Link roundup, week 31/2012

• Predicting the future, Part 1: What is predictive analytics?

  tags: datascience datamining facts

  • discover hidden patterns in data that the human expert may not see
  • Part 1 of a four part series on predictive analytics
  • descriptive analytics lets us know what  happened in the past, predictive analytics focuses on what will happen  next
  • business intelligence. It allows us to make decisions based  on statistics obtained from historical data
  • Expert knowledge is based on experience
  • data-driven knowledge, as its name suggests, is based  upon data
  • supervised because, during training,  data is presented to a predictive model with the input data and the  desired output or outcome
  • neural networks, support vector machines, and decision trees
  • only  presented with the input data
  • also use unsupervised learning
  • Clustering
  • historical  data in search of features that you could use
  • After we build a predictive model, we need to validate it
  • PMML (Predictive Model Markup Language) exists that allows  predictive models to easily move between different systems
  • IBM SPSS Statistics to build and validate a  predictive model
  • upload it into a scoring engine such as the Zementis ADAPA
  • a hundred or so records containing data for customers that  churned in the past may not be enough. If not enough data is used for  training, a model may not be able learn or worse, it may over fit.
  • garbage in, garbage out
  • Data by itself does not translate to predictive value. Good data does.
  • book by Duda,  Hart, and Stock entitled Pattern Classification
  • fraud detection
  • healthcare
  • recommend products and  services
  • GPS mobile device data to predict traffic
  • data from sensors is a clear way towards helping to ensure safety
  • Dr. Alex Guazzelli is the VP of Analytics at Zementis
  • co-authored the book PMML in Action

• Predicting the future, Part 2: Predictive modeling techniques

  tags: datascience datamining facts

  • we are accumulating data on an exponential scale
  • 90 percent of the data available today was created just in the past  two years
  • neural networks (NNs), clustering, support vector machines (SVMs), and  association rules
  • extract value from historical data obtained from people and  sensors
  • predict the risk of  customer churn or defection, in case of people data, or the risk of  machinery breakdown, in case of sensor data
  • compute  a score or risk by implementing a regression
  • Support vector machines (SVMs)
  • decision trees are the most commonly used predictive modeling  technique
  • association rules can be used  to discover that people who buy diapers and milk, also buy beer
  • give a NN too few hidden nodes, it may not learn the  mapping function between the input fields and the target. Too many nodes  and it will over fit
  • many models can be combined together in  what is called a model ensemble
  • scores from all models are computed and the  final prediction is determined by a voting mechanism or the  average
  • Black-box
  • techniques  that are not capable of explaining their reasoning
  • NNs and SVMs fall into this category
  • If it outputs a high risk of churn for a  particular customer, it will not be able to tell us why
  • technique that clearly pinpoints the reasons for its decisions.  Scorecards fit such a criteria very well
  • Based on regression models,  scorecards are a popular technique used by financial institutions to  assess risk
  • decision trees are easy to explain and understand
  • black-box modeling  techniques are hard to explain, the models themselves should not be
  • representing data pre-processing as well as predictive models  is now straightforward with PMML, the Predictive Model Markup Language
  • represented as a PMML file, a predictive  model can be moved right away from the scientist’s desktop, where it was  developed, to the operational environment, where it is put to work
  • predictive solutions are only now experiencing a boom in  all industries, due to the advent of: 1) big data derived from people and  sensors; 2) cost-efficient processing platforms such as Cloud- and  Hadoop-based; and 3) PMML
  • this data gives us the  potential to transform the world into a smarter world

• Predicting the future, Part 3: Create a predictive solution

  tags: datascience datamining facts

  • I will use IBM SPSS Statistics to  illustrate many of these phases
  • phases involved in the making of  a predictive solution, from data pre-processing all the way to its  operational deployment
  • structured data encompasses fields such as  customer age, gender, and number of purchases in the last month
  • Unstructured data may be represented by a comment the same  customer provided as feedback
  • all ages lower than  21 may be binned together into a student category
  • all ages  higher than 55 may be binned together into a retiree category
  • Category  worker may then be assigned to everyone 21 to 55 years old
  • if customer A is 25 years old
  • three distinct fields: student, worker, and  retiree, which would be mapped to 0, 1, and 0
  • continuous fields may need to be normalized
  • text mining to identify attrition cues in  comments
  • statistical packages allow  users to select an option to pre-process the data automatically
  • 1) Balance speed and accuracy; 2) Optimize for speed;  3) Optimize for accuracy; and 4) Customize analysis
  • neural networks (NNs) and support vector  machines
  • Decision Trees and Scorecards, are also capable  of explaining the reasoning
  • I will focus  in this article on NNs
  • split your input into factors and covariates.  Factors represent categorical input
  • Covariates represent continuous variables
  • one reserved for model training, and the other for testing
  • option for  Automatic architecture selection
  • you want the model to have a low  rate of false-positives (FP) and false-negatives (FN), which implies a  high rate of true-positives (TP) and true-negatives (TN)
  • cost associated with a FN is usually very different from the cost  associated with a FP
  • ROC (Receiver  Operating Characteristic) curve
  • graphical representation of the true positive rate  (sensitivity) versus the false positive rate (one minus specificity) for a  binary classifier as its discrimination threshold varies
  • area under the ROC curve
  • scores need to be translated into business decisions
  • benefit from two types of knowledge: expert  and data-driven

• Predicting the future, Part 4: Put a predictive solution to work

  tags: datascience datamining facts

  • Predictive Model Markup Language (PMML) has completely changed this  scenario
  • a predictive solution needs to successfully bridge  the gap between two very different worlds. I call these, planets  Predicta and Engira
  • Planet Predicta is populated by  data scientists with expertise in statistics, data mining, and language  skills such as Perl and Python. Planet Engira, on the other hand, is  populated by IT engineers with expertise in Java™, .NET, C, SQL
  • phases involved  in the making of a predictive solution, from data pre-preprocessing and  model building all the way to post-processing of model scores. PMML is  able to represent all these phases in a single file
  • include multiple models or a model ensemble
  • we can use it to unveil the secrecy and the black box  feeling
  • R project
  • At Zementis, we have created a PMML-based predictive analytics decision  management platform called ADAPA
  • ADAPA  lives on the operational side
  • Universal PMML  Plug-in (UPPI) for in-database scoring and for Hadoop

Posted from Diigo. The rest of my favorite links are here.