CH8. 신용카드 거래에 대한 그래프 분석(frac=0.4)

graph
Author

김보람

Published

April 4, 2023

ref

신용카드 거래에 대한 그래프 분석

  • 신용카드 거래 그래프 생성

  • 그래프에서 속성 및 커뮤니티 추출

  • 사기 거래 분류에 지도 및 비지도 머신러닝 알고리즘 적용

import pandas as pd

import os
import math
import numpy as np
import networkx as nx
import matplotlib.pyplot as plt
%matplotlib inline

default_edge_color = 'gray'
default_node_color = '#407cc9'
enhanced_node_color = '#f5b042'
enhanced_edge_color = '#cc2f04'

샘플 = 0.4

import pandas as pd
df = pd.read_csv("fraudTrain.csv")
df = df[df["is_fraud"]==0].sample(frac=0.40, random_state=42).append(df[df["is_fraud"] == 1])
df.head()
Unnamed: 0 trans_date_trans_time cc_num merchant category amt first last gender street ... lat long city_pop job dob trans_num unix_time merch_lat merch_long is_fraud
669418 669418 2019-10-12 18:21 4.089100e+18 fraud_Haley, Jewess and Bechtelar shopping_pos 7.53 Debra Stark F 686 Linda Rest ... 32.3836 -94.8653 24536 Multimedia programmer 1983-10-14 d313353fa30233e5fab5468e852d22fc 1350066071 32.202008 -94.371865 0
32567 32567 2019-01-20 13:06 4.247920e+12 fraud_Turner LLC travel 3.79 Judith Moss F 46297 Benjamin Plains Suite 703 ... 39.5370 -83.4550 22305 Television floor manager 1939-03-09 88c65b4e1585934d578511e627fe3589 1327064760 39.156673 -82.930503 0
156587 156587 2019-03-24 18:09 4.026220e+12 fraud_Klein Group entertainment 59.07 Debbie Payne F 204 Ashley Neck Apt. 169 ... 41.5224 -71.9934 4720 Broadcast presenter 1977-05-18 3bd9ede04b5c093143d5e5292940b670 1332612553 41.657152 -72.595751 0
1020243 1020243 2020-02-25 15:12 4.957920e+12 fraud_Monahan-Morar personal_care 25.58 Alan Parsons M 0547 Russell Ford Suite 574 ... 39.6171 -102.4776 207 Network engineer 1955-12-04 19e16ee7a01d229e750359098365e321 1361805120 39.080346 -103.213452 0
116272 116272 2019-03-06 23:19 4.178100e+15 fraud_Kozey-Kuhlman personal_care 84.96 Jill Flores F 639 Cruz Islands ... 41.9488 -86.4913 3104 Horticulturist, commercial 1981-03-29 a0c8641ca1f5d6e243ed5a2246e66176 1331075954 42.502065 -86.732664 0

5 rows × 23 columns

df["is_fraud"].value_counts()
0    417028
1      6006
Name: is_fraud, dtype: int64
df["is_fraud"].value_counts()/len(df)
0    0.985803
1    0.014197
Name: is_fraud, dtype: float64

- 이분그래프

def build_graph_bipartite(df_input, graph_type=nx.Graph()):
    df=df_input.copy()
    mapping={x:node_id for node_id, x in enumerate(set(df["cc_num"].values.tolist()+\
                                                      df["merchant"].values.tolist()))}
    
    df["from"]=df["cc_num"].apply(lambda x:mapping[x])  #엣지의 출발점
    df["to"]=df["merchant"].apply(lambda x:mapping[x])  #엣지의 도착점
    
    df = df[['from', 'to', "amt", "is_fraud"]].groupby(['from','to']).agg({"is_fraud":"sum","amt":"sum"}).reset_index()
    df["is_fraud"]=df["is_fraud"].apply(lambda x:1 if x>0 else 0)
    
