100+ Machine Learning interview questions and answers 2024

Decision tree is a supervised machine learning method that shows a map of possible outcomes of related decisions, similar to a tree structure using datasets divided into smaller subsets as the decision tree develops.

The Naive Bayes is called naive because, as a supervised learning algorithm, it makes assumptions by applying the Bayes Theorem that all attributes are independent of each other.

The Bayes' theorem for Naive Bayes Method is stated as follows:

P(A|B) = P(B|A) * P(A) / P(B)

Where,
P(A|B) is the posterior probability of class (target) given predictor (attribute)
P(B|A) is the likelihood which is the probability of predictor given class
P(A) is the prior probability of class
P(B) is the prior probability of predictor

This is when memory is allocated as soon as the array is declared at compilation time. This allocation offers more efficiency and speed in operation as there is no overhead. However, once allocated, the assigned memory space cannot be deallocated during compilation. Static memory is automatically deallocated when the program terminates, and the memory is returned to the operating system.

Memory allocation is said to be dynamic if the allocation occurs at runtime whenever a new node is added. Nowadays, modern operating systems use dynamic memory allocation.

DMA is useful in scenarios such as:

• You don't know the amount of memory needed beforehand.
• You want data structures without any upper limit
• For efficient use of memory space

Precision and Recall are two commonly used metrics in the field of ML to evaluate the performance of a binary classification model.

Precision measures how many of the items predicted to belong to the positive class are actually positive. It is defined as the ratio of true positives (TP) to the total number of predicted positives.

Precision = TP / (TP + FP)

Recall, on the other hand, measures how many of the actual positive items the model is able to identify. It is defined as the ratio of TP to the total number of actual positives.

Recall = TP / (TP + FN)

Outliers are errors or extreme values that differ from the data in a set. They can impact the accuracy of your result if not detected through the outlier detection tools. Some popular tools to discover outliers include Z-score, Scatter plot, and Box plot.

This mathematical technique is used to transform a function of time into a function of frequency. It is necessary for ML and deep learning because a convolution in the time domain is a multiplication in the frequency domain.

Inductive logic programming (ILP) is a part of AI that uses logic programming and serves as a uniform representation for examples, background knowledge, and hypotheses. It involves developing algorithms and systems that can automatically learn logic programs from input-output examples, which can then be used to make predictions on new input data.
We can apply it to various domains, including software engineering, natural language processing, and bioinformatics.

Kernel SVM stands for Kernel Support Vector Machine. In SVM, a kernel is a function that aids in problem-solving. They provide shortcuts to help you avoid doing complicated math. The amazing thing about kernel is that it allows us to go to higher dimensions and execute smooth computations. Kernel SVMs can work with a variety of kernel functions, including linear, polynomial, and radial basis function (RBF) kernels, among others.

These are some types of clustering:

Fuzzy clustering: Each data point can belong to multiple clusters.

: It divides a set of n observations into k clusters. Used in situations when one doesn’t have existing group labels.

Hierarchical clustering: It is an algorithm that groups similar objects into segments called clusters. Here, each cluster differs from the others, and the objects within each cluster are the same.

Density-based clustering: It clusters data based on their density and hence requires uniform density within the cluster. Here density drops n=between clusters.

Reinforcement learning is a type of ML where an agent interacts with an environment and learns to make decisions based on feedback from the environment in the form of rewards or penalties. Unlike supervised and unsupervised learning, the agent is not given labeled data but instead learns from its own experiences through trial and error. The objective of reinforcement learning is to find an optimal policy that maximizes the cumulative reward obtained by the agent over time.

Some methods to improve inference time are as follows:

• Cache predictions
• Knowledge distillation
• Reduction of parameters by pruning
• Model size reduction
• Parallel computing

Content-based filtering recommends items to a user based on their past behavior, preferences, and interests. For example, if a user has shown interest in action movies, then the content-based filtering algorithm will recommend more action movies to that user.

Collaborative filtering recommends items to a user based on the preferences of other users who have similar tastes. For example, if a user has watched several action movies, the collaborative filtering algorithm will recommend other action movies that were also watched by users who have similar movie preferences.

Deductive learning is the process of using conclusions to form observations, while inductive learning is using observations to draw conclusions.

Deductive learning is a top-down approach where a general principle or theory is used to make specific predictions about observations or new data. The process involves starting with a general statement, or hypothesis, and then testing it through observation and experimentation to confirm or reject the hypothesis.

Inductive learning, on the other hand, is a bottom-up approach where specific observations or data are used to form a general principle or theory. This approach involves observing patterns and regularities in the data, and then using those observations to form a hypothesis or general principle.

We use inductive learning when there is little or no data available to derive conclusions. On the contrary, some theories are available in the deductive learning approach.

Data mining is the process of using structured data to extract information or find anomalies, correlations, and patterns within large datasets to predict outcomes using machine learning algorithms.

But Machine Learning is the development of algorithms to provide abilities to processors to learn without being programmed.

