Analyzing Golden State Warriors' passing network using GraphFrames in Spark

转载自：http://opiateforthemass.es/articles/analyzing-golden-state-warriors-passing-network-using-graphframes-in-spark/

Yuki Katoh (yukiegosapporo@gmail.com)
发表于

Databricks recently announced GraphFrames, awesome Spark extension to implement graph processing using DataFrames.I performed graph analysis and visualized beautiful ball movement network of Golden State Warriors using rich data provided by NBA.com’s stats

Passes received & made

The league’s MVP Stephen Curry received the most passes and the team’s MVP Draymond Green provides the most passes.We’ve seen most of the offense start with their pick & roll or Curry’s off-ball cuts with Green as a pass provider.

via GIPHY

inDegree
id inDegree  
CurryStephen   3993
GreenDraymond   3123
ThompsonKlay   2276
LivingstonShaun   1925
IguodalaAndre   1814
BarnesHarrison   1241
BogutAndrew   1062
BarbosaLeandro   946
SpeightsMarreese   826
ClarkIan   692
RushBrandon   685
EzeliFestus   559
McAdooJames Michael   182
VarejaoAnderson   67
LooneyKevon   22
outDegree
id outDegree  
GreenDraymond   3841
CurryStephen   3300
IguodalaAndre   1896
LivingstonShaun   1878
BogutAndrew   1660
ThompsonKlay   1460
BarnesHarrison   1300
SpeightsMarreese   795
RushBrandon   772
EzeliFestus   765
BarbosaLeandro   758
ClarkIan   597
McAdooJames Michael   261
VarejaoAnderson   94
LooneyKevon   36

Label Propagation

Label Propagation is an algorithm to find communities in a graph network.The algorithm nicely classifies players into backcourt and frontcourt without providing label!

name label  
Thompson, Klay   3
Barbosa, Leandro   3
Curry, Stephen   3
Clark, Ian   3
Livingston, Shaun   3
Rush, Brandon   7
Green, Draymond   7
Speights, Marreese   7
Bogut, Andrew   7
McAdoo, James Michael   7
Iguodala, Andre   7
Varejao, Anderson   7
Ezeli, Festus   7
Looney, Kevon   7
Barnes, Harrison   7

PageRank can detect important nodes (players in this case) in a network.It’s no surprise that Stephen Curry, Draymond Green and Klay Thompson are the top three.The algoritm detects Shaun Livingston and Andre Iguodala play key roles in the Warriors’ passing games.

name pagerank  
Curry, Stephen   2.17
Green, Draymond   1.99
Thompson, Klay   1.34
Livingston, Shaun   1.29
Iguodala, Andre   1.21
Barnes, Harrison   0.86
Bogut, Andrew   0.77
Barbosa, Leandro   0.72
Speights, Marreese   0.66
Clark, Ian   0.59
Rush, Brandon   0.57
Ezeli, Festus   0.48
McAdoo, James Michael   0.27
Varejao, Anderson   0.19
Looney, Kevon   0.16 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
library(networkD3)

setwd('/Users/yuki/Documents/code_for_blog/gsw_passing_network')
passes <- read.csv("passes.csv")
groups <- read.csv("groups.csv")
size <- read.csv("size.csv")

passes$source <- as.numeric(as.factor(passes$PLAYER))-1
passes$target <- as.numeric(as.factor(passes$PASS_TO))-1
passes$PASS <- passes$PASS/50

groups$nodeid <- groups$name
groups$name <- as.numeric(as.factor(groups$name))-1
groups$group <- as.numeric(as.factor(groups$label))-1
nodes <- merge(groups,size[-1],by="id")
nodes$pagerank <- nodes$pagerank^2*100


forceNetwork(Links = passes,
 Nodes = nodes,
 Source = "source",
 fontFamily = "Arial",
 colourScale = JS("d3.scale.category10()"),
 Target = "target",
 Value = "PASS",
 NodeID = "nodeid",
 Nodesize = "pagerank",
 linkDistance = 350,
 Group = "group", 
 opacity = 0.8,
 fontSize = 16,
 zoom = TRUE,
 opacityNoHover = TRUE)

Here is a network visualization using the results of above.

