Mining Facebook for Feelings
0 likes615 views
The document discusses mining Facebook data to analyze feelings. It finds over 100,000 instances each of "happy" and "proud" feelings. Machine learning classifiers were able to accurately classify feelings into basic categories like joy, anger, fear with up to 87% accuracy. The classifiers also classified feelings in terms of arousal and valence, or positive and negative feelings. A test set of Facebook data was manually assessed to check if automated classification of feelings matched how human coders interpreted the expressed feelings, with agreement of around 80% between coders and the automated system.
Mining Facebook for Feelings
- 1. Mining Facebook for Feelings Daniel Hardt Copenhagen Business School 9 June 2015 Daniel Hardt Mining Facebook for Feelings
- 2. Data: Facebook Feelings Daniel Hardt Mining Facebook for Feelings
- 3. Arousal and Valence: Data Daniel Hardt Mining Facebook for Feelings
- 4. Five Basic Feelings: Data Animated Excited 155291 Pumped 2979 Surprised 752 Amused 14993 Joy Happy 114259 Wonderful 54691 Awesome 22351 Super 5794 Great 55180 Fantastic 3596 Delighted 805 Satisfied 1349 Content 628 Hopeful 21399 Angry Angry 12680 Pissed 3851 Annoyed 16839 Frustrated 1145 Disappointed 2534 Disgusted 1566 Fearful Worried 3274 Scared 2075 Anxious 1002 Shocked 1391 Confused 3904 Empowered Determined 29850 Confident 2341 Accomplished 6570 Proud 31363 Daniel Hardt Mining Facebook for Feelings
- 5. Classifier Basic Feelings (5-way classification) Classifier: MaxEnt Training Accuracy: .87 Testing Accuracy: .75 (10-fold validation) Arousal (2-way classification) Classifier: MaxEnt Training Accuracy: .99 Testing Accuracy: .80 (10-fold validation) Valence (2-way classification) Classifier: MaxEnt Training Accuracy: .99 Testing Accuracy: .83 (10-fold validation) Daniel Hardt Mining Facebook for Feelings
- 6. Two-D Classification: Valence and Arousal Daniel Hardt Mining Facebook for Feelings
- 7. Two-D Classification: Comparisons Daniel Hardt Mining Facebook for Feelings
- 8. Feeling Meter: Manual Assessment Test Set: 160 examples from different sources Manual Task: Order Feelings Expressed (1 is most expressed, 5 least; 0 not expressed at all) Results: Binary Decision – is feeling expressed or not? (Ignore examples where 1st coder notes no feelings expressed – leaves 92 examples) Agreement on Feelings Expressed 1st coder vs 2nd coder: 0.797385620915033 366 out of 459 in 92 cases 1st coder vs System: 0.734204793028322 337 out of 459 in 92 cases Daniel Hardt Mining Facebook for Feelings