers, so a customer waits only if all tellers are busy. In System B, a customer may wait even if other tellers are idle. Thus, expected wait time is lower in System A.

3. Compare variance of wait time

Highlight that System A reduces variance because the queue is pooled; a slow customer only blocks one teller, not an entire queue. In System B, variance is higher due to random assignment to queues.

4. Discuss fairness

Argue that System A is fairer because it follows first-come-first-served order; in System B, a customer might be stuck behind a slow customer while later arrivals get served faster in other queues.

5. Conclude and mention practical considerations

Conclude that System A is preferable for customers. Acknowledge that System B might be used for other reasons (e.g., teller specialization, perceived control) but those are not customer-centric.

Key Points to Mention

  • Queueing theory: single queue with multiple servers (M/M/c) vs. multiple queues (M/M/1) with random assignment.
  • Expected wait time is lower in single queue due to pooling of resources.
  • Variance of wait time is lower in single queue, leading to more consistent service.
  • Fairness: single queue ensures first-come-first-served; multiple queues can lead to queue jumping and perceived unfairness.
  • Real-world examples: banks, airports, call centers often use single queue for better customer experience.
  • Trade-offs: multiple queues may reduce teller idle time if customers have different service needs, but that's not the customer's primary concern.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

Describe the distribution of adult male heights in the US, and separately, the distribution of adult female heights.

Product Analytics & Metrics
Author's notes

Basically a freebie.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by stating that adult heights in a homogeneous population follow an approximately normal distribution, then provide the mean and standard deviation for each sex based on CDC/NHANES data. Explain that the distributions differ in both center and spread, with males taller on average and slightly more variable. Finally, discuss practical implications for product analytics, such as setting height-based thresholds or segmenting users.

Pro tip: Mention that the distributions overlap substantially, so height alone is a weak classifier of sex; this shows you understand effect sizes and avoid overgeneralizing from group differences.

1. State the distributional form

Explain that adult heights within each sex are approximately normally distributed, citing biological and statistical reasons (many small additive genetic and environmental factors).

2. Provide parameters for males

Give the mean and standard deviation for US adult males, e.g., mean ≈ 69 inches (175 cm), SD ≈ 3 inches (7.6 cm), based on CDC/NHANES data.

3. Provide parameters for females

Give the mean and standard deviation for US adult females, e.g., mean ≈ 63.5 inches (161 cm), SD ≈ 2.5 inches (6.4 cm), noting the lower mean and slightly smaller spread.

4. Compare and contrast

Highlight that the two distributions overlap significantly (e.g., many women are taller than many men) and that the male distribution has a slightly larger variance.

5. Relate to product analytics

Discuss how these distributions might inform product decisions, such as setting height-based defaults or analyzing height-related user behavior, while cautioning against assuming normality for all subgroups.

Key Points to Mention

  • Approximate normality of height within each sex
  • Mean and standard deviation values from reliable sources (e.g., CDC/NHANES)
  • Difference in means (males taller by about 5-6 inches)
  • Difference in standard deviations (males slightly more variable)
  • Substantial overlap between the two distributions
  • Practical implications for product analytics and segmentation

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

If you pool men and women together and ignore sex entirely, what does the combined height distribution look like?

Product Analytics & Metrics
Author's notes

I said bimodal and sketched the logic verbally.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying that the combined distribution is a mixture of two normal distributions (male and female heights) with different means and variances. Explain that the resulting shape is typically bimodal if the means are sufficiently separated relative to the standard deviations, but may appear unimodal if they overlap significantly. Use the specific parameters of male and female height distributions to illustrate the shape, and discuss the implications for analysis.

Pro tip: Acknowledge that in practice, the combined distribution might look unimodal due to overlapping ranges, but statistically it's a mixture; this shows you understand both the theoretical and practical aspects. Also, mention that ignoring sex can lead to misleading conclusions in product analytics, such as missing subgroup differences.

1. Clarify the distributions

State that male and female heights each follow approximately normal distributions with different means and standard deviations. For example, adult male heights ~ N(175 cm, 7 cm) and female heights ~ N(162 cm, 6 cm) (or similar).

2. Define the mixture distribution

Explain that the combined distribution is a mixture of these two normal distributions, with weights proportional to the population proportions of men and women (e.g., 50-50 or based on context).

3. Analyze modality

Discuss that the mixture can be bimodal if the difference in means is large relative to the standard deviations (e.g., difference > 2*SD). Otherwise, it may appear unimodal with a wider spread. Use the specific numbers to determine which case applies.

4. Describe the shape

If bimodal, describe two peaks corresponding to male and female modes, with a dip in between. If unimodal, describe a single peak with greater variance than either subgroup. Mention that the overall variance is a combination of within-group and between-group variance.

5. Relate to analytics implications

Explain why this matters: ignoring sex can mask important differences, affect segmentation, and lead to biased insights. For example, in product analytics, pooling without considering sex might hide distinct usage patterns.

Key Points to Mention

  • Mixture of two normal distributions
  • Bimodality depends on separation of means relative to standard deviations
  • Overall variance includes between-group variance
  • Population proportions as mixture weights
  • Practical implications: Simpson's paradox, subgroup analysis
  • Visualization: histogram or density plot would show shape

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q4

On a social network like LinkedIn, what does the distribution of number of connections per user look like? Is it symmetric, left-skewed, or right-skewed?

