Data Scientist Interview Questions

A p-value determines the results equal to or more than the results achieved under a specific hypothesis when the imaginary null hypothesis is correct. It is the measure of the probability and indicates the probability that the observed difference occurred by chance.

• Low p-value i.e., values <0.05 represents that the null hypothesis can be ignored, and data is not likely with true null.
• High p-value i.e., values >0.05 means the null hypothesis can’t be ignored and data is like with true null.
• P-value=0.05 indicates null hypothesis can go either way.

This is one of the most commonly asked data scientist questions which if answered correctly can increase your chances of getting hired. It is impossible to do data analysis on a large volume of data at a given time, especially on larger datasets. It is mandatory to take some data samples that can represent the whole data and then perform an analysis on it. While doing this, the sample data we are taking must be taken in a way that truly covers the whole dataset. This process is known as Sampling.

Categories of techniques used for sampling

• Probability Sampling Techniques: Simple Random Sampling, Stratified Sampling, Clustered Sampling.
• Non-Probability Sampling Techniques: Convenience Sampling, Snowball, and Quota Sampling.

When researchers have to make a decision regarding which participant to study, selection bias occurs in that case. It is associated with the research where participant selection is not random and is also known as selection effort.

Types of Selection Bias.

• Sampling Bias: Some members of a dataset have fewer chances of getting selected than others, which results in a biased sample and hence causes an error known as sampling bias.
• Time Interval: If we reach any extreme value, we can stop the trials early. But if all variables are similar, the variable with the high variance has more chance of achieving the extreme value.
• Data: When specific data is picked randomly, and the agreed criteria are not followed.
• Attrition: Loss of the participants is known as attrition.

Logistic regression which is also known as the logit model is a technique that predicts binary outcomes from the linear combinations of predictor variables.

Example: Let’s suppose we want to predict the results of elections for a political leader. We will assume whether he is going to win or not. Therefore, the outcome is binary i.e. win (1) or loss (0). But the input will be a combination of various linear variables like money spent on advertisement, their past work history, etc.

This data scientist interview question gives an idea to the interviewer if you are familiar with the algorithms and machine learning.
 An algorithm is a well-defined procedure to resolve any issue and frequent changes should not be made to an algorithm on a regular basis as it won’t be well-defined anymore. It also brings various problems to the other existing algorithms.

Therefore, it should be updated in the below cases.

• It is fine to make changes if you want the model to evolve.
• It is necessary to update the algorithm in the case underlying data sources are changing.
• In the case of non-stationarity.
• One of the major reasons which required updating the algorithm is its underperformance and lack of efficiency.
   

As dirty data often results in poor and incorrect output which can have damaging effects, it is very important to do data cleaning to have correct and relevant information.

• Cleaned data highly increases the accuracy of a model and gives good predictions.
• It results in increased speed and efficiency of an application.
• Data cleaning helps a user to identify any high-risk issues and helps them to fix them.
• It maintains data consistency and helps in removing duplicates.
• Data cleaning can also increase the data quality.

To identify the missing values the criteria is to find the variables with the missing values. Suppose a pattern is identified. Concentrating on it could give you interesting and meaningful observations. However, if in case no patterns are identified, we can replace the missing values with mean or median values or we can simply ignore them.

If the variable is categorical, we assigned the default value to the mean, minimum, and maximum. The missing value is then assigned to that default value.
If for a variable, 80% of the values are missing, then instead of treating the missing values we would drop that variable.

Welcome to the first real-life data scientist interview question. Grab this opportunity by showing your understanding of the recommendation engine. These recommendations are based on a recommendation engine which is accomplished with collaborative filtering. Collaborative filtering determines the behavior of other users, their purchase history in terms of reviews, ratings, and selection, etc. These engines make a prediction on what a customer can buy based on the preferences of other customers. Item features are unknown in this algorithm.

• Monitor: To determine a model’s accuracy, constant monitoring is needed. When we make any changes, we want to check if those changes are going to affect anything. And this needs to be monitored constantly to ensure if the changes are doing what they are supposed to do.
• Evaluation:  Evaluation of any model is necessary to determine if a new algorithm is needed.
• Comparison: To check which model performance is the best, a comparison between models is needed.
• Rebuilding: To achieve a best-performing model, re-built on the current state of data is required.