ML Interview Questions

What are some differences between a linked list and an array?

An array is an ordered collection of objects. A linked list is a series of objects with pointers that direct how to process them sequentially. An array assumes that every element has the same size, unlike the linked list. A linked list can more easily grow organically: an array has to be pre-defined or re-defined for organic growth. Shuffling a linked list involves changing which points direct where — meanwhile, shuffling an array is more complex and takes more memory.

How would you handle an imbalanced dataset?

An imbalanced dataset is when you have, for example, a classification test and 90% of the data is in one class. That leads to problems: an accuracy of 90% can be skewed if you have no predictive power on the other category of data! Here are a few tactics to get over the hump: 1- Collect more data to even the imbalances in the dataset. 2- Resample the dataset to correct for imbalances. 3- Try a different algorithm altogether on your dataset. What's important here is that you have a keen sense for what damage an unbalanced dataset can cause, and how to balance that.

What is Bayes' Theorem? How is it useful in a machine learning context?

Bayes' Theorem gives you the posterior probability of an event given what is known as prior knowledge. Mathematically, it's expressed as the true positive rate of a condition sample divided by the sum of the false positive rate of the population and the true positive rate of a condition. Say you had a 60% chance of actually having the flu after a flu test, but out of people who had the flu, the test will be false 50% of the time, and the overall population only has a 5% chance of having the flu. Would you actually have a 60% chance of having the flu after having a positive test? Bayes' Theorem says no. It says that you have a (.6 * 0.05) (True Positive Rate of a Condition Sample) / (.6*0.05)(True Positive Rate of a Condition Sample) + (.5*0.95) (False Positive Rate of a Population) = 0.0594 or 5.94% chance of getting a flu. Bayes' Theorem is the basis behind a branch of machine learning that most notably includes the Naive Bayes classifier. That's something important to consider when you're faced with machine learning interview questions.

What's the trade-off between bias and variance?

Bias is error due to erroneous or overly simplistic assumptions in the learning algorithm you're using. This can lead to the model underfitting your data, making it hard for it to have high predictive accuracy and for you to generalize your knowledge from the training set to the test set. Variance is error due to too much complexity in the learning algorithm you're using. This leads to the algorithm being highly sensitive to high degrees of variation in your training data, which can lead your model to overfit the data. You'll be carrying too much noise from your training data for your model to be very useful for your test data. The bias-variance decomposition essentially decomposes the learning error from any algorithm by adding the bias, the variance and a bit of irreducible error due to noise in the underlying dataset. Essentially, if you make the model more complex and add more variables, you'll lose bias but gain some variance — in order to get the optimally reduced amount of error, you'll have to tradeoff bias and variance. You don't want either high bias or high variance in your model.

When should you use classification over regression?

Classification produces discrete values and dataset to strict categories, while regression gives you continuous results that allow you to better distinguish differences between individual points. You would use classification over regression if you wanted your results to reflect the belongingness of data points in your dataset to certain explicit categories (ex: If you wanted to know whether a name was male or female rather than just how correlated they were with male and female names.)

What is deep learning, and how does it contrast with other machine learning algorithms?

Deep learning is a subset of machine learning that is concerned with neural networks: how to use backpropagation and certain principles from neuroscience to more accurately model large sets of unlabelled or semi-structured data. In that sense, deep learning represents an unsupervised learning algorithm that learns representations of data through the use of neural nets.

Why is "Naive" Bayes naive?

Despite its practical applications, especially in text mining, Naive Bayes is considered "Naive" because it makes an assumption that is virtually impossible to see in real-life data: the conditional probability is calculated as the pure product of the individual probabilities of components. This implies the absolute independence of features — a condition probably never met in real life. As a Quora commenter put it whimsically, a Naive Bayes classifier that figured out that you liked pickles and ice cream would probably naively recommend you a pickle ice cream.

We know that one hot encoding increasing the dimensionality of a data set. But, label encoding doesn't. How ?

Don't get baffled at this question. It's a simple question asking the difference between the two. Using one hot encoding, the dimensionality (a.k.a features) in a data set get increased because it creates a new variable for each level present in categorical variables. For example: let's say we have a variable 'color'. The variable has 3 levels namely Red, Blue and Green. One hot encoding 'color' variable will generate three new variables as Color.Red, Color.Blue and Color.Green containing 0 and 1 value. In label encoding, the levels of a categorical variables gets encoded as 0 and 1, so no new variable is created. Label encoding is majorly used for binary variables.

You are working on a classification problem. For validation purposes, you've randomly sampled the training data set into train and validation. You are confident that your model will work incredibly well on unseen data since your validation accuracy is high. However, you get shocked after getting poor test accuracy. What went wrong?

In case of classification problem, we should always use stratified sampling instead of random sampling. A random sampling doesn't takes into consideration the proportion of target classes. On the contrary, stratified sampling helps to maintain the distribution of target variable in the resultant distributed samples also.

What cross-validation technique would you use on a time series dataset?

Instead of using standard k-folds cross-validation, you have to pay attention to the fact that a time series is not randomly distributed data — it is inherently ordered by chronological order. If a pattern emerges in later time periods for example, your model may still pick up on it even if that effect doesn't hold in earlier years! You'll want to do something like forward chaining where you'll be able to model on past data then look at forward-facing data. fold 1 : training [1], test [2] fold 2 : training [1 2], test [3] fold 3 : training [1 2 3], test [4] fold 4 : training [1 2 3 4], test [5] fold 5 : training [1 2 3 4 5], test [6]

How is KNN different from k-means clustering?

K-Nearest Neighbors is a supervised classification algorithm, while k-means clustering is an unsupervised clustering algorithm. While the mechanisms may seem similar at first, what this really means is that in order for K-Nearest Neighbors to work, you need labeled data you want to classify an unlabeled point into (thus the nearest neighbor part). K-means clustering requires only a set of unlabeled points and a threshold: the algorithm will take unlabeled points and gradually learn how to cluster them into groups by computing the mean of the distance between different points. The critical difference here is that KNN needs labeled points and is thus supervised learning, while k-means doesn't — and is thus unsupervised learning.

What are the last machine learning papers you've read?

