LE5 - Modeling I - Clustering / Association Rules

supervised

• goal is predict outcome for new data, input data is provided to the model along with the output in the supervised learning. you know up front the type of results to expect. the output is predictet by the supversied learning model

unsupervised

• goal is get insights from large volumes of new data. the ml itself deteermines what is different or intetesting from the dataset. hiden patterns in the data can be found using the unsupervised learning model

• unsupervised

• collect elements into segments with similar characteristics

• partitional algorithms

  • k means clustering, model based clustering

  • needs a fix k before

  • start with random partitioning

  • sensitive to initialization

  • fast and efficient

  • problems when clusters are different size and outliers

• hierarchical algorithms

  • bottom up, top down

  • no particulare numver of clusters

  • computational complex in time and space

  • pros and cons depend on method: sensitive to outliers, handling sizes

• density based algorithms

  • can handle clusters of different shapes and sizes

  • resistant to noice

  • dont work if density varies a lot and high dimensional data

• model based

  • bases on probabilistic model (like K means)

  • clustering by EM (Expectation Maximazation)

  • need number of cluster before start

• hierarchical

  • Building a Dendogram (trees)

    • agglomerative -> bottom up -> each obs. starts in its own cluster -> merged as one move up hierarchy

    • divisive -> top down -> all obs. start in one cluster -> splits moves down hierarchy

• detect outliers

• segment customers

• find like-minded users

• analyze social networks

• midicine and biology

• high bias

  • oversimplifying

  • underfitting

• variance

  • overcomplex

  • overfitting

when the model suffers from

• high bias

  • the avg response of the model is far from the true value and this is called underfitting

• high variance

  • this is usually result of its inability to generalize well beyong the training data and this is called overfitting

• Build a model that achieves a balance between bias and variance -> combined error at minimum

• confidence

  • measures of the quality of a given rule

  • Support(X∪Y)​ // Support(X)

    
    

• Support

  • tells us which proportion of transactions from a dataset include items from both LHS and RHS

  • Support(X)=Gesamtzahl der TransaktionenAnzahl der Transaktionen // die X enthalten​

    
    

• Confidence

  • expresses which proportion of transactions include items from LHS also include items from RHS

  • Lift = Support(X∪Y) // Support(x) * Support (y)

• Identify user behaviour by finding associations and correlations between different items in a basket

• LHS (lefthand side) -> RHS[support, confidence) (righthand side)

• supervised

• analyze historical data

• generate model to predict future

• e.g. decision trees, neural network, svm

1. select the most dicriminatory feature

2. split the entire set into subsets using the feature

3. recursively find the most significant feature for each subset

• neural networks

• bayesian networks

• decision trees

• support vector machines

• genetic algorithm

1. create root node and select splitting attribute

2. add branch to root node for each split candidate value und label

3. take following iterative steps

   1. classify data by applying split value

   2. if stopping points is reached, then create leaf node and label it -> otherwise build another subtree

• based on attributes of instances

• branch for each value

• e.g.: Cusomer buy tablet?

• Attribute value description

  • same attributes must describe each example and muste have fixed number of values

• predefined classes

  • samples attributes must already be defined

• discrete classes

  • classes must be sharply delineated

• sufficient examples

  • enough test cases are needed to distinguish valid patterns from chance occurences