Flashcards
DataAnalyticsAIMachineLearningPart2_Amia22
Cards in this deck (57)
- 1What are classification methods used forUsed for predictions, classification and categorization ==>
Predicts probabilities that an outcome (dependent) variable belongs to a particular class ( independent variable) - 2why is a classfication mehtod called Logistic Regression?NOT a regression model --> output is categorical --> classfication model but called regression because the mathematical formula similar to regression models
- 3How does Logistic regression work?Uses simplistic logistic function to fit data to a sigmoid output with goal to maximise the maximum likelhood estimation
Output may be binary or multiple class - 4what are the limitations of logistic regression models?cannot handle large number of features, can't be used for situations where the relationship btw input and output variables is not constant, best for binary outcomes , at risk for overfitting when insufficient data.
- 5What is a Naive Bayes Classifier?uses bayes theorem to perform classification based on probability - assumes independence btw features (naive), assigns current instance probabilities for K-cluster of potential outcomes -predict binary, categorical or numerical data
- 6what type of model do you use with data that has a guassian distribution ?Gaussian Naive Bayes method
- 7What is a support Vector Machine model used for?determining the optimal boundary btw 2 classes in a multidimensional space i.e. 2 features, boundary = line
3 features , boundary is plane
>3 features - boundary = hyperplane - 8what is support vector?data point closest to the boundary ( hardest to classify)
- 9what does training a SVM do?optimize the boundary by maximizing the margin btw support vectors of different classes - the focus on the borderlines is unique i.e. outliers are automatically ignored
- 10what type of input are SVM's require?Numerical - even categorical variables need to be converted to numeric .
- 11what type of output does SVM have?binary or multiclass classification -- new instances are classified based on location with respect to the hyperplane
- 12what algorithms typical outperform SVM algorithms?Random Forest and Gradient boosting machines
- 13T/F K-nearest Neighbor = K-means ClusteringFALSE
- 14what type of model is K-nearest neighbor?It is an INSTANCE-BASED method, i.e. NO MODEL - knowledge is stored in the structure of the mapped data (i.e. training data not discarded)
- 15what are the hyperparameters of teh K -nearest neighbor algorithm?Distance Function
K - 16what does a K-nearest neighbor algorithm do?For each new instance X, the algorithm finds K training instances with closest distance to X and returns the majority result i.e. the class for classification problems or mean for regression problems
- 17What is the most common supervised classification model?Decision Tree - goal is to classify a set of training instances accurately
- 18how does a decision tree workeach node is a feature that is examined - if vlaue of feature is above threshold, follow one branch, otherwise the other,
Internal node = 1 feature - independent variable
Leaf node = Dependent variable outcome class - 19what is growing the tree ?training a decision tree model
- 20what are the two evaluation methods used for decision trees?Gini impurity
Entropy - 21what is a clear advantage of decision trees?each node is explainable as if-then- cutoffs
- 22what are the limitations of decision trees?susceptible to overfitting (high variance), small variations in data can cause branches to be created - need to limit nodes in a branch or prevent nodes from being added unless statistically significant increase in purity (pruning the decision tree)
- 23what are ensemble models?Groups (ensembles) of >1 model to improve performance
- 24what are the common uses of ensemble methods?used on output of high variance and low bias, group simple models may outperform single complex model, can be trained in parallel or in sequence, used for both regression and classification
- 25what are methods create diversity in parallel ensembles trained together?Bagging
Random subspaces - 26What is Bagging for ensemble methods?train each model against random subset of training data --> Bootstrap aggregating , reduces chance for overfitting, trains large number of relatively unconstrained models in parallel, combines all models together to smooth out predictions, bootstrap each model in the ensemble of models and base outcomes on total # of outcomes or aggregate probability of each prediction
- 27what is random subspaces for ensemble methods?train each model with entire training data set but use random subsets of features per model - good when # of features is high
- 28what is a common ensemble modelRandom Forrest Model
- 29How does a random forest work?ensemble of randomly selected decision trees that run in parallel, if they are all different, (uncorrelated) it runs better. Can be used for classification or regression and uses both baggin and random subspaces as metrics
- 30what are the 3 main hyperparameters of the random forest that need to be set before running the model?Node Size, # trees allowed, # features sampled (note that trees are usually kept short)
- 31what happens when Random Forest is used for Classification ?output of each model is aggregated by voting and each vote is equally weighted (i.e. not weighted)
Hard voting = most frequent class is selected
Soft Vote - average probability for each class calculated average probability select class with highest average probability - 32What is A sequential Ensemble method?each short decision tree is run sequentially with subsequent interations focusing on mistakes of the one before it and constructed using Boosting
- 33what is boosting?attempts to improve predictive flexibility of models by training a large # of constrained (weak) models and combine all constrained models into a strong learner model with "weighted" voting
- 34what are pros and cons of sequential ensemble methods?pro- no data preprocessing required, handles missing data, good performance, pwoerful, reduces high bias (inaccuracy)
con- may cause overfitting, requires intense computation - 35What is AdaBoost (adaptive Boosting)misclassified data for each algorithm has weight increased in order to increase the chance it is classified correctly in next tree - used for binary classification - quick and easy to use, works with large data, takes longer to train may be hard to interpret
- 36what is Gradient Boosting?operates similar to AdaBoost but model trains on the errors of the previously run model . The residual errors are calculated using a gradient descent algorithm to reduce the loss function
- 37what is a clustering method?unsupervised model to discover patterns /relationships in data
- 38what is Exclusive (hard) clustering?an instance can only belong to a single cluster i.e. K -means clustering
- 39what is Fuzzy(soft) clusteringAn instance can have more than 1 cluster
- 40what is Hierarchical clustering?used for bioinformatics and genetics --> genetic profiles , treatments and their outcomes - grouped on similarities and differences
- 41what is agglomerative clustering?more common , bottom up approach , each instance is a cluster and successively merged with most similar other cluster
EACH FEATURE of the Instance population = Dimension - 42what is divisive clustering ?top down approach, not common , entire data is cluster and divided into two clusters based no similarities and differences, cycle repeated until all instances are in separate clusters
- 43what is probabilistic clustering?instances are clustered based on probability that they belong to a particular distribution-
Guassian mixture model is used to determine which guassian distribution an instance belongs to - 44What is K-means Clusteringinstances are assigned to manually defined K - number of clusters based on similarity --> random assign to the clusters--> compute distances of all instances in the cluster to the centroid using a distance metric, --> move instances closest to teh centroid of the K cluster to that Kcluster --> recalculate the centroid position --> repeat until centroids stop moving location
- 45in K means clustering, what does a high k vs low k meanhigher K = more granularity
lower K = less granularity - 46what is the limitation of K means Clustering?if K is not set appropriately for the # of instances and population characteristics, you wont get proper clustering and values of some features (dimensions) may be correlated or redundant - leads to increase noise and no improved clustering
- 47What are association rules?Unsupervised method Rule-based method to find relationships btw independent variables - generate associations not causation
must be categorical data, convert numerical data by way of discretization - 48What is Market Basket Analysis?association rules algorithm that finds association rules for events i.e. customers who bought x also bought y where X and Y are independent
- 49What is an Apriori Algorithm ?uses HASH tree to count item sets navigating through data set in breadth first manner?
ie. for detecting adverse drug reactions - 50When are Dimensionality reduction methods used?in the pre-processing stage rather than as a model on its own
- 51How does a Dimensionality reduction method work and what is purpose?Methods rank the importance of features (dimensions) with goal ot reduce the number of features to the most important ones - deals with redundancy which can cause problems with data
- 52what is principle component analysis?used to identify unimportant features or select first several dimensions (highest ranked) features selection only
- 53How does PCA (principle component analysis work?)principle component = axis through the data that is a function of contribution of the variability in a population - aligns with maximum variance in a population and each principle component has to be orthogonal (at right angleS) with all other principle components > than 3 can not be visualized 1Principel component per dimension
- 54what is the input for PCA ?requires continuous metric dimensions (features) but variations can use categorical variables
- 55what is singular vlaue decomposition (SVD)?A = USVT where
A = Matrix
Uand V are orthogonal matrices
S is the diagonal matrix - 56What is an autoencoder ?uses neural network to compress data then recreate data as output
usually part of an ANN that uses unsupervised learning to reduce the data dimensions - 57—
Information
- Description
- informaticsBoards DAtaAnalytics Flashcards on DataAnalyticsAIMachineLearningPart2_Amia22, created by Michael Riben on 16/10/2022.
- Created by
- Michael Riben
- Views
- 0
- Rating