Graph Neural Net, Recommender System 학습 기록

Written by JunPyo Park

그래프 딥러닝, 추천 시스템 공부한 내용 실습을 위해 생성한 repo

Lecture¶

CS224W: Machine Learning with Graphs / Stanford / Winter 2021¶

Course Info

CS224W | Home
Lectures

CS224W: Machine Learning with Graphs | 2021 | Lecture 1.1 - Why Graphs

Books¶

Graph Neural Network¶

Graph Representation Learning Textbook(CS224W 교재)

Yao Ma, Jiliang Tang - Deep Learning on Graphs-Cambridge University Press (2021)

Recommender System¶

Charu C. Aggarwal (auth.) - Recommender Systems_ The Textbook-Springer International Publishing (2016)

Matrix and Tensor Factorization Techniques for Recommender Systems by Panagiotis Symeonidis, Andreas Zioupos (auth.)

Network Theory¶

Albert-Laszlo Barabasi - Network Science-Cambridge University Press (2016)

Graph Recommender System Papers¶

Matrix_Factorization_Techniques_for_Recommender_Systems
BPR: Bayesian Personalized Ranking from Implicit Feedback
Neural Collaborative Filtering (NCF)
Neural Graph Collaborative Filtering (NGCF)
LightGCN
UltraGCN

GRL 기초 논문¶

첫 실습 Cocktail DB Visualizer Project¶

Networkx 모듈을 활용하여 네트워크 만들기 실습

연습용으로 사용할 DB¶

IBA(International Bartenders Association - IBA) 칵테일 제조 레시피 DB

https://github.com/JunPyoPark/cocktails

집에 쌓아둔 술들을 잘 활용하기(?) 위한 선택

cocktails/src/data/cocktails.json 파일을 보면 칵테일 레시피 조합이 json 형태로 DB화 되어 있음

ingredient를 노드로 사용할 예정이며 특정 레시피에 재료가 같이 들어가 있으면 서로 엣지 연결

import networkx as nx
import requests, json
import numpy as np

Read json from recipe DB¶

json_path = 'https://raw.githubusercontent.com/JunPyoPark/cocktails/master/src/data/cocktails.json'
url = requests.get(json_path)
text = url.text
data = json.loads(text)

len(data) # 80개의 레시피

80

data[0]

{'iba': True,
 'name': 'Vesper',
 'colors': '#D88317',
 'glass': 'martini',
 'category': 'Before Dinner Cocktail',
 'ingredients': [{'unit': 'cl', 'amount': 6, 'ingredient': 'Gin'},
  {'unit': 'cl', 'amount': 1.5, 'ingredient': 'Vodka'},
  {'unit': 'cl', 'amount': 0.75, 'ingredient': 'Lillet Blonde'}],
 'garnish': 'Lemon twist',
 'preparation': 'Shake and strain into a chilled cocktail glass.'}

ingredient_list = data[0]['ingredients']
ingredient_list

[{'unit': 'cl', 'amount': 6, 'ingredient': 'Gin'},
 {'unit': 'cl', 'amount': 1.5, 'ingredient': 'Vodka'},
 {'unit': 'cl', 'amount': 0.75, 'ingredient': 'Lillet Blonde'}]

Create Graph(Network) from recipe DB¶

import networkx as nx
from itertools import combinations 
G = nx.Graph()

total_ingredients = []
for recipe_num in range(len(data)):
    recipe = data[recipe_num]
    ingredients = recipe['ingredients']
    
    ingredient_list = []
    for ingredient in ingredients:
        try:
            ingredient_list.append(ingredient['ingredient'])
        except:
            pass
    
    # add nodes
    G.add_nodes_from(ingredient_list)
    
    # add edges
    edges = list(combinations(ingredient_list, 2))
    G.add_edges_from(edges)

Visualize¶

import matplotlib.pyplot as plt

노드의 Degree를 계산하여 시각화 할 때 사용합니다. (연결이 많이 된 노드를 크게 그리기 위함)

dig = dict(nx.degree(G)) # degree of network
d = np.array(list(dig.values()))

pos=nx.spring_layout(G,k=2) # positions for all nodes
plt.figure(figsