Explore topic-wise InterviewSolutions in .

Tuning strategies are used to find the right set of hyperparameters. Hyperparameters are those properties that are fixed and model-specific before the model is TESTED or trained on the dataset. Both the grid search and random search tuning strategies are optimization techniques to find efficient hyperparameters.

• Grid Search:
  • Here, every combination of a preset list of hyperparameters is tried out and evaluated.
  • The search pattern is similar to searching in a grid where the values are in a matrix and a search is performed. Each parameter set is tried out and their accuracy is tracked. after every combination is tried out, the model with the highest accuracy is chosen as the best one.
  • The main drawback here is that, if the number of hyperparameters is increased, the technique suffers. The number of evaluations can increase exponentially with each increase in the hyperparameter. This is called the problem of dimensionality in a grid search.

• Random Search:
  • In this technique, random combinations of hyperparameters set are tried and evaluated for finding the best solution. For optimizing the search, the function is tested at random configurations in parameter space as shown in the image below.
  • In this method, there are increased chances of finding optimal parameters because the pattern followed is random. There are chances that the model is trained on optimized parameters without the need for aliasing.
  • This search works the best when there is a lower number of dimensions as it takes less time to find the right set.

Data Science is a very vast field and comprises many topics like Data MINING, Data Analysis, Data Visualization, Machine Learning, Deep Learning, and most importantly it is laid on the foundation of MATHEMATICAL concepts like Linear Algebra and Statistical analysis. Since there are a LOT of pre-requisites for becoming a good professional Data Scientist, the perks and benefits are very big. Data Scientist has become the most sought job role these days. 

Looking for a comprehensive course on Data Science: Check out our Offerings.

Useful Resources:

• Best Data Science Courses
• Data Scientist Salary
• Data Science Resume
• Data Analyst: Career Guide
• Tableau Interview
• Additional Technical Interview Questions

The PROCESS of DIMENSIONALITY reduction CONSTITUTES reducing the number of features in a dataset to avoid overfitting and reduce the variance. There are mostly 4 advantages of this process:

• This reduces the storage space and time for model execution.
• Removes the ISSUE of multi-collinearity thereby improving the parameter interpretation of the ML model.
• Makes it easier for visualizing DATA when the dimensions are reduced.
• Avoids the curse of increased dimensionality.

To identify this, we perform a hypothesis test as below:
According to the null hypothesis, the coin is unbiased if the probability of HEAD flipping is 50%. According to the alternative hypothesis, the coin is biased and the probability is not equal to 500. Perform the below steps:

• Flip coin 500 times
• Calculate p-value.
• Compare the p-value against the alpha -> result of two-tailed test (0.05/2 = 0.025). FOLLOWING two cases MIGHT occur:
  • p-value > alpha: Then null hypothesis holds good and the coin is unbiased.
  • p-value < alpha: Then the null hypothesis is REJECTED and the coin is biased.

The method of regularization ENTAILS the addition of PENALTIES to different parameters in the machine learning model for reducing the freedom of the model to avoid the issue of overfitting.
There are various regularization methods available such as linear model regularization, Lasso/L1 regularization, etc. The linear model regularization applies penalty over COEFFICIENTS that multiplies the predictors. The Lasso/L1 regularization has the feature of shrinking some coefficients to zero, thereby making it ELIGIBLE to be removed from the model.

LINEAR regression is done under the following assumptions:

• The sample data used for MODELING represents the entire population.
• There EXISTS a linear relationship between the X-axis variable and the mean of the Y variable.
• The residual variance is the same for any X values. This is called homoscedasticity
• The observations are independent of one another.
• Y is DISTRIBUTED normally for any value of X.

Extreme VIOLATIONS of the above assumptions lead to redundant results. Smaller violations of these result in greater variance or bias of the estimates.

If the features number is greater than observations then doing DIMENSIONALITY REDUCTION IMPROVES the SVM (Support VECTOR MODEL).

In Banking fields: Lending loans are the main sources of income to the BANKS. But if the repayment rate isn’t good, then there is a risk of huge LOSSES INSTEAD of any profits. So GIVING out loans to customers is a gamble as banks can’t risk losing good customers but at the same time, they can’t AFFORD to acquire bad customers. This case is a classic example of equal importance in false positive and false negative scenarios.

