MusicMood - Machine Learning in Automatic Music Mood Prediction Based on Song Lyrics

This project is about building a music recommendation system for users who want to listen to happysongs. Such a system can not only be used to brighten up one's mood on a rainy weekend; especially in hospitals, other medical clinics, or public locations such as restaurants, the MusicMood classifier could be used to spread positive mood among people.

Transcript

1. MusicMood Machine Learning in Automatic Music Mood Prediction Based on Song Lyrics Sebastian Raschka December 10, 2014
2. Music Mood Prediction • We like to listen to music [1][2] • Digital music libraries are growing • Recommendation system for happy music (clinics, restaurants ...) & genre selection [1] Thomas Schaefer, Peter Sedlmeier, Christine Sta ̈dtler, and David Huron. The psychological functions of music listening. Frontiers in psychology, 4, 2013. [2] Daniel Vaestfjaell. Emotion induction through music: A review of the musical mood induction procedure. Musicae Scientiae, 5(1 suppl):173–211, 2002. 

3. Predictive Modeling Unsupervised learning Supervised learning Regression Classification Clustering Reinforcement learning Ranking Hidden Markov models DBSCAN on a toy dataset Naive Bayes on Iris (after LDA)
4. Supervised Learning In a Nutshell
5. Feature Extraction Feature Selection Dimensionality Reduction Normalization Raw Data Collection Pre-processing Sampling Test Dataset Training Dataset Training Learning Algorithms Post-Processing Cross Validation Final Classification/ Regression Model New Data Pre-processing Refinement Prediction Split Supervised Learning Sebastian Raschka 2014 Missing Data Prediction-error Metrics Model Selection Hyperparameter optimization This work is licensed under a Creative Commons Attribution 4.0 International License. Supervised Learning - A Quick Overview
6. MusicMood - The Plan
7. http://labrosa.ee.columbia.edu/millionsong/ The Dataset
8. Lyrics available? Lyrics in English? Sampling 200 songs for validation 1000 songs for training http://lyrics.wikia.com/Lyrics_Wiki Python NLTK
9. Mood Labels Downloading mood labels from Last.fm Manual labeling based on lyrics and listening • Dark topic (killing, war, complaints about politics, ...) • Artist in sorrow (lost love, ...) sad if ...
11. Word Clouds happy: sad:
12. A Short Introduction to Naive Bayes Classification
13. Naive Bayes - Why? • Small sample size, can outperform the more powerful alternatives [1] • "Eager learner" (on-line learning vs. batch learning) • Fast for classification and re-training • Success in Spam Filtering [2] • High accuracy for predicting positive and negative classes in a sentiment analysis of Twitter data [3] [1] Pedro Domingos and Michael Pazzani. On the optimality of the simple bayesian classifier under zero-one loss. Machine learning, 29(2-3):103–130, 1997. [2] Mehran Sahami, Susan Dumais, David Heckerman, and Eric Horvitz. A bayesian approach to filtering junk e-mail. In Learning for Text Categorization: Papers from the 1998 workshop, volume 62, pages 98–105, 1998. [3] Alec Go, Richa Bhayani, and Lei Huang. Twitter sentiment classification using distant supervision. CS224N Project Report, Stanford, pages 1–12, 2009. 

14. Bayes Classifiers It’s All About Posterior Probabilities objective function: maximize the posterior probability
15. The Prior Probability Maximum Likelihood Estimate (MLE)
16. The Effect of Priors on the Decision Boundary
17. Class-Conditional Probability Maximum Likelihood Estimate (MLE) "chance of observing feature given that it belongs to class " "
18. Evidence just a normalization factor, can be omitted in decision rule:
19. Naive Bayes Models Gaussian Naive Bayes for continuous variables
20. Naive Bayes Models Multi-variate Bernoulli Naive Bayes for binary features
21. Naive Bayes Models Multinomial Naive Bayes
22. Naive Bayes and Text Classification
23. The Bag of Words Model Feature Vectors
24. Tokenization and N-grams “a swimmer likes swimming thus he swims”
25. Stemming and Lemmatization Porter Stemming Lemmatization
26. Stop Word Removal
27. Term and Frequency
28. Term Frequency - Inverse Document Frequency (Tf-idf)
29. • TP = true positive (happy predicted as happy) • FP = false positive (sad predicted as happy) • FN = false negative (happy predicted as sad) Grid Search and 10-fold Cross Validation to Optimize F1
30. K-Fold Cross Validation
31. 10-Fold Cross Validation After Grid Search (final model)
32. 10-fold Cross Validation (mean ROC) Multinomial vs Multi-variate Bernoulli Naive Bayes
33. 10-fold Cross Validation (mean ROC) Multinomial Naive Bayes & Hyperparameter Alpha
34. 10-fold Cross Validation (mean ROC) Multinomial Naive Bayes & Vocabulary Size
35. 10-fold Cross Validation (mean ROC) Multinomial Naive Bayes & Document Frequency Cut-off
36. 10-fold Cross Validation (mean ROC) Multinomial Naive Bayes & N-gram Sequence Length
37. Contingency Tables of the Final Model training test
38. Live Demo http://sebastianraschka.com/ Webapps/musicmood.html
39. Future Plans • Growing a list of mood labels (majority rule). • Performance comparisons of different machine learning algorithms. • Genre prediction and selection based on sound.
40. Thank you!