Keeping up with the latest scientific literature on machine learning is a must if you want to demonstrate interest in a machine learning position. This overview of deep learning in Nature by the scions of deep learning themselves (from Hinton to Bengio to LeCun) can be a good reference paper and an overview of what's happening in deep learning — and the kind of paper you might want to cite.

How do you think Google is training data for self-driving cars?

Machine learning interview questions like this one really test your knowledge of different machine learning methods, and your inventiveness if you don't know the answer. Google is currently using recaptcha to source labeled data on storefronts and traffic signs. They are also building on training data collected by Sebastian Thrun at GoogleX — some of which was obtained by his grad students driving buggies on desert dunes!

What cross validation technique would you use on time series data set? Is it k-fold or LOOCV?

Neither. In time series problem, k fold can be troublesome because there might be some pattern in year 4 or 5 which is not in year 3. Resampling the data set will separate these trends, and we might end up validation on past years, which is incorrect. Instead, we can use forward chaining strategy with 5 fold as shown below: -fold 1 : training [1], test [2] -fold 2 : training [1 2], test [3] -fold 3 : training [1 2 3], test [4] -fold 4 : training [1 2 3 4], test [5] -fold 5 : training [1 2 3 4 5], test [6] where 1,2,3,4,5,6 represents "year".

What's the F1 score? How would you use it?

The F1 score is a measure of a model's performance. It is a weighted average of the precision and recall of a model, with results tending to 1 being the best, and those tending to 0 being the worst. You would use it in classification tests where true negatives don't matter much.

What's the "kernel trick" and how is it useful?

The Kernel trick involves kernel functions that can enable in higher-dimension spaces without explicitly calculating the coordinates of points within that dimension: instead, kernel functions compute the inner products between the images of all pairs of data in a feature space. This allows them the very useful attribute of calculating the coordinates of higher dimensions while being computationally cheaper than the explicit calculation of said coordinates. Many algorithms can be expressed in terms of inner products. Using the kernel trick enables us effectively run algorithms in a high-dimensional space with lower-dimensional data.

Q38- How would you approach the "Netflix Prize" competition?

The Netflix Prize was a famed competition where Netflix offered $1,000,000 for a better collaborative filtering algorithm. The team that won called BellKor had a 10% improvement and used an ensemble of different methods to win. Some familiarity with the case and its solution will help demonstrate you've paid attention to machine learning for a while.

You've built a random forest model with 10000 trees. You got delighted after getting training error as 0.00. But, the validation error is 34.23. What is going on? Haven't you trained your model perfectly?

The model has overfitted. Training error 0.00 means the classifier has mimiced the training data patterns to an extent, that they are not available in the unseen data. Hence, when this classifier was run on unseen sample, it couldn't find those patterns and returned prediction with higher error. In random forest, it happens when we use larger number of trees than necessary. Hence, to avoid these situation, we should tune number of trees using cross validation.

How do you ensure you're not overfitting with a model?

This is a simple restatement of a fundamental problem in machine learning: the possibility of overfitting training data and carrying the noise of that data through to the test set, thereby providing inaccurate generalizations. There are three main methods to avoid overfitting: 1- Keep the model simpler: reduce variance by taking into account fewer variables and parameters, thereby removing some of the noise in the training data. 2- Use cross-validation techniques such as k-folds cross-validation. 3- Use regularization techniques such as LASSO that penalize certain model parameters if they're likely to cause overfitting.

You are given a data set. The data set has missing values which spread along 1 standard deviation from the median. What percentage of data would remain unaffected? Why?

This question has enough hints for you to start thinking! Since, the data is spread across median, let's assume it's a normal distribution. We know, in a normal distribution, ~68% of the data lies in 1 standard deviation from mean (or mode, median), which leaves ~32% of the data unaffected. Therefore, ~32% of the data would remain unaffected by missing values.

What's your favorite algorithm, and can you explain it to me in less than a minute?

This type of question tests your understanding of how to communicate complex and technical nuances with poise and the ability to summarize quickly and efficiently. Make sure you have a choice and make sure you can explain different algorithms so simply and effectively that a five-year-old could grasp the basics!

Which data visualization libraries do you use? What are your thoughts on the best data visualization tools?

What's important here is to define your views on how to properly visualize data and your personal preferences when it comes to tools. Popular tools include R's ggplot, Python's seaborn and matplotlib, and tools such as Plot.ly and Tableau.

Is it possible capture the correlation between continuous and categorical variable? If yes, how?

Yes, we can use ANCOVA (analysis of covariance) technique to capture association between continuous and categorical variables.

Considering the long list of machine learning algorithm, given a data set, how do you decide which one to use?

You should say, the choice of machine learning algorithm solely depends of the type of data. If you are given a data set which is exhibits linearity, then linear regression would be the best algorithm to use. If you given to work on images, audios, then neural network would help you to build a robust model. If the data comprises of non linear interactions, then a boosting or bagging algorithm should be the choice. If the business requirement is to build a model which can be deployed, then we'll use regression or a decision tree model (easy to interpret and explain) instead of black box algorithms like SVM, GBM etc. In short, there is no one master algorithm for all situations. We must be scrupulous enough to understand which algorithm to use.

Write a basic Machine Learning program to check the accuracy of a model, by importing any dataset using any classifier?

#importing dataset import sklearn from sklearn import datasets iris = datasets.load_iris() X = iris.data Y = iris.target #splitting the dataset from sklearn.cross_validation import train_test_split X_train, Y_train, X_test, Y_test = train_test_split(X,Y, test_size = 0.5) #Selecting Classifier my_classifier = tree.DecisionTreeClassifier() My_classifier.fit(X_train, Y_train) predictions = my_classifier(X_test) #check accuracy From sklear.metrics import accuracy_score print accuracy_score(y_test, predictions)

What is convex hull ? (Hint: Think SVM)

In case of linearly separable data, convex hull represents the outer boundaries of the two group of data points. Once convex hull is created, we get maximum margin hyperplane (MMH) as a perpendicular bisector between two convex hulls. MMH is the line which attempts to create greatest separation between two groups.

How do you map nicknames (Pete, Andy, Nick, Rob, etc) to real names?

