1. The Machine Learning Landscape

TERM

• Training set : The examples that system uses to learn
• Training instance(sample) : EachU training example
• Accuracy : The particular performance measure
• Data Mining : Applying ML techiques to dig into large amounts of data can help discover patterns that were not immediately apparent
• Labels : training data you feed to the algorithm includes the desired solution

(예측하고자 하는 항목 =+= 정답지)

• Learning rate : how fast they should adapt to changing data

CHAPTER 1) The Machine Learning Landscape

: clarifying what ML is & why we may want to use it.

1. What?

2. Why Use?

• automatically learn → much shorter / easier to maintain / most likely more accurate

• ML can help humans learn

3. Types

• whether or not they are trained with human supervision

: supervised / unsupervised / semisupervised / Reinforcement Learning

• whether or not they can learn incrementally on the fly

: online / batch learning

• simply comparing new data points to known data points

vs detect patterns in the training data and build a predictive model

: instance-based / model-based Learning

A) Supervised / Unsupervised Learning

: classified according to the amount and type of supervision they get during training.

• Supervised Learning(지도 학습)

****

⇒ classification(분류)

⇒ regression(회귀) : predict a target numeric value given a set of features(=attribute)(predictors)

-Most important SL Algorithm

: k-Nearest Neighbors / Linear Regression / Logistic Regression / Support Vector Machines/Decision Trees and Random Forests / Neural networks

• Unsupervised Learning(비지도 학습)

-Most important SL Algorithm

• Clustering(K-Means/DBSCAN/Hierarchical Cluster Analysis(HCA))

: detect groups of similar features. / find connections without help

-hierarchical clustering Algorithm : subdivide each group into smaller groups

• Anomaly(변칙) detection and novelty detection(One-class SVM/Isolation Forest)

: : detecting unusual transactions to prevent fraud, catching manufactoring defects, automatically removing outliers from db.

+=+Novelty detection : see only normal data

• Visualization and Dimensionality Reduction

(Principal Component Analysis(PCA)/Kernel PCA/Locally-Linear Embedding(LLE),t-distributed Stochastic Neighbor Embedding(t-SNE))

: -Visualization Algorithm : output a 2D or 3D representation of data that can easily be plotted.

+=+Dimensionality reduction : simplify the data without losing too much information.

⇒ run much faster, less disk&memory space, perfom better

+=+Feature extraction : merge several correlated features into one.

• Asscoiation rule learning(Apriori/Eclat)

: dig into large amounts of data and discover interesting relations between attributes.

• Semisupervised Learning(반지도 학습)

• Reinforcement Learning(강화 학습)

: must learn by itself what is the best strategy = policy

cf) Alphago

B) Batch / Online Learning

: whether or not the system can learn incrementally(점진적으로) from a stream of incoming data

• Batch Learning

Offline Learning : First the system is trained, and hen it is launched into production and runs without learning anymore

⇒ if, ++ new data ⇒ train a new version of the system from scratch on the full dataset, then stop the old system and replace it with the new one

😓DIsadvantages😓

⇒ take a lot of time/computing resources/money ⇒ done offline

⇒ if amount of data is huge, may be IMPOSSIBLE

😀Advantages😀

⇒ simple and often works fine

• Online Learning

: mini-batches (mini batches 뭉탱이 sequentially)

-out-of-learning

😀Advantages😀

⇒ fast and cheap

⇒ change rapidly & autonomously

⇒ save a huge amount of space.

😓DIsadvantages😓

⇒ High Learning rate → rapidly adapt to new data, BUT quickly forget the old data

⇒ Low Learning rate → learn more slowly, BUT less sensitive to noise

⇒ if bad data is fed → clients will notice that performance will gradually decline.

C) Instance-Based / Model-Based Learning

: How they generalize

• Instance-based Learning

: certainly not the best.

KNN

• Model-based Learning

model selection → utility function=fitness function(measure how good my model is) / define cost function

→ study the data

→ select a model

→ trained it (the learning algorithm searched for the model parameter values that minimize a cost function)

→apply the model to make predictions on new cases(=inference), hoping that this model will generalize well.

4. Main Challenges of Machine Learning

: Bad Algorithm / Bad data

1) Insufficient Quantity(양) of Training Data

: Even for very simple problems, typically need thousands of examples.

2) Nonrepresentative Training Data

: training data be representative of the new cases you want to generalize to

⇒ if not , unlikey to make accurate predictions

sample is small → sampling noise

sample is large → sampling bias /ex) Roosvelt_Literary Digest

3) Poor-Quality Data

: will make it harder for the system to detect the underlying patterns → less likely to perform well.

→ 多 spend time cleaning up your training data.

4) Irrelevant Features

: “Garbage in, Garbage out”

-feature engineering : coming up with a good set of features to train on(enough relevant features & not too many irrelevant)

-feature selection : selection the most useful features to train on among existing features.

-Feature extraction : combing existing features to produce a more useful one

-creating new features by gathering new data

5) Overfitting the Training Data

: the model performs well on the training data, but it does not generalize well.

: it happens when the model is too complex relative to the amount and noises of the training data.

: find the right balance : fitting the training data perfectly & keeping the model simple enough to ensure that it will generalize well.

-To simplify the model by selecting one with fewer parameters

-To gather more training data

-To reduce the noise

• Regularization : Constraining a model to make it simpler and reduce the risk of overfitting.

cf) Training error(LOW) & Generalization error(HIGH) ⇒ OVERFITTING

6) Underfitting the Training Data

: occurs when the model is too simple to learn the underlying structure of the data.

-Selecting a more powerful model, with more parameters

-Feeding better features to the learning algorithm( feature engineering)

-Reducing the constraints on the model.