1
Q
What does NLP stand for?
A
Natural Language Processing
NLP is the process of turning messy human language into something machines can analyze and learn from.
2
Q
Why is NLP considered challenging?
A
- Language is unstructured
- Meaning includes context, tone, sarcasm, and word combinations
These factors make it difficult for machines to understand human language accurately.
3
Q
What is the Bag of Words / TF-IDF technique used for in NLP?
A
Counts how often words appear
Common words like ‘and’ are less important, while rare words are more significant.
4
Q
What does tokenization do?
A
Splits text into words or sentences and removes punctuation
It can break meaning, as in the case of names.
5
Q
What is the purpose of stop word removal?
A
Removes boring words like ‘the’, ‘and’, ‘to’
This makes text smaller and faster to process, but care must be taken not to remove meaning.
6
Q
What is stemming in NLP?
A
Chops words down to a base form
For example, ‘faster’ becomes ‘fast’.
7
Q
What is the difference between stemming and lemmatization?
A
- Stemming: Chops words to a base form
- Lemmatization: Maps words to their real root
Lemmatization keeps meaning better, e.g., ‘coding’, ‘coded’, ‘codes’ → ‘code’.
8
Q
What do recommender systems do?
A
Suggest things you’ll probably like
They are used by platforms like Netflix, Amazon, and Spotify.
9
Q
What is Collaborative Filtering?
A
Based on your past behavior
Example: ‘You liked X, so try Y’. It improves over time but struggles with new users.
10
Q
What is Content-Based Filtering?
A
Based on who you are and what the item is
It uses factors like age, gender, preferences, and item features.
11
Q
True or false: NLP helps machines extract meaning from human language.
A
TRUE
It involves both simple methods that count words and advanced methods that try to understand context.
12
Q
What is the final takeaway regarding NLP and recommender systems?
A
NLP helps machines extract meaning; cleaned text feeds into ML models for recommendations
Human-level language understanding remains a significant AI research goal.
13
Q
What is the definition of Machine learning?
A
A data-driven approach that uses algorithms to learn patterns and relationships from data without being explicitly programmed
Machine learning enables systems to improve their performance on tasks through experience.
14
Q
In machine learning, what does the developer provide to the algorithm?
A
- Data
- An objective
The algorithm uses this information to learn and create a model.
15
Q
What is created after the algorithm is trained on the provided data?
A
A model
The trained model is used for predicting behaviors and outputs.
16
Q
True or false: The trained model in machine learning can be used for decision-making on unseen data.
A
TRUE
This capability allows for predictions based on new inputs.
17
Q
List some practical applications of machine learning.
A
- E-mail spam detection
- Customer Churn
- Text Sentiment Analysis
- Fraud Detection
- Real-time Ads
- Recommendation Engine
These applications demonstrate the versatility of machine learning across various industries.
18
Q
What is the first step in the supervised ML project workflow in scikit-learn?
A
Data
The workflow follows a loop: Data → split → fit model → predict → (evaluate) → save model.
19
Q
What libraries are imported to handle data in a supervised ML project?
A
- numpy
- pandas
These libraries are essential for data manipulation and analysis.
20
Q
In the context of supervised ML, what does X represent?
A
Features (inputs)
X = df.drop(‘species’, axis=1) represents measurements of the flower.
21
Q
In the context of supervised ML, what does y represent?
A
Target (what we predict)
y = df[‘species’] indicates the flower species being predicted.
22
Q
What is the purpose of the train_test_split function?
A
- Model learns from train set
- Model is tested on unseen data from test set
It is crucial for evaluating model performance.
23
Q
What does test_size=0.2 indicate in the train_test_split function?
A
80% train, 20% test
This defines the proportion of data used for training versus testing.
24
Q
What does random_state=101 ensure in the train_test_split function?
A
Reproducibility
It allows for the same split every time, yielding consistent results.
25
What is created when you execute **model = DecisionTreeClassifier()**?
A model object
## Footnote
This does NOT train the model yet; it merely specifies the algorithm to be used.