-- This problem can be solved in n number of ways. Let's assume that you're given a data set containing 1000s of twitter interactions. You will begin by studying the relationship between two people by carefully analyzing the words used in the tweets. -- This kind of problem statement can be solved by implementing Text Mining using Natural Language Processing techniques, wherein each word in a sentence is broken down and co-relations between various words are found. -- NLP is actively used in understanding customer feedback, performing sentimental analysis on Twitter and Facebook. Thus, one of the ways to solve this problem is through Text Mining and Natural Language Processing techniques.

What is A/B Testing?

-A/B is Statistical hypothesis testing for randomized experiment with two variables A and B. It is used to compare two models that use different predictor variables in order to check which variable fits best for a given sample of data. -Consider a scenario where you've created two models (using different predictor variables) that can be used to recommend products for an e-commerce platform. -A/B Testing can be used to compare these two models to check which one best recommends products to a customer.

What are collinearity and multicollinearity?

-Collinearity occurs when two predictor variables (e.g., x1 and x2) in a multiple regression have some correlation. - Multicollinearity occurs when more than two predictor variables (e.g., x1, x2, and x3) are inter-correlated.

what do you understand by eigenvecctors and eigenvalues?

-Eigenvectors: Eigenvectors are those vectors whose direction remains unchanged even when a linear transformation is performed on them. -Eigenvalues: Eigenvalue is the scalar that is used for the transformation of an Eigenvector.

Running a binary classification tree algorithm is quite easy. But do you know how the tree decides on which variable to split at the root node and its succeeding child nodes?

-Measures such as, Gini Index and Entropy can be used to decide which variable is best fitted for splitting the Decision Tree at the root node. -We can calculate Gini as following: Calculate Gini for sub-nodes, using the formula - sum of square of probability for success and failure (p^2+q^2). -Calculate Gini for split using weighted Gini score of each node of that split -Entropy is the measure of impurity or randomness in the data, (for binary class): Entropy = -p log2p - q log2 a here p and q is the probability of success and failure respectively in that node -entropy is zero when a node is homogeneous and is maximum when both the classes are present in a node at 50 - 50%. To sum it up, the entropy must be as low as possible in order to decide whether or not a variable is suitable as the root node

Name an example where ensemble techniques might be useful

Ensemble techniques use a combination of learning algorithms to optimize better predictive performance. They typically reduce overfitting in models and make the model more robust (unlikely to be influenced by small changes in the training data). You could list some examples of ensemble methods, from bagging to boosting to a "bucket of models" method and demonstrate how they could increase predictive power.

While working on a data set, how do you select important variables? Explain your methods.

Following are the methods of variable selection you can use: 1. Remove the correlated variables prior to selecting important variables 2. Use linear regression and select variables based on p values 3. Use Forward Selection, Backward Selection, Stepwise Selection 4. Use Random Forest, Xgboost and plot variable importance chart 5. Use Lasso Regression 6. Measure information gain for the available set of features and select top n features accordingly.

Do you suggest that treating a categorical variable as continuous variable would result in a better predictive model?

For better predictions, categorical variable can be considered as a continuous variable only when the variable is ordinal in nature.

How would you predict who will renew their subscription next month? What data would you need to solve this? What analysis would you do? Would you build predictive models? If so, which algorithms?

- Let's assume that we're trying to predict renewal rate for Netflix subscription. So our problem statement is to predict which users will renew their subscription plan for the next month. - Next, we must understand the data that is needed to solve this problem. In this case, we need to check the number of hours the channel is active for each household, the number of adults in the household, number of kids, which channels are streamed the most, how much time is spent on each channel, how much has the watch rate varied from last month, etc. Such data is needed to predict whether or not a person will continue the subscription for the upcoming month. - After collecting this data, it is important that you find patterns and correlations. For example, we know that if a household has kids, then they are more likely to subscribe. Similarly, by studying the watch rate of the previous month, you can predict whether a person is still interested in a subscription. Such trends must be studied. - The next step is analysis. For this kind of problem statement, you must use a classification algorithm that classifies customers into 2 groups: - Customers who are likely to subscribe next month - Customers who are not likely to subscribe next month - Would you build predictive models? Yes, in order to achieve this you must build a predictive model that classifies the customers into 2 classes like mentioned above. - Which algorithms to choose? You can choose classification algorithms such as Logistic Regression, Random Forest, Support Vector Machine, etc. - Once you've opted the right algorithm, you must perform model evaluation to calculate the efficiency of the algorithm. This is followed by deployment.

We have two options for serving ads within Newsfeed: 1 - out of every 25 stories, one will be an ad 2 - every story has a 4% chance of being an ad For each option, what is the expected number of ads shown in 100 news stories? If we go with option 2, what is the chance a user will be shown only a single ad in 100 stories? What about no ads at all?

- The expected number of ads shown in 100 new stories for option 1 is equal to 4 (100/25 = 4). - Similarly, for option 2, the expected number of ads shown in 100 new stories is also equal to 4 (4/100 = 1/25 which suggests that one out of every 25 stories will be an ad, therefore in 100 new stories there will be 4 ads) - Therefore for each option, the total number of ads shown in 100 new stories is 4. - The second part of the question can be solved by using Binomial distribution. Binomial distribution takes three parameters: -- The probability of success and failure, which in our case is 4%. -- The total number of cases, which is 100 in our case. -- The probability of the outcome, which is a chance that a user will be shown only a single ad in 100 stories - p(single ad) = (0.96)^99*(0.04)^1 (note: here 0.96 denotes the chance of not seeing an ad in 100 stories, 99 denotes the possibility of seeing only 1 ad, 0.04 is the probability of seeing an ad once in 100 stories ) - In total, there are 100 positions for the ad. Therefore, 100 * p(single ad) = 7.03%

There's a game where you are asked to roll two fair six-sided dice. If the sum of the values on the dice equals seven, then you win $21. However, you must pay $5 to play each time you roll both dice. Do you play this game? And in the follow-up: If he plays 6 times what is the probability of making money from this game?

