Exam 2 Data Analytics

use cases- cluster similar documents

cluster news articles

use cases- geospatial clustering

cluster people by space and time i.e. groups of states interested in watching the same shows

when the event is equally likely to occur and not occur, the odds is

exactly 1

parameters for random forest

Number of decision trees to train in parallel: -higher trees give better performance but takes longer -must make trade-off based on computing power and performance Number of variabels to choose for splitting at each node: -must be less than total number of IVs -recommend and default value is sqrt(M) -example: if there are 100 IVs, then m=10 variables are randomly chosen at each node for splitting

odds vs probability

Odds: # of ways event can occur / # of ways event cannot occur. Probability: # of ways event can occur / total # of outcomes. example: you have a box filled with 5 red balls, 3 blue balls, 2 yellow balls (10 balls total) -the probability of taking a blue ball is 3/10 -odds of taking a blue ball is 3/7

steps in supervised predictive modeling

*1. Obtain data:* -data is info about the problem that you're working on -includes both inputs and outputs *2. Data pre-processing* -use appropriate data type for each column -code categorial variables -handle missing values -feature selection (select best features that are most relevant to your prediction) *3. Split processed data into training and testing data into training and testing data* -randomly split into training and testing -generally proportion of training data is greater than the testing data example: 75% training and 25% testing *4. Learn from training dataset* -use the known inputs and outputs in the training data to train a model to uncover the relationship between them *5. Make predictions using testing dataset* -use ONLY the inputs in the testing dataset to predict the output *6. Evaluate machine learning model* -compare the actual output and the predicted output -use performance measure to evaluate model -measures depend based on model task --for classification: accuracy, AUC, sensitivity, specificity --for regression: MSE, MAD, etc.

initialization: choose the number of clusters

*Empirical method*: - # of clusters: k=sqrt of N/2 for a dataset of N points - i.e. if n=200, then k=10 Other methods: -Elbow method: keep increasing the cluster until your decrease in error is exponential (once it starts becoming linear)

evaluating k-means cluster

*External*: employ criteria not inherent to the dataset -compare a clustering against priot or expert-specified knowledge using certain clustering quality measure *Internal*: unsupervised, criteria derived from data itself -evaluate the goodness of a clustering by considerign how well the clusters are seperated, and how compact the clusters are, -SSE, Modularity, Silhouette coefficient

maximum likelihood estimator

- Given the input variables, max likelihood estimator chooses coefficients (slope and intercept) to max the probability of the observed pattern of classes - To max likelihood, we must choose slope and intercept to max the following equation Ln(k1) + Ln (k2) +....Ln (kn) where ki is the likelihood of sample i being in a class *you find the slope and intercept using maximum likelihood estimator* NOT sum of square errors

when do you stop an algorithm?

- a branch with entropy of 0 is leaf node - a branch with fewer than a certain sample -have already split on all variables

unsupervised learning association

-*rule based machine learning* method for discovering interesting relations between variables in large databases -rules do not extract an individuals preference, rather find relationships between set of elements of every distinct transaction applications: -market basket analysis: takes recipts from all customers from a store and analyze all patterns of buying -web intrustion detection

real life applications of clustering

-Neflix divides its 93 million users around the world into 1,300 "taste communities" -AI in dating apps -Southwest airlines uses big data to deliver excellent customer service

confusion matrix for more than 2 classes

-you calculate the sensitivity for each category -calculate precision for each category -all values are between 0 and 1 -higher values preferred for each above measures

random chosen variable subset at each node for best split

-a problem contains a total of M variables -to ensure randomization, for each node of a tree: 1. randomly select 'm' variables (m << M) for splitting 2. evaluate information gain for each of the selected 'm' variables 3. choose the variable with max information gain for splitting -repeat the above steps at each node for all the decision trees that you are training

what is clustering

-a way of grouping *similar* objects together based on data describing the object samples i.e. similarity is measured using some criteria -a form of *unsupervised learning* - deals with finding a structure in unlabled data i.e. data has input variables but does not include examples of expected outcome/output -method of *data exploration* - a way of looking for patterns or structure in the data that are of interest

mean squared error

-also measures dispersion of erros, but larger errors get penalized more due to squarring -computed as average of square of errors of all time periods MSE= square all errors and add up divide by total number of errors

advantages of using random forest

-avoids overfitting -less variance compared to a single decision tree -provides good prediciton accuracy -useful for learning complex non-linear relationships

converting proabability to class label

-class label refers to output class or levels of categories in your DV (typically 2 categories) -output of classification models (logistic regression) is probabilities for each output class example: if our output class is presence or absence of diabetes, then logistic regression will give: 1. probability of presence of diabetes= p 2. probability of absence of diabetes= q -we choose a threshold between 0 and 1 to convert probability to class label example: if we set threshold for presence of diabetes as 0.7 then when p >= 0.7, we classify that individual as diabetic. Non diabetic when p is < 0.7

