Скачать презентацию Netflix Prize and Heritage Health Prize Philip Chan Скачать презентацию Netflix Prize and Heritage Health Prize Philip Chan

e65befd9e81d0ec2e86625b3bc075004.ppt

  • Количество слайдов: 43

Netflix Prize and Heritage Health Prize Philip Chan Netflix Prize and Heritage Health Prize Philip Chan

Cash Prizes to Stimulate Research n Ansari X Prize for Private Spaceflight (2004) [$10 Cash Prizes to Stimulate Research n Ansari X Prize for Private Spaceflight (2004) [$10 M] n 100 km above earth twice within 2 weeks n DAPRA Grand Challenge (2005) [$2 M] n autonomous vehicle: 131 miles in 10 hours n Archon X Prize for Genomics (2006) [$10 M] n map 100 human genomes in 10 days

Cash Prizes to Stimulate Research n Netflix Prize (2006) [$1 M] n Recommend movies Cash Prizes to Stimulate Research n Netflix Prize (2006) [$1 M] n Recommend movies with 10% improvement n Heritage Health Prize (2011) [$3 M] n Days in hospital next year with 0. 4 error

Netflix Prize netflixprize. com Netflix Prize netflixprize. com

Netflix Prize n Task n Given customer ratings on some movies n Predict customer Netflix Prize n Task n Given customer ratings on some movies n Predict customer ratings on other movies n If John rates n “Mission Impossible” a 5 n “Over the Hedge” a 3, and n “Back to the Future” a 4, n how would he rate “Harry Porter”, … ? n Performance n Error rate (accuracy)

Cash Award n Grand Prize n $1 M n 10% improvement n by 2011 Cash Award n Grand Prize n $1 M n 10% improvement n by 2011 (in 5 years) n Progress Prize n $50 K per year n 1% improvement

Intellectual Property n Netflix has a non-exclusive license to the algorithm n Authors tell Intellectual Property n Netflix has a non-exclusive license to the algorithm n Authors tell the world what the algorithm is

Participation n 51 K contestants n 41 K teams n 186 countries Participation n 51 K contestants n 41 K teams n 186 countries

Leader Board n Started on Oct 2, 2006 n Improvement by the top algorithm Leader Board n Started on Oct 2, 2006 n Improvement by the top algorithm n after a week: ~0. 9% n after two weeks: ~4. 5% n after a month: ~5% n after a year: ~8. 4% n after two years: ~9. 4% n July 26, 2009 (less than 3 years): 10%

Winner n Bell. Kor’s Pragmatic Chaos n 7 members n Merger of 3 teams Winner n Bell. Kor’s Pragmatic Chaos n 7 members n Merger of 3 teams n Bell. Kor § AT&T Labs, USA & Yahoo! Research, Israel n Pragmatic. Theory § telecommunications, Canada n Big. Chaos § started a company, Austria n A combination of different algorithms

Runner-up n The Ensemble n ~30 members n “last-minute” merger n teams had 30 Runner-up n The Ensemble n ~30 members n “last-minute” merger n teams had 30 days to beat the first team that crossed the 10% threshold same accuracy n behind by 20 minutes! n

Heritage Health Prize heritagehealthprize. com Heritage Health Prize heritagehealthprize. com

Health Care n 71 M individuals admitted to US hospitals each year n Unnecessary Health Care n 71 M individuals admitted to US hospitals each year n Unnecessary admissions cost $30 B

Heritage Provider Network n Has a network of doctors in California n Can we Heritage Provider Network n Has a network of doctors in California n Can we identify earlier those most at risk and ensure they get the treatment they need? n Can we reduce unnecessary hospitalizations?

Heritage Health Prize n Launch n http: //www. youtube. com/watch? v=Gu. Z 8 nkpyg. Heritage Health Prize n Launch n http: //www. youtube. com/watch? v=Gu. Z 8 nkpyg. As n Given patient data n Predict how many days a patient will spend in a hospital in the next year n The prediction helps develop strategies to reduce emergencies and hence hospitalizations

Grand Prize n $3 M n At most 0. 4 in error (~0. 5 Grand Prize n $3 M n At most 0. 4 in error (~0. 5 day) n By Apr 4, 2013 [2 years] n $500 K Consolation Prize n not below 0. 4 error

Milestone Prizes n top 2 performers at each milestone n Aug 31, 2011 n Milestone Prizes n top 2 performers at each milestone n Aug 31, 2011 n $30 K, $20 K n Feb 13, 2012 n $50 K, $30 K n http: //www. youtube. com/watch? v=pkmk. Nn. Gyih. Y n Sep 4, 2012 n $60 K, $40 K

Performance of Algorithms n Prediction Error Rate (RMSLE) n where n real = log Performance of Algorithms n Prediction Error Rate (RMSLE) n where n real = log ( actual # of days + 1 ) n prediction = log ( predicted # of days + 1 ) n Prediction error threshold = 0. 4 (~0. 5 day)

Intellectual Property n Exclusive license to Sponsor n and participant’s own use n Algorithms Intellectual Property n Exclusive license to Sponsor n and participant’s own use n Algorithms not previously published n Use of data sets is for the competition only n written consent for other purposes