    G=nx.from_edgelist(df[["from","to"]].values, create_using=graph_type)
    
    nx.set_edge_attributes(G, {(int(x["from"]),int(x["to"])):x["is_fraud"] for idx, x in df[["from","to","is_fraud"]].iterrows()}, "label")  #엣지 속성 설정,각 속성의 사기 여부부 
    
    nx.set_edge_attributes(G,{(int(x["from"]),int(x["to"])):x["amt"] for idx,x in df[["from","to","amt"]].iterrows()}, "weight") # 엣지 속성 설정, 각 엣지의 거래 금액

    return G
  • 판매자, 고객에게 node 할당
G_bu = build_graph_bipartite(df, nx.Graph(name="Bipartite Undirect"))
  • 무향 그래프 작성

- 삼분그래프

def build_graph_tripartite(df_input, graph_type=nx.Graph()):
    df=df_input.copy()
    mapping={x:node_id for node_id, x in enumerate(set(df.index.values.tolist() + #set으로 중복 제거
                                                       df["cc_num"].values.tolist() +
                                                       df["merchant"].values.tolist()))}
    df["in_node"]= df["cc_num"].apply(lambda x: mapping[x])
    df["out_node"]=df["merchant"].apply(lambda x:mapping[x])
    
        
    G=nx.from_edgelist([(x["in_node"], mapping[idx]) for idx, x in df.iterrows()] +\
                        [(x["out_node"], mapping[idx]) for idx, x in df.iterrows()], create_using=graph_type)
    
    nx.set_edge_attributes(G,{(x["in_node"], mapping[idx]):x["is_fraud"] for idx, x in df.iterrows()}, "label")
     
    nx.set_edge_attributes(G,{(x["out_node"], mapping[idx]):x["is_fraud"] for idx, x in df.iterrows()}, "label")
    
    nx.set_edge_attributes(G,{(x["in_node"], mapping[idx]):x["amt"] for idx, x in df.iterrows()}, "weight")
    
    nx.set_edge_attributes(G,{(x["out_node"], mapping[idx]):x["amt"] for idx, x in df.iterrows()}, "weight")
    
    
    return G
  • 판매자, 고객, 거래에 노드 할당
G_tu = build_graph_tripartite(df, nx.Graph())
for G in [G_bu, G_tu]:
    print(nx.number_of_nodes(G))
1636
424670

사기 탐지를 위한 지도 및 비지도 임베딩

  • 트랜잭션 간선으로 표기

  • 각 간선을 올바른 클래스(사기 또는 정상)으로 분류

지도학습

from sklearn.utils import resample

df_majority = df[df.is_fraud==0]
df_minority = df[df.is_fraud==1]

df_maj_dowsampled = resample(df_majority,
                             n_samples=len(df_minority),
                             random_state=42)

df_downsampled = pd.concat([df_minority, df_maj_dowsampled])

print(df_downsampled.is_fraud.value_counts())
G_down = build_graph_bipartite(df_downsampled)
1    6006
0    6006
Name: is_fraud, dtype: int64