Before citing instances, LET us understand what are false positives and false negatives.

• False Positives are those cases that were wrongly identified as an event even if they were not. They are called Type I errors.
• False Negatives are those cases that were wrongly identified as non-events despite being an event. They are called Type II errors.

Some examples where false positives were important than false negatives are:

• In the medical field: Consider that a lab report has predicted cancer to a patient even if he did not have cancer. This is an example of a false positive error. It is dangerous to start chemotherapy for that patient as he doesn’t have cancer as starting chemotherapy would lead to damage of healthy cells and might even actually lead to cancer.
• In the e-commerce field: Suppose a company DECIDES to start a campaign where they GIVE $100 gift vouchers for purchasing $10000 worth of items without any minimum purchase conditions. They ASSUME it would result in at least 20% profit for items sold above $10000. What if the vouchers are given to the customers who haven’t purchased anything but have been mistakenly marked as those who purchased $10000 worth of products. This is the case of false-positive error.

We know that there are TWO types of coins - fair and double-headed. HENCE, there are two possible ways of choosing a coin. The first is to choose a fair coin and the second is to choose a coin having 2 heads.

P(selecting fair coin) = 999/1000 = 0.999
P(selecting double headed coin) = 1/1000 = 0.001

Using Bayes RULE,

So, the answer is 0.7531 or 75.3%.

We KNOW that,PROBABILITY of finding ATLEAST 1 SHOOTING star in 15 min = P(sighting in 15min) = 30% = 0.3Hence, Probability of not sighting any shooting star in 15 min = 1-P(sighting in 15min) = 1-0.3 = 0.7 Probability of not finding shooting star in 1 hour = 0.7^4 = 0.1372Probability of finding atleast 1 shooting star in 1 hour = 1-0.1372 = 0.8628

So the probability is 0.8628 = 86.28%

Random forest is BETTER than multiple DECISION trees as random forests are much more robust, accurate, and lesser PRONE to OVERFITTING as it is an ensemble METHOD that ensures multiple weak decision trees learn strongly.

There are different techniques to correct/balance imbalanced data. It can be done by increasing the sample numbers for MINORITY CLASSES. The number of samples can be decreased for those classes with extremely high data points. Following are some approaches followed to balance data:

• Use the right evaluation metrics: In cases of imbalanced data, it is very important to use the right evaluation metrics that provide valuable information. 
  
  • Specificity/Precision: Indicates the number of selected instances that are relevant.
  • Sensitivity: Indicates the number of relevant instances that are selected.
  • F1 score: It represents the harmonic mean of precision and sensitivity.
  • MCC (Matthews correlation coefficient): It represents the correlation coefficient between observed and predicted binary classifications.
  • AUC (Area Under the Curve): This represents a relation between the true positive rates and false-positive rates.

For EXAMPLE, consider the below graph that illustrates training data:

Here, if we measure the accuracy of the model in terms of getting "0"s, then the accuracy of the model would be very high -> 99.9%, but the model does not guarantee any valuable information. In such cases, we can apply different evaluation metrics as stated above.

• Training Set Resampling: It is also possible to balance data by working on getting different datasets and this can be achieved by resampling. There are two approaches followed under-sampling that is used based on the use case and the requirements:
  • Under-sampling This balances the data by reducing the size of the abundant class and is used when the data quantity is sufficient. By performing this, a new dataset that is balanced can be retrieved and this can be used for further modeling.
  • Over-sampling This is used when data quantity is not sufficient. This method balances the dataset by trying to increase the samples size. Instead of getting rid of extra samples, new samples are generated and introduced by employing the methods of repetition, bootstrapping, etc.
• Perform K-fold cross-validation correctly: Cross-Validation needs to be applied properly while USING over-sampling. The cross-validation should be done before over-sampling because if it is done later, then it would be LIKE overfitting the model to get a specific result. To avoid this, resampling of data is done repeatedly with different ratios. 

