Questions et reponses d'entretien

Data science is a multidisciplinary field that uses scientific methods, processes, algorithms, and systems to extract insights and knowledge from structured and unstructured data. It combines expertise from various domains such as statistics, mathematics, computer science, and domain-specific knowledge to analyze and interpret complex data sets.

The primary goal of data science is to uncover hidden patterns, correlations, and trends in data that can be used to make informed decisions and predictions. Data scientists use a variety of tools and techniques, including statistical analysis, machine learning, data visualization, and big data technologies, to extract meaningful information from large and diverse data sets.

Data science examples in business include processes such as aggregating a customer's email address, credit card information, social media handles, and purchase identifications in order to identify trends in their behavior.

Example:

Creating a new feature 'days_since_last_purchase' for a customer churn prediction model.

Example:

Converting categorical variables like 'color' into binary vectors (e.g., red: [1, 0, 0], blue: [0, 1, 0], green: [0, 0, 1]).

Example:

Supervised learning: Classification tasks like spam detection. Unsupervised learning: Clustering similar customer profiles.

Example:

A complex polynomial regression model fitting the training data perfectly but performing poorly on test data.

Example:

K-fold cross-validation divides data into k subsets; each subset is used for both training and validation in different iterations.

Example:

A linear regression model might have high bias if it oversimplifies a complex problem, while a high-degree polynomial may have high variance.

Example:

A model with a higher Area Under the ROC Curve (AUC-ROC) is generally considered better at distinguishing between classes.

Example:

If the p-value is 0.05, there is a 5% chance of observing the data if the null hypothesis is true.

Example:

A Random Forest model combining predictions from 100 decision trees to enhance accuracy and reduce overfitting.

Example:

A Bagged decision tree ensemble, where each tree is trained on a different bootstrap sample of the data.

Example:

In fraud detection, precision is crucial to minimize the number of false positives, i.e., legitimate transactions flagged as fraudulent.

Example:

Segmenting customers based on purchasing behavior to identify marketing strategies for different groups.

Example:

There may be a correlation between ice cream sales and drownings, but ice cream consumption does not cause drownings.

Example:

Testing two different website designs (A and B) to determine which leads to higher user engagement.

Example:

Increasing model complexity may reduce bias but increase variance, leading to overfitting.

Example:

For a binary classification problem, a confusion matrix might look like: [[TN, FP], [FN, TP]].

Example:

In high-dimensional spaces, data points are more spread out, and distance metrics become less meaningful.

Example:

L1 regularization (Lasso) penalizes the absolute values of coefficients, encouraging sparsity in feature selection.

Example:

Adjusting the learning rate and the number of hidden layers in a neural network to maximize accuracy.

Example:

In a neural network for image classification, cross-entropy loss penalizes incorrect predictions with low probabilities.