use cases- recommendation engines

-cluster similar movies and tv shows together

disadvantages of using random forest

-complex and time consuming to train model

decision trees for prediction

-decision trees can be used for prediciting continuous output - instead of using info gain criterion, we will aim to minimize the SSE at each split

unsupervised learning clustering

-determine the internal grouping in a set of unlabled data -help profile attributes of different groups -evaluation --external: experts review the results --internal: measure within group similarity (this is what we use in this class)

advantages to decision trees

-easy to understand -non parametric method-no assumption about the space distribution and classifier structure

strengths of k-means

-fast and computationally efficient -simple and easy to implement -easy to interpret -measurable and efficient in large data collection

choosing machine learning algorithm

-interpretability -performance -computational requirement

common clustering algorithms

-k-means (will be covered) -hierarchial (will NOT be covered) -density based

mean absolute deviation

-measures dispersion of prediction errors -computed as average of absolute value of errors of all time periods -error is predicted value - actual value MAD= average of absolute values of errors

objective function of k-means

-minimize total intra-cluster variance (or squared error function) -centroid of a cluster: average of all training data points in that cluster

getting odds ratio from logistic regression coefficients

-odds ratio can be calculted from the coefficient values (Bi) in the logistic regression -odds ratio for variable xi is e^Bi -in other words, the exponent of the coefficient of a variable will give the odds ratio for that variable -if coefficient of variable x is 0.25, then the odds ratio is e^0.25 = 1.2840

k-means clustering

-partitional clustering approach -each cluster is associated with a *centroid* (center point) --a centroid is the mean of data points in a cluster --does not have to be one of the original data points -each point is assigned to the cluster with the closest centroid --closeness is measured using a distance metric -number of clusters, K, must be specified

real world applications of using random forest

-predicting and detecting Alzheimer's disease at hospitals -predict whether a patient will miss the appointment -predict economic impact of adverse weather events -fraud detection by banks -click-through rate by online retailers

disadvantages to decision trees

-prone to overfitting: model learns the noise along with the pattern

limitations of k-means

-sensitive to initial condition. Different initial condition ma produce different results of cluster. The algorithm may be trapped in the local optimum. -sensitive to outliers -K-means has problems when clusters are of differing: sizes, densities, and non-globular shapes

random forests

-supervised algorithm machine learning for classification and regression -combines output from multiple decision trees to make predictions -overcomes the drawback of overfitting in decision trees

supervised learning prediction

-task of approximating a mapping function (f) from input variables (x) to *continuous output variables (y)* -continuous output variable is a real-value, such as an integer or floating point value -input variables can be real valued or discrete -regression problem requires the prediction of a quantity

supervised learning classification

-task of approximating a mapping function (f) from input variables (x) to *discrete output variables (y)* i.e. discrete outputs: MU can win or lose, colors like red or blue, patient has disease or no disease -requires the output to be classified into two or more classes -output variables often called labels or categories -problems with 2 classes is caled *binary classification problem*, and problem with more than 2 classes is called a *multi-class classification problem* -common for classification models to predict a continuous value as the probability of a given example belonging to each output class i.e. a patient may be assigned the probablities of 0.1 as being diabetic and 0.9 as being not diabetic. we can categorize it as not diabetic due to a high likelihood

choosing the right attribute for classification

1. compute the information gain of splitting the dataset by each of the predictors (IVs) in the dataset 2. choose the attribute with the LARGEST information gain 3. divide the dataset by it's branches to construct a decision tree, these steps are repeated until you reach stopping criteria

important properties of probability

1. probability of getting heads when you toss a coin? 0.5 or 1/2 2. what is the probability of getting 2 when you roll a dice? 1/6 3. probability of getting 7 when you roll a dice? 0 4. probability of getting heads or tails when you toss a coin? 1 5. probability of getting 1, 2, 3, 4, 5, or 6 when you roll a die? 1 *Property 1: Probability ranges from 0 to 1* *Property 2: Sum of probabilities of all possible events equals 1*

k-means clustering algorithm steps

1. select K points as initial centroids 2. repeat 3. form K clusters by assigning all points to the closest centroid. 4. recompute the centroid of each cluster. 5. until the centroids don't change

