*Using K-Means to Cluster ISFL Defensive Players (and interactive charts!) – S24

~~wonderful_art~~ · (This post was last modified: 10-02-2020, 01:54 PM by Tesla.)

Using K-Means to Cluster ISFL Defensive Players (and interactive charts!) – S24

Hey all, so in follow up to this article clustering offensive players in the ISFL using K-Means, I’ve completed the task and clustered ISFL defensive players with the same algorithm. It uses the exact same mechanism as the last one, and as before here's a link to the Google Colab notebook I used to run the code.

NEW – I also created a simple Dash by Plotly application hosted on Heroku for you to play with the 3D graphs in a more interactive manner. Check it out here and you can move the charts to investigate the position of your choice in more depth.

Next, I’ll look to add kickers and OL to the charts and produce the same results for DSFL, considering that’s where I’ll be playing next season anyway. I’ve also started work on some deep learning with dense neural networks to predict wins based on individual player statistics. The hope there is that the model will be able to produce expected-wins or some equivalent advanced metric. It’s a work in progress, but QBs are done and I’ll release that sometime next week. Other interesting ideas in the cooker is regarding data around drafts, as it seems that time is just starting for the league.

Since this is now all defensive positions and a lot more of them, lets get into it.

[Image: mnv8LS1.png]

Defensive End

2 - 103.0 Tck - 5.0 TFL - 2.0 Sck - 6.0 PD - Howyanow, W. - PHI
2 - 94.0 Tck - 4.0 TFL - 10.0 Sck - 5.0 PD - Jones, D. - OCO
5 - 63.0 Tck - 17.0 TFL - 8.0 Sck - 0.0 PD - Kearse, J. - HON
5 - 60.0 Tck - 18.0 TFL - 7.0 Sck - 0.0 PD - Blackstone, I. - COL
4 - 58.0 Tck - 15.0 TFL - 6.0 Sck - 0.0 PD - Egghands, T. - AUS
4 - 54.0 Tck - 21.0 TFL - 5.0 Sck - 0.0 PD - Hole, M. – BAL
4 - 53.0 Tck - 14.0 TFL - 7.0 Sck - 0.0 PD - Isbest, C. - AUS
4 - 50.0 Tck - 15.0 TFL - 5.0 Sck - 0.0 PD - Karpaasi, M. - SAR
4 - 46.0 Tck - 15.0 TFL - 4.0 Sck - 0.0 PD - Bass, L. - SJS
4 - 43.0 Tck - 19.0 TFL - 4.0 Sck - 0.0 PD - VonPebbles, M. - CHI
4 - 38.0 Tck - 18.0 TFL - 3.0 Sck - 0.0 PD - Eller ®, A. - BAL
4 - 31.0 Tck - 17.0 TFL - 3.0 Sck - 0.0 PD - Works, S. - CHI
1 - 56.0 Tck - 9.0 TFL - 10.0 Sck - 0.0 PD - Alexander, N. - YKW
1 - 54.0 Tck - 10.0 TFL - 6.0 Sck - 0.0 PD - Ryeu, S. - AZ
1 - 53.0 Tck - 10.0 TFL - 9.0 Sck - 0.0 PD - Tu'inukuafe, B. - NO
1 - 52.0 Tck - 6.0 TFL - 8.0 Sck - 0.0 PD - Ocewilder, S. - OCO
1 - 50.0 Tck - 5.0 TFL - 10.0 Sck - 0.0 PD - Noble Jr., L. - HON
1 - 46.0 Tck - 6.0 TFL - 7.0 Sck - 0.0 PD - Claxton, S. - AUS
1 - 44.0 Tck - 10.0 TFL - 6.0 Sck - 0.0 PD - Cho, J. - YKW
1 - 42.0 Tck - 10.0 TFL - 6.0 Sck - 0.0 PD - Goku, S. – SAR
1 - 26.0 Tck - 3.0 TFL - 7.0 Sck - 0.0 PD - de Pengu ®, A. - COL
0 - 40.0 Tck - 10.0 TFL - 3.0 Sck - 0.0 PD - Om, T. - PHI
0 - 38.0 Tck - 4.0 TFL - 0.0 Sck - 0.0 PD - Sanchez, I. - COL
0 - 38.0 Tck - 4.0 TFL - 1.0 Sck - 0.0 PD - Zbornak ®, D. – PHI
0 - 36.0 Tck - 11.0 TFL - 1.0 Sck - 0.0 PD - Montain ®, A. - HON
0 - 30.0 Tck - 11.0 TFL - 2.0 Sck - 0.0 PD - Chestnut III ®, W. - SJS
0 - 23.0 Tck - 0.0 TFL - 2.0 Sck - 0.0 PD - Smart, M. – SAR
3 - 38.0 Tck - 18.0 TFL - 5.0 Sck - 0.0 PD - Ulysses, B. - NO

