sfu-cl-lab / yeti-thesis-project Goto Github PK

Want to make sure we have valid raw data. Can use birthdate to identify players.

@chaostewart I see on nhl.com more stats for each player, like shots, shots percentage, face-off win percentage. Shouldn't we be using that data? @liuyejia

Evaluate using Shucker's metric (i.e. correlation between predicted ranking and actual ranking according to number of games).

• For each player, export their leaf node
• For each leaf, export its equation (currently manual)
• For each player, apply the equation to get a predicted score (number of games played, probability of playing > 0 games).
• use predicted score to rank players. Compute spearman rank correlation with ranking according to observed metric

Can we

• upload NBA data to github?
• upload NBA + NHL data to the prague relational repository? Or first the database server?

Hi

Please bind the ~/.ssh/id_rsa to your github configuration so that you don't get publickey error

using linear model: which components contribute most to predictive difference from average member of cluster?

Some notes for the conference paper

• I see in your thesis there has been discussion about grouping NBA players into positions. The tree in a sense discovers groupings of positions. e.g.

For basketball,
many clustering approaches focus on defining appropriate roles or positions for a
player.

• meet with Max and Galen about learning model trees

• explain why plus-minus is important in the tree

• submit to KDD workshop?

Hi Yeti,

thanks for doing this work. It looks like an interesting result, especially that SVM does so much better. A few suggestions.

• how about neural net? Deep neural net?

• In your report, please include a brief description of the dataset, maybe some sample lines from the data file. Also a description of what you want to predict.

• How does this relate to Wilson's result?

• How does this relate to Shuckers' work? I guess we should get a script for computing the correlations with TOI performance as he did.

let's find out which players form the AHL set also appeared in the NHL
https://github.com/sfu-cl-lab/Yeti-Thesis-Project/blob/master/AHL_seasons_players/AHL%20Data%20-%20Full%20Set.csv

1. why are there CSS ranks missing (e.g. in chao_draft.join_skater_and_season_stats_10_years_view we go from 249 to 242)?
2. why are there 668 players at maximum rank in chao_draft.norm_dataset_for_lmt but only 666 players with CSS rank null in chao_draft.join_skater_and_season_stats_10_years_CSS_null?

plus learning tree to support linear equations (see R, Quinlan M5)

link to Shucker's paper: https://github.com/sfu-cl-lab/Yeti-Thesis-Project/blob/master/papers/1559-Draft-by-Numbers.pdf

1. Schucker used Spearman rank correlation as a measure of association and predictive ability. Can I calculate this correlation by following the instructions as follows: https://statistics.laerd.com/statistical-guides/spearmans-rank-order-correlation-statistical-guide.php ???
2. Schucker used data from 3 consecutive (1998, 1999, and 2000) NHL drafts to build the model, then predicted each players performance in the 2 subsequent (2001 and 2002) out of sample years. Should we do it in the same way?
3. Table 4 and Table 5 in Schucker's paper display results for "Drafted Players" & "All Players" respectively. Their definition of "drafted players" & "all players" is a bit of unclear to me.
   @oschulte

Hi @oschulte Oliver,
Connecting to our datebase in WEKA is non-trivial. I uploaded an instruction file for this particular task. Please check out: https://github.com/sfu-cl-lab/Yeti-Thesis-Project/blob/master/How%20to%20connect%20to%20MySql%20database%20in%20WEKA.md

I can't make the connection work on my loptop for now due two multi-hops. Neither can I export databases I need (chao_draft or ckm_and_exception_mining) to my local. Thus, (if i can't come up with a better solution in the next 24 hours) I suggest, to be able to use WEKA in our next meeting, the easiest way is to have you install it on your desktop in your office. OR come to my desktop in the lab.

Thank you for your understanding!

1. check the websites mentioned by Shuckers to find the CSS rank for missing players
2. if we can't find it, is that because they are low ranked (late pick)?
3. How did we get Cescin? This means that for the European players, we have a different ranking. Cescin is converted European rank. I wonder if we can find the original scouting ranking for Europeans.

@chaostewart : I've added instructions for how to transfer databases to Bugagoo

https://github.com/sfu-cl-lab/group-sharing/wiki/Database-Servers

We tested these once, not 100% sure about them. Galen has worked with Bugaboo too.

Could try using a discrete prediction. Probably with three classes:

1. NHL player with > 0 games
2. NHL player with 0 < 160 games (whatever the official treshold is)
3. NHL player with > 160 games

Problem 1 seems especially natural, compared to the somewhat artificial threshold of 160.

If we implement Shucker's ranking metric #4 then we can evaluate both regression trees and decision trees in terms of how they rank players. For that we may have to combine decision trees with logistic regression - I wonder if that exists? In R perhaps?

• for junior leagues
• maybe NHL too?

• write outline
• make result figures with d_i^2
• make plots

Player stats crawled from eliteprospects.com are saved as table chao_draft.elite_prospects_skaters_stats_1998_2008_original. There are ~150 of them have the position as 'W' or 'F'. How do we determine if they are 'L', 'R' or 'C'? Note that even their 'Shoots" information is also know, left or right-handed shoots does not determine a player's position as 'L' or 'R'.

• Euro 2016 data (see google drive, Maity email)
• Canadian Women's soccer (Peter Chow-White)
• dataset from Russ (?? source)