  • 무작위 언더샘플링 사용

  • 소수 클래스(사기거래)이 샘플 수 와 일치시키려고 다수 클래스(정상거래)의 하위 샘플을 가져옴

  • 데이터 불균형을 처리하기 위해서

from sklearn.model_selection import train_test_split


train_edges, test_edges, train_labels, test_labels = train_test_split(list(range(len(G_down.edges))), 
                                                                      list(nx.get_edge_attributes(G_down, "label").values()), 
                                                                      test_size=0.20, 
                                                                      random_state=42)
edgs = list(G_down.edges)
train_graph = G_down.edge_subgraph([edgs[x] for x in train_edges]).copy()
train_graph.add_nodes_from(list(set(G_down.nodes) - set(train_graph.nodes)))
  • 데이터 8:2 비율로 학습 검증
pip install node2vec
Collecting node2vec
  Downloading node2vec-0.4.6-py3-none-any.whl (7.0 kB)
Requirement already satisfied: joblib<2.0.0,>=1.1.0 in /home/coco/anaconda3/envs/py38/lib/python3.8/site-packages (from node2vec) (1.2.0)
Collecting gensim<5.0.0,>=4.1.2
  Downloading gensim-4.3.1-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (26.5 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 26.5/26.5 MB 71.8 MB/s eta 0:00:0000:0100:01
Collecting tqdm<5.0.0,>=4.55.1
  Downloading tqdm-4.65.0-py3-none-any.whl (77 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 77.1/77.1 kB 18.6 MB/s eta 0:00:00
Collecting networkx<3.0,>=2.5
  Downloading networkx-2.8.8-py3-none-any.whl (2.0 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 2.0/2.0 MB 90.4 MB/s eta 0:00:00
Requirement already satisfied: numpy<2.0.0,>=1.19.5 in /home/coco/anaconda3/envs/py38/lib/python3.8/site-packages (from node2vec) (1.24.2)
Requirement already satisfied: scipy>=1.7.0 in /home/coco/anaconda3/envs/py38/lib/python3.8/site-packages (from gensim<5.0.0,>=4.1.2->node2vec) (1.10.1)
Collecting smart-open>=1.8.1
  Downloading smart_open-6.3.0-py3-none-any.whl (56 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 56.8/56.8 kB 13.6 MB/s eta 0:00:00
Installing collected packages: tqdm, smart-open, networkx, gensim, node2vec
  Attempting uninstall: networkx
    Found existing installation: networkx 3.0
    Uninstalling networkx-3.0:
      Successfully uninstalled networkx-3.0
Successfully installed gensim-4.3.1 networkx-2.8.8 node2vec-0.4.6 smart-open-6.3.0 tqdm-4.65.0
Note: you may need to restart the kernel to use updated packages.
from node2vec import Node2Vec
from node2vec.edges import HadamardEmbedder, AverageEmbedder, WeightedL1Embedder, WeightedL2Embedder

node2vec_train = Node2Vec(train_graph, weight_key='weight')
model_train = node2vec_train.fit(window=10)
Generating walks (CPU: 1): 100%|██████████| 10/10 [00:03<00:00,  2.62it/s]
  • Node2Vec 알고리즘 사용해 특징 공간 구축
from sklearn.ensemble import RandomForestClassifier 
from sklearn import metrics 

classes = [HadamardEmbedder, AverageEmbedder, WeightedL1Embedder, WeightedL2Embedder]
for cl in classes:
    embeddings_train = cl(keyed_vectors=model_train.wv) 
    # 벡터스페이스 상에 edge를 투영.. 

    train_embeddings = [embeddings_train[str(edgs[x][0]), str(edgs[x][1])] for x in train_edges]
    test_embeddings = [embeddings_train[str(edgs[x][0]), str(edgs[x][1])] for x in test_edges]
    
    rf = RandomForestClassifier(n_estimators=1000, random_state=42) 
    rf.fit(train_embeddings, train_labels); 

    y_pred = rf.predict(test_embeddings)
    print(cl)
    print('Precision:', metrics.precision_score(test_labels, y_pred)) 
    print('Recall:', metrics.recall_score(test_labels, y_pred)) 
    print('F1-Score:', metrics.f1_score(test_labels, y_pred)) 
<class 'node2vec.edges.HadamardEmbedder'>
Precision: 0.7451737451737451
Recall: 0.16538131962296487
F1-Score: 0.270687237026648
<class 'node2vec.edges.AverageEmbedder'>
Precision: 0.691699604743083
Recall: 0.7497857754927164
F1-Score: 0.7195723684210527
<class 'node2vec.edges.WeightedL1Embedder'>
Precision: 0.7142857142857143
Recall: 0.029991431019708654
F1-Score: 0.0575657894736842
<class 'node2vec.edges.WeightedL2Embedder'>
Precision: 0.64
Recall: 0.027420736932305057
F1-Score: 0.052588331963845526

  • Node2Vec 알고리즘 사용해 각 Edge2Vec 알고리즘으로 특징 공간 생성

  • sklearn 파이썬 라이브러리의 RandomForestClassifier은 이전 단계에서 생성한 특징에 대해 학습

  • 검증 테스트 위해 정밀도, 재현율, F1-score 성능 지표 측정

비지도학습

  • k-means 알고리즘 사용

  • 지도학습과의 차이점은 특징 공간이 학습-검증 분할을 안함.