Notes

To name the clusters: 2 – Pass Hawks (high PDs) , 5 – Pressure Kings (High TFL and Sacks), 4 – Edge Rushers (High TFL to mediocre Sacks), 1 – Blitz Kings (High sacks relative to over metrics), 0 – Scrimmagers (Low tackles, mediocre TFL, low Sacks), 3 – Ulysses
Ulysses is on a category to their own, with too few tackles to be in cluster 1, but too many tackles for loss and sacks for cluster 0. Not shown in the above, Ulysses also had 2 safeties, which is a huge differentiator for DE clustering.
Both Kearse and Blackstone also had safeties. Those two and Ulysses were the only ones to register a safety, and would separate themselves into a node of their own on that dimension. Ulysses just had too few tackles to be grouped there.
Howyanow and Jones both had such high tackle numbers to be forced out of the other cluster nodes and into the high tackle node. Interestingly both were also the only ones to post PDs, which would greatly separate that node from the others.

Defensive Tackle

3 - 68.0 Tck - 3.0 TFL - 13.0 Sck - 0.0 PD - Edd, B. - SAR
3 - 59.0 Tck - 4.0 TFL - 10.0 Sck - 0.0 PD - Miller, P. - AZ
3 - 58.0 Tck - 5.0 TFL - 13.0 Sck - 0.0 PD - Eagle, R. - OCO
3 - 55.0 Tck - 0.0 TFL - 13.0 Sck - 0.0 PD - Thumper, B. – COL
3 - 50.0 Tck - 3.0 TFL - 8.0 Sck - 0.0 PD - Money, P. - YKW
3 - 50.0 Tck - 4.0 TFL - 9.0 Sck - 0.0 PD - Steel ®, H. - NO
0 - 52.0 Tck - 0.0 TFL - 5.0 Sck - 0.0 PD - Brooks ®, D. - CHI
0 - 49.0 Tck - 2.0 TFL - 4.0 Sck - 0.0 PD - Rikiya, M. - NO
0 - 48.0 Tck - 2.0 TFL - 6.0 Sck - 0.0 PD - Montain, T. – AUS
0 - 47.0 Tck - 1.0 TFL - 4.0 Sck - 0.0 PD - Cloudera ®, Y. – SJS
0 - 45.0 Tck - 4.0 TFL - 4.0 Sck - 0.0 PD - Clemente, C. - SAR
0 - 42.0 Tck - 3.0 TFL - 5.0 Sck - 0.0 PD - Allen, O. – SJS
0 - 39.0 Tck - 1.0 TFL - 4.0 Sck - 0.0 PD - Kroustis, G. - YKW
0 - 36.0 Tck - 1.0 TFL - 3.0 Sck - 0.0 PD - Annastesia, N. - BAL
0 - 31.0 Tck - 3.0 TFL - 2.0 Sck - 0.0 PD - Miller II, C. - AZ
4 - 47.0 Tck - 5.0 TFL - 8.0 Sck - 0.0 PD - Masters, S. - AZ
4 - 46.0 Tck - 6.0 TFL - 5.0 Sck - 0.0 PD - Smirh, J. - CHI
4 - 40.0 Tck - 8.0 TFL - 6.0 Sck - 0.0 PD - Wheezer ®, C. - OCO
1 - 30.0 Tck - 1.0 TFL - 1.0 Sck - 0.0 PD - Ka'ana'ana ®, B. - HON
1 - 22.0 Tck - 1.0 TFL - 2.0 Sck - 0.0 PD - Gronlie, T. - PHI
1 - 15.0 Tck - 1.0 TFL - 0.0 Sck - 0.0 PD - McTackle, R. - AUS
1 - 12.0 Tck - 1.0 TFL - 0.0 Sck - 0.0 PD - Bean, S. - BAL
1 - 5.0 Tck - 0.0 TFL - 1.0 Sck - 0.0 PD - Culbertson, R. – PHI
2 - 29.0 Tck - 2.0 TFL - 3.0 Sck - 1.0 PD - Boucher, B. – OCO

