*Using K-Means to Cluster ISFL Offensive Players - S24 - Printable Version +- [DEV] ISFL Forums (http://dev.sim-football.com/forums) +-- Forum: Community (http://dev.sim-football.com/forums/forumdisplay.php?fid=5) +--- Forum: Media (http://dev.sim-football.com/forums/forumdisplay.php?fid=37) +---- Forum: Graded Statistical Analysis (http://dev.sim-football.com/forums/forumdisplay.php?fid=153) +---- Thread: *Using K-Means to Cluster ISFL Offensive Players - S24 (/showthread.php?tid=25715)

*Using K-Means to Cluster ISFL Offensive Players - S24 - wonderful_art - 09-17-2020 Using K-Means to Cluster ISFL Offensive Players – S24 Hey all, I’m wonderful_art. This place seems really, really cool, and as a stats / data science nerd, I instantly gravitated to the index and the statistics held there. With the season ended, I figured I’d try my hand at grabbing some stuff and seeing what I could do with it for fun. This project was using K-means clustering to group players into segments within their position. The result had some interesting takeaways vs just directly looking at the index and sorting by different stats which seemed like a pain. While you review the data below, remember that the cluster number doesn’t mean anything (its just referring to that node, not a ranking). So ‘cluster 3’ is not worse than ‘cluster 0’ … they are just different. That said, you’ll notice some clusters have objectively better stats than others, and I tried to explain them where possible. I used a Google Colab notebook and Python to do the analysis. For anyone else who is in / into data science, I used pandas’ read_html function which made it super easy to just pull this stuff directly from the html pages. Some normalization later (x - µ / σ) and I put it into scikit learn’s K-means algorithm which spit out the following clusters. If you want to check out the Google Colab notebook, here's the link. The visualizations are interactive under the Visualizations header. Apologies for the poorly commented code, as I mostly sped through it.  Next I’m planning on doing defenses and clustering them together. I’m also looking at pulling NFL players for creating the clusters based on NFL stats and then predicting the NFL cluster for ISFL players. Want to also do similar work for DSFL players. And I’m saving these for next season, where we can see how players move in clusters. Perhaps later on I'll deploy to heroku as well for more interactivity.  