59-minute beginner-friendly tutorial on text classification in WEKA; all text changes to numbers and categories after 1-2, so 3-5 relate to many other data analysis (not specifically text classification) using WEKA. 5 main sections: 0:00 Introduction (5 minutes) 5:06 TextToDirectoryLoader (3 minutes) 8:12 StringToWordVector (19 minutes) 27:37 AttributeSelect (10 minutes) 37:37 Cost Sensitivity and Class Imbalance (8 minutes) 45:45 Classifiers (14 minutes) 59:07 Conclusion (20 seconds) Some notable sub-sections: - Section 1 - 5:49 TextDirectoryLoader Command (1 minute) - Section 2 - 6:44 ARFF File Syntax (1 minute 30 seconds) 8:10 Vectorizing Documents (2 minutes) 10:15 WordsToKeep setting/Word Presence (1 minute 10 seconds) 11:26 OutputWordCount setting/Word Frequency (25 seconds) 11:51 DoNotOperateOnAPerClassBasis setting (40 seconds) 12:34 IDFTransform and TFTransform settings/TF-IDF score (1 minute 30 seconds) 14:09 NormalizeDocLength setting (1 minute 17 seconds) 15:46 Stemmer setting/Lemmatization (1 minute 10 seconds) 16:56 Stopwords setting/Custom Stopwords File (1 minute 54 seconds) 18:50 Tokenizer setting/NGram Tokenizer/Bigrams/Trigrams/Alphabetical Tokenizer (2 minutes 35 seconds) 21:25 MinTermFreq setting (20 seconds) 21:45 PeriodicPruning setting (40 seconds) 22:25 AttributeNamePrefix setting (16 seconds) 22:42 LowerCaseTokens setting (1 minute 2 seconds) 23:45 AttributeIndices setting (2 minutes 4 seconds) - Section 3 - 28:07 AttributeSelect for reducing dataset to improve classifier performance/InfoGainEval evaluator/Ranker search (7 minutes) - Section 4 - 38:32 CostSensitiveClassifer/Adding cost effectiveness to base classifier (2 minutes 20 seconds) 42:17 Resample filter/Example of undersampling majority class (1 minute 10 seconds) 43:27 SMOTE filter/Example of oversampling the minority class (1 minute) - Section 5 - 45:34 Training vs. Testing Datasets (1 minute 32 seconds) 47:07 Naive Bayes Classifier (1 minute 57 seconds) 49:04 Multinomial Naive Bayes Classifier (10 seconds) 49:33 K Nearest Neighbor Classifier (1 minute 34 seconds) 51:17 J48 (Decision Tree) Classifier (2 minutes 32 seconds) 53:50 Random Forest Classifier (1 minute 39 seconds) 55:55 SMO (Support Vector Machine) Classifier (1 minute 38 seconds) 57:35 Supervised vs Semi-Supervised vs Unsupervised Learning/Clustering (1 minute 20 seconds) Classifiers introduces you to six (but not all) of WEKA's popular classifiers for text mining; 1) Naive Bayes, 2) Multinomial Naive Bayes, 3) K Nearest Neighbor, 4) J48, 5) Random Forest and 6) SMO. Each StringToWordVector setting is shown, e.g. tokenizer, outputWordCounts, normalizeDocLength, TF-IDF, stopwords, stemmer, etc. These are ways of representing documents as document vectors. Automatically converting 2,000 text files (plain text documents) into an ARFF file with TextDirectoryLoader is shown. Additionally shown is AttributeSelect which is a way of improving classifier performance by reducing the dataset. Cost-Sensitive Classifier is shown which is a way of assigning weights to different types of guesses. Resample and SMOTE are shown as ways of undersampling the majority class and oversampling the majority class. Introductory tips are shared throughout, e.g. distinguishing supervised learning (which is most of data mining) from semi-supervised and unsupervised learning, making identically-formatted training and testing datasets, how to easily subset outliers with the Visualize tab and more... Update March 24, 2014: Some people asked where to download the movie review data. It is named Polarity_Dataset_v2.0 and shared on Bo Pang's Cornell Ph.D. student page http://www.cs.cornell.edu/People/pabo/movie-review-data/ (Bo Pang is now a Senior Research Scientist at Google)

• published: 28 May 2013
• views: 40061