- The first condition states that if the sum of the values on the 2 dices is equal to 7, then you win $21. But for all the other cases you must pay $5. - First, let's calculate the number of possible cases. Since we have two 6-sided dices, the total number of cases => 6*6 = 36. - Out of 36 cases, we must calculate the number of cases that produces a sum of 7 (in such a way that the sum of the values on the 2 dices is equal to 7) - Possible combinations that produce a sum of 7 is, (1,6), (2,5), (3,4), (4,3), (5,2) and (6,1). All these 6 combinations generate a sum of 7. - This means that out of 36 chances, only 6 will produce a sum of 7. On taking the ratio, we get: 6/36 = 1/6 - So this suggests that we have a chance of winning $21, once in 6 games. - So to answer the question if a person plays 6 times, he will win one game of $21, whereas for the other 5 games he will have to pay $5 each, which is $25 for all five games. Therefore, he will face a loss because he wins $21 but ends up paying $25.

A jar has 1000 coins, of which 999 are fair and 1 is double headed. Pick a coin at random, and toss it 10 times. Given that you see 10 heads, what is the probability that the next toss of that coin is also a head?

- There are two ways of choosing a coin. One is to pick a fair coin and the other is to pick the one with two heads. - Probability of selecting fair coin = 999/1000 = 0.999 - Probability of selecting unfair coin = 1/1000 = 0.001 - Selecting 10 heads in a row = Selecting fair coin * Getting 10 heads + Selecting an unfair coin - P (A) = 0.999 * (1/2)^10 = 0.999 * (1/1024) = 0.000976 - P (B) = 0.001 * 1 = 0.001 - P( A / A + B ) = 0.000976 / (0.000976 + 0.001) = 0.4939 - P( B / A + B ) = 0.001 / 0.001976 = 0.5061 - Probability of selecting another head = P(A/A+B) * 0.5 + P(B/A+B) * 1 = 0.4939 * 0.5 + 0.5061 = 0.7531

how would you screen for outliers and what shoul dyou do if you find one?

1. Boxplot: A box plot represents the distribution of the data and its variability. The box plot contains the upper and lower quartiles, so the box basically spans the Inter-Quartile Range (IQR). One of the main reasons why box plots are used is to detect outliers in the data. Since the box plot spans the IQR, it detects the data points that lie outside this range. These data points are nothing but outliers. 2. Probabilistic and statistical models: Statistical models such as normal distribution and exponential distribution can be used to detect any variations in the distribution of data points. If any data point is found outside the distribution range, it is rendered as an outlier. 3. Linear models: Linear models such as logistic regression can be trained to flag outliers. In this manner, the model picks up the next outlier it sees. 4. Proximity-based models: An example of this kind of model is the K-means clustering model wherein, data points form multiple or 'k' number of clusters based on features such as similarity or distance. Since similar data points form clusters, the outliers also form their own cluster. In this way, proximity-based models can easily help detect outliers. - If your data set is huge and rich then you can risk dropping the outliers. - However, if your data set is small then you can cap the outliers, by setting a threshold percentile. For example, the data points that are above the 95th percentile can be used to cap the outliers. - Lastly, based on the data exploration stage, you can narrow down some rules and impute the outliers based on those business rules.

What's a Fourier transform?

A Fourier transform is a generic method to decompose generic functions into a superposition of symmetric functions. Or as this more intuitive tutorial puts it, given a smoothie, it's how we find the recipe. The Fourier transform finds the set of cycle speeds, amplitudes and phases to match any time signal. A Fourier transform converts a signal from time to frequency domain — it's a very common way to extract features from audio signals or other time series such as sensor data.

Running a binary classification tree algorithm is the easy part. Do you know how does a tree splitting takes place i.e. how does the tree decide which variable to split at the root node and succeeding nodes?

A classification trees makes decision based on Gini Index and Node Entropy. In simple words, the tree algorithm find the best possible feature which can divide the data set into purest possible children nodes. Gini index says, if we select two items from a population at random then they must be of same class and probability for this is 1 if population is pure. We can calculate Gini as following: 1. Calculate Gini for sub-nodes, using formula sum of square of probability for success and failure (p^2+q^2). 2. Calculate Gini for split using weighted Gini score of each node of that split Entropy is the measure of impurity as given by (for binary class): Entropy = -p log2p -q log2q Here p and q is probability of success and failure respectively in that node. Entropy is zero when a node is homogeneous. It is maximum when a both the classes are present in a node at 50% - 50%. Lower entropy is desirable

What's the difference between a generative and discriminative model?

A generative model will learn categories of data while a discriminative model will simply learn the distinction between different categories of data. Discriminative models will generally outperform generative models on classification tasks.

Describe a hash table.

A hash table is a data structure that produces an associative array. A key is mapped to certain values through the use of a hash function. They are often used for tasks such as database indexing.

How would you evaluate a logistic regression model?

A logistic regression is said to provide a better fit to the data if it demonstrates an improvement over a model with fewer predictors. This is performed using the likelihood ratio test, which compares the likelihood of the data under the full model against the likelihood of the data under a model with fewer predictors.

How would you implement a recommendation system for our company's users?

A lot of machine learning interview questions of this type will involve implementation of machine learning models to a company's problems. You'll have to research the company and its industry in-depth, especially the revenue drivers the company has, and the types of users the company takes on in the context of the industry it's in. https://stackoverflow.com/questions/6302184/how-to-implement-a-recommendation-system#6302223

Rise in global average temperature led to decrease in number of pirates around the world. Does that mean that decrease in number of pirates caused the climate change?

After reading this question, you should have understood that this is a classic case of "causation and correlation". No, we can't conclude that decrease in number of pirates caused the climate change because there might be other factors (lurking or confounding variables) influencing this phenomenon. Therefore, there might be a correlation between global average temperature and number of pirates, but based on this information we can't say that pirated died because of rise in global average temperature.

How would you simulate the approach AlphaGo took to beat Lee Sidol at Go?

AlphaGo beating Lee Sidol, the best human player at Go, in a best-of-five series was a truly seminal event in the history of machine learning and deep learning. The Nature paper above describes how this was accomplished with "Monte-Carlo tree search with deep neural networks that have been trained by supervised learning, from human expert games, and by reinforcement learning from games of self-play

After spending several hours, you are now anxious to build a high accuracy model. As a result, you build 5 GBM models, thinking a boosting algorithm would do the magic. Unfortunately, neither of models could perform better than benchmark score. Finally, you decided to combine those models. Though, ensembled models are known to return high accuracy, but you are unfortunate. Where did you miss?

