Data Scientist & ML Engineer English Essentials

ML basics

model

A function learned from data that makes predictions on new inputs.

feature

An input variable the model uses to make a prediction.

label / target

The correct answer the model is trying to predict.

training

The process of fitting a model to data by adjusting its parameters.

inference

Using a trained model to make predictions in production.

supervised learning

Learning from labelled examples; unsupervised has no labels.

overfitting

When a model memorises the training data and fails on new data.

underfitting

When a model is too simple to capture the patterns in the data.

hyperparameter

A setting you choose before training (learning rate, tree depth), not learned from data.

epoch

One full pass over the training dataset.

gradient descent

The optimisation method that nudges parameters to reduce error.

loss function

The number measuring how wrong the model’s predictions are.

regularization

Techniques that penalise complexity to reduce overfitting.

embedding

A dense numeric vector representing something like a word or user.

fine-tuning

Adapting a pre-trained model to a specific task with extra training.

Evaluation

accuracy

The fraction of predictions that are correct — misleading on imbalanced data.

precision

Of the items flagged positive, how many really were positive.

recall

Of the truly positive items, how many the model caught.

F1 score

The harmonic mean of precision and recall, balancing the two.

AUC / ROC

A threshold-independent measure of how well a classifier ranks positives above negatives.

confusion matrix

A table of true/false positives and negatives.

baseline

A simple reference model your model must beat to be worth deploying.

cross-validation

Splitting data several ways to estimate performance more reliably.

train/test split

Holding out data to test on examples the model never saw.

bias–variance trade-off

Balancing systematic error against sensitivity to the training data.

A/B test

Comparing two versions on live traffic to measure real impact.

RMSE / MAE

Common error metrics for regression — how far predictions are from actuals.

leakage

When information from the future or the target sneaks into training, inflating scores.

Data & pipelines

pipeline

An automated sequence that ingests, transforms and serves data or predictions.

ETL

Extract, Transform, Load — moving and reshaping data into a warehouse.

feature store

A central system for storing and serving features consistently to models.

data drift

When the input data distribution shifts away from what the model trained on.

concept drift

When the relationship between inputs and the target changes over time.

data warehouse

A central store optimised for analytical queries over large datasets.

feature engineering

Creating useful input variables from raw data.

imputation

Filling in missing values in the data.

normalization / scaling

Putting features on a comparable numeric range.

labeling

Annotating data with the correct answers for supervised training.

batch vs streaming

Processing data in scheduled chunks vs continuously as it arrives.

reproducibility

Being able to rerun an experiment and get the same result.

Presenting findings

statistical significance

Confidence that a result is unlikely to be due to chance.

p-value

The probability of seeing the result if there were truly no effect.

confidence interval

A range that likely contains the true value, with a stated confidence.

correlation vs causation

Two things moving together doesn’t mean one causes the other.

effect size

How big a difference is, not just whether it’s statistically detectable.

sample size

How many observations the conclusion is based on.

distribution

How values are spread across a range.

outlier

A data point far from the rest that can distort results.

hypothesis

A testable statement about what the data will show.

interpretability

How easily a human can understand why a model made a prediction.

Key phrases for presenting results & explaining models

• The data suggests a clear correlation, but I’d caution against reading causation into it.
• The model beats the baseline on F1, though precision dipped slightly.
• This result is statistically significant at p < 0.05, with a 95% confidence interval of 2–4%.
• We’re seeing data drift — the input distribution has shifted since we trained.
• Recall matters more than precision here, because missing a fraud case is costly.
• I suspect leakage — the score looks too good, so let me audit the features.
• In plain terms, the model is most confident when the customer has a long history.
• Let me caveat this: the sample size is small, so treat the numbers as directional.
• We optimised for recall at the cost of more false positives — that’s the trade-off the business asked for.
• The headline number is encouraging, but the confidence interval is wide.
• To translate that into impact: this should reduce churn by roughly two percentage points.
• The model is overfitting — training accuracy is high but it falls apart on the holdout set.
• I’d recommend we A/B test this before rolling it out to everyone.
• These are early results from a single experiment, not a final recommendation.
• The feature importance tells us location is the strongest signal by far.
• Happy to go deeper on the methodology, but the takeaway is the lift is real and repeatable.
• We retrained after detecting drift, and performance recovered to the previous level.
• Let me walk you through the assumptions before we get to the conclusion.
• Yesterday: finished the feature engineering. Today: tuning hyperparameters. No blockers.
• Quick question before I proceed: do we care more about accuracy or interpretability for this use case?