K-means example

Pizza outlet location: -New pizza chain wants to establish 3 outlets in como -Based on market research, the company has identified potential customer locations -Task: What would be the best locationcof these 3 outlets so that they can serve their customers effectively? **input/sample dataset=potential customer locations value of k=3 Step 1: randomly locate 3 clusters Step 2: form 3 clusters by assigning customer locations to the nearest outlet (and also compute the distance between each customer to each of the 3 chosen locations to assign them to the location with the shortest distance to them) Step 3: Change cluster center (outlet) to the centroid (average) of it's assigned customer locations, and then repeat step 2. Some customers will change the cluster location. Repeat steps 2 and 3 until stopping criterion is reached. Stopping criterion: No change in clusters

decision tree terminology

ROOT node: top most decision node which corresponds to the best predictor. Most important input variable. SPLITTING: process of dividing a node into two or more sub-nodes Decision node: when a sub-node splits into further sub-nodes Leaf/terminal node: nodes do not split at all (typically this is the prediction of the outcome)

ensemble methods

Use multiple algorithms to obtain better predictive performance than could be obtained from any of the algorithms by itself -in many real life problems, a single decision tree may not perform well (but it's very fast) so use multiple trees to create an ensemble method. -all algorithms must learn different info 1. Boostrapping 2. random chosen variable subset at each node for best split

use cases - image segmentation

break the image into meaningful or perceptually similar regions i.e. self driving cars

why is logistic regression necessary

cannot use simple or multiple regression to predict binary dependent variables simple regression line assumptions are violated (not normally distributed), also leads to some negative values which are not accurate, and errors do not have constant variance (violates line assumptions again)

Machine Learning

a subset of AI that uses algorithms to learn from and make predictions about data without being explicitly programmed -One domain -allow computers to evolve behaviors based on empirical data -it is necessary for the computers to acquire knowledge just like intelligence requires knowledge

predictive analytics

analysis of *historical* info (and external data) to find patterns and predict future outcomes -history tends to repeat itself examples: -Will MU win the next football match -Price of Google stock tomorrow? -Disease risk of an individual 10 years from now? -Patters are established using machine learning, a subset of artificial intelligence

artificial intelligence (AI)

any technique that enables computers to mirror human intelligence, using logic, if-then rules, decision trees, and machine learning (including deep integration) Examples: -Google home -Alexa -Self driving cars -Many domains Other subsets of AI: -computer vision -natural language processing

decision tree for predicion steps

at each node: 1. fit a simple regression model for each IV over all possibly binary splits 2. choose a variable that minimizes the SSE for further splitting 3. if stopping criterion is reach - STOP, or else go to step 1 stopping criteria -if argest decrease in SSE is less than some threshold - a branch with fewer than a certain sample - if all the points in the node have the same value for the IV

goal of clustering?

finding groups of objects such that the -data within a cluster will be similar to one another -data across clusters is different from one another intra-cluster distances are minimized (distance between groups within one cluster) inter-cluster distances are maximized (distance between two clusters)

when the event is more likely to occur, then odds is

greater than 1

types of clusters

hard clustering: -objects belong exclusively to one cluster i.e. k-means clustering soft clustering: -objects can belong to multuple clsuters -small degree of association i.e. fuzzy clustering

simple regression

has to be used when the output is continuous

if outcome is all continuous and is not a class or a category,

it is a prediction problem

when the event is less likely to occur, then odds is

less than 1

linear vs logistic model

linear: y= bo + b1x linear is straight line logistic: p= 1/ 1+e^(b0+b1x) logstic model is S shape

logistic regression model

ln[P/(1-P)] = βo + β1x1 +β2x2 ... logistic regression model= p= 1/ (1+ e ^ -(βo + β1x1 +β2x2 ...) p denotes the probability that a dependent variable is 1 relationship between binary DV on IV p is between 0 and 1 if p denotes the probability that a DV is 1, then q denotes the probability that a DV is 0 and is given: q = 1-p

logistic function

looks like a big S and will transform any value into the 0 to 1 range can have errors

logistic regression

many business problems/decisions deal with understanding the probability associated with certain behaviors or events these events are mostly dichotomous (binary) variable example: win or lose a game, die or survive a surgery, if a stock will go up or down the next day in this situation, the analyst must predict a binary dependent variable from a set of IVs one of the fundamental machine learning approach for classification of binary dependent variable is logistic regression

evaluating supervised regression algorithm

mean absolute errors: sum of absolute differences between predictions and actual values mean squared error: sum of squares of the differences between predictions and actual values

information gain

measures how well a given variable (attribute) separates the training examples according to the output class. Information gain uses the notion of entropy, commonly used in information theory. based on the decrease in entropy after a dataset is split on an attribute. Information gain= expected reduction of entropy Gain (S,X)= Entropy(S) -Entropy(S,X)

entropy of an input variable

measures the entropy of a variable with respect to the output class. Entropy (S,X)= probability of the class x the entropy of that class