Here are the results, a graph and some interesting notes per group: Quarterbacks - 4 Clusters 3 - 5033.0 Yards - 53.9 Pct - 32.0 TDs - 19.0 Int - Banks II, D. - SAR 3 - 4152.0 Yards - 55.7 Pct - 23.0 TDs - 17.0 Int - Cue, J. - AZ 1 - 4225.0 Yards - 61.5 Pct - 21.0 TDs - 11.0 Int - Armstrong, F. - OCO 1 - 4192.0 Yards - 59.0 Pct - 22.0 TDs - 9.0 Int - Cole, E. - AUS 1 - 4002.0 Yards - 62.5 Pct - 31.0 TDs - 9.0 Int - McDummy, W. - COL 1 - 3974.0 Yards - 57.6 Pct - 20.0 TDs - 9.0 Int - O'Donnell, G. - CHI 2 - 3708.0 Yards - 58.5 Pct - 21.0 TDs - 8.0 Int - Jack, M. - SJS 2 - 3605.0 Yards - 56.0 Pct - 22.0 TDs - 8.0 Int - Slothlisberger ®, B. - NO 2 - 3077.0 Yards - 57.5 Pct - 19.0 TDs - 8.0 Int - Skywalker ®, L. - HON 4 - 3524.0 Yards - 57.5 Pct - 15.0 TDs - 13.0 Int - Phoenix, B. - PHI 0 - 3524.0 Yards - 53.0 Pct - 25.0 TDs - 16.0 Int - Jack ®, C. - YKW 0 - 3414.0 Yards - 51.7 Pct - 22.0 TDs - 17.0 Int - Fujiwara, C. – BAL Quarterback If I had to name the clusters, I think this is the trend I’m seeing: 3 – Risk it for the Biscuit, 1 – Efficient, 2 – Getting it Done, 4 – Pheonix, and 0 – Work in Progress As alluded above, Pheonix is in his own cluster. Too few TDs and INTs to be put in with Jack and Fujiwara. Despite having the second-highest yards, Armstrong is not in cluster 3 with Banks II and Cue. Most likely this is due to having significantly lower interceptions than the other two (and significantly fewer TDs than Banks). Great seasons by Slothlisberger and Skywalker to be in cluster 2 for rookie seasons (assuming that’s what the R stands for). Runningbacks - 4 Clusters 2 - 1233.0 Yards - 299.0 Att - 12.0 TDs - 4.1 Avg - Nakamura, T. - OCO 2 - 1161.0 Yards - 286.0 Att - 8.0 TDs - 4.1 Avg - Torenson, S. - PHI 2 - 1051.0 Yards - 273.0 Att - 11.0 TDs - 3.8 Avg - Lackson, J. - SJS 2 - 1048.0 Yards - 259.0 Att - 8.0 TDs - 4.0 Avg - Mendonca, M. - AUS 2 - 993.0 Yards - 242.0 Att - 10.0 TDs - 4.1 Avg - Ayers ®, N. - HON 2 - 910.0 Yards - 218.0 Att - 10.0 TDs - 4.2 Avg - Skyline, A. - YKW 2 - 881.0 Yards - 200.0 Att - 7.0 TDs - 4.4 Avg - Owens, A. – COL 3 - 1161.0 Yards - 253.0 Att - 3.0 TDs - 4.6 Avg - Toriki, M. - NO 0 - 967.0 Yards - 234.0 Att - 6.0 TDs - 4.1 Avg - Jones, K. - AUS 0 - 952.0 Yards - 264.0 Att - 7.0 TDs - 3.6 Avg - Williams, D. - BAL 0 - 907.0 Yards - 245.0 Att - 4.0 TDs - 3.7 Avg - Kane-Maika'i, K. - HON 0 - 872.0 Yards - 241.0 Att - 6.0 TDs - 3.6 Avg - Cardrissian, R. - SJS 0 - 808.0 Yards - 208.0 Att - 6.0 TDs - 3.9 Avg - Tirtawidjaja, J. - CHI 0 - 783.0 Yards - 201.0 Att - 6.0 TDs - 3.9 Avg - Gump, F. - NO 0 - 744.0 Yards - 192.0 Att - 6.0 TDs - 3.9 Avg - Fukushu ®, M. - BAL 0 - 697.0 Yards - 182.0 Att - 4.0 TDs - 3.8 Avg - Petrongolo ®, J. - AZ 0 - 694.0 Yards - 188.0 Att - 6.0 TDs - 3.7 Avg - Dotson, F. - PHI 0 - 574.0 Yards - 138.0 Att - 7.0 TDs - 4.2 Avg - Back, R. - SAR 0 - 468.0 Yards - 127.0 Att - 5.0 TDs - 3.7 Avg - Keppler ®, Z. - AZ 1 - 608.0 Yards - 137.0 Att - 2.0 TDs - 4.4 Avg - Gilbert, R. - COL 1 - 368.0 Yards - 98.0 Att - 3.0 TDs - 3.8 Avg - McZeal ®, K. - OCO 1 - 42.0 Yards - 9.0 Att - 0.0 TDs - 4.7 Avg - Skuff, R. – BAL Runningback I don’t really want to name the clusters, but I’ll attempt to here. 2 – Studs, 