Question 1
Difficulty: medium
Tell me about a data science project where you had to clean messy data before analysis. What did you do?
Sample answer
In my last project, I worked with customer transaction data that had missing values, inconsistent date formats, and duplicate records. I started by profiling the dataset to understand where the biggest quality issues were. Then I standardized the date fields, removed duplicates based on business rules, and handled missing values differently depending on the column. For numeric fields, I used median imputation when the missingness was small and likely random, but for categorical fields I created an explicit 'unknown' category when that made more sense. I also checked for outliers to make sure they were real and not data entry errors. After cleaning, I documented every transformation so the process could be repeated. What I learned was that data cleaning is not just a preparation step—it often reveals important patterns about the business process itself, and it can change how you frame the problem before modeling even begins.
Question 2
Difficulty: easy
How would you explain the difference between precision and recall to a non-technical stakeholder?
Sample answer
I would explain it using a simple example that matters to the business. If we were building a fraud detection model, precision tells us how many of the transactions we flagged as fraud were actually fraud. High precision means we are not bothering the team with too many false alarms. Recall, on the other hand, tells us how many of the real fraud cases we managed to catch. High recall means we are missing fewer bad transactions. I usually say that precision is about being careful when raising an alert, while recall is about being thorough in finding the problem. Then I connect it back to the business goal, because the right balance depends on the cost of false positives versus false negatives. For example, in fraud detection, missing fraud may be more expensive than investigating a few extra alerts, so recall may matter more. Keeping the explanation grounded in the business context usually makes it much easier to understand.
Question 3
Difficulty: medium
Describe a time when you had to choose between two modeling approaches. How did you decide?
Sample answer
In one project, I had to predict customer churn, and I was deciding between logistic regression and a random forest model. Logistic regression was attractive because it was simple, easy to explain, and gave a strong baseline. The random forest performed a little better on validation data, but it was harder to interpret and more expensive to maintain. I compared them using cross-validation, looked at both ROC-AUC and recall, and also checked calibration because the business wanted usable probabilities, not just classifications. I also considered the audience: the retention team wanted to understand why a customer was likely to churn, not just see a score. In the end, I recommended logistic regression with careful feature engineering because the performance gap was small and the interpretability was a major advantage. I think good model selection is not just about the highest metric—it is about the best tradeoff between performance, explainability, and operational fit.
Question 4
Difficulty: easy
How do you handle missing data in a dataset?
Sample answer
I usually start by asking why the data is missing, because the reason matters as much as the amount. If values are missing completely at random and the number is small, I may use a simple approach like median or mode imputation. If the missingness is more systematic, I look deeper. For example, if a field is only missing for a specific user segment or time period, that may signal an issue in the collection process. In those cases, I might add a missingness indicator, use a more thoughtful imputation method, or even leave it missing if the model can handle it. I also consider whether the feature is important enough to justify more advanced treatment. I try not to overcomplicate things early on, but I also avoid blindly filling values just to make the data look clean. My goal is to preserve as much signal as possible without introducing bias or hiding patterns that could matter to the model.
Question 5
Difficulty: easy
What steps would you take to build your first baseline machine learning model for a new problem?
Sample answer
My first goal would be to understand the business question clearly, then translate it into a machine learning task. After that, I would define the target variable and make sure the labeling makes sense. I would do basic data exploration to understand the feature distributions, missing values, and any obvious leakage risks. Once the data is ready, I would split it properly into train, validation, and test sets, making sure the split reflects the real use case, especially if it is time-based. For the baseline model, I would usually start with something simple like logistic regression, linear regression, or a decision tree depending on the problem type. The reason is that a baseline gives you a reference point and helps you understand whether more complex models are actually worth it. I would evaluate it using the most relevant metric, document the assumptions, and then use the results to guide the next iteration. I think the baseline is valuable because it creates structure and prevents wasted effort.
Question 6
Difficulty: medium
Tell me about a time you found an error in your analysis or code. What did you do?