entropy

measures the impurity of a collection of examples it depends on the distribution of the random variable p Entropy= -plog2p - qlog2q if S is the collection of training examples.. If sample is completely homogeneous, then Entropy is 0. (or if all outcomes are yes, with zero no's, entropy is 0) If sample is equally divided between yes and no's, entropy=1 (max value an entropy can have). If we have 9 yes, and 5 no's, the probability of yes would be 9/14 and probability of no would be 5/14. So the entropy= -9/14 log2 (9/14) - 5/14 log2 (5/14) = 0.94

unlike probability value

odds value varies from 0 to infinity

concept of odds in logistic regression

odds: probability of an event occuring divided by the probability of that event NOT occurring example: probability of having a disease is 0.25. What are the odds of having a disease? -probability of having a disease = 0.25 -probability of not having a disease = 0.75 -odds (having disease)= 0.25/0.75 = 0.33 interpretation: for every 0.33 individual who have a disease there is 1 who does not have a disease. OR for every 33 individuals who have a disease, there are 100 individuals who do NOT have the disease Probability of having the disease is 0.25, what are the odds of NOT having a disease? 0.75/0.25 =3 interpretation: for every 3 individuals who does NOT have a disease, there is 1 who has a disease

types of clustering methods

partitional clustering: -devision of objects into non-overlapping clusters such that each object is exactly in one cluster ex: k-means clustering hierarchial clustering: -set of nested cluster organized as a hierarchial tree (start with one big cluster and goes into smaller custers, or the other way around) -degree of association with each cluster is assigned ex: birch

why clustering?

pattern detection -group related documents for browsing -group genes and proteins that have similar functionality -group stocks with similar price fluctuations summariazation -reduce the size of large data sets

Logistic regression examples

prediction demographic behavior whether a person will or will not subscribe to a magazine predicting whether a person will or will not respond to a direct mail campaign predicting whether or not a cell phone customer will "churn" by end of year and switch to another carrier

supervised learning

problem of developing a model using historical data (both INPUTS and OUTPUTS) to make a prediction on new data, where we do not have the answer -described as the mathematical problem of approximating a mapping function (f) from input variables (x) to output variables (y) y~f(x) -called supervised learning because process of an algorithm learning from the training dataset can be thought of as a teacher supervising the learning process -*majority of practical machine learning uses supervised learning*

decision tree example

problem: predict whether a person cheats income tax input data: refund, material status, and taxable income output: cheat or no cheat

odds ratio

ratio of two odds measure of association between y and x example: -probability of male having disease is 0.6 -probability of female having disease is 0.2 solution: odds ratio (OR)= odds of male having disease/odds of female having disease Odds of male having disease= 0.6/0.4 = 1.5 Odds of female having disease= 0.2/0.8=0.25 odds ratio (OR) = 1.5/0.25=6 interpretation= for every 6 males who have the disease, there is 1 female who will get the disease

science behind netflix's popularity?

recommendations

bootsrapping

sampling with replacement from original dataset draw sample populations from the super population but replacing them each time before taking another sample

logistic regression for classification

set of IVs (ex: x1, x2, x3) have coefficients associated with the iVs (ex: b1, b2, b3) use them to predict an outcome variable which has two categories DV or outcome variable is categoric- it has categories

machine learning techniques

supervised learning: -develop predictive model based on both input and output data -you know input and ouput data to know patterns unsupervised learning -group and interpret data based only on input data -only given inputs

the odds ratio tells us

the relationship between an IV and a DV

if odds ratio is greater than 1

then the relationship between that IV and DV is positive

if odds ratio is less than 1

then the relationship between the IV and DV is negative

random forest for regression

to find final prediction=take average between all tree's outcomes

random forest for classification

to make final prediction, use the majority of the outcomes

decision trees

type of supervised algorithm that forms a tree structure that breaks up a dataset into smaller and smaller subsets while positioned top down key benefits: can use for classification as well as prediction problems. Also can provide a set of rules and insights. High interpretability.

confusion matrix

used to compute the performance measures for classification algorithms accuracy= sum of all correct predictions/total samples or accuracy = (true positive + true negative)/ (true positive +true negative +false positive +false negative) sensitivity= the proportion of acutual positives that have correctly identified as positives or sensitivity= true positive/ (true positive + false negative) precision= true positive/ (true positive + false positive) F1 score = 2 x (precision x recall/precision + recall) -recall is sensitivity value

unsupervised learning

where you only have the INPUT data (x) and no output variables -goal: to model the underlying structure or distribution in the dta in order to learn more about the data -these are called unsupervised learning because unlike supervised learning above there is no correct answers and no teacher -algorithms are left to their own devises to discover and present the interesting structure in the data -no right answer, just group them to establish a pattern