As we know, ensemble learners are based on the idea of combining weak learners to create strong learners. But, these learners provide superior result when the combined models are uncorrelated. Since, we have used 5 GBM models and got no accuracy improvement, suggests that the models are correlated. The problem with correlated models is, all the models provide same information. For example: If model 1 has classified User1122 as 1, there are high chances model 2 and model 3 would have done the same, even if its actual value is 0. Therefore, ensemble learners are built on the premise of combining weak uncorrelated models to obtain better predictions.

You are given a data set. The data set contains many variables, some of which are highly correlated and you know about it. Your manager has asked you to run PCA. Would you remove correlated variables first? Why?

Chances are, you might be tempted to say No, but that would be incorrect. Discarding correlated variables have a substantial effect on PCA because, in presence of correlated variables, the variance explained by a particular component gets inflated. For example: You have 3 variables in a data set, of which 2 are correlated. If you run PCA on this data set, the first principal component would exhibit twice the variance than it would exhibit with uncorrelated variables. Also, adding correlated variables lets PCA put more importance on those variable, which is misleading.

What is the difference between covariance and correlation?

Correlation is the standardized form of covariance Covariances are difficult to compare. For example: if we calculate the covariances of salary ($) and age (years), we'll get different covariances which can't be compared because of having unequal scales. To combat such situation, we calculate correlation to get a value between -1 and 1, irrespective of their respective scale.

How is kNN different from kmeans clustering?

Don't get mislead by 'k' in their names. You should know that the fundamental difference between both these algorithms is, kmeans is unsupervised in nature and kNN is supervised in nature. kmeans is a clustering algorithm. kNN is a classification (or regression) algorithm. kmeans algorithm partitions a data set into clusters such that a cluster formed is homogeneous and the points in each cluster are close to each other. The algorithm tries to maintain enough separability between these clusters. Due to unsupervised nature, the clusters have no labels. kNN algorithm tries to classify an unlabeled observation based on its k (can be any number ) surrounding neighbors. It is also known as lazy learner because it involves minimal training of model. Hence, it doesn't use training data to make generalization on unseen data set.

You've got a data set to work having p (no. of variable) > n (no. of observation). Why is OLS as bad option to work with? Which techniques would be best to use? Why?

In such high dimensional data sets, we can't use classical regression techniques, since their assumptions tend to fail. When p > n, we can no longer calculate a unique least square coefficient estimate, the variances become infinite, so OLS cannot be used at all. To combat this situation, we can use penalized regression methods like lasso, LARS, ridge which can shrink the coefficients to reduce variance. Precisely, ridge regression works best in situations where the least square estimates have higher variance. Among other methods include subset regression, forward stepwise regression.

Explain machine learning to me like a 5 year old.

It's simple. It's just like how babies learn to walk. Every time they fall down, they learn (unconsciously) & realize that their legs should be straight and not in a bend position. The next time they fall down, they feel pain. They cry. But, they learn 'not to stand like that again'. In order to avoid that pain, they try harder. To succeed, they even seek support from the door or wall or anything near them, which helps them stand firm. This is how a machine works & develops intuition from its environment.

Explain the difference between L1 and L2 regularization.

L2 regularization tends to spread error among all the terms, while L1 is more binary/sparse, with many variables either being assigned a 1 or 0 in weighting. L1 corresponds to setting a Laplacean prior on the terms, while L2 corresponds to a Gaussian prior.

You came to know that your model is suffering from low bias and high variance. Which algorithm should you use to tackle it? Why?

Low bias occurs when the model's predicted values are near to actual values. In other words, the model becomes flexible enough to mimic the training data distribution. While it sounds like great achievement, but not to forget, a flexible model has no generalization capabilities. It means, when this model is tested on an unseen data, it gives disappointing results. In such situations, we can use bagging algorithm (like random forest) to tackle high variance problem. Bagging algorithms divides a data set into subsets made with repeated randomized sampling. Then, these samples are used to generate a set of models using a single learning algorithm. Later, the model predictions are combined using voting (classification) or averaging (regression). Also, to combat high variance, we can: 1. Use regularization technique, where higher model coefficients get penalized, hence lowering model complexity. 2. Use top n features from variable importance chart. May be, with all the variable in the data set, the algorithm is having difficulty in finding the meaningful signal.

OLS is to linear regression. Maximum likelihood is to logistic regression. Explain the statement.

OLS and Maximum likelihood are the methods used by the respective regression methods to approximate the unknown parameter (coefficient) value. In simple words, Ordinary least square(OLS) is a method used in linear regression which approximates the parameters resulting in minimum distance between actual and predicted values. Maximum Likelihood helps in choosing the the values of parameters which maximizes the likelihood that the parameters are most likely to produce observed data.

Explain prior probability, likelihood and marginal likelihood in context of naiveBayes algorithm?

Prior probability is nothing but, the proportion of dependent (binary) variable in the data set. It is the closest guess you can make about a class, without any further information. For example: In a data set, the dependent variable is binary (1 and 0). The proportion of 1 (spam) is 70% and 0 (not spam) is 30%. Hence, we can estimate that there are 70% chances that any new email would be classified as spam. Likelihood is the probability of classifying a given observation as 1 in presence of some other variable. For example: The probability that the word 'FREE' is used in previous spam message is likelihood. Marginal likelihood is, the probability that the word 'FREE' is used in any message.

Q1. You are given a train data set having 1000 columns and 1 million rows. The data set is based on a classification problem. Your manager has asked you to reduce the dimension of this data so that model computation time can be reduced. Your machine has memory constraints. What would you do? (You are free to make practical assumptions.)

