Data Visualization Interview Questions

Sample Answer

Use an overlaid distribution view, a log-scaled histogram or density, plus a quantile plot showing p50, p90, p99 for both weeks. The tail is the business risk, so you need an encoding that makes high latencies visible rather than compressing them into a mean. A log $x$ axis is justified because latency spans orders of magnitude, and it prevents the tail from flattening into a line. If you must pick one chart under time pressure, choose a quantile dot plot with two colors, it makes regression at p99 instantly readable.

Sample Answer

You could do a correlation heatmap or a sorted coefficient bar chart from a simple logistic regression. The bar chart wins here because you can rank effects, show directionality, and label units, while a heatmap invites misreads and hides magnitude. Put attributes on the $y$ axis, estimated effect on conversion on the $x$ axis, center at 0, and use horizontal bars with confidence intervals. If attributes have different units, standardize inputs or plot effects for meaningful deltas like $+1$ review star, $+10\%$ price, so the encoding stays comparable.

Sample Answer

First, decide what the business question is: compare plans, prioritize countries, or estimate total impact. Next, map each country to a row and use a dot plot with three points per row, Current, Plan A, Plan B, on an $x$ axis of ARPU, this makes deltas easy to see without bars dominating. Then encode user count as point size or add a right-side column showing users, so you do not confuse ARPU changes with revenue contribution. Finally, if leadership wants total revenue, compute $\text{revenue} = \text{ARPU} \times \text{users}$ and show a separate ranked bar chart of incremental revenue by country to avoid mixing metrics on one axis.

Sample Answer

You could join payments to users as is, or you could first reduce users to one row per user with a clear effective dating rule. The raw join is faster but it will duplicate transactions when a user has multiple profile rows, which inflates revenue bars. The one row per user approach wins here because it preserves the transaction grain, for example pick the latest country as of paid_at using a window function. After that, normalize currency to a single unit before aggregating, and decide whether refunds subtract revenue or are excluded, then group by week and country.

Sample Answer

First, I check whether duration_seconds is derived or raw, and I recompute it as $\text{ended\_at} - \text{started\_at}$ when both timestamps exist. Next, I handle missing ended_at by either dropping in progress sessions or censoring them at the extract time, but I label that choice because it changes the left tail. Then I remove or fix impossible negatives by investigating clock skew or swapped fields, and I cap extreme outliers with a rule like winsorizing at the 99.5th percentile so the binning is not dominated. Finally, I confirm the grain is one row per session_id and I use consistent bin widths so comparisons across days mean something.

Sample Answer

First, you define the primary decision, like where to invest supply and ranking work, then you pick a default view that answers it for most users, usually global plus top markets. Next, you separate global filters from local interactions: global for date, product surface, experiment arm, local for country and segment so users do not accidentally create empty states. Then you design drilldowns in a fixed order, KPI, driver decomposition, segment comparison, and finally row level examples, each step preserving context and showing the active filter set. Finally, you add guardrails: pinned defaults, filter reset, and a warning when sample size drops below a threshold so performance and interpretation stay reliable.

Sample Answer

This question is checking whether you can prevent KPI theater by making definitions, denominators, and mix effects explicit. You should define conversion as $$\text{Bookings} / \text{Unique searchers}$$ and display the denominator trend alongside the rate so a change in who is searching is visible. Add a mix adjusted conversion metric, like reweighting to a fixed country and trip length distribution, and show raw vs adjusted side by side. Include a definition tooltip that states inclusion rules, like excluding canceled stays or bot traffic, and a driver view that decomposes change into volume, mix, and within segment performance.

Sample Answer

The standard move is to keep the landing page decision focused: a small set of North Star and guardrail metrics with clear deltas and a single default cut. But here, audience needs diverge, execs want status and exceptions, analysts want diagnostics, so you split the experience into a clean summary plus guided drill paths. Use progressive disclosure: top shows overall viewing, completion, and retention, then drill into title, cohort, device, and region. Add an exceptions module that auto surfaces the largest movers so execs get signals without scanning every chart.

Sample Answer

This question is checking whether you can separate signal from confounders, and communicate confidence clearly without hiding behind jargon. You lead with a confidence statement tied to a method, for example, "After controlling for day of week and the campaign window, the lift is about 2 to 3%, and it is outside normal variance." You show one visual that decomposes the series, like actual versus expected with the campaign shaded, plus a simple uncertainty band. You close with what you would do next to de risk the decision, like holdout measurement, cohort slicing, and a date by which you will confirm causality.

Sample Answer

The standard move is to lead with the global KPI impact and a clear recommendation. But here, segment concentration matters because the average can hide that most users saw no change, and the business risk is in overgeneralizing. You show the overall lift, then immediately pivot to a contribution chart that makes it obvious where the lift comes from, plus guardrail metrics for the rest of the population. You recommend a scoped rollout, for example, target the high value segment first, run a broader experiment with pre set stop loss thresholds, and define what evidence would justify global shipping.

Sample Answer

The standard move is a funnel chart with treatment vs control at each step and a lift column. But here, guardrails matter because a win on an early metric can mask harm later, so you should add confidence intervals for each step and highlight the primary metric vs guardrails with clear labels. Include a delta funnel, for each stage show $Δ$ with CI, and explicitly call out multiplicity if you are scanning many steps. If retention is measured over time, add a cohort retention curve by variant with shaded uncertainty bands to show whether the difference persists.

Sample Answer

Get this wrong in production and you ship a pricing change based on confounding, then revenue drops when the campaign ends. The right call is to visualize the treatment effect by country alongside a timeline of external events, and separate in experiment estimates from observational correlations. Show a forest plot of country level effects with CIs, and flag the outlier country with an annotation about the campaign start date. In your narrative, you say the aggregate lift is not trustworthy until you rerun with the campaign excluded, extend the test, or use a difference in differences style check if assignment was compromised.