Data science tests offered

What does a data science interview question really test when it asks for a definition? Most of the time, it is checking whether the candidate can connect a term to a decision, a model, or a result.

A weak answer sounds like a glossary. A useful answer shows how the idea fits into work.

Start with the job of the data

Data science work often begins with a question. A team gathers data from one or more sources, names the outcome it wants to explain, and then identifies the inputs that may affect it.

That simple shape appears in many interview questions. A good answer often moves in this order: find the data, define the question, name the variables, look for patterns, fit a model, and then judge whether the result answers the original question.

Here is a small example. Suppose a team wants to study whether annual temperature changes in India are linked to pollution and deforestation. Temperature is the outcome. Pollution and deforestation are the predictors. The point is not the example itself. The point is that the question decides which variable plays which role.

Know the basic variable types

Interview questions often use plain words where the meaning matters.

A categorical variable groups observations into names or classes. Species is a simple case. So is blood type. There is no built-in order.

A binary variable has two classes. Alive or dead is one example. Yes or no is another.

A nominal variable is a category with no order, even if numbers appear in it. Zip codes fit here. So do color labels.

An ordinal variable has order. Education level and race placement are common examples. The categories rank higher or lower, but the gaps between them are not equal.

A continuous variable can take any value in a range. Temperature and height are familiar cases.

A discrete variable is countable. Number of customers is a good example.

These labels matter because they shape the model choice and the summary you use. A count behaves differently from a temperature reading. A ranked label behaves differently from a true measurement.

Measurement is not the same as truth

A common test topic is measurement error. That is the gap between what was recorded and what was real.

Two other words show up a lot. Validity asks whether the instrument measures the right thing. Reliability asks whether it gives stable results under similar conditions.

Those ideas are easy to mix up. A scale can be reliable and still be wrong by a fixed amount. It can keep giving the same bad answer. That is a reliability problem solved only if the tool is also valid for the job.

Correlation, experiments, and variation

Some interview questions ask how people study cause and effect.

A correlational setup watches what happens in the world without changing anything. It is useful, but it does not prove cause. An experimental setup changes one thing and watches the result.

That difference matters because variation can come from two places. Some variation is systematic. Some comes from noise. If an experiment changes temperature, that is systematic variation. If a participant gets sick, that is unsystematic variation.

Randomization helps here. It spreads hidden differences across groups so one factor does not quietly explain the result.

Summary measures are small tools with real weight

Questions about mean, median, mode, and range look basic. They are basic. They are also useful.

The mode is the most common value. The median is the middle value after sorting. The mean is the arithmetic average. The range is the max minus the min.

These measures answer different questions. The mean shifts when extreme values appear. The median resists outliers better. A trimmed mean sits between them because it drops the extremes before averaging.

Here is a simple example. Imagine five salaries: 30, 32, 34, 36, and 200. The mean is pulled up hard by the 200. The median stays near the center. That is why interviewers often ask which summary fits the data shape.

Standard deviation and variance also matter. Variance is the average squared spread from the mean. Standard deviation is its square root, so it returns to the original unit. A larger value means the data are more spread out.

Sample, population, and the limits of certainty

Interview questions often shift from raw data to inference. That is the move from sample to population.

A sample is the smaller set actually observed. A population is the larger group the result is meant to speak about. Estimation uses the sample to guess the population.

Standard error appears here too. It describes how much a sample mean tends to vary as a statistic. A smaller standard error means the mean is a tighter estimate of the population mean.

Confidence intervals build on that idea. A 95% interval is a range built by a rule that would capture the true mean about 95% of the time across repeated samples. It is not a promise about one specific interval. It is a statement about the method.

Bias and accuracy are separate as well. A result can be close to the truth on average and still be noisy. It can also be consistently off target. Bias is the steady miss. Accuracy is closeness to the true value.

Hypotheses and errors

Many interview questions use the language of hypothesis testing.

The null hypothesis says there is no effect. The alternative says there is an effect. The p-value tells how surprising the observed data would be if the null were true.

That p-value is compared with a cutoff such as alpha. Alpha is the chance of rejecting a true null. That is a Type I error.

A Type II error is the other mistake. It means failing to reject a false null.

These ideas sound abstract until they are tied to a decision. A test does not prove a claim in full. It gives a rule for weighing evidence under uncertainty.

Model fit and learning type

A model can underfit, overfit, or fit well.

Underfitting means the model is too simple for the pattern. Overfitting means it memorizes noise as if it were signal. A good fit captures the useful pattern without chasing every odd point.

Machine learning questions also ask about learning paradigms. Supervised learning uses labeled data and predicts a target. Unsupervised learning looks for structure in unlabeled data. Semi-supervised learning mixes both.

A classification model is a supervised model that predicts a category. A feature is one measurable input used by the model. Binary classification has two classes. Multi-class classification has more than two.

What a clear interview answer sounds like

A strong answer names the term, then shows its use.

If asked about an ordinal variable, a clear answer might say it is a category with order, like education level, and that the spacing between levels is not equal. If asked about standard error, a clear answer might say it shows how much a sample mean tends to move from sample to sample. If asked about overfitting, a clear answer might say the model learned noise too well and may fail on new data.

That kind of reply does two things at once. It proves the definition is known. It also proves the concept is usable.

I find that interview prep gets better when the learner stops memorizing word pairs and starts matching each term to a job. The term is the label. The job is the real test.

The next useful step is plain. Take one common question, such as “What is the difference between mean and median?” and answer it in one sentence, one example, and one warning about when the measure misleads. That is the kind of small, steady practice The Dravelo Field Notes is built around: one practical technical idea, one learning decision, and one useful network resource each edition.