Processing a high dimensional data on a limited memory machine is a strenuous task, your interviewer would be fully aware of that. Following are the methods you can use to tackle such situation: 1. Since we have lower RAM, we should close all other applications in our machine, including the web browser, so that most of the memory can be put to use. 2. We can randomly sample the data set. This means, we can create a smaller data set, let's say, having 1000 variables and 300000 rows and do the computations. 3.To reduce dimensionality, we can separate the numerical and categorical variables and remove the correlated variables. For numerical variables, we'll use correlation. For categorical variables, we'll use chi-square test. 4. Also, we can use PCA and pick the components which can explain the maximum variance in the data set. 5. Using online learning algorithms like Vowpal Wabbit (available in Python) is a possible option. 6. Building a linear model using Stochastic Gradient Descent is also helpful. 7. We can also apply our business understanding to estimate which all predictors can impact the response variable. But, this is an intuitive approach, failing to identify useful predictors might result in significant loss of information.

How is a decision tree pruned?

Pruning is what happens in decision trees when branches that have weak predictive power are removed in order to reduce the complexity of the model and increase the predictive accuracy of a decision tree model. Pruning can happen bottom-up and top-down, with approaches such as reduced error pruning and cost complexity pruning. Reduced error pruning is perhaps the simplest version: replace each node. If it doesn't decrease predictive accuracy, keep it pruned. While simple, this heuristic actually comes pretty close to an approach that would optimize for maximum accuracy.

Define precision and recall

Recall is also known as the true positive rate: the amount of positives your model claims compared to the actual number of positives there are throughout the data. Precision is also known as the positive predictive value, and it is a measure of the amount of accurate positives your model claims compared to the number of positives it actually claims. It can be easier to think of recall and precision in the context of a case where you've predicted that there were 10 apples and 5 oranges in a case of 10 apples. You'd have perfect recall (there are actually 10 apples, and you predicted there would be 10) but 66.7% precision because out of the 15 events you predicted, only 10 (the apples) are correct.

When does regularization becomes necessary in Machine Learning?

Regularization becomes necessary when the model begins to ovefit / underfit. This technique introduces a cost term for bringing in more features with the objective function. Hence, it tries to push the coefficients for many variables to zero and hence reduce cost term. This helps to reduce model complexity so that the model can become better at predicting (generalizing).

Do you have research experience in machine learning?

Related to the last point, most organizations hiring for machine learning positions will look for your formal experience in the field. Research papers, co-authored or supervised by leaders in the field, can make the difference between you being hired and not. Make sure you have a summary of your research experience and papers ready — and an explanation for your background and lack of formal research experience if you don't.

Write an SQL query that makes recommendations using the pages that your friends liked. Assume you have two tables: a two-column table of users and their friends, and a two-column table of users and the pages they liked. It should not recommend pages you already like.

SELECT f.user_id, l.page_id FROM friend f JOIN like l ON f.friend_id = l.user_id WHERE l.page_id NOT IN (SELECT page_id FROM like WHERE user_id = f.user_id)

What is the difference between supervised and unsupervised machine learning?

Supervised learning requires training labeled data. For example, in order to do classification (a supervised learning task), you'll need to first label the data you'll use to train the model to classify data into your labeled groups. Unsupervised learning, in contrast, does not require labeling data explicitly.

Explain how a ROC curve works.

The ROC curve is a graphical representation of the contrast between true positive rates and the false positive rate at various thresholds. It's often used as a proxy for the trade-off between the sensitivity of the model (true positives) vs the fall-out or the probability it will trigger a false alarm (false positives).

What's the difference between probability and likelihood?

The answer depends on whether you are dealing with discrete or continuous random variables. So, I will split my answer accordingly. I will assume that you want some technical details and not necessarily an explanation in plain English. Discrete Random Variables Suppose that you have a stochastic process that takes discrete values (e.g., outcomes of tossing a coin 10 times, number of customers who arrive at a store in 10 minutes etc). In such cases, we can calculate the probability of observing a particular set of outcomes by making suitable assumptions about the underlying stochastic process (e.g., probability of coin landing heads is p and that coin tosses are independent). Denote the observed outcomes by O and the set of parameters that describe the stochastic process as θ. Thus, when we speak of probability we want to calculate P(O|θ). In other words, given specific values for θ, P(O|θ) is the probability that we would observe the outcomes represented by O . However, when we model a real life stochastic process, we often do not know θ . We simply observe O and the goal then is to arrive at an estimate for θ that would be a plausible choice given the observed outcomes O. We know that given a value of θ the probability of observing O is P(O|θ). Thus, a 'natural' estimation process is to choose that value of θ that would maximize the probability that we would actually observe O. In other words, we find the parameter values θ that maximize the following function: L(θ|O)=P(O|θ) L(θ|O) is called the likelihood function. Notice that by definition the likelihood function is conditioned on the observed O and that it is a function of the unknown parameters θ. Continuous Random Variables In the continuous case the situation is similar with one important difference. We can no longer talk about the probability that we observed O given θ because in the continuous case P(O|θ)=0 . Without getting into technicalities, the basic idea is as follows: Denote the probability density function (pdf) associated with the outcomes O as: f(O|θ). Thus, in the continuous case we estimate θ given observed outcomes O by maximizing the following function: L(θ|O)=f(O|θ) In this situation, we cannot technically assert that we are finding the parameter value that maximizes the probability that we observe O as we maximize the PDF associated with the observed outcomes O.

'People who bought this, also bought...' recommendations seen on amazon is a result of which algorithm?

The basic idea for this kind of recommendation engine comes from collaborative filtering. Collaborative Filtering algorithm considers "User Behavior" for recommending items. They exploit behavior of other users and items in terms of transaction history, ratings, selection and purchase information. Other users behaviour and preferences over the items are used to recommend items to the new users. In this case, features of the items are not known.

What do you understand by Bias Variance trade off?

The error emerging from any model can be broken down into three components mathematically. Following are these component : Err(x) = Bias^2 + Variance + Irreducible Error Bias error is useful to quantify how much on an average are the predicted values different from the actual value. A high bias error means we have a under-performing model which keeps on missing important trends. Variance on the other side quantifies how are the prediction made on same observation different from each other. A high variance model will over-fit on your training population and perform badly on any observation beyond training.

Both being tree based algorithm, how is random forest different from Gradient boosting algorithm (GBM)?