Box plots and histograms are both visualizations used for showing data DISTRIBUTIONS for efficient COMMUNICATION of information.
Histograms are the bar CHART representation of information that represents the frequency of numerical variable values that are useful in estimating probability distribution, variations and outliers.
Boxplots are used for communicating different aspects of data distribution where the shape of the distribution is not seen but still the insights can be gathered. These are useful for comparing multiple CHARTS at the same time as they take less space when COMPARED to histograms.

Kernel functions are GENERALIZED dot PRODUCT functions USED for the COMPUTING dot product of VECTORS xx and yy in high dimensional feature space. Kernal trick method is used for solving a non-linear problem by using a linear classifier by transforming linearly inseparable data into separable ones in higher dimensions.

The test set is USED to test or EVALUATE the performance of the TRAINED model. It evaluates the PREDICTIVE power of the model.
The VALIDATION set is part of the training set that is used to select parameters for avoiding model overfitting.

Statistical analyses are classified based on the number of variables processed at a GIVEN time.

ROC (Receiver Operating Characteristic) curve is a graphical REPRESENTATION of the contrast between false-positive rates and true positive rates at DIFFERENT thresholds. The curve is used as a PROXY for a trade-off between sensitivity and specificity.

The ROC curve is created by plotting values of true positive rates (TPR or sensitivity) against false-positive rates (FPR or (1-specificity)) TPR represents the PROPORTION of observations correctly predicted as positive out of overall positive observations. The FPR represents the proportion of observations incorrectly predicted out of overall negative observations. Consider the example of medical testing, the TPR represents the rate at which people are correctly TESTED positive for a particular disease.

The impact of missing values can be known after IDENTIFYING what kind of VARIABLES have the missing values.

• If the data ANALYST finds any pattern in these missing values, then there are chances of finding meaningful insights.
• In case of patterns are not FOUND, then these missing values can either be IGNORED or can be replaced with default values such as mean, minimum, maximum, or median values.
• If the missing values belong to categorical variables, then they are assigned with default values. If the data has a normal distribution, then mean values are assigned to missing values.
• If 80% values are missing, then it depends on the analyst to either replace them with default values or drop the variables.

Yes! A CATEGORICAL variable is a variable that can be assigned to two or more CATEGORIES with no definite category ORDERING. ORDINAL variables are similar to categorical variables with proper and clear ordering defines. So, if the variable is ordinal, then treating the categorical VALUE as a continuous variable will result in better predictive models.

To IDENTIFY the extent of missing VALUES, we first have to identify the variables with the missing values. Let us say a pattern is identified. The analyst should now concentrate on them as it could lead to interesting and meaningful insights. HOWEVER, if there are no patterns identified, we can SUBSTITUTE the missing values with the median or mean values or we can simply ignore the missing values. 

If the variable is categorical, the default value to the mean, MINIMUM, and maximum is assigned. The missing value is assigned to the default value. If we have a distribution of data coming, for normal distribution, we give the mean value.

If 80% of the values are missing for a particular variable, then we would drop the variable instead of treating the missing values.

While using a dataset in data science or machine learning algorithms, it so happens that not all the variables are necessary and useful to build a model. Smarter feature selection methods are required to avoid REDUNDANT models to increase the efficiency of our model. Following are the three main methods in feature selection:

• Filter Methods:
  • These methods pick up only the intrinsic properties of features that are measured via univariate statistics and not cross-validated performance. They are straightforward and are generally faster and require less computational resources when compared to wrapper methods.
  • There are VARIOUS filter methods such as the Chi-Square test, Fisher’s Score method, Correlation Coefficient, Variance Threshold, Mean Absolute Difference (MAD) method, Dispersion Ratios, etc.

• Wrapper Methods:
  • These methods need some sort of method to search greedily on all possible feature subsets, access their quality by learning and evaluating a classifier with the feature.
  • The selection technique is built upon the machine learning algorithm on which the given dataset needs to FIT.
  • There are three types of wrapper methods, they are:
    • Forward Selection: Here, one feature is tested at a time and new features are added until a good fit is obtained.
    • Backward Selection: Here, all the features are tested and the non-fitting ones are eliminated one by one to see while checking which works better.
    • Recursive Feature Elimination: The features are recursively checked and evaluated how well they perform.
  • These methods are generally computationally INTENSIVE and require high-end resources for analysis. But these methods usually lead to better predictive models having higher accuracy than filter methods.

