Everything for Data Scientist & ML Engineers
Data scientists and ML engineers need to explain models to non-technical stakeholders as often as they read research papers. This hub covers ML vocabulary, the grammar of hedged, evidence-based reporting, and every relevant interview and blog resource — plus a link into the full AI/ML vocabulary cluster for deeper study.
Vocabulary sets
Grammar & writing
Interview prep
Related deep-dive hub
Blog articles (3)
- Claude API English: Tool Use and Extended Thinking Vocabulary
Master the English vocabulary of the Anthropic Claude API — tool use, extended thinking, streaming, and prompt engineering terms explained for IT professionals.
- LangChain English: LCEL and RAG Pipeline Vocabulary
Learn the English vocabulary used in LangChain development — LCEL chains, RAG pipelines, retrievers, memory, and agent vocabulary explained in professional context.
- OpenAI Assistants API English: Threads, Runs, and Vector Stores
Learn the English vocabulary of the OpenAI Assistants API — threads, runs, vector stores, tool calls, and streaming responses explained for IT professionals.
Other role hubs
Explore more
Browse every exercise category, or search the full site.
Frequently Asked Questions
What's the difference between feature engineering and feature selection, specifically in a time series context?
Feature engineering involves creating new features from existing ones – for example, calculating rolling averages or lagged values in time series data. Feature selection focuses on choosing the most relevant subsets of original features. In time series, you might engineer features like exponentially weighted moving averages to capture trends and then select those that best predict future values based on statistical significance.
I'm building a recommendation system; how do I handle cold start problems effectively with collaborative filtering?
Cold start occurs when you have little or no interaction data for new users or items. Techniques like content-based filtering (using item metadata) can provide initial recommendations based on similarity to existing user preferences. Hybrid approaches, combining both collaborative and content-based methods, often mitigate cold start issues by leveraging available information.
Can you explain the concept of 'drift' in model monitoring for deployed ML systems?
Drift is the change in data characteristics over time which impacts model accuracy. It can be feature drift – changes in individual features' distributions – or concept drift – a shift in the relationship between input features and the target variable. Monitoring these drifts allows for proactive model retraining to maintain performance.
What is A/B testing and how do I set up an experiment to evaluate different ML models?
A/B testing is a controlled experiment used to evaluate different versions of a product or feature. For ML models, you'd split your user base into two groups and have each group interact with a different model. Metrics like AUC, precision, or recall are compared between the groups using statistical tests.
What is Bayesian Optimization for hyperparameter tuning and how does it differ from grid search?
Bayesian Optimization utilizes a surrogate model to predict the performance of different hyperparameter combinations. It efficiently explores the search space by balancing exploration (trying new values) and exploitation (refining existing good ones), unlike grid search's brute-force approach.
Explain how you would implement data validation within a machine learning pipeline?
Implementing data validation ensures your ML pipeline receives clean and reliable input. You'd define rules for each data field (e.g., numerical ranges, categorical values) and use libraries like Great Expectations or Pandas to automatically check against these rules.
What is the purpose of a 'shadow deployment' in deploying a new ML model?
Shadow deployments are a low-risk way to test new ML models in a real-world environment. By comparing the shadow model's output to the live model, you can identify potential issues and refine the model without impacting users.
What is Differential Privacy and why is it important when working with sensitive data?
Differential Privacy is a technique that protects user data by adding carefully calibrated noise to calculations. It ensures that an adversary cannot determine whether a particular individual's data was included in the dataset, preserving privacy while maintaining useful statistical insights.
How can I use version control (like Git) effectively for my ML code and models?
Version control is essential for managing ML projects. Track changes to your code, data preprocessing steps, and model parameters. Use Git branches for experimentation and tags for releasing stable versions of your models.
What are the key differences between online learning and batch learning in machine learning?
Batch learning trains models on complete datasets, while online learning updates them iteratively with each new data point. Batch learning is computationally intensive but provides a stable model; online learning adapts dynamically but can be sensitive to noisy data.