26
What happens during the **fit** step of the model?
Model learns patterns in X_train
## Footnote
This is where the actual training occurs using the training data.
27
What does the **predict** method do in a trained model?
Predict labels
## Footnote
It outputs an array of predicted species based on the test data.
28
What is the output of **model.predict_proba(X_test)**?
Confidence for each class
## Footnote
It provides probabilities for each predicted class.
29
What is the **golden rule** for live data when using a model?
Same columns, same format, same preprocessing
## Footnote
This ensures compatibility with the trained model.
30
What command is used to **save the model** after training?
joblib.dump(model, 'my_first_ml_model.pkl')
## Footnote
This allows for future use without needing to retrain.
31
True or false: The lesson covers evaluation metrics, hyperparameter tuning, and data cleaning.
FALSE
## Footnote
These topics are mentioned as aspects that will be covered later.
32
What is the **standard scikit-learn supervised ML workflow**?
* split data
* fit a model
* predict
* save the model
## Footnote
This pattern repeats forever in ML.
33
What is the purpose of **importing packages** in the ML workflow?
* tools for data handling
* visuals
* ML models + utilities
## Footnote
This step is considered boilerplate and does not involve learning yet.
34
What is the **rule** for deciding features vs target in a dataset?
* X = what the model sees
* y = what the model learns to predict
## Footnote
This decision depends on the business/problem goal, not the dataset.
35
What is the purpose of the **train_test_split** function in supervised learning?
* Train set: model learns patterns
* Test set: model is tested on unseen data
## Footnote
Key parameters include test_size=0.2 and random_state=101 for reproducibility.
36
What does creating the model with **model = DecisionTreeClassifier()** signify?
This does NOT train anything
## Footnote
You’re just choosing which algorithm to use.
37
What happens during the **fit** step in the ML workflow?
The model sees feature patterns and learns how they map to species
## Footnote
This is the actual training step.
38
What does **model.predict(X_test)** output?
* Array of predicted species
* Confidence for each class
## Footnote
Example: [0.0, 1.0, 0.0] indicates 100% confidence for versicolor.
39
What is the **golden rule** when predicting on live data?
* Same columns
* Same order
* Same preprocessing as training data
## Footnote
This ensures the model works correctly with new inputs.
40
What is the purpose of **saving the model** using joblib?
* Don’t retrain every time
* Use model in apps, APIs, dashboards
## Footnote
Load later with loaded_model = joblib.load('my_first_ml_model.pkl').
41
What are **Parameters** in machine learning?
* Learned from data
* Examples: tree splits, weights, coefficients
## Footnote
Parameters are the internal variables that the model learns during training.
42
What are **Hyperparameters** in machine learning?
* Set by YOU
* Examples: n_estimators, max_depth
## Footnote
Hyperparameters are external configurations that govern the training process.
43
What does **GridSearchCV** do?
Tunes hyperparameters
## Footnote
It automates the process of finding the best hyperparameter combinations using cross-validation.
44
Before using GridSearchCV, what was the common issue with training models?
Trained one model with default settings
## Footnote
There was uncertainty about whether the default settings were optimal.
45
After implementing GridSearchCV, how do you train models?
* Train many models
* Different hyperparameter combinations
* Using cross-validation
## Footnote
This approach allows for a systematic evaluation of model performance.
46
What is the **test set** in machine learning?
NEVER touched during tuning
## Footnote
The test set is used solely for evaluating the final model's performance.
47
What is the **workflow** when using GridSearchCV?
* Split data
* Define pipeline
* Define hyperparameter grid
* GridSearchCV (fit MANY models)
* Select best model
* Evaluate on test set
## Footnote
This workflow includes an additional step for hyperparameter tuning.
48
What is the purpose of a **pipeline** in machine learning?
* Ensures no data leakage
* Makes GridSearchCV possible
* Treats preprocessing + model as ONE object
## Footnote
Pipelines streamline the process and maintain data integrity.
49
What does the **param_grid** define in GridSearchCV?