The ROC curve is important because it is a visual representation of how well a model can distinguish between two classes, and it can be used to compare different models. The area under the curve (AUC) is a measure of how well the model performs, with a higher AUC indicating a better model. Additionally, the shape of the curve can indicate whether a model is biased towards one class or another.

Overfitting occurs in ML when the model is too complex or has too many parameters relative to the amount of data that is available. This causes the model to fit the noise of the data rather than the underlying patterns, resulting in poor generalization and an inability to accurately predict on previously unseen data.

The following are some functions of unsupervised learning:

• Dimensionality reduction
• Finding clusters of data
• Anomaly deduction

Some functions of supervised learning as follows:

• Regression
• Feature selection
• Classification
• Time series forecasting

Two components of Bayesian logic are as follows:

Prior distribution: This is the information or beliefs that you have about the world before observing any new data. Prior knowledge is typically represented as a probability distribution over possible states of the world.

Likelihood principle: This is a function that describes the probability of observing some data given a particular state of the world. The likelihood function is typically derived from scientific theories, empirical observations, or expert opinions.

A recommendation engine can be seen as a system to predict users’ likes and interests and recommend products that fit their tastes. Data used in this system can be derived from user ratings of movies and songs and search engine history.

ML algorithms need numeric inputs. It requires converting categorical values into factors to get the latter. The functions factor () and as.factor () are used to perform categorical conversion into factor.

No. Elements can be stored anywhere in linked lists. A linked list contains nodes where each node is composed of a data field and a reference (link) to the next node.

Regularization is a technique used to prevent overfitting in ML models. Overfitting occurs when a model is too complex and fits the training data too closely, resulting in poor performance on new, unseen data. Regularization techniques add a penalty term to the model's objective function, which encourages the model to find simpler solutions by reducing the magnitude of the coefficients or parameters in the model. This reduces the model's ability to fit the noise in the training data, resulting in better generalization performance on new data. Some popular regularization techniques include L1 (lasso) and L2 (ridge) regularization, which differ in the way they penalize the coefficients.

Decision tree algorithms are favored for their interpretability, scalability, and accuracy. They can be used to classify data points quickly and accurately, and are great for dealing with large datasets. They are also simple to understand, making them great for explaining complex models to non-technical stakeholders.

The exploding gradient problem occurs when large error gradients gather in training, resulting in large changes in the neural network weights. It can result in NaN values due to the values of weights being too large.

If your model is simple with only a few parameters, it will have high bias and low variance. With many parameters, it will have low bias and high variance. One needs to find the right balance without overfitting and underfitting the data. This is why there is a tradeoff between bias and variance. Bias-variance tradeoff is important as it helps us know the limitations of our model and how to improve on it.

This is an evaluation metric in ML that sums up the predictive performance of a model by combining two competing metrics, namely, Precision and Recall. In binary classification, we assume F1 score to measure a model’s accuracy because it is the weighted average of precision and recall scores.

Scores of F1 range from 0 to 1, with 1 being the best score and 0 being the worst. These F1 scores are used in information retrieval to see how well a model recovers results and how the model performs.

We usually call it the Loss function when calculating loss considering only a single point. However, we call it the Cost function when calculating the sum of errors for multiple data.
In simpler terms, the cost function sums the difference for the training database, while the loss function shows the difference between the actual and predicted values for a single record.

Three simple strategies can be used to handle outliers, and these are:

• Dropping them
• Marking them as outliers and including them as a feature
• Transforming the feature to reduce the effect of the outlier

Random forest is an ensemble learning method that combines multiple decision trees to make predictions. Each decision tree is built using a different subset of the training data and a random subset of the features. This randomness helps to prevent overfitting and improve the generalization of the model.

Random forest works in the following steps:

1. Take up a sample size from the training data
2. Start with a particular node
3. Run the algorithm from the start node viz: If the number of observations is smaller than node size, then stop Select random variables Spot the variable that is best at splitting the observations Divide the observations into two nodes Call step ‘a’ on each of these two nodes

During the process of building each decision tree, the algorithm splits the data at each node based on the feature that provides the best split. This split is determined using a metric such as information gain or Gini impurity.

Once all the decision trees are built, they are used together to make predictions. The idea is that each tree has its strengths and weaknesses, but by combining them, the weaknesses are averaged out, and the strengths are amplified. This ensemble approach is often more accurate than using a single decision tree.

Two ensemble techniques that can be used include Bagging and Boosting:

Bagging ((Bootstrap Aggregating): It is a technique where multiple independent models are trained on different subsets of the training data, and their predictions are aggregated to form the final prediction. The idea behind bagging is to reduce the variance of the model by averaging over multiple models.

Boosting: This is an iterative ensemble technique where the models are trained sequentially, and each subsequent model is trained on the instances where the previous model performed poorly. The idea behind boosting is to reduce the bias of the model by focusing on the misclassified instances

The threshold for a classifier can be derived using precision-recall Curves and ROC Curves. In other instances, a grid search can be used to adjust the threshold for finding the best value.

