Practical Data Analysis in Python

Practical Data Analysis in Python Hilary Mason @hmason www.hilarymason.com hilary@path101.com

Data is ubiquitous. The ability and tools to use it are not.

(Focused) Data == Intelligence

Data Analysis on the Web Data items change rapidly. Data items are not independent. There’s a lot of semi-structured data around. There’s a LOT of data around. == Too many problems, few tools, and few experts.

Entity Disambiguation This is important.

Entity Disambiguation This is important. Company disambiguation is a very common problem – Are “Microsoft”, “Microsoft Corporation”, and “MS” the same company? This is a hard problem.

Classification Document classification. Image recognition. Topic recognition.

Recommendation Systems Product recommendations. Disease predictions. Behavior analysis.

IEEE Tag Clustering immunity ultrasound medical imaging medical devices thermoelectric devices fault-tolerant circuits low power devices

Python for Data Analysis import why_python_is_awesome Python is readable. Easy to transition from Matlab or R. Numerical computing support. Growing set of machine learning libraries.

Libraries NLTK (Natural Language Toolkit) – www.nltk.org mlpy (Machine Learning PY) – mlpy.fbk.eu numpy & scipy – scipy.org

An EC2 AMI provisioned with all of the toys you need: http://blog.infochimps.org/2009/02/06/start- hacking-machetec2-released/ MachetEC2

Supervised Classification Text Feature Extractor Trained Classifier Spam Not Spam Training Data Feature Extractor

Data: Tweets Hand-classified. For example, some spam: | don't disrespect me. I just wanted yall to get a head start so don't feel bad when I have more followers in two days. http://xyyx.eu/a1ha | | oh yay more new followers..hiii...if u want go to http://xyyx.eu/a1hb | | My friend made this new tool to get more twitter followers, http://xyyx.eu/a1ht | | Yes, Twitter is doing some Follower/Following count corrections. Get it back at: http://xyyx.eu/a1h8 | | man if i see one more person cry about losing followers!!! http://xyyx.eu/a1h4 |

Features def document_features(self, document): document_words = set(document) features = {} for word in self.word_features: features['contains(%s)' % word] = (word in document_words) return features Break tweets into lists of relevant words.

Naïve Bayesian Classifer P(A|B) = the conditional probability of A given B http://yudkowsky.net/rational/bayes http://blog.oscarbonilla.com/2009/05/visualizin g-bayes-theorem/ classifier = nltk.NaiveBayesClassifier.train(train_set)

Classifer Accuracy Use a hand-classified test set to see the accuracy of the classifier: nltk.classify.accuracy(classifier, test_set)

Feature Relevance contains(') = True not_s : spam = 53.6 : 1.4 contains(") = True not_s : spam = 32.2 : 1.1 contains(#) = True not_s : spam = 22.0 : 1.0 contains(!) = True not_s : spam = 10.8 : 1.0 contains(*) = True spam : not_s = 7.4 : 1.0 contains(=) = True not_s : spam = 5.5 : 1.0 contains(i) = False spam : not_s = 5.2 : 1.0 contains(?) = True not_s : spam = 2.4 : 1.0 contains(:) = True spam : not_s = 2.3 : 1.0 contains(&) = True not_s : spam = 1.8 : 1.0 contains(;) = True not_s : spam = 1.6 : 1.0 contains($) = True spam : not_s = 1.5 : 1.0 contains(u) = True spam : not_s = 1.5 : 1.0 contains(2.0) = False not_s : spam = 1.4 : 1.0 contains(saw) = False not_s : spam = 1.4 : 1.0 contains(noble) = False not_s : spam = 1.4 : 1.0 contains(sound) = False not_s : spam = 1.3 : 1.0 contains(approach) = False not_s : spam = 1.3 : 1.0 contains(finally) = False not_s : spam = 1.3 : 1.0 contains(more) = False spam : not_s = 1.3 : 1.0

Kitchen Sink wash, rinse, repeat

Results 90% accuracy on spam tweets – not bad! Other possibilities: categorization – what do you tweet about? human vs bot? which celebrity tweeter are you?

Practical Data Analysis in Python

More Related Content

Viewers also liked

Similar to Practical Data Analysis in Python

More from Hilary Mason

Recently uploaded

Practical Data Analysis in Python

Editor's Notes