Recommended
More Related Content
Viewers also liked
Viewers also liked (11)
Similar to Mike_Yung
Similar to Mike_Yung (20)
Mike_Yung
- 4. Evaluating the Model Test Set Accuracy 88.1 87.8 88.0 Precision 62.8 61.9 57.7 Recall 25.4 21.6 37.1 Target Metric: Precision False Positive Rate True Positive Rate ROC Curve (Test)
- 5. Interpreting the Model Logistic Regression Model Variable eCoefficient Leader 2.26 Student Gov 1.69 Varsity Sport 1.58 Sports Captain 1.29 Award 1.22 Community Service 1.21 SAT Score 1.0003 SAT Times Taken 0.39 How to Interpret? If you aren’t already in a leadership position, taking one will more than double your odds of being admitted *Note: this is only a subset of all variables used in the model
- 7. What’s in an Essay? Family Music/Arts Culture Sports Personal Story Science Career mother, father, family, parent, sister, ... music, play, piano, perform, theater, ... culture, world, language, travel, chinese, ... team, game, coach, player, season, ... feel, think, friend, love, life, moment, ... research, science, technology, math,... work, success, career, educate, community,... Topic Distribution of a Sample Essay
- 8. A 2-D Representation of College Essays Cultural Focus Career Driven Scientific Focus Personality Driven Step Vectorize Text Topic Modeling Visualize in 2D Clustering Method TF-IDF NMF PCA K-Means *Each point represents a college/university
- 9. Cultural Focus Career Driven Scientific Focus Personality Driven *Each point represents a college/university A 2-D Representation of College Essays Bowdoin Skidmore Wellesley Amherst Middlebury CalTech MIT CMU Purdue Arizona State Michigan State Cal State LB SD State Step Vectorize Text Topic Modeling Visualize in 2D Clustering Method TF-IDF NMF PCA K-Means
- 10. A 2-D Representation of College Essays Cultural Focus Career Driven Scientific Focus Personality Driven *Each point represents a college/university Step Vectorize Text Topic Modeling Visualize in 2D Clustering Method TF-IDF NMF PCA K-Means STEM State Schools Liberal Arts Mixed Ivy League
- 11. Final Thoughts • Limitation of data – Only enough to model ‘top school’ admittance – With more data: • School-level model • Graduate school model • Explore deeper feature-engineering – Interaction effects (e.g. Varsity*Captain) – Deeper effects (e.g. Hispanic student leading an African-American society) • Refine topic modeling with LDA
- 14. Train Set Logistic Regression underfits Random Forest overfits Accuracy 92.4 91.2 94.2 Precision 93.1 91.2 90.1 Recall 45.7 37.2 63.3 Target Metric: Precision False Positive Rate True Positive Rate ROC Curve (Train) Appendix 2: ROC Curve for Train Set
- 15. Appendix 3: NMF Visualized Step Vectorize Text Topic Modeling Analyze Method TF-IDF NMF (Fit) NMF (Transform) Words ≈ x Topics Essays Essays Topics Words New Matrix (Topic vs. Essay) Semantic Reference (Word vs. Topic) Old Matrix (Word vs. Essay) Visualization of NMF
- 16. Step Vectorize Text Topic Modeling Analyze Method TF-IDF NMF (Fit) NMF (Transform) Topics Words Non-Negative Matrix Factorization Semantic Reference (Word vs. Topic) mother father family parent sister …. music play piano perform theater … culture world language travel american … team game coach player season … … research science computer technology math … work education career success community … Family Music/Arts Culture Sports Personal/Story Science Career Appendix 4: NMF Semantic Reference Table