• Embedded Methods:
  • Embedded methods constitute the advantages of both filter and wrapper methods by including feature interactions while maintaining reasonable computational costs.
  • These methods are ITERATIVE as they take each model iteration and carefully extract features contributing to most of the training in that iteration.
  • Examples of embedded methods: LASSO Regularization (L1), Random Forest Importance.

While running an algorithm on any data, to gather proper insights, it is very much necessary to have correct and clean data that contains only relevant information. Dirty data most often results in poor or incorrect insights and predictions which can have damaging effects.

For example, while launching any big campaign to market a product, if our data analysis tells us to target a product that in reality has no demand and if the campaign is launched, it is bound to fail. This results in a loss of the company’s revenue. This is where the importance of having proper and clean data comes into the picture.

• Data Cleaning of the data coming from different SOURCES helps in data transformation and results in the data where the data scientists can work on.
• Properly cleaned data increases the accuracy of the MODEL and provides very good predictions.
• If the dataset is very large, then it becomes cumbersome to run data on it. The data cleanup step takes a lot of TIME (around 80% of the time) if the data is HUGE. It cannot be incorporated with running the model. Hence, cleaning data before running the model, results in increased speed and efficiency of the model.
• Data cleaning helps to identify and fix any structural issues in the data. It also helps in removing any duplicates and helps to maintain the consistency of the data.

The following diagram represents the advantages of data cleaning:

Selection Bias happens in cases where there is no RANDOMIZATION specifically achieved while PICKING a part of the dataset for analysis. This bias tells that the sample analyzed does not represent the WHOLE POPULATION meant to be analyzed.

• For example, in the below image, we can see that the sample that we SELECTED does not entirely represent the whole population that we have. This helps us to question whether we have selected the right data for analysis or not.

We do not want to update and make changes to an algorithm on a regular basis as an algorithm is a well-defined step procedure to solve any problem and if the steps keep on updating, it cannot be said well defined anymore. ALSO, this brings in a LOT of problems to the systems ALREADY implementing the algorithm as it BECOMES difficult to bring in continuous and regular changes. So, we should update an algorithm only in any of the following cases:

• If you want the MODEL to evolve as data streams through infrastructure, it is fair to make changes to an algorithm and update it accordingly.
• If the underlying data source is changing, it almost becomes necessary to update the algorithm accordingly.
• If there is a case of non-stationarity, we may update the algorithm.
• One of the most important reasons for updating any algorithm is its underperformance and lack of efficiency. So, if an algorithm lacks efficiency or underperforms it should be either replaced by some better algorithm or it must be updated.

Generally, we follow the below steps:

• The first STEP is to thoroughly understand the business requirement/problem
• Next, explore the given data and analyze it carefully. If you FIND any data missing, get the requirements clarified from the business.
• Data CLEANUP and preparation step is to be performed next which is then used for modelling. Here, the missing values are found and the variables are transformed.
• Run your model against the data, build meaningful visualization and analyze the results to get meaningful insights.
• Release the model implementation, and track the results and PERFORMANCE over a specified period to analyze the usefulness.
• Perform cross-validation of the model.

Check out the list of data ANALYTICS projects.

Although these two terms are used for establishing a relationship and dependency between any two random variables, the following are the differences between them:

• <STRONG>Correlation: This TECHNIQUE is used to measure and estimate the quantitative relationship between two variables and is measured in terms of how strong are the variables related.
• Covariance: It represents the extent to which the variables change together in a cycle. This explains the systematic relationship between pair of variables where CHANGES in one affect changes in another variable.

Mathematically, consider 2 random variables, X and Y where the means are represented as μX{"detectHand":false} and μY{"detectHand":false} RESPECTIVELY and standard deviations are represented by σX{"detectHand":false} and σY{"detectHand":false} respectively and E represents the EXPECTED value operator, then:

• covarianceXY = E[(X-μX{"detectHand":false}),(Y-μY{"detectHand":false})]
• correlationXY = E[(X-μX{"detectHand":false}),(Y-μY{"detectHand":false})]/(σX{"detectHand":false}σY{"detectHand":false})
  so that

Based on the above formula, we can deduce that the correlation is dimensionless whereas covariance is represented in units that are obtained from the multiplication of units of two variables.

