here - MIT Sloan Sports Analytics Conference
Transcription
here - MIT Sloan Sports Analytics Conference
“Win at Home and Draw Away”: Automa5c Forma5on Analysis Highligh5ng the Differences in Home and Away Team Behaviors Alina Bialkowski, Patrick Lucey, Peter Carr, Yisong Yue and Iain Ma;hews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esearch, Pittsburgh Individual Analy?cs CVPR 2014 Submission #1090. CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE. eled the shot chart. In contr shots may happen within th appears to lead to a blurring two factors have coefficient than the short-range shot fac the fourth factor could be d basket. In our modeling appr spond to a near-term shootin our approach; it may be mor Figure 7: Top Row: Depict location. Bottom Row: Def See Figure 8 for the affiniti Sports Analy5cs Team A naly?cs #1090 CVPR ? 486 487 488 489 490 491 492 493 PG 494 PG 495 496 4-‐4-‐2 SF es of our prediction results. Top: Examples of ball owner prediction in soccer. Black trajectories indicate the past passing last four time steps, yellow circle shows the predicted ball owner while blue circle shows the ground truth. Bottom: SF C 497 498 499 C 500 501 502 4-‐2-‐3-‐1 SG SG PF PF Team Analy5cs in Con5nuous Sports Analysis is very difficult: • Continuous and lowscoring 1) What formation they are playing? • Time-varying 2) How they are playing? (sitting back or pressing) • Annotating tactics and playing style is subjective • Dynamic and complex environment What we want: How to Compare Two Plays? x = x = Play2 Play1 Sterling Sturridge Cou=nho Henderson Gerrard Cissoko Toure Skrtel Mignolet Play 1 Suarez Flanagan Sterling Sturridge Cou=nho Henderson Gerrard Cissoko 10 players have Toure Skrtel 3,628,800 permutations Mignolet Play 2 Suarez Flanagan x,y ball x,y ball x,y Mignolet x,y Mignolet x,y Cissoko x,y Cissoko x,y Toure x,y Toure x,y Flanagan x,y Flanagan x,y Gerrard -‐ x,y Gerrard x,y Coutinho x,y Coutinho x,y Henderson x,y Henderson x,y Sterling x,y Sterling x,y Sturridge x,y Sturridge x,y Suarez x,y Suarez 0 0 0 0 0 0 = 0 0 0 0 100 0 0 100 0 Subs5tu5ons and Role-‐Swaps x = x = Play2 Play1 Sterling Sturridge Cou=nho Henderson Gerrard Cissoko Toure Skrtel Mignolet Play 1 Suarez Sturridge Suarez Flanagan Cou=nho Allen x,y Allen Henderson Gerrard Cissoko Toure Skrtel Mignolet Play 2 Sterling Flanagan x,y ball x,y ball x,y Mignolet x,y Mignolet x,y Cissoko x,y Cissoko x,y Toure x,y Toure x,y Flanagan x,y Flanagan x,y Gerrard -‐ x,y Gerrard x,y Coutinho x,y Coutinho x,y Henderson x,y Henderson x,y Sterling x,y Suarez x,y Sturridge x,y Sturridge x,y Suarez x,y Sterling 0 0 0 0 0 0 = 0 0 0 80 100 0 60 Use Role Representa5on x = x = Play2 Play1 ST LW ST LW ACM RW ACM LCM LCM RCM LB RCM RB LCB GK Play 1 RCB RW LB RB LCB RCB GK Play 2 x,y ball x,y ball x,y Mignolet GK x,y Mignolet GK x,y Cissoko LB x,y Cissoko LB x,y Toure LCB x,y Toure LCB x,y Flanagan RB x,y Flanagan RB x,y Gerrard LCM -‐ x,y Gerrard LCM x,y Coutinho RCM x,y Coutinho RCM x,y Henderson ACM x,y Henderson ACM x,y Sterling LW x,y Suarez LW x,y Sturridge ST x,y Sturridge ST x,y Suarez RW x,y Sterling RW 0 0 0 0 0 0 = 0 0 0 0 100 0 0 100 0 Automa5cally Finding Forma5ons from Data • • Goal: Find Forma5ons Directly From Tracking Data (en$re season from Prozone) Solu5on: A forma=on is a set of roles of a team Need to find the set of roles which have minimum overlap • Applica5ons: • • Allows us to do global team behavior analysis -‐> Case Study: Home Advantage • We can then assign a player a role at each frame -‐ > retrieval EM Algorithm for Finding Roles Procedure: 1) Choose ini?al ordering of player posi?ons 2) Ini?alize set of roles: Find mean of player posi?ons 3) Assign a role to each player in each frame according to ini?al set of roles 4) Re-‐calculate the mean of updated roles 5) Calculate the change in means 6) If change > threshold, con?nue steps 3-‐5 un?l convergence Examples Iden5ty vs Role Match Summariza5on Case Study: Home Advantage Home Field Advantage Sport Soccer % won of home games Sport MLS 69.1% Basketball Serie A (Italy) League League % won of home games NCAA (college) 69.1% 67.0% NBA 62.7% Central America 65.2% WNBA 61.7% La Liga (Spain) 65.0% Cricket International cricket 60.1% South America 63.6% Hockey NHL 59.0% EPL (UK) 63.1% Football NCAA (college) 63.1% Europe 61.0% NFL 57.6% Asia/Africa 60.0% MLB 54.1% Baseball Moskowitz and Wertheim (2011), “Scorecas=ng” Case Study: Win at Home and Draw Away • • • Used an en?re season of Prozone data which had player tracking data for an en?re season (home and away matches) Captured player posi?ons at 10fps and also had ball event data Every ball event with ?me-‐ stamps and loca?on Home Away p-value 1.61 per game 1.10 per game p < 0.0001 Shots 15.42 per game 12.18per game p < 0.0001 Goals 1.57 per game 1.21 per game p < 0.0001 Passing 451 per game 436 per game p = 0.1483 Shooting Accuracy (On vs Off-Target) 41.7% 42.4% p = 0.5046 Points (3=win, 1=draw, 0=loss) Forward-‐Third Bias A B C D E −5 −5 xx1010 55 44 F G H I J Home 2214.1% 33 11 K L M N O Away 00 11.8% −1−1 p<0.0001 −2−2 P Q R S T −3−3 −4−4 −5−5 Forma5ons Across the Season Team A Home Games Team A Away Games Forma5ons Across the Season Home = Red Away = Blue Overall Team Movement Team A All Games -‐ Unnormalized Overall Team Posi5on When the team has possession ! When the opposi=on has possession Summary: Home Advantage • • • • • • Big discrepancy in outcomes between home and away performances More shots, more goals, same passing and shoo?ng proficiency Heavy possession bias in the forward third Teams tend to play the same forma?on At home they press higher up the field Can our analysis be used for other team sports? Analyzing Team Behaviors in Other Sports Measure Team Behavior in Basketball Using 2012-‐2013 STATS SportsVU NBA data, we analyzed team behavior for 3-‐point shots • Open shots correlated with more defensive movement and defensive role-‐swaps • • Significant for all 13 teams analyzed Po ste r !!! Take Home Points 1) Team Behaviors are difficult to measure due to problems of permuta?ons 2) Using roles to represent a forma?on nullifies this problem 3) We use forma?on analysis to inves?gate the home advantage • More possession forward third, press higher up the pitch, li;le difference in forma?on 4) Our approach can be used for analysis of other team sports and for tasks like retrieval