3 – Toriki, 0 – Dependable, 1 – Handcuffs Once again, a player is in their own cluster, this time Toriki. The New Orleans back has the yardage and average yardage for cluster 2, but fewer TDs than many of the runningbacks in cluster 0, so he’s been singled out as an edge-case. How Owens is clustered in group 2 and not group 0 is a mystery – he must’ve been right on the boundary between those two cluster nodes. Only his average (a mere 0.3 higher than Jones) separates him from the average cluster 0 back. What a rookie season for Ayers. Wide Receivers - 8 Clusters 7 - 1406.0 Yards - 75.0 No - 7.0 TDs - 18.7 Avg - Thomas-Fox, D. - SJS 7 - 1346.0 Yards - 84.0 No - 8.0 TDs - 16.0 Avg - Donut, S. - AZ 7 - 1316.0 Yards - 78.0 No - 9.0 TDs - 16.9 Avg - Gaines, N. - AUS 7 - 1312.0 Yards - 63.0 No - 8.0 TDs - 20.8 Avg - Witheblock, M. - SAR 7 - 1271.0 Yards - 67.0 No - 7.0 TDs - 19.0 Avg - Gordon, R. - SAR 4 - 1349.0 Yards - 88.0 No - 5.0 TDs - 15.3 Avg - O'Leary, S. - CHI 4 - 1208.0 Yards - 81.0 No - 6.0 TDs - 14.9 Avg - Vuxta, R. - PHI 4 - 1206.0 Yards - 77.0 No - 7.0 TDs - 15.7 Avg - Trunks, F. - OCO 4 - 1164.0 Yards - 76.0 No - 4.0 TDs - 15.3 Avg - Podolak, J. - HON 4 - 1145.0 Yards - 69.0 No - 5.0 TDs - 16.6 Avg - Barker, E. - NO 4 - 1034.0 Yards - 71.0 No - 5.0 TDs - 14.6 Avg - Jeeta, E. - AUS 4 - 998.0 Yards - 65.0 No - 4.0 TDs - 15.4 Avg - Passmann, T. - AZ 0 - 1103.0 Yards - 68.0 No - 7.0 TDs - 16.2 Avg - Sakura ®, K. - YKW 0 - 1080.0 Yards - 68.0 No - 13.0 TDs - 15.9 Avg - Lim, W. - COL 0 - 1014.0 Yards - 60.0 No - 8.0 TDs - 16.9 Avg - Kaepercolin, N. - BAL 0 - 866.0 Yards - 53.0 No - 8.0 TDs - 16.3 Avg - Quinn, A. - BAL 0 - 828.0 Yards - 52.0 No - 7.0 TDs - 15.9 Avg - Hood, T. - SJS 0 - 646.0 Yards - 41.0 No - 9.0 TDs - 15.8 Avg - Jenkins ®, L. - COL 3 - 831.0 Yards - 58.0 No - 3.0 TDs - 14.3 Avg - Kingston, J. - COL 3 - 777.0 Yards - 52.0 No - 4.0 TDs - 14.9 Avg - Rodriguez, B. - CHI 3 - 744.0 Yards - 52.0 No - 4.0 TDs - 14.3 Avg - Swift, N. - SJS 3 - 646.0 Yards - 41.0 No - 5.0 TDs - 15.8 Avg - Snyder ®, S. - NO 3 - 575.0 Yards - 36.0 No - 4.0 TDs - 16.0 Avg - Cash Jr. ®, S. - YKW 3 - 554.0 Yards - 40.0 No - 3.0 TDs - 13.8 Avg - Varga, N. - PHI 3 - 515.0 Yards - 33.0 No - 1.0 TDs - 15.6 Avg - Kross, C. - BAL 3 - 486.0 Yards - 29.0 No - 4.0 TDs - 16.8 Avg - Bodenhamer ®, B. - NO 6 - 711.0 Yards - 41.0 No - 4.0 TDs - 17.3 Avg - Panda, F. - PHI 5 - 649.0 Yards - 60.0 No - 6.0 TDs - 10.8 Avg - Alexander, W. - HON 5 - 620.0 Yards - 44.0 No - 6.0 TDs - 14.1 Avg - Despacito Jr. ®, G. - OCO 5 - 500.0 Yards - 42.0 No - 6.0 TDs - 11.9 Avg - Jackson, A. - YKW 5 - 479.0 Yards - 55.0 No - 4.0 TDs - 8.7 Avg - Clay, J. - HON 1 - 545.0 Yards - 45.0 No - 2.0 TDs - 12.1 Avg - Mongo, H. - OCO 1 - 431.0 Yards - 41.0 No - 2.0 TDs - 10.5 Avg - Videl-San ®, N. - AUS 1 - 399.0 Yards - 33.0 No - 2.0 TDs - 12.1 Avg - Bur'Berry, J. - AZ 1 - 356.0 Yards - 34.0 No - 4.0 TDs - 10.5 Avg - Scabbia ®, L. - CHI 1 - 331.0 Yards - 31.0 No - 4.0 TDs - 10.7 Avg - Lee, M. - YKW 1 - 330.0 Yards - 25.0 No - 2.0 TDs - 13.2 Avg - Tidwell, R. - AUS 1 - 130.0 Yards - 11.0 No - 2.0 TDs - 11.8 Avg - Oberwald, G. - CHI 2 - 57.0 Yards - 8.0 No - 0.0 TDs - 7.1 Avg - Z, D. - HON 2 - 17.0 Yards - 2.0 No - 0.0 TDs - 8.5 Avg - Z, D. - OCO 2 - 7.0 Yards - 1.0 No - 0.0 TDs - 7.0 Avg - Z, D. – AZ Wide Receiver Let’s give it a shot. 7 – studs, 4 – all yardage all day, 0 – TD hawks, 3 – dependable, 6 – Panda, 5 – 600 x 6 group (+ Clay), 1 – WR3, 2 – Z, D. I’m assuming Z, D. played for HON, OCO and AZ this past season, so they get cluster 2 all to themselves. Panda also has their own category, from the looks of it too high of an average to be placed into cluster 3 or cluster 5 (plus they didn’t have 600 yards x 6 TDs). As mentioned, I love cluster 5. 3/4 players have 600-something yards and 6 TDs. No wonder they’re clustered together. And then there’s Clay, also put into Group 5 since he has too many TDs and receptions for Group 1. Sakura and Jenkins had a great rookie season. Tight Ends - 4 Clusters 2 - 1064.0 Yards - 106.0 No - 5.0 TDs - 10.0 Avg - Evans, H. - AZ 2 - 716.0 Yards - 70.0 No - 5.0 TDs - 10.2 Avg - McCormick, A. - NO 1 - 582.0 Yards - 61.0 No - 1.0 TDs - 9.5 Avg - Blocksdale Jr., A. - PHI 1 - 506.0 Yards - 61.0 No - 3.0 TDs - 8.3 Avg - Gelbman, T. - CHI 1 - 462.0 Yards - 50.0 No - 2.0 TDs - 9.2 Avg - Lewandowski, J. - BAL 1 - 390.0 Yards - 44.0 No - 3.0 TDs - 8.9 Avg - Rechs ®, Z. - YKW 3 - 493.0 Yards - 64.0 No - 3.0 TDs - 7.7 Avg - Boyd, C. - SAR 3 - 407.0 Yards - 52.0 No - 3.0 TDs - 7.8 Avg - Sauce, E. - COL 0 - 416.0 Yards - 62.0 No - 0.0 TDs - 6.7 Avg - McDavid, L. - SJS 0 - 215.0 Yards - 36.0 No - 0.0 TDs - 6.0 Avg - Hayes, V. - CHI 0 - 211.0 Yards - 33.0 No - 0.0 TDs - 6.4 Avg - Arrabiata, D. - PHI 0 - 132.0 Yards - 17.0 No - 0.0 TDs - 7.8 Avg - Larson, P. - AUS Tight Ends Here we go: 2 – All stars, 1 – Slot TEs, 3 – Goal line TEs, 0 – No TDs TEs Although McCormick was placed with Evans (the 5 TDs and otherwise relatively similar score in yards [compared to the mean and std] most likely), Evans had a standout season (106 receptions!) There’s a subtle drop from the average yard per reception in cluster 1 to cluster 3 that is most likely the result of this cluster change. With very similar TD and avg lines, no wonder Boyd and Sauce were clustered together. (As mentioned, if you want to check out the visualizations in 3D interactivity, click here and scroll down till you see the chart you want. You might have to run cells, which you can RE: Using K-Means to Cluster ISFL Offensive Players - S24 - Pat - 09-17-2020 My god this article is a stat nerd's wet dream. Keep writing these.