Plots can be used to check for the normality of datasets, and some of these checks are:

• Kolmogorov-Smirnov Test
• Shapiro-Wilk Test
• Anderson-Darling Test
• D’Agostino Skewness Test

Parametric models have limited parameters and only need to know the model's parameters to predict new data, whereas non-parametric models have unlimited parameters and have more flexibility in new data prediction.

No. Logistic regression is a binary classifier; thus, it cannot be applied to more than two classes. However, it can be employed in multinomial logistic regression of solving multi-class classification problems.

The sigmoid function is a mathematical function that maps any input value to a value between 0 and 1 and it is used for binary classification. On the other hand, the softmax function is used for multi-classification, and the sum of the probabilities will be 1.

Some of the methods that you can use to avoid overfitting in machine learning are:

• Cross validation
• Custom feature selection
• Data augmentation
• Using larger datasets
• Data simplification
• Ensembling
• Regularization

t depends on the situation. A false negative is risky in fields like medicine when a virus is scanned and reported absent when it is present. A false positive is risky in situations like spam email detection because it can classify an important email as spam.

The bagging algorithm can be used to split the data into subgroups sampling replicated from random data. After this split, the random data is utilized to establish rules using a training algorithm. Thereafter, the polling technique is used to combine all the predicted outcomes of the model.

Genetic programming is similar to Evolutionary algorithm. It is a subset of evolutionary algorithms that implement the model's parameter, random mutation, crossover, fitness function, and various layers of evolution to resolve a user-defined task.

Through repeated generations of evaluation and modification, genetic programming can evolve increasingly complex and sophisticated programs that are optimized for the task at hand. This approach can be used for a wide range of applications, from artificial intelligence and machine learning to robotics and optimization problems.

Some classification methods SVM can handle are as follows:

• Binary classification
• Multiclass classification
• Multilabel classification

Instance-based learning algorithm is sometimes referred to as lazy learning algorithm because it does not require any training data or generalization of the data. Instead, it uses the data the algorithm is given and stores it in memory for use in future predictions. The algorithm does not actively learn from the data, but rather relies on the stored instances to make predictions. Thus, its learning process can be seen as "lazy" as compared to other algorithms that actively learn from the data.

Pruning is the process of removing unnecessary nodes in a decision tree in order to reduce the complexity and improve the generalization accuracy of the model. It is used to prevent overfitting. Pruning helps in reducing the complexity of the decision trees by removing the nodes that do not have a great impact on the prediction accuracy.

Regularization is a technique used to prevent overfitting by adding a penalty term to the cost function. The penalty term is designed to discourage the model from assigning too much importance to any single feature or parameter, and encourages the model to distribute importance across all features.

Converting categorical variables into factors is necessary because many ML algorithms cannot work with categorical variables in their raw form. ML algorithms work with numerical values, and factors are a way of representing categorical variables as numerical values.

For a better predictive model, the variable has to be ordinal in nature.

The confusion matrix is usually a summary of predictions on a classification model that enables us to visualize the performance of an algorithm or model. It measures the performance of an algorithm or model.

This happens when there is a power transformation of non-normal dependent variables into normal variables. It involves a lambda parameter which, when set to zero, indicates that this transform is equivalent to log-transform.

Usually, missing or corrupted values happen when no data is available for certain variables; this can happen due to incomplete data entry, lost files, etc. Some functions that identify, drop, and replace missing or corrupted values are IsNull(), dropna( ), and Fillna() functions in Pandas.

Cross validation divides data into three parts which are training, testing, and validation data. Data is further divided into k numbers of sub-parts, and the model is trained on k-1 of those sub-parts and then tested on the remaining fold. This process is repeated k times, with each of the k folds used exactly once for testing. The final performance metric is the average of the performance metrics from each of the k test sets.

• The first condition states that if the sum of the values on the 2 dice is equal to 7, then you win $50. But for all the other cases, you must pay $4.
• To begin with, let's estimate the number of possible cases if we throw the two dice. The total number of cases is 6*6 = 36.
• In these 36 cases, how many produce a sum of 7?
• Possible combinations that produce a sum of 7 are (1,6), (2,5), (3,4), (4,3), (5,2) and (6,1)..
• This indicates that out of 36 chances, 6 will produce a sum of 7. The ratio now is 6/36 = 1/6
• So this suggests that we have a chance of winning $50, once in every 6 games.
• So to answer the question, if I play 6 times, I will win $50 for one game, whereas for the other 5 games I will have to pay $4 each, which is $20 for all five games. Therefore, I will be at a winning advantage if I win $50 but end up paying $20. Statistically, I’ll be making a profit of $30 every 6 games.

This is a recommendation based on collaborative filtering. Collaborative filtering compares users' activities on a website and recommends similar actions to users with similar interests or behavior.

To solve high variance problems, we will use regularization techniques to rectify this or select the higher features in the feature importance graph to train the model.

For low bias, we could use the bagging algorithm to divide the data into sub-parts using randomized sampling. Thereafter, we can create models with these samples using just one algorithm. Finally, we join the model prediction using voting classification or, perhaps, averaging.