The fundamental difference is, random forest uses bagging technique to make predictions. GBM uses boosting techniques to make predictions. In bagging technique, a data set is divided into n samples using randomized sampling. Then, using a single learning algorithm a model is build on all samples. Later, the resultant predictions are combined using voting or averaging. Bagging is done is parallel. In boosting, after the first round of predictions, the algorithm weighs misclassified predictions higher, such that they can be corrected in the succeeding round. This sequential process of giving higher weights to misclassified predictions continue until a stopping criterion is reached. Random forest improves model accuracy by reducing variance (mainly). The trees grown are uncorrelated to maximize the decrease in variance. On the other hand, GBM improves accuracy my reducing both bias and variance in a model.

How can we use your machine learning skills to generate revenue?

This is a tricky question. The ideal answer would demonstrate knowledge of what drives the business and how your skills could relate. For example, if you were interviewing for music-streaming startup Spotify, you could remark that your skills at developing a better recommendation model would increase user retention, which would then increase revenue in the long run. https://www.slideshare.net/dmc500hats/startup-metrics-for-pirates-long-version

Pick an algorithm. Write the psuedo-code for a parallel implementation.

This kind of question demonstrates your ability to think in parallelism and how you could handle concurrency in programming implementations dealing with big data. Take a look at pseudocode frameworks such as Peril-L and visualization tools such as Web Sequence Diagrams to help you demonstrate your ability to write code that reflects parallelism.

What do you think of our current data process?

This kind of question requires you to listen carefully and impart feedback in a manner that is constructive and insightful. Your interviewer is trying to gauge if you'd be a valuable member of their team and whether you grasp the nuances of why certain things are set the way they are in the company's data process based on company- or industry-specific conditions. They're trying to see if you can be an intellectual peer. Act accordingly.

Which is more important to you- model accuracy, or model performance?

This question tests your grasp of the nuances of machine learning model performance! Machine learning interview questions often look towards the details. There are models with higher accuracy that can perform worse in predictive power — how does that make sense? Well, it has everything to do with how model accuracy is only a subset of model performance, and at that, a sometimes misleading one. For example, if you wanted to detect fraud in a massive dataset with a sample of millions, a more accurate model would most likely predict no fraud at all if only a vast minority of cases were fraud. However, this would be useless for a predictive model — a model designed to find fraud that asserted there was no fraud at all! Questions like this help you demonstrate that you understand model accuracy isn't the be-all and end-all of model performance.

You are working on a time series data set. You manager has asked you to build a high accuracy model. You start with the decision tree algorithm, since you know it works fairly well on all kinds of data. Later, you tried a time series regression model and got higher accuracy than decision tree model. Can this happen? Why?

Time series data is known to posses linearity. On the other hand, a decision tree algorithm is known to work best to detect non - linear interactions. The reason why decision tree failed to provide robust predictions because it couldn't map the linear relationship as good as a regression model did. Therefore, we learned that, a linear regression model can provide robust prediction given the data set satisfies its linearity assumptions.

After analyzing the model, your manager has informed that your regression model is suffering from multicollinearity. How would you check if he's true? Without losing any information, can you still build a better model?

To check multicollinearity, we can create a correlation matrix to identify & remove variables having correlation above 75% (deciding a threshold is subjective). In addition, we can use calculate VIF (variance inflation factor) to check the presence of multicollinearity. VIF value <= 4 suggests no multicollinearity whereas a value of >= 10 implies serious multicollinearity. Also, we can use tolerance as an indicator of multicollinearity. But, removing correlated variables might lead to loss of information. In order to retain those variables, we can use penalized regression models like ridge or lasso regression. Also, we can add some random noise in correlated variable so that the variables become different from each other. But, adding noise might affect the prediction accuracy, hence this approach should be carefully used.

You have been asked to evaluate a regression model based on R², adjusted R² and tolerance. What will be your criteria?

Tolerance (1 / VIF) is used as an indicator of multicollinearity. It is an indicator of percent of variance in a predictor which cannot be accounted by other predictors. Large values of tolerance is desirable. We will consider adjusted R² as opposed to R² to evaluate model fit because R² increases irrespective of improvement in prediction accuracy as we add more variables. But, adjusted R² would only increase if an additional variable improves the accuracy of model, otherwise stays same. It is difficult to commit a general threshold value for adjusted R² because it varies between data sets. For example: a gene mutation data set might result in lower adjusted R² and still provide fairly good predictions, as compared to a stock market data where lower adjusted R² implies that model is not good.

How is True Positive Rate and Recall related? Write the equation.

True Positive Rate = Recall. Yes, they are equal having the formula (TP/TP + FN).

What's the difference between Type I and Type II error?

Type I error is a false positive, while Type II error is a false negative. Briefly stated, Type I error means claiming something has happened when it hasn't, while Type II error means that you claim nothing is happening when in fact something is. A clever way to think about this is to think of Type I error as telling a man he is pregnant, while Type II error means you tell a pregnant woman she isn't carrying a baby.

What do you understand by Type I vs Type II error ?

Type I error is committed when the null hypothesis is true and we reject it, also known as a 'False Positive'. Type II error is committed when the null hypothesis is false and we accept it, also known as 'False Negative'. In the context of confusion matrix, we can say Type I error occurs when we classify a value as positive (1) when it is actually negative (0). Type II error occurs when we classify a value as negative (0) when it is actually positive(1).

You are given a data set consisting of variables having more than 30% missing values? Let's say, out of 50 variables, 8 variables have missing values higher than 30%. How will you deal with them?

We can deal with them in the following ways: 1. Assign a unique category to missing values, who knows the missing values might decipher some trend 2. We can remove them blatantly. 3. Or, we can sensibly check their distribution with the target variable, and if found any pattern we'll keep those missing values and assign them a new category while removing others.

I know that a linear regression model is generally evaluated using Adjusted R² or F value. How would you evaluate a logistic regression model?

We can use the following methods: 1. Since logistic regression is used to predict probabilities, we can use AUC-ROC curve along with confusion matrix to determine its performance. 2. Also, the analogous metric of adjusted R² in logistic regression is AIC. AIC is the measure of fit which penalizes model for the number of model coefficients. Therefore, we always prefer model with minimum AIC value. 3. Null Deviance indicates the response predicted by a model with nothing but an intercept. Lower the value, better the model. Residual deviance indicates the response predicted by a model on adding independent variables. Lower the value, better the model.