Notes

Try to name the clusters again: 3 – Studs, 0 – Low event, 4 – High event, 1 – Low spectrum, 2 – Boucher.
Boucher gets their own category on DT, as the PD sets them apart from the entire field and cluster 1.
Montain would be right on the edge of cluster 3 and cluster 0 – most likely just 1 more sack and/or 5 more tackles and it would’ve pushed them into the cluster 3 node.
Cluster 0 and cluster 4 are in the same ballpark for tackles, but a climb in sacks and TFL for the latter moves them to a slightly different node.
Great rookie season by Steel making cluster 3.

Linebacker

3 - 158.0 Tck - 0.0 TFL - 3.0 Sck - 1.0 PD - 1.0 Int - Kennedy, D. - BAL
3 - 142.0 Tck - 0.0 TFL - 2.0 Sck - 5.0 PD - 1.0 Int - Hockhertz, M. - CHI
6 - 129.0 Tck - 3.0 TFL - 4.0 Sck - 5.0 PD - 0.0 Int - Edmond, H. - AUS
6 - 128.0 Tck - 1.0 TFL - 1.0 Sck - 6.0 PD - 0.0 Int - Bode, Q. - NO
6 - 125.0 Tck - 3.0 TFL - 3.0 Sck - 5.0 PD - 0.0 Int - Virtanen, M. – YKW
6 - 120.0 Tck - 0.0 TFL - 1.0 Sck - 5.0 PD - 1.0 Int - Koniecpolski, S. – AZ
6 - 106.0 Tck - 0.0 TFL - 2.0 Sck - 4.0 PD - 0.0 Int - Fisher, G. - SJS
1 - 128.0 Tck - 3.0 TFL - 3.0 Sck - 5.0 PD - 3.0 Int - Arianlacher, M. - HON
1 - 120.0 Tck - 1.0 TFL - 0.0 Sck - 7.0 PD - 2.0 Int - Dillmiballs, F. – SAR
1 - 107.0 Tck - 2.0 TFL - 3.0 Sck - 6.0 PD - 2.0 Int - Gabagool, T. - COL
1 - 103.0 Tck - 2.0 TFL - 2.0 Sck - 13.0 PD - 3.0 Int - Banks, J. - HON
4 - 117.0 Tck - 2.0 TFL - 13.0 Sck - 29.0 PD - 2.0 Int - De Ville, D. - BAL
4 - 112.0 Tck - 1.0 TFL - 8.0 Sck - 26.0 PD - 2.0 Int - McCringleberry, H. – CHI
4 - 88.0 Tck - 0.0 TFL - 8.0 Sck - 23.0 PD - 1.0 Int - Stephens, W. - COL
5 - 106.0 Tck - 7.0 TFL - 9.0 Sck - 4.0 PD - 1.0 Int - Suzuki, S. – PHI
5 - 74.0 Tck - 6.0 TFL - 5.0 Sck - 4.0 PD - 0.0 Int - Zelos ®, J. - PHI
2 - 116.0 Tck - 0.0 TFL - 9.0 Sck - 20.0 PD - 3.0 Int - Quaid, D. - YKW
2 - 106.0 Tck - 1.0 TFL - 10.0 Sck - 20.0 PD - 3.0 Int - Mango-Panda, H. – SAR
2 - 95.0 Tck - 1.0 TFL - 11.0 Sck - 17.0 PD - 1.0 Int - Brackenridge, T. - SJS
7 - 106.0 Tck - 2.0 TFL - 14.0 Sck - 13.0 PD - 1.0 Int - Wilf, G. – AZ
7 - 98.0 Tck - 2.0 TFL - 9.0 Sck - 16.0 PD - 0.0 Int - Dissector, U. - NO
7 - 97.0 Tck - 2.0 TFL - 6.0 Sck - 11.0 PD - 0.0 Int - Kackpoo, H. - HON
7 - 93.0 Tck - 2.0 TFL - 10.0 Sck - 13.0 PD - 0.0 Int - Porter, G. - AUS
0 - 76.0 Tck - 1.0 TFL - 3.0 Sck - 1.0 PD - 0.0 Int - Deck, I. – OCO
0 - 64.0 Tck - 1.0 TFL - 3.0 Sck - 7.0 PD - 1.0 Int - Joestar ®, J. - OCO
0 - 10.0 Tck - 0.0 TFL - 2.0 Sck - 0.0 PD - 0.0 Int - Summers, H. - AUS