The following image graphically shows the difference between correlation and covariance:

Cross-Validation is a Statistical technique used for improving a model’s PERFORMANCE. Here, the model will be TRAINED and tested with rotation using different samples of the training dataset to ensure that the model performs WELL for unknown data. The training data will be SPLIT into various groups and the model is run and VALIDATED against these groups in rotation.

The most commonly used techniques are:

• K- Fold method
• Leave p-out method
• Leave-one-out method
• Holdout method

Depending on the size of the dataset, we FOLLOW the below WAYS:

• In case the datasets are small, the MISSING values are substituted with the mean or average of the remaining data. In pandas, this can be done by USING mean = df.mean() where df represents the pandas dataframe representing the dataset and mean() calculates the mean of the data. To substitute the missing values with the calculated mean, we can use df.fillna(mean).
• For LARGER datasets, the rows with missing values can be removed and the remaining data can be used for data prediction.

Tensorflow is a very famous library in DEEP learning. The reason is PRETTY simple actually. It provides C++ as well as Python APIs which makes it very easier to work on. ALSO, TensorFlow has a fast COMPILATION speed as COMPARED to Keras and Torch (other famous deep learning libraries). Apart from that, Tenserflow supports both GPU and CPU computing devices. Hence, it is a major success and a very popular library for deep learning.

P-value is a number that ranges from 0 to 1. In a hypothesis test in statistics, the p-value helps in telling us how strong the results are. The claim that is kept for EXPERIMENT or trial is called NULL Hypothesis.

• A low p-value i.e. p-value less than or EQUAL to 0.05 INDICATES the strength of the results against the Null Hypothesis which in turn means that the Null Hypothesis can be rejected. 
• A high p-value i.e. p-value greater than 0.05 indicates the strength of the results in FAVOUR of the Null Hypothesis i.e. for the Null Hypothesis which in turn means that the Null Hypothesis can be accepted.

• EXPLODING GRADIENTS: Let us say that you are training an RNN. Say, you saw exponentially growing error gradients that accumulate, and as a result of this, very large updates are made to the neural network model weights. These exponentially growing error gradients that update the neural network weights to a great extent are called Exploding Gradients.
• Vanishing Gradients: Let us say again, that you are training an RNN. Say, the slope became too small. This problem of the slope becoming too small is called Vanishing Gradient. It causes a major increase in the training time and causes POOR performance and extremely low accuracy.

Auto-encoders are learning NETWORKS. They TRANSFORM inputs into outputs with MINIMUM possible ERRORS. So, basically, this means that the output that we want should be almost equal to or as close as to input as follows. 

Multiple layers are added between the input and the output layer and the layers that are in between the input and the output layer are smaller than the input layer. It received unlabelled input. This input is encoded to reconstruct the input LATER.

A computational GRAPH is also known as a “Dataflow Graph”. Everything in the famous deep learning LIBRARY TensorFlow is based on the computational graph. The computational graph in Tensorflow has a network of nodes where each node operates. The nodes of this graph represent OPERATIONS and the edges represent TENSORS.

This APPROACH can be understood with the FAMOUS example of the wine seller. Let us say that there is a wine seller who has his own shop. This wine seller purchases wine from the dealers who sell him the wine at a low cost so that he can sell the wine at a high cost to the customers. Now, let us say that the dealers whom he is purchasing the wine from, are SELLING him fake wine. They do this as the fake wine costs way less than the original wine and the fake and the real wine are indistinguishable to a normal CONSUMER (customer in this case). The shop owner has some friends who are wine experts and he sends his wine to them every time before keeping the stock for sale in his shop. So, his friends, the wine experts, give him feedback that the wine is probably fake. Since the wine seller has been purchasing the wine for a long time from the same dealers, he wants to make sure that their feedback is right before he complains to the dealers about it. Now, let us say that the dealers also have got a tip from somewhere that the wine seller is suspicious of them.

