• Measure, Measure, Measure
• Garbage in, Garbage out
• Correlation is not Causation
• More Data Beats Cleverer Algorithms
• Algorithms that do better with more data are more interesting
• Independent Sources Of data add new signals
• Feature Engineering is the key to being a good data scientist
• How do machines and Human interplay in Big Data?
• Learn many models ‐ ensembles
• Outliers are always interesting..
MEASURE, MEASURE, MEASURE
• Have a Hypothesis
• Create a metric to determine if hypothesis is correct
• Build a solution that can be measured
If you can not measure it you can not improve it – Lord Kelvin
MORE DATA BEATS CLEVERER ALGORITHMS
• Adding IMDB data For Netflix prize
• Adding Protein Expression Data or Patient Data to Gene Expression Data
• Bag of Words Approach for Word Sense Disambiguation
WORD SENSE DISAMBIGUATION
• Sloping Land Alongside a river or a lake. It typically has thick vegetation growing..
• A financial institution that takes deposits from some customers and gives loans to others who require the
To disambiguate in typical sentences look for co‐occurrences of words with words in definition.
Unsupervised Learning. Bootstrap a model.
The pilot landed the plane on the Hudson River amongst several boats and an appreciative audience
cheered from the banks of the river.
He issued a check and took it to the bank so he could transfer money.
Can look for frequent co‐occurrences with each sense of the word (boats and check respectively) and build
a larger bag of words in which to disambiguate.
Can not expect arbitrarily complex models to be learned by the computer
CITYY 1 LAT. CITY 1 LNG. CITY 2 LAT. CITY 2 LNG. DRIVABLE?
123.24 46.71 121.33 47.34 Yes
123.24 56.91 121.33 55.23 Yes
123.24 46.71 121.33 55.34 No
123.24 46.71 130.99 47.34 No
DISTANCE (MI.) DRIVABLE?
OF HUMANS AND MACHINES
• Partnership is important
• Aha moment and the strategy comes from humans..
• Machines do the hard work of calculating fast and do not tire
• Maybe some day Machines will be able to do more than they are asked to do explicitly.. Today Explicit
Instructions are the norm..
ENSEMBLES ‐ OUTLIERS ARE NOT INTERESTING – FOR
• Learn many models from random subsets of training data
• Effect of outliers is reduced on a majority of the models
• Random Forests
OUTLIERS ARE ALWAYS INTERESTING FOR RANKING
• You have to be so good that they can not ignore you
• My personal thesis: Average in everything is boring. Be
outstanding in something.
• Outliers along some dimension always have interesting
information – whenever you are combining multiple
variables to come up with one global rank
• Job Interviews!
UNKNOWN UNKNOWNS – VERY INTERESTING TO A
BUSINESS – OUTLIERS