Notes

Whew a lot to break down here. I’ll give the clusters a shot to name, but there’s lots of moving parts. 3 – High Tackle TD Returners, 6 – All Around LB, 1 – Interception Kings, 4 – Studs, 5 – Run Stuffers, 2 – Pocket Collapsers with reach, 7 – Pocket Collapsers with less reach, 0 – Low Tackle
Not shown in the above list, but alluded to in the cluster name, both Kennedy and Hockhertz had TDs and so differentiated themselves from the field into cluster 3.
The difference between cluster 7 and cluster 2 is very minute. The latter has slightly higher PD numbers and higher interception totals, but I’ve to assume that the difference in Euclidean distance for Brackenridge and Wilf to both clusters were almost identical.
The LB list is one of the most interesting so far, as each cluster seems to have its strengths. You can mostly see the differential in one category that the players lead in. Fisher is interesting in cluster 6, as they otherwise have high tackle numbers but nothing exceptional in the other categories. Cluster 6 might have been a catchall, central node in the dimensional spectrum.
Only two rookies? Just Zelos and Joestar with the R-tag. Zelos in the run stuffer cluster 5, with a substantially higher TFL. Joestar made cluster 0 for low tackles as a differentiator, but with 7 PDs and 1 Interception he’d just need to increase the tackle total to make the other clusters.

Cornerback

3 - 127.0 Tck - 0.0 TFL - 0.0 Sck - 5.0 PD - 0.0 Int - Shields, T. – CHI
3 - 122.0 Tck - 1.0 TFL - 0.0 Sck - 12.0 PD - 0.0 Int - Krause, M. - SJS
3 - 119.0 Tck - 2.0 TFL - 0.0 Sck - 9.0 PD - 1.0 Int - Mendoza, C. - AUS
0 - 122.0 Tck - 3.0 TFL - 2.0 Sck - 9.0 PD - 3.0 Int - Quin, A. - AZ
1 - 116.0 Tck - 0.0 TFL - 2.0 Sck - 12.0 PD - 4.0 Int - Crooks ®, L. - BAL
1 - 87.0 Tck - 0.0 TFL - 0.0 Sck - 23.0 PD - 4.0 Int - Waters, J. – HON
1 - 77.0 Tck - 0.0 TFL - 0.0 Sck - 22.0 PD - 3.0 Int - Purchase, L. - COL
1 - 75.0 Tck - 0.0 TFL - 0.0 Sck - 14.0 PD - 2.0 Int - Firestorm-Fjord, O. - CHI
1 - 73.0 Tck - 0.0 TFL - 0.0 Sck - 17.0 PD - 2.0 Int - Stackinpaper, B. - BAL
1 - 69.0 Tck - 0.0 TFL - 0.0 Sck - 17.0 PD - 3.0 Int - Rector, D. - NO
1 - 65.0 Tck - 0.0 TFL - 0.0 Sck - 18.0 PD - 3.0 Int - Kehla, Z. - AZ
1 - 62.0 Tck - 0.0 TFL - 0.0 Sck - 16.0 PD - 3.0 Int - Knight, L. - SAR
1 - 61.0 Tck - 0.0 TFL - 0.0 Sck - 16.0 PD - 3.0 Int - Scarlett, D. - AZ
6 - 83.0 Tck - 0.0 TFL - 0.0 Sck - 20.0 PD - 0.0 Int - Avery, E. – BAL
6 - 72.0 Tck - 0.0 TFL - 0.0 Sck - 20.0 PD - 1.0 Int - Leonard, M. – YKW
6 - 63.0 Tck - 0.0 TFL - 0.0 Sck - 18.0 PD - 0.0 Int - Harris, C. - CHI
6 - 62.0 Tck - 0.0 TFL - 0.0 Sck - 17.0 PD - 2.0 Int - Adams, X. - SAR
6 - 60.0 Tck - 0.0 TFL - 0.0 Sck - 13.0 PD - 1.0 Int - Oles JR, T. - SJS
6 - 59.0 Tck - 1.0 TFL - 0.0 Sck - 17.0 PD - 0.0 Int - Scott, R. - COL
6 - 56.0 Tck - 0.0 TFL - 0.0 Sck - 18.0 PD - 2.0 Int - Jackson ®, R. - HON
6 - 55.0 Tck - 0.0 TFL - 0.0 Sck - 19.0 PD - 2.0 Int - Abernathy, K. - OCO
6 - 52.0 Tck - 0.0 TFL - 0.0 Sck - 19.0 PD - 2.0 Int - Snuggles, J. - YKW
6 - 52.0 Tck - 0.0 TFL - 0.0 Sck - 21.0 PD - 2.0 Int - PickSix, H. - SJS
6 - 51.0 Tck - 0.0 TFL - 0.0 Sck - 16.0 PD - 2.0 Int - Beerus, L. - OCO
2 - 76.0 Tck - 1.0 TFL - 11.0 Sck - 4.0 PD - 1.0 Int - Andrews, J. - NO
5 - 75.0 Tck - 0.0 TFL - 0.0 Sck - 16.0 PD - 5.0 Int - Witten, A. - NO
5 - 74.0 Tck - 0.0 TFL - 0.0 Sck - 20.0 PD - 6.0 Int - Lavelle Jr., D. - AUS
4 - 73.0 Tck - 1.0 TFL - 3.0 Sck - 10.0 PD - 5.0 Int - Booker, B. - PHI
7 - 66.0 Tck - 0.0 TFL - 3.0 Sck - 1.0 PD - 0.0 Int - McGullager ®, T. - AUS

