Identify N nearest neighbors to player's season statistics

player_similarity.py

"""
An outline for identifying N players with similar summary statistics to a player of interest.
A list of candidate players is compiled by filtering on position and end-of-season age.
Summary statistics are scaled to z-scores, which are the inputs to the machine learning model.
This algorithm uses K-Nearest Neighbor, but other algorithms (eg K-Means Clustering) can be substituted.

(c) 2015 Soccermetrics Research LLC
This code is licensed under the terms of the MIT License (http://choosealicense.com/licenses/mit/)
"""
from sklearn.neighbors import NearestNeighbors
from scipy.spatial.distance import cosine
import numpy as np


def get_statistical_fields(position):
    """
    Compile list of statistical categories to be retrieved from database
    given player position.
    """
    # metrics consistent among all players
    all_metric_list = ('games_started', 'games_subbed', 'minutes', 'yellows', 'reds')
    # metrics for field players
    field_metric_list = ('goals_total', 'goals_headed', 'goals_freekick', 'goals_in_area',
                         'goals_out_area', 'goals_penalty', 'penalties', 'winners', 'assists',
                         'deadball_assists', 'shots', 'fouls')
    # metrics for goalkeepers
    gk_metric_list = ('wins', 'draws', 'losses', 'goals_allowed', 'shutouts', 'shots_allowed')

    if position != POSITION['Goalkeeper']:
        stat_table = 'field_stats_list'
        metrics = all_metric_list + field_metric_list
    else:
        stat_table = 'goalkeeper_stats_list'
        metrics = all_metric_list + gk_metric_list
    

def identify_similar_player(player, competition, season, N=20):
    """
    Identify N nearest neighbors to player's statistical performance
    controlling for players of same age and position
    """
    def mydist(x,y):
        # internal function
        # compute cosine distance between two points
        return cosine(x,y)
        
    # retrieve player position and age
    player_position, player_age = get_position_and_age(player, season)
    
    # get statistical fields that are relevant to player position
    # field players and goalkeepers have statistical categories unique to them
    stat_table, stat_metrics = get_statistical_fields(player_position)
    
    # retrieve player z-scores for each metric --> player statistical record
    # some metrics do not exist for player --> these are zeroed out
    player_record = get_statistical_record(player, competition, season, metrics)
    
    # retrieve candidate list of players who match position and age
    # player, competition, season IDs
    candidate_player_list = create_candidate_list(stat_table, player_position, player_age)
    
    # create candidate player statistical record --> feature matrix
    training_list = [
        get_statistical_record(candidate_player, candidate_competition, candidate_season, metrics)
        for candidate_player, candidate_competition, candidate_season in candidate_player_list
    ]

    # train a K-Nearest Neighbor model on candidate player statistical records
    # N neighbors, BallTree algorithm, cosine similarity distance
    # for more details see http://scikit-learn.org/stable/modules/generated/sklearn.neighbors.NearestNeighbors.html
    X = np.array(training_list)
    nbrs = NearestNeighbors(n_neighbors=N, algorithm='ball_tree', metric='pyfunc', func=mydist).fit(X)
    
    # find the K-neighbors to the player's statistical record
    distances, indices = nbrs.kneighbors(player_record)
    
    return distances, indices