So, in this situation, the dealers will try their best to sell the fake wine WHEREAS the wine seller will try his best to identify the fake wine. Let us see this with the help of a diagram shown below:

From the image above, it is clear that a noise vector is entering the generator (dealer) and he generates the fake wine and the discriminator has to distinguish between the fake wine and real wine. This is a Generative Adversarial Network (GAN).

In a GAN, there are 2 main components viz. Generator and Discrminator. So, the generator is a CNN that keeps producing images and the discriminator tries to identify the real images from the fake ones. 

In the human brain, DIFFERENT neurons are present. These neurons combine and perform various tasks. The Neural Network in deep learning tries to imitate human brain neurons. The neural network learns the patterns from the data and uses the knowledge that it gains from various patterns to predict the OUTPUT for new data, without any human assistance.

A perceptron is the simplest neural network that CONTAINS a SINGLE neuron that performs 2 functions. The first function is to perform the weighted sum of all the inputs and the second is an activation function.

There are some other neural networks that are more complicated. Such networks consist of the following three layers:

• Input Layer: The neural network has the input layer to receive the input.
• Hidden Layer: There can be multiple hidden layers between the input layer and the output layer. The initially hidden layers are used for detecting the low-level patterns whereas the further layers are responsible for combining output from previous layers to find more patterns.
• Output Layer: This layer outputs the prediction.

An EXAMPLE neural network image is shown below:

What will you do? Have you experienced such an issue before?

In such types of questions, we first need to ask what ML model we have to train. After that, it depends on whether we have to train a model based on Neural Networks or SVM.

The steps for Neural Networks are given below:

• The Numpy array can be USED to load the ENTIRE data. It will never store the entire data, rather just create a mapping of the data.
• Now, in order to get some desired data, pass the index into the NumPy Array.
• This data can be used to pass as an INPUT to the neural network maintaining a small batch size.

The steps for SVM are given below:

• For SVM, small data SETS can be obtained. This can be done by dividing the big data set.
• The SUBSET of the data set can be obtained as an input if using the partial fit function.
• Repeat the step of using the partial fit method for other subsets as well.

Now, you may describe the situation if you have faced such an issue in your projects or working in machine learning/ data science.

In the above DIAGRAM, we can see that the thin lines MARK the distance from the classifier to the closest data points (darkened data points). These are called support vectors. So, we can DEFINE the support vectors as the data points or vectors that are nearest (closest) to the hyperplane. They AFFECT the position of the hyperplane. Since they support the hyperplane, they are KNOWN as support vectors.

RMSE: RMSE stands for Root Mean Square Error. In a linear REGRESSION MODEL, RMSE is used to test the PERFORMANCE of the machine learning model. It is used to evaluate the data spread around the line of best fit. So, in simple words, it is used to measure the deviation of the residuals.

RMSE is CALCULATED using the formula:

• Yi is the actual value of the output variable.
• Y(Cap) is the predicted value and,
• N is the number of data points.

MSE: Mean Squared Error is used to find how close is the line to the actual data. So, we make the difference in the distance of the data points from the line and the difference is squared. This is done for all the data points and the submission of the squared difference divided by the total number of data points gives us the Mean Squared Error (MSE).

So, if we are taking the squared difference of N data points and dividing the sum by N, what does it mean? Yes, it represents the average of the squared difference of a data point from the line i.e. the average of the squared difference between the actual and the predicted values. The formula for finding MSE is given below:

• Yi is the actual value of the output variable (the ith data point)
• Y(cap) is the predicted value and,
• N is the total number of data points.

So, RMSE is the square root of MSE.

• Time series data can be thought of as an extension to linear regression which uses terms LIKE autocorrelation, movement of averages for summarizing historical data of y-axis variables for predicting a better future.
• Forecasting and prediction is the MAIN goal of time series problems where accurate predictions can be MADE but sometimes the UNDERLYING reasons might not be known.
• Having Time in the problem does not necessarily mean it becomes a time series problem. There should be a relationship between target and time for a problem to become a time series problem.
• The observations close to one another in time are EXPECTED to be similar to the ones far away which provide accountability for seasonality. For instance, today’s weather would be similar to tomorrow’s weather but not similar to weather from 4 months from today. Hence, weather prediction based on past data becomes a time series problem.