Transcript 24-wu
Movie Info Web Search & Classification Frankie Wu CS224N Final Project Spring 2008 Movie Info Search & ClassificationMotivation • Monetary Reward! Netflix Prize Contest • $50,000 Incremental Prizes (annual) • $1,000,000 Grand Prize • Goal: predict how users will rate movies based on how they have rated other movies and how other users have rated all movies • Only Movie Info Given: Title and Year • Assumption: users will rate similar movies similarly • What is similar? One Possibility: Cast and Crew • Why not just use IMDB or Amazon’s DVD database? • Whole system must be commercially usable by Netflix. • Even barred from using Netflix movie database (oddly). Movie Info Search & ClassificationGeneral Approach • Data Collection • Spider the web and collect web pages based on the movie title and year. • Hand annotate data to create training and test sets. • All new code. • Classification • Maximum Entropy Markov Model (MEMM) classifier to learn relative weights of hand-designed features on training set. • Viterbi decoder to find optimal label sequences on test set (and eventually “real” unannotated data). • Code starting point: CS224N PA3. Movie Info Search & ClassificationData Collection • Yahoo! Web Search API to search web • • • • • Java program harness 100 movies (first 100 of 17700 Netflix list) 50 web pages per movie (or fewer if unavailable) Save HTML files locally Replace with own web crawler in production system • Data Annotation • Hand build information files for the 100 movies • ACTOR, DIRECTOR, SCREENWRITER, PRODUCER, COMPOSER • Programmatically annotate the 5000 movie web pages (imperfect) Movie Info Search & ClassificationClassification: Breadth vs. Depth • Initially wanted to use 80x50 files for the training set and 20x50 files for the test set. • Too much training data—computationally impractical. • Which is the better compromise? • Breadth: 80 movies x 10 files = 800 • Depth: 10 movies x 50 files = 500 • Speed: Depth faster than Breadth, 5m to 8m (expected) • Accuracy: Depth F-measure ~3x better than Breadth (surprising?) Movie Info Search & ClassificationClassification: Features • Features Hand Built • • • • • Word and Previous Label (a la PA3) Bigrams and Trigrams Name-Shaped Words (initial caps) Name-Shaped Bigrams and Trigrams Nearby strings: star, act, direct, produc, compos • Individual Feature Contribution • Determined by turning off features one at a time • Best and worst features? Still being determined at the time of this writing. Movie Info Search & ClassificationResults • Best results at the time of this writing: • • • • ACTOR: precision: recall: f-measure: 60.0% (161/268) 2.5% (161/6476) 4.8% • In general, disappointing result. • Highly skewed toward better precision than recall. • Likely due to extreme variance in data format— virtually free form.