Data Sets n Training and validation data sets n For participants to design algorithms Data Sets n Training and validation data sets n For participants to design algorithms n Feedback data set n For calculating standings on Leaderboard n Scoring data set n For determining winners for prizes n http: //www. heritagehealthprize. com/c/hhp/Data

Data (in CSV format) n Members Data (113 K members) n Claims Data (2. Data (in CSV format) n Members Data (113 K members) n Claims Data (2. 7 M claims) n Drug Count Data (818 K prescriptions) n Lab Count Data (361 K labs) n Outcome Data (76 K in Y 2, 71 K in Y 3) n Target (71 K in Y 4 for prediction) n Total ~264 MB (including other files)

Members Data n Member. ID n Age. At. First. Claim n Sex Members Data n Member. ID n Age. At. First. Claim n Sex

Claims Data n Member. ID n Provider. ID n Vendor ID n PCP (Primary Claims Data n Member. ID n Provider. ID n Vendor ID n PCP (Primary care physician) ID n Year n Specialty (of physician/vendor? ) n Place. Svc (place of service) n office, outpatient hospital, inpatient hospital, … n Pay. Delay (between service and payment)

Claims Data [continued] n Length. Of. Stay (in hospital) n DSFS (days since first Claims Data [continued] n Length. Of. Stay (in hospital) n DSFS (days since first claim) n Primary. Condition. Group (diagnostic categories) n Charlson. Index (affect of diseases on illness) n Procedure. Group (intervention categories) n Sup. LOS (supplement to Length. Of. Stay) n 1 if Lenght. Of. Stay is NULL because of deidentificaiton

Drug Count Data n Member. ID n Year n DSFS (Days since first service) Drug Count Data n Member. ID n Year n DSFS (Days since first service) n Drug. Count (unique prescription drugs)

Lab Count Data n Member Id n Year n DSFS (Days since first service) Lab Count Data n Member Id n Year n DSFS (Days since first service) n Lab. Count (unique lab or pathology tests)

Outcome Data n Member. ID n Days. In. Hospital_Y 2 (claims in Y 1) Outcome Data n Member. ID n Days. In. Hospital_Y 2 (claims in Y 1) n ie, Predict Y 2 based on Y 1 n Days. In. Hospital_Y 3 (claims in Y 2) n Claimed. Truncated n 1 if members with “truncated” claims

Using Other Data? n Yes n Freely available to anyone (public source) n URL Using Other Data? n Yes n Freely available to anyone (public source) n URL needs to be published to the forum n Except for n demographic, socioeconomic or clinical information about the members

Naive Algorithms n For predicting the number of Days in Hospital in the next Naive Algorithms n For predicting the number of Days in Hospital in the next year n Posted as “benchmarks” on the Leaderboard

Always Predict 15 (max) n Everyone goes to the hospital for at least 15 Always Predict 15 (max) n Everyone goes to the hospital for at least 15 days

Always Predict 15 (max) n Everyone goes to the hospital for at least 15 Always Predict 15 (max) n Everyone goes to the hospital for at least 15 days n RMSLE = 2. 628062 n 550+% over threshold

Always Predict Zero n no one goes to the hospital Always Predict Zero n no one goes to the hospital

Always Predict Zero n no one goes to the hospital n RMSLE = 0. Always Predict Zero n no one goes to the hospital n RMSLE = 0. 522226 n 31% over threshold

Predict Random Values n between 0 and 15 Predict Random Values n between 0 and 15

Predict Random Values n between 0 and 15 n RMSLE = 0. 752297 n Predict Random Values n between 0 and 15 n RMSLE = 0. 752297 n 88% over threshold

Always Predict Average n Average ~= 0. 209179 Always Predict Average n Average ~= 0. 209179

Always Predict Average n Average ~= 0. 209179 n RMSLE = 0. 486459 n Always Predict Average n Average ~= 0. 209179 n RMSLE = 0. 486459 n 22% over threshold

Leader Board n Competition started on Apr 4, 2011 with partial data n All Leader Board n Competition started on Apr 4, 2011 with partial data n All data were released on June 4, 2011 n Sep 9, 2011

Leader Board n Competition started on Apr 4, 2011 with partial data n All Leader Board n Competition started on Apr 4, 2011 with partial data n All data were released on June 4, 2011 n Sep 9, 2011 n RMSLE: 0. 456384 n ~14. 1% over threshold n Aug 29, 2012 n RMSLE: 0. 450426 n ~12. 6% over threshold

Teams n Sep 9, 2011 n 914 teams n 6021 entries n Aug 29, Teams n Sep 9, 2011 n 914 teams n 6021 entries n Aug 29, 2012 n 1292 teams

Considerations n Accurate Prediction n algorithms n Efficiency n time n space Considerations n Accurate Prediction n algorithms n Efficiency n time n space

Teams n Form your own teams n www. heritagehealthprize. com n Join my team Teams n Form your own teams n www. heritagehealthprize. com n Join my team n CSE 4403 Independent Study n CSE 5801 Independent Research

THANK YOU www. heritagehealthprize. com THANK YOU www. heritagehealthprize. com