Hyperparameter values to try
## Footnote
It specifies the hyperparameters and their respective values for tuning.
50
What does the **cv** parameter represent in GridSearchCV?
Number of cross-validation folds
## Footnote
It determines how many subsets the training data will be divided into for validation.
51
What happens when you call **grid.fit(X_train, y_train)**?
* Trains models
* Validates models
* Scores models
* Averages results
## Footnote
This process involves multiple iterations based on the defined hyperparameters and cross-validation.
52
What does **grid.cv_results_** provide?
A dictionary of results
## Footnote
It contains information about hyperparameters tried, mean validation scores, and rankings.
53
How do you obtain the **best hyperparameters** from GridSearchCV?
grid.best_params_
## Footnote
This returns the hyperparameter combination that yielded the best performance.
54
What does **grid.best_estimator_** return?
The entire trained pipeline
## Footnote
This includes the model with the best hyperparameters, ready for deployment or evaluation.
55
What does a big train–test performance gap indicate?
Overfitting
## Footnote
This suggests that the model learned the training data very well but generalizes poorly to unseen data.
56
True or false: GridSearchCV automates hyperparameter tuning using cross-validation.
TRUE
## Footnote
It simplifies the process of optimizing model performance through systematic evaluation.
57
What is the objective of using **GridSearchCV** in this context?
Use GridSearchCV with a custom scoring metric to optimise a binary classification model based on business priorities, not just accuracy
## Footnote
Focus on recall for malignant cancer.
58
What are the key **differences** between regression and classification metrics?
* Regression: R², MAE, MSE
* Classification: accuracy, recall, precision, F1
* Regression: One output
* Classification: Multiple classes
* Regression: No class priority
* Classification: Class priority matters
## Footnote
In classification, you must decide what matters more.
59
In the business context, what does **Class 0** represent?
malignant
## Footnote
Class 1 represents benign.
60
What is the client decision regarding **malignant tumors**?
Missing a malignant tumour is worse than falsely flagging a benign one
## Footnote
This decision drives the optimization for recall on class 0.
61
What is the first step in the **workflow** for using GridSearchCV?
Load & split data
## Footnote
Use the breast cancer dataset for binary classification.
62
What does the **pipeline** in GridSearchCV help prevent?
Data leakage
## Footnote
It keeps preprocessing and model together.
63
Why is **accuracy** not enough as a scoring metric in classification?
Accuracy can lie
## Footnote
A model could predict 'benign' always and still achieve high accuracy while failing to identify malignant cases.
64
What is the purpose of using **make_scorer** in GridSearchCV?
To define a custom scoring metric focused on recall for malignant cases
## Footnote
Example: scoring = make_scorer(recall_score, pos_label=0).
65
What happens when you run **GridSearchCV** with the defined parameters?
Multiple models are trained and recall is computed for malignant cases
## Footnote
Scores are averaged per hyperparameter.
66
What does the output of GridSearchCV provide regarding **n_estimators**?
Mean recall scores for different n_estimators
## Footnote
Example output: n_estimators=50 → mean ≈ 0.86.
67
How do you obtain the **best model** from GridSearchCV?
pipeline = grid.best_estimator_
## Footnote
This gives you the full pipeline with the best hyperparameters.
68
What should you evaluate after obtaining the best model?
Confusion matrix + report
## Footnote
Check recall for malignant and precision trade-off.
69
What defines **model success** in this context?
Business rules, not code
## Footnote
Example: If client threshold ≥ 90% → acceptable.
70
What is the final mental model for **GridSearchCV** in classification?
Same workflow as regression, but scoring metric defines what 'best' means
## Footnote
High recall often lowers precision (trade-off).
71
What are some **real-life applications** of this approach?
* Build medical ML systems
* Detect fraud
* Catch churn
* Flag risky users
## Footnote
Focus on optimizing decisions rather than just training models.
72
What is the first step in the **workflow** for binary classification?
Split data
## Footnote
This initiates the process of preparing data for model training and evaluation.
73
What is the structure of a **pipeline** in the workflow?