Product Analytics & MetricsProduct Sense & Ideation
Author's notes

Right-skewed, power-law-ish.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by reasoning about the underlying mechanisms that drive connection growth, such as preferential attachment and network effects, to infer the shape of the distribution. Then, explicitly state that the distribution is right-skewed (positively skewed) and explain why, using examples like a few power users with thousands of connections while most have modest numbers.

Pro tip: Acknowledge that while the overall distribution is right-skewed, it may not be a perfect power law; factors like LinkedIn's connection limits and user behavior can create a log-normal-like shape. This shows nuance and avoids oversimplification.

1. Identify the metric and population

Clarify that we're analyzing the number of connections per user across all LinkedIn users, including active and inactive accounts.

2. Consider generative mechanisms

Discuss how network effects and preferential attachment lead to a 'rich get richer' phenomenon, where highly connected users attract more connections.

3. Determine skewness

Conclude that the distribution is right-skewed because a small fraction of users have a very large number of connections, pulling the mean above the median.

4. Validate with real-world context

Mention that this pattern is common in social networks and aligns with observed data, such as LinkedIn's own analytics or public research.

Key Points to Mention

  • Right-skewed (positively skewed) distribution
  • Preferential attachment / rich get richer
  • Power law or log-normal distribution
  • Mean > median due to outliers
  • Network effects and virality
  • LinkedIn's connection limit (30,000) as a potential cap

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q5

For that connections distribution, how do the mean, median, and mode compare? And roughly what factors would determine the scale of the mean?

Product Analytics & Metrics
Author's notes

Mode less than median less than mean, classic right-skew ordering.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying what 'connections distribution' refers to (e.g., number of connections per user) and the population (e.g., all LinkedIn members). Then describe the typical shape of such a distribution (right-skewed) and explain how mean, median, and mode relate, and finally discuss factors that influence the scale of the mean.

Pro tip: Acknowledge that the distribution is likely heavy-tailed and that the mean is sensitive to outliers, so it's often more informative to report median and percentiles. Also, mention that the mean can be influenced by platform design choices (e.g., connection limits) and user behavior.

1. Clarify the distribution

Confirm that we're talking about the distribution of number of connections per user on LinkedIn. Note that it's likely right-skewed with a long tail of highly connected users.

2. Compare mean, median, mode

Explain that in a right-skewed distribution, mean > median > mode. The mode is likely a small number (e.g., 0 or a low number) because many users have few connections, while the mean is pulled up by power users.

3. Identify factors affecting the mean

List factors that determine the scale of the mean: user activity level, tenure on platform, network effects, platform policies (e.g., connection limits), and presence of bots or inactive accounts.

4. Discuss implications

Explain why this matters for product analytics: the mean may not represent the typical user, so median or percentiles are often better for understanding typical behavior. The mean is useful for total network size or resource planning.

Key Points to Mention

  • Right-skewed distribution with a long tail
  • Mean > median > mode in right-skewed data
  • Mode likely at 0 or a small number
  • Mean influenced by outliers (power users)
  • Factors: user tenure, activity, platform limits, bots
  • Median and percentiles better for typical user experience

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q6

Are L1-regularized (lasso) and L2-regularized (ridge) estimators biased? Why or why not?

Technical Trade-offsAlgorithms & Data Structures
Author's notes

Both are biased.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Clarify that both L1 and L2 regularized estimators are biased by design, trading increased bias for reduced variance to improve generalization. Explain the mechanisms: L1 introduces bias through soft-thresholding and sparsity, while L2 shrinks coefficients proportionally. Conclude that the bias is intentional and often beneficial, especially in high-dimensional settings.

Pro tip: Mention that the bias-variance tradeoff is central here: regularization deliberately increases bias to decrease variance, and the optimal amount of regularization is typically chosen via cross-validation. This shows you understand the practical implications beyond just the mathematical definition.

1. Define bias in estimation

State that an estimator is biased if its expected value does not equal the true parameter. For regularized estimators, the penalty term intentionally shifts estimates away from the ordinary least squares (OLS) solution.

2. Explain L2 (ridge) bias

Describe how L2 regularization adds a penalty proportional to the square of coefficients, shrinking them towards zero but not exactly to zero. This shrinkage introduces bias that increases with the regularization parameter λ.

3. Explain L1 (lasso) bias

Describe how L1 regularization adds a penalty proportional to the absolute value of coefficients, leading to soft-thresholding and sparse solutions. This also introduces bias, and for large λ, coefficients can be set exactly to zero.

4. Discuss the bias-variance tradeoff

Emphasize that the bias is deliberate: by accepting some bias, we reduce variance and often achieve lower mean squared error, especially when multicollinearity or high dimensionality is present.

5. Conclude with practical implications

Summarize that both estimators are biased, but the bias is a feature, not a bug. The choice between L1 and L2 depends on whether sparsity or shrinkage is desired, and the amount of bias is controlled by λ.

Key Points to Mention

  • Bias is defined as E[θ̂] - θ; regularized estimators do not satisfy E[θ̂] = θ.
  • L2 regularization shrinks coefficients proportionally, introducing bias but never setting them exactly to zero.
  • L1 regularization performs soft-thresholding, introducing bias and enabling feature selection by setting some coefficients to zero.
  • The bias increases with the regularization parameter λ; λ=0 recovers the unbiased OLS estimator.
  • The bias-variance tradeoff: regularization increases bias to reduce variance, often lowering overall mean squared error.
  • Cross-validation is commonly used to select λ, balancing bias and variance for optimal predictive performance.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.