Notes

Let’s give naming the clusters a shot: 3 – High Tackle, 0 – Quin, 1 – Interception Masters, 6 – Classical CB, 2 – Andrews, 5 – Interception Experts, 4 – Booker, 7 – McGullager
In hindsight, too many clusters here. This could’ve been collapsed by a couple, as I set the minimum number during the Bayesian optimization. Oh well.
Let’s break down the individual clusters. Quin is the first – his TFL differentiator pushed him out of cluster 1. Andrews is next – his sack differentiator is the reason he’s not in cluster 6 (plus far too few PDs). Booker has too few sacks to be with Andrews, but too few PDs to be in cluster 6 (and a high amount of interceptions). McGullager is similar to Booker, but has too few interceptions to be clustered there. As a result, we’ve isolated these individual edge-cases into their own groups.
Not shown, but Booker also has 2 return TDs, which would push him further into his own category on that dimension. Cluster 5 (Witten and Lavelle Jr.) also both have a return TD, which helps to further isolate that node.
Great rookie season by Crooks.

Safeties

7 - 133.0 Tck - 3.0 TFL - 1.0 Sck - 8.0 PD - 0.0 Int - Eriksen, W. - PHI
4 - 132.0 Tck - 2.0 TFL - 0.0 Sck - 10.0 PD - 2.0 Int - Sly ®, G. - COL
4 - 125.0 Tck - 2.0 TFL - 0.0 Sck - 7.0 PD - 3.0 Int - Hexagon ®, L. – NO
4 - 107.0 Tck - 2.0 TFL - 1.0 Sck - 14.0 PD - 2.0 Int - Sexjoke, A. - AUS
1 - 114.0 Tck - 0.0 TFL - 0.0 Sck - 10.0 PD - 2.0 Int - Nkiah, J. – SAR
1 - 107.0 Tck - 1.0 TFL - 1.0 Sck - 13.0 PD - 2.0 Int - Biggums, T. – OCO
1 - 95.0 Tck - 1.0 TFL - 2.0 Sck - 8.0 PD - 1.0 Int - Valdyr, M. - YKW
1 - 71.0 Tck - 0.0 TFL - 0.0 Sck - 14.0 PD - 2.0 Int - D'Attoria V, R. - PHI
0 - 110.0 Tck - 0.0 TFL - 8.0 Sck - 1.0 PD - 1.0 Int - Jadakiss, D. - AZ
0 - 95.0 Tck - 0.0 TFL - 4.0 Sck - 2.0 PD - 0.0 Int - Smoothie, E. - PHI
0 - 87.0 Tck - 1.0 TFL - 6.0 Sck - 2.0 PD - 1.0 Int - Beat, N. – SAR
0 - 77.0 Tck - 0.0 TFL - 7.0 Sck - 2.0 PD - 1.0 Int - MacGregor, G. – BAL
0 - 63.0 Tck - 0.0 TFL - 6.0 Sck - 1.0 PD - 1.0 Int - Vegeta, P. - OCO
3 - 84.0 Tck - 1.0 TFL - 2.0 Sck - 8.0 PD - 5.0 Int - Dawkins, S. - HON
3 - 79.0 Tck - 0.0 TFL - 4.0 Sck - 4.0 PD - 3.0 Int - Cross, M. – PHI