* Scaling
* Feature selection
* Model (e.g., RandomForestClassifier)
## Footnote
The pipeline organizes the steps for data processing and model training.
74
What is included in the **hyperparameter grid**?
* Parameters to tune
* Example: n_estimators = [10, 20]
## Footnote
Hyperparameter tuning is essential for optimizing model performance.
75
How does scoring differ between **binary** and **multiclass** classification?
Binary: make_scorer(metric, pos_label=…)
Multiclass: make_scorer(metric, labels=[class_of_interest], average=None)
## Footnote
Different scoring methods are required to evaluate models based on the type of classification.
76
In the context of **GridSearchCV**, what does the workflow involve?
* Train multiple CV folds for each hyperparameter combination
* Compute metric for the target class
* Pick combination with highest score
## Footnote
This process ensures thorough evaluation of hyperparameter settings.
77
What do **grid.best_params_** and **grid.best_estimator_** represent?
* grid.best_params_: best hyperparameters
* grid.best_estimator_: trained best pipeline
## Footnote
These attributes provide access to the optimal settings and model after grid search.
78
What is crucial for evaluating results in terms of business requirements?
* Precision
* Recall
* F1
## Footnote
The chosen metric should align with the client's specific needs for a particular class.
79
What does **PCA** stand for?
Principal Component Analysis
## Footnote
PCA is a technique used to reduce the dimensionality of data while preserving as much variance as possible.
80
What is the **big idea** behind PCA?
Squish many features into fewer, smarter features while keeping as much important information (variance) as possible
## Footnote
This allows for a more manageable dataset while retaining essential information.
81
What is the trade-off when using PCA?
* Fewer dimensions
* Faster models
* Better visuals
* Components are not directly interpretable
## Footnote
While PCA simplifies data, it can make interpretation of components challenging.
82
When should you **use PCA**?
* Lots of numeric features
* Features are correlated
* Visualise high-dimensional data
* Speed up models
* Reduce noise
* Help clustering or classification
## Footnote
PCA is beneficial in scenarios where data complexity needs to be managed.
83
When should you **avoid PCA**?
* Must explain features to stakeholders
* Feature meaning matters more than performance
## Footnote
In cases where interpretability is crucial, PCA may not be suitable.
84
What is the first step in the PCA process outlined in the notebook?
Load data
## Footnote
The dataset used is the breast cancer dataset with 30 numeric features.
85
What is the target variable in the breast cancer dataset?
* 0 = malignant
* 1 = benign
## Footnote
The target variable indicates the diagnosis of the cancer.
86
Why is it important to **clean and scale** the data before applying PCA?
PCA is distance-based; different scales can break PCA
## Footnote
Cleaning and scaling ensure that PCA functions correctly.
87
What is the result of cleaning and scaling the breast cancer dataset?
341 rows × 30 features → NumPy array
## Footnote
This prepares the data for PCA application.
88
What is the key result when testing PCA with 30 components?
* Component 0 → 43.7%
* Component 1 → 18.5%
* Component 2 → 10.3%
## Footnote
The first three components explain approximately 72% of the information.
89
What is the rule of thumb for deciding how many components to keep in PCA?
80–95% variance is usually enough
## Footnote
This guideline helps in determining the optimal number of components.
90
What is the final dimensionality reduction result of the breast cancer dataset using PCA?
30 features → 7 components
## Footnote
This reduction retains 91.37% variance.
91
What does the output of PCA look like after dimensionality reduction?
x_PCA shape: (341, 7)
## Footnote
Each row represents the same observation in a new coordinate system.
92
What is the **visualisation magic** achieved after applying PCA?
Much clearer separation with just 2 numbers containing 62% of the data’s info
## Footnote
PCA allows for better visual analysis of data.
93
What is the mental model to remember about PCA?
PCA = rotate + compress
## Footnote
Components represent directions of maximum variation, and explained variance indicates usefulness.
94
What are the amazing benefits of PCA?