Sample answer
I once built an analysis for a product team and noticed that one metric looked unexpectedly strong compared to previous reports. Instead of presenting it immediately, I double-checked the data pipeline and found that I had accidentally included records from a future time window because of a filtering mistake. Once I identified the issue, I fixed the query, reran the analysis, and compared the corrected results with the original output to understand the impact. The corrected numbers were still useful, just less dramatic. I then explained the issue clearly to the team, along with what changed and why the new result was more reliable. I think being transparent was important because it showed that the mistake was caught early and corrected responsibly. That experience reinforced a habit I still use: when something looks too good or too unusual, I slow down and verify the assumptions before I trust the result. It saved me from making a misleading recommendation.
Question 7
Difficulty: hard
How would you approach an imbalanced classification problem?
Sample answer
I would first check how imbalanced the classes are and whether the imbalance reflects the real-world situation or a data issue. Then I would choose evaluation metrics carefully, because accuracy is often misleading in these cases. I usually look at precision, recall, F1-score, PR-AUC, and the confusion matrix depending on the problem. On the modeling side, I might try class weights, oversampling, undersampling, or a combination of those, but I would be careful not to distort the validation process. If the data is time-based or user-based, I would make sure the sampling strategy does not leak information. I would also look at threshold tuning because sometimes the model is fine, but the default decision threshold is not aligned with the business cost. For example, in a rare-event detection setting, catching more positives may matter more than maximizing overall accuracy. My general approach is to focus on the metric and decision process that actually reflect the business need, not just the class distribution.
Question 8
Difficulty: easy
Why do you want to work as a Junior Data Scientist instead of a more general analyst role?
Sample answer
I enjoy analysis, but I am especially interested in roles where I can move from describing what happened to predicting what might happen next and testing ideas systematically. What draws me to junior data science is the combination of problem solving, experimentation, and technical depth. I like working with data to find patterns, but I also want to build models, validate assumptions, and think about how results perform in real use cases. A general analyst role often focuses more on reporting and business insights, which is valuable, but I am looking for a position where I can grow stronger in statistics, machine learning, and Python-based workflows. I also like that junior data science involves both technical work and communication, because the best model is only useful if the team understands it and can act on it. I am excited by the chance to learn from experienced teammates while contributing practical work that supports real decisions.
Question 9
Difficulty: medium
How do you decide whether a model is good enough to ship?
Sample answer
I decide based on a mix of performance, reliability, interpretability, and business value. First, I check whether the model meets the core metric target and whether it improves meaningfully over the baseline. Then I look at stability across validation folds or time splits to see whether the performance is consistent. If the model is going into a real product or workflow, I also consider how easy it is to explain, monitor, and maintain. Sometimes a slightly less accurate model is the better choice if it is more robust or easier for stakeholders to trust. I also think about failure modes: what happens when the input distribution shifts, or when the model makes the wrong prediction? If the risk is high, I want more safeguards before shipping. I prefer to involve the business partner in the decision so we agree on what “good enough” means upfront. For me, a model is ready when it is accurate, understandable, and practical enough to support a real decision reliably.
Question 10
Difficulty: hard
If a product manager asked you to predict customer lifetime value, how would you approach the problem?
Sample answer
I would start by clarifying exactly how the business wants to use the prediction. Customer lifetime value can mean different things depending on the horizon, the revenue definition, and whether the goal is targeting, budgeting, or segmentation. Once that is clear, I would define the target carefully and make sure it is measured consistently. Then I would explore the data for useful signals such as purchase frequency, recency, average order value, engagement patterns, and customer tenure. I would also think about censoring, because some customers have not had enough time to show their true long-term value. For modeling, I might begin with a baseline regression or a simple probabilistic approach before trying more complex methods. I would validate using a time-aware split so the model reflects future prediction, not just historical fit. Finally, I would present the output in a way the product team can actually use, such as value bands or prioritized segments, rather than only raw predictions. That keeps the work tied to action.