3 - 63.0 Tck - 0.0 TFL - 3.0 Sck - 6.0 PD - 5.0 Int - Stein ®, P. - SJS
6 - 79.0 Tck - 1.0 TFL - 3.0 Sck - 0.0 PD - 0.0 Int - Blanco ®, B. – AZ
6 - 69.0 Tck - 0.0 TFL - 2.0 Sck - 4.0 PD - 1.0 Int - Riley, C. - BAL
6 - 68.0 Tck - 0.0 TFL - 2.0 Sck - 5.0 PD - 1.0 Int - Bob ®, B. - OCO
6 - 65.0 Tck - 0.0 TFL - 5.0 Sck - 3.0 PD - 1.0 Int - Hellzapoppin, J. - YKW
6 - 65.0 Tck - 1.0 TFL - 0.0 Sck - 3.0 PD - 1.0 Int - Dookmarriot ®, L. - YKW
6 - 63.0 Tck - 0.0 TFL - 4.0 Sck - 2.0 PD - 1.0 Int - Thornton, B. – HON
6 - 62.0 Tck - 0.0 TFL - 4.0 Sck - 6.0 PD - 1.0 Int - Green, W. - BAL
6 - 62.0 Tck - 0.0 TFL - 4.0 Sck - 1.0 PD - 1.0 Int - Richards ®, E. - SAR
6 - 60.0 Tck - 1.0 TFL - 6.0 Sck - 2.0 PD - 1.0 Int - Lanier ®, B. - SJS
5 - 75.0 Tck - 0.0 TFL - 12.0 Sck - 6.0 PD - 3.0 Int - Kirkby, T. - COL
5 - 74.0 Tck - 0.0 TFL - 10.0 Sck - 6.0 PD - 4.0 Int - Parker, P. - COL
5 - 73.0 Tck - 0.0 TFL - 8.0 Sck - 5.0 PD - 3.0 Int - MacAndrew, H. - CHI
5 - 70.0 Tck - 1.0 TFL - 8.0 Sck - 6.0 PD - 3.0 Int - Blaylock, M. – NO
5 - 66.0 Tck - 0.0 TFL - 7.0 Sck - 6.0 PD - 2.0 Int - d'Alcott, R. - CHI
2 - 23.0 Tck - 0.0 TFL - 0.0 Sck - 2.0 PD - 0.0 Int - Okura, T. - AUS
2 - 1.0 Tck - 0.0 TFL - 0.0 Sck - 0.0 PD - 0.0 Int - Z, D. – YKW

Notes

Lets name the clusters: 7 – Eriksen, 4 – Run Stuffers, 1 – Warriors, 0 – Blitzers, 3 – Playmakers, 6 – All Around Safeties, 5 – Slot Safeties, 2 – Low Tackles
Eriksen gets a category to themselves as they have high tackles, but not enough in other categories to be added to the cluster 4 or 1.
Warriors for cluster 1 refers to the battles for the ball that happen on each throw, as they put up serious pass deflection numbers.
Gotta love cluster 3 and those interception numbers.
What a rookie season for a number of safeties, foremost Sly and Hexagon.

**Hordle** · 09-20-2020, 07:25 PM

Great work again. All you smart kids and your fancy graphs.

~~wonderful_art~~ · 09-20-2020, 07:26 PM

(09-20-2020, 07:25 PM)Hordle Wrote: Great work again. All you smart kids and your fancy graphs.

thanks man Heart

Billybolo53 · 11-24-2020, 12:15 PM

how did i miss this, very cool stuff