In k-means or kNN, we use euclidean distance to calculate the distance between nearest neighbors. Why not manhattan distance ?

We don't use manhattan distance because it calculates distance horizontally or vertically only. It has dimension restrictions. On the other hand, euclidean metric can be used in any space to calculate distance. Since, the data points can be present in any dimension, euclidean distance is a more viable option. Example: Think of a chess board, the movement made by a bishop or a rook is calculated by manhattan distance because of their respective vertical & horizontal movements.

You have built a multiple regression model. Your model R² isn't as good as you wanted. For improvement, your remove the intercept term, your model R² becomes 0.8 from 0.3. Is it possible? How?

Yes, it is possible. We need to understand the significance of intercept term in a regression model. The intercept term shows model prediction without any independent variable i.e. mean prediction. The formula of R^2 = 1 - E(y-y')^2/E(y-ymean)^2 where y' is predicted value. When intercept term is present, R² value evaluates your model wrt. to the mean model. In absence of intercept term (ymean), the model can make no such evaluation, with large denominator, E(y-y')^2/E(y)62 equation's value becomes smaller than actual, resulting in higher R^2.

Q2. Is rotation necessary in PCA? If yes, Why? What will happen if you don't rotate the components?

Yes, rotation (orthogonal) is necessary because it maximizes the difference between variance captured by the component. This makes the components easier to interpret. Not to forget, that's the motive of doing PCA where, we aim to select fewer components (than features) which can explain the maximum variance in the data set. By doing rotation, the relative location of the components doesn't change, it only changes the actual coordinates of the points. If we don't rotate the components, the effect of PCA will diminish and we'll have to select more number of components to explain variance in the data set.

When is Ridge regression favorable over Lasso regression?

You can quote ISLR's authors Hastie, Tibshirani who asserted that, in presence of few variables with medium / large sized effect, use lasso regression. In presence of many variables with small / medium sized effect, use ridge regression. Conceptually, we can say, lasso regression (L1) does both variable selection and parameter shrinkage, whereas Ridge regression only does parameter shrinkage and end up including all the coefficients in the model. In presence of correlated variables, ridge regression might be the preferred choice. Also, ridge regression works best in situations where the least square estimates have higher variance. Therefore, it depends on our model objective.

How do you handle missing or corrupted data in a dataset?

You could find missing/corrupted data in a dataset and either drop those rows or columns, or decide to replace them with another value. In Pandas, there are two very useful methods: isnull() and dropna() that will help you find columns of data with missing or corrupted data and drop those values. If you want to fill the invalid values with a placeholder value (for example, 0), you could use the fillna() method.

You are assigned a new project which involves helping a food delivery company save more money. The problem is, company's delivery team aren't able to deliver food on time. As a result, their customers get unhappy. And, to keep them happy, they end up delivering food for free. Which machine learning algorithm can save them?

You might have started hopping through the list of ML algorithms in your mind. But, wait! Such questions are asked to test your machine learning fundamentals. This is not a machine learning problem. This is a route optimization problem. A machine learning problem consist of three things: 1. There exist a pattern. 2. You cannot solve it mathematically (even by writing exponential equations). 3. You have data on it. Always look for these three factors to decide if machine learning is a tool to solve a particular problem.

What evaluation approaches would you work to gauge the effectiveness of a machine learning model?

You would first split the dataset into training and test sets, or perhaps use cross-validation techniques to further segment the dataset into composite sets of training and test sets within the data. You should then implement a choice selection of performance metrics: here is a fairly comprehensive list. You could use measures such as the F1 score, the accuracy, and the confusion matrix. What's important here is to demonstrate that you understand the nuances of how a model is measured and how to choose the right performance measures for the right situations.

You are given a data set on cancer detection. You've build a classification model and achieved an accuracy of 96%. Why shouldn't you be happy with your model performance? What can you do about it?

f you have worked on enough data sets, you should deduce that cancer detection results in imbalanced data. In an imbalanced data set, accuracy should not be used as a measure of performance because 96% (as given) might only be predicting majority class correctly, but our class of interest is minority class (4%) which is the people who actually got diagnosed with cancer. Hence, in order to evaluate model performance, we should use Sensitivity (True Positive Rate), Specificity (True Negative Rate), F measure to determine class wise performance of the classifier. If the minority class performance is found to to be poor, we can undertake the following steps: 1. We can use undersampling, oversampling or SMOTE to make the data balanced. 2. We can alter the prediction threshold value by doing probability caliberation and finding a optimal threshold using AUC-ROC curve. 3. We can assign weight to classes such that the minority classes gets larger weight. 4. We can also use anomaly detection.

What are your favorite use cases of machine learning models?

https://www.quora.com/What-are-the-typical-use-cases-for-different-machine-learning-algorithms-For-instance-under-what-typical-conditions-would-one-prefer-to-use-one-over-the-other-without-having-tested-the-accuracy-of-learning

Where do you usually source datasets?

https://www.springboard.com/blog/free-public-data-sets-data-science-project/ Machine learning interview questions like these try to get at the heart of your machine learning interest. Somebody who is truly passionate about machine learning will have gone off and done side projects on their own, and have a good idea of what great datasets are out there. If you're missing any, check out Quandl for economic and financial data, and Kaggle's Datasets collection for another great list.

Why is naive Bayes so 'naive' ?

naive Bayes is so 'naive' because it assumes that all of the features in a data set are equally important and independent. As we know, these assumption are rarely true in real world scenario.

bagging and boosting similarities and differences:

similarities: both are ensemble methods to get N learns from 1 learner, both generate several training data sets by random sampling, both make the final decision by taking the average of N learners, both are good at reducing variance and proving higher scalability difference: while they are built independently for bagging, boosting tries to add new models that do well where previous models fall. -only boosting determines weight for the data to top the scales in favour of the most difficult cases -is an equally average for bagging and a weighted average for boosting more weight in those with better performance on training data -only boosting tries to reduce bias. on the other hand, bagging may solve the problem of over fititng, while boosting can increase it