* Reduces dimensionality
* Requires scaled numeric data
* Outputs components, not features
* Useful for visualisation, preprocessing, noise reduction, ML performance
## Footnote
PCA is a powerful tool in data analysis and machine learning.
95
In **NLP**, what is the key shift compared to normal ML?
Features = text
## Footnote
Text is not numbers; ML models only understand numbers.
96
What is the main goal of **NLP**?
Turning messy human language into clean numeric features
## Footnote
Everything else (pipelines, CV, GridSearch) stays familiar.
97
What type of classifier is being built in this notebook?
A spam classifier for SMS messages
## Footnote
Input: 'WIN a free ticket now!!!' Output: spam or ham.
98
What are the steps in the **NLP workflow**?
* Load data
* Split train/test
* Clean text
* Convert text → numbers
* Train models
* Compare performance
## Footnote
The magic happens in steps 3 & 4.
99
What are the columns in the **dataset**?
* message → feature
* label → target (spam/ham)
## Footnote
After sampling: Smaller dataset for faster learning.
100
In the **train/test split**, what do X_train and y_train represent?
* X_train → messages
* y_train → labels
## Footnote
Same logic as usual. No surprises here.
101
Why is **text cleaning** necessary in NLP?
Raw text is chaotic
## Footnote
Issues include uppercase/lowercase differences, punctuation noise, and inconsistent formatting.
102
What does the custom transformer **text_cleaning** do?
* Lowercases text
* Removes punctuation
## Footnote
Normalizes language so the model doesn’t panic.
103
What is the core NLP trick for converting text to numbers?
Feature extraction
## Footnote
This includes CountVectorizer and TF-IDF Transformer.
104
What does **CountVectorizer** do?
* Tokenises text into words
* Counts word frequency
* Removes English stop words
## Footnote
Example: 'I love free pizza' → ['love', 'free', 'pizza'] → [1, 1, 1].
105
What is the purpose of the **TF-IDF Transformer**?
Penalises common words and boosts rare, meaningful words
## Footnote
Makes spam words pop (free, win, cash).
106
Why does **TF-IDF** beat raw counts?
'the' ≠ useful; 'win' = very useful
## Footnote
It helps identify important words in messages.
107
What are the steps in the **full NLP pipeline**?
* Text cleaning
* Tokenisation (CountVectorizer)
* TF-IDF weighting
* Model
## Footnote
This pipeline allows passing raw text to get predictions without manual preprocessing.
108
Which models were tested in the notebook?
* SGDClassifier
* LinearSVC
## Footnote
These models are designed for high-dimensional sparse data and work well with TF-IDF.
109
Why are tree models not used in NLP?
Trees hate sparse text matrices
## Footnote
This makes them less effective for text classification.
110
What does **HyperparameterOptimizationSearch** do?
Builds multiple pipelines and runs GridSearchCV
## Footnote
Here, only default hyperparameters are tested for model comparison.
111
What was observed in the **results interpretation**?
* SGDClassifier performed best
* LinearSVC very close behind
## Footnote
Both are linear models that love TF-IDF features.
112
What is the core takeaway regarding **NLP success**?
NLP success depends more on feature extraction than fancy models
## Footnote
Key tools include CountVectorizer and TfidfTransformer.
Data Analytics
flashcards
Decks in class (27)
# Cards
-
Productivity & Growth Mindset82
-
What Is Data Analytics69
-
Data Cleansing26
-
Data Storage43
-
Data Transformation30
-
Data Management43
-
ETL & SAP31
-
Data Governance25
-
Data Modelling22
-
Business Requirements55
-
Project Coordination20
-
Analysing Problems & Developing Solutions46
-
TeamWork35
-
Stakeholder Management & Emotional Intelligence80
-
Exploratory Data Analytics Techniques31
-
Design Requirements67
-
Probability and Statistics129
-
Python- Libraries-Streamlit26
-
ML134
-
SQL11
-
Power BI30
-
Image & Text Analysis28
-
Neural-network Algorithms.67
-
ML Model Analysis96
-
ML OPs70
-
Implement AI Model112
-
Ethics113