nod2vec_unsup = Node2Vec(G_down, weight_key='weight')
unsup_vals = nod2vec_unsup.fit(window=10)
Generating walks (CPU: 1): 100%|██████████| 10/10 [00:04<00:00,  2.32it/s]
  • 다운샘플링 절차에 전체 그래프 알고리즘 계산
from sklearn.cluster import KMeans

classes = [HadamardEmbedder, AverageEmbedder, WeightedL1Embedder, WeightedL2Embedder]
true_labels = [x for x in nx.get_edge_attributes(G_down, "label").values()]

for cl in classes:
    embedding_edge = cl(keyed_vectors=unsup_vals.wv) 

    embedding = [embedding_edge[str(x[0]), str(x[1])] for x in G_down.edges()]
    kmeans = KMeans(2, random_state=42).fit(embedding)
    
    
    nmi = metrics.adjusted_mutual_info_score(true_labels, kmeans.labels_)
    ho = metrics.homogeneity_score(true_labels, kmeans.labels_)
    co = metrics.completeness_score(true_labels, kmeans.labels_)
    vmeasure = metrics.v_measure_score(true_labels, kmeans.labels_)
    
    print(cl)
    print('NMI:', nmi)
    print('Homogeneity:', ho)
    print('Completeness:', co)
    print('V-Measure:', vmeasure)
/home/coco/anaconda3/envs/py38/lib/python3.8/site-packages/sklearn/cluster/_kmeans.py:870: FutureWarning: The default value of `n_init` will change from 10 to 'auto' in 1.4. Set the value of `n_init` explicitly to suppress the warning
  warnings.warn(
/home/coco/anaconda3/envs/py38/lib/python3.8/site-packages/sklearn/cluster/_kmeans.py:870: FutureWarning: The default value of `n_init` will change from 10 to 'auto' in 1.4. Set the value of `n_init` explicitly to suppress the warning
  warnings.warn(
/home/coco/anaconda3/envs/py38/lib/python3.8/site-packages/sklearn/cluster/_kmeans.py:870: FutureWarning: The default value of `n_init` will change from 10 to 'auto' in 1.4. Set the value of `n_init` explicitly to suppress the warning
  warnings.warn(
/home/coco/anaconda3/envs/py38/lib/python3.8/site-packages/sklearn/cluster/_kmeans.py:870: FutureWarning: The default value of `n_init` will change from 10 to 'auto' in 1.4. Set the value of `n_init` explicitly to suppress the warning
  warnings.warn(
<class 'node2vec.edges.HadamardEmbedder'>
NMI: 0.04336246478827236
Homogeneity: 0.0383178531539123
Completeness: 0.05011351404941123
V-Measure: 0.043428985282038965
<class 'node2vec.edges.AverageEmbedder'>
NMI: 0.11206093720015026
Homogeneity: 0.10817496918905492
Completeness: 0.11635805522328385
V-Measure: 0.11211739628609738
<class 'node2vec.edges.WeightedL1Embedder'>
NMI: 0.16558117117175825
Homogeneity: 0.16557714823761863
Completeness: 0.16568764408717976
V-Measure: 0.16563237773404058
<class 'node2vec.edges.WeightedL2Embedder'>
NMI: 0.1349652677966787
Homogeneity: 0.1337881599748603
Completeness: 0.1362723387302234
V-Measure: 0.13501882386803338

- NMI(Normalized Mutual Information)

  • 두 개의 군집 결과 비교

  • 0~1이며 1에 가까울수록 높은 성능

- Homogeneity

  • 하나의 실제 군집 내에서 같은 군집에 속한 샘플들이 군집화 결과에서 같은 군집에 속할 비율

  • 1에 가까울수록 높은 성능

- Completeness

  • 하나의 예측 군집 내에서 같은 실제 군집에 속한 샘플들이 군집화 결과에서 같은 군집에 속할 비율

  • 0~1이며 1에 가까울수록 높은 성능

- V-measure

  • Homogeneity와 Completeness의 조화 평균

  • 0~1이며 1에 가까울수록 높은 성능

  • 비지도 학습에 이상치 탐지 방법

  • k-means/LOF/One-class SVM 등이 있다.. 한번 같이 해보자.

  • 조금씩 다 커졌넹..

- 지도학습에서 정상거래에서 다운샘플링을 했는데

만약, 사기거래에서 업샘플링을 하게되면 어떻게 될까?