  • Node size: pagerank

  • Node color: community

  • Link width: passes received & made

Calling API

I used the endpoint playerdashptpass and saved data for all the players in the team into local JSON files.The data is about who passed how many times in 2015-16 season

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
# GSW player IDs
playerids = [201575,201578,2738,202691,101106,2760,2571,203949,203546,
203110,201939,203105,2733,1626172,203084]

# Calling API and store the results as JSON
for playerid in playerids:
 os.system('curl "http://stats.nba.com/stats/playerdashptpass?'
 'DateFrom=&'
 'DateTo=&'
 'GameSegment=&'
 'LastNGames=0&'
 'LeagueID=00&'
 'Location=&'
 'Month=0&'
 'OpponentTeamID=0&'
 'Outcome=&'
 'PerMode=Totals&'
 'Period=0&'
 'PlayerID={playerid}&'
 'Season=2015-16&'
 'SeasonSegment=&'
 'SeasonType=Regular+Season&'
 'TeamID=0&'
 'VsConference=&'
 'VsDivision=" > {playerid}.json'.format(playerid=playerid))

JSON -> Panda’s DataFrame

Then I combined all the individual JSON files into a single DataFrame for later aggregation.

1
2
3
4
5
6
7
8
9
10
raw = pd.DataFrame()
for playerid in playerids:
 with open("{playerid}.json".format(playerid=playerid)) as json_file:
 parsed = json.load(json_file)['resultSets'][0]
 raw = raw.append(
 pd.DataFrame(parsed['rowSet'], columns=parsed['headers']))

raw = raw.rename(columns={'PLAYER_NAME_LAST_FIRST': 'PLAYER'})

raw['id'] = raw['PLAYER'].str.replace(', ', '')

Prepare vertices and edges

You need a special data format for GraphFrames in Spark, vertices and edges.Vertices are lis of nodes and IDs in a graph.Edges are the relathionship of the nodes. You can pass additional features like weight but I couldn’t find out a way to utilize there features well in later analysis. A workaround I took below is brute force and not even a proper graph operation but works (suggestions/comments are very welcome).

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# Make raw vertices
pandas_vertices = raw[['PLAYER', 'id']].drop_duplicates()
pandas_vertices.columns = ['name', 'id']

# Make raw edges
pandas_edges = pd.DataFrame()
for passer in raw['id'].drop_duplicates():
 for receiver in raw[(raw['PASS_TO'].isin(raw['PLAYER'])) &
 (raw['id'] == passer)]['PASS_TO'].drop_duplicates():
 pandas_edges = pandas_edges.append(pd.DataFrame(
 {'passer': passer, 'receiver': receiver
 .replace( ', ', '')}, 
 index=range(int(raw[(raw['id'] == passer) &
 (raw['PASS_TO'] == receiver)]['PASS'].values))))

pandas_edges.columns = ['src', 'dst']

Graph analysis

Bring the local vertices and edges to Spark and let it spark.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
vertices = sqlContext.createDataFrame(pandas_vertices)
edges = sqlContext.createDataFrame(pandas_edges)

# Analysis part
g = GraphFrame(vertices, edges)
print("vertices")
g.vertices.show()
print("edges")
g.edges.show()
print("inDegrees")
g.inDegrees.sort('inDegree', ascending=False).show()
print("outDegrees")
g.outDegrees.sort('outDegree', ascending=False).show()
print("degrees")
g.degrees.sort('degree', ascending=False).show()
print("labelPropagation")
g.labelPropagation(maxIter=5).show()
print("pageRank")
g.pageRank(resetProbability=0.15, tol=0.01).vertices.sort(
 'pagerank', ascending=False).show()

Visualise the network

When you run gsw_passing_network.py in my github repo, you have passes.csv, groups.csv and size.csv in your working directory.I used networkD3 package in R to make a cool interactive D3 chart.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
library(networkD3)

setwd('/Users/yuki/Documents/code_for_blog/gsw_passing_network')
passes <- read.csv("passes.csv")
groups <- read.csv("groups.csv")
size <- read.csv("size.csv")

passes$source <- as.numeric(as.factor(passes$PLAYER))-1
passes$target <- as.numeric(as.factor(passes$PASS_TO))-1
passes$PASS <- passes$PASS/50

groups$nodeid <- groups$name
groups$name <- as.numeric(as.factor(groups$name))-1
groups$group <- as.numeric(as.factor(groups$label))-1
nodes <- merge(groups,size[-1],by="id")
nodes$pagerank <- nodes$pagerank^2*100


forceNetwork(Links = passes,
 Nodes = nodes,
 Source = "source",
 fontFamily = "Arial",
 colourScale = JS("d3.scale.category10()"),
 Target = "target",
 Value = "PASS",
 NodeID = "nodeid",
 Nodesize = "pagerank",
 linkDistance = 350,
 Group = "group", 
 opacity = 0.8,
 fontSize = 16,
 zoom = TRUE,
 opacityNoHover = TRUE)

The full codes are available on github.