15 Most Common Data Science Interview Questions

Ans. Almost every supervised Machine Learning model is represented mathematically as y = f(x), where, y is the output variable and x is the dependent input feature. f(x) is a generalized representation of a Machine Learning model, that takes in input features(x_i = x₁, x₂, x₃….,x_n) and gives out predicted output (ŷ).

Ans. The bootstrap method is a resampling technique used to estimate statistics on a population by sampling a dataset with replacement. It is used in applied Machine Learning to estimate the skill of machine learning models when making predictions on data not included in the training data.

In Machine Learning, bootstrap sampling is used in an ensemble algorithm called bootstrap aggregation or Bagging. It helps avoid overfitting and improves the stability of a Machine Learning algorithm.

For example, the Bootstrap Aggregation algorithm helps create multiple different models from a single training dataset. A very popular use case is predicting election results for an area using sample data instead of taking election data of the entire population.

Ans. Overfitting is the phenomenon that occurs when we fit a huge amount of data into a Machine Learning model (which might include noisy and inaccurate data) or also when a model fits exactly to its training data points such that it starts giving wrong predictions. The model becomes ‘overconfident’ and treats data points outside its knowledge to be wrong or inaccurate. Statistically, Overfitting occurs due to low bias and high variance. Overfitting can be reduced by using a linear model for linear data, pruning unnecessary nodes in a Decision tree to reduce its complexity, training with more data where necessary etc.

Underfitting occurs when a Machine Learning model neither fits on the training data nor generalizes to the new data to test on. The model fails to recognize the underlying trends of the data. It usually happens when we have very fewer data to build an accurate model, also when we try to build a linear model with fewer non-linear data. Statistically, underfitting occurs due to high bias and low variance. Underfitting can be reduced by using other Machine Learning algorithms, fitting more data, removing noise from data, increasing the number of features by performing feature engineering etc.

Ans. Euclidean distance: Euclidean distance between two points is the length of the line that joins the two points. Consider two points A(x₁, y₁) and B(x₂, y₂) in the cartesian plane. The Euclidean distance between them is represented as:

Manhattan distance: Manhattan distance between two points in an n-dimensional space is the sum of the distances in each dimension. Consider two points A(x₁, y₁) and B(x₂, y₂) in the cartesian plane. The Manhattan distance between them is represented as:

Minkowski distance: Minkowski distance is a generalized form of Euclidean and Manhattan distances. For two points A(x₁, x₂) and B(y₁, y₂) the Minkowski between them is:

Ans. Bias is the difference between the actual and predicted output. In n-dimensional space, bias can be the distance between the true value points and the predicted value points. A high bias means the model is streamlining the common data points and hence such a simplified approach leads to incorrect prediction and a huge distance gap between true and predicted values. Such bias introduces underfitting.

On the other hand, variance is how varied predictions our model is doing. In a space where data points are widely scattered, the model tries to learn every data point so perfectly that instead of recognizing patterns, it becomes more like a forcefully memorizing them. Memorization will only lead to good performance as long as it is predicting training data’s outputs. As soon as we give new unknown data to predict, it will give an error. Such behaviour occurs when the model tries to fit the noise and fluctuations (variance). It introduces overfitting.

So bias and variance are opposites to each other, and the trade-off is simply computing the balance between bias and variance so that error is minimised. Whenever we have a low bias, variance increases and vice versa. So, we have to model in a way that bias and variance both stay balanced and neither overfitting nor underfitting occurs. Such kind of tuning is called Bias Variance Tradeoff.

Ans. Machine Learning algorithms cannot work on raw text data. They are to be converted to numerical data. Vectorization is a method of converting text data to numerical vectors so that the text can be easily analyzed or consumed by the Machine Learning algorithm.

In NLP, Vectorization is used to map words or phrases from vocabulary to a corresponding vector of real numbers to enable word predictions, word similarities/semantics and grammar checks.

Ans. a. Correlation coefficients range from -1 to 1 and covariance values range from -∞ to +∞.

b. Correlation measures how two variables are related to each other and by how much. Covariance measures whether a variation in one variable results in a variation in another variable.

c. Changing the scale of variables affects covariance i.e. if the variables are changed using similar or different constants, the calculated covariance between the variables changes. But in the case of correlation, this doesn’t happen. Correlation is not affected by a change of scale.

Ans. A naive Bayes classifier presumes that a particular feature of a class is unrelated to any other feature, given the class variable. Simply, Naive Bayes assumes that the variables are independent of each other. This assumption may or may not be correct as it falls under the probabilistic classifiers in statistics. This is the reason it’s called Naive.

Advantages

1. Extremely easy to implement and follows a simple probabilistic approach (Bayes’ Theorem).

2. Works great in small datasets, and can classify stuff even when no data is there!

3. Massively used for spam-filtering and specially text-classification tasks ( as it works better when attributes/features in data are independent of each other).

Disadvantages

1. Works poorly with numerical data, as numerical data is mostly normally distributed.

2. The ‘naive’ approach to considering attributes independent of each other, doesn’t often work well in complex problems.

Ans. A decision tree is a tree-like structure that contains leaves, nodes and edges. It has a root node, parent nodes and child nodes. Each parent node is split into one or more child nodes. The terminal nodes or leaf nodes represent the output variables.

In the decision tree, ID3 stands for Iterative Dichotomiser 3. Dichotomiser means dividing. The algorithm is used for feature selection in a decision tree by the top-down greedy approach (builds the tree from the top and selects the best feature at each step iteratively using Entropy/Information Gain). Each parent node is divided into one or more child nodes iteratively till no more division is possible.

Ans. Entropy: It is the measure of impurity of a variable. It describes how well a node is split. The higher the entropy of a node, the higher will be its impurity. For the best split, it is always suggested to choose nodes with low entropy. Entropy is calculated by –

Gini Impurity: Like Entropy, Gini impurity also measures the impurity of a variable or how good a node is split. In both, Entropy and Gini impurity, a node that has multiple classes is considered impure and a node having a single class is pure.

If we compare both the methods then Gini Impurity is more efficient than entropy in terms of computing power. Training time when using the Entropy criterion is much higher because Entropy is calculated using logarithms. Hence, Gini impurity must be preferred over Entropy.

Ans. Leaf nodes in a Decision tree are the end or terminal or child nodes. These nodes cannot be split any further. Hence, they are the output classes or categories like YES or NO; GOOD, BAD and SATISFACTORY etc.

Ans. a. Decision tree has a broad chance of getting overfitted. This generally happens when the dataset is huge, has many features or when we try to fit more data into it. An overfitted model will give wrong predictions for unseen data.

b. It tends to become more complex when we add more data to it. Upon adding new data points, the nodes get calculated again and the tree is restructured. This reduces the efficiency of the model. Hence, for time-series analysis, the Decision tree is not used. Random forest is used instead.

c. Decision tree is vulnerable to noisy datasets. It tries to fit all the data including the noisy data, and later tends to give wrong predictions.

Ans. Information gain tells us how good a split between a parent node and its children nodes is. The higher the Information gain, the better is the split. High entropy indicates that data is uniform and low entropy means data is more scattered or distributed. Hence, with decreasing entropy, Information gain increases. Information gain is mathematically represented as:

Here, E_parent is the entropy of the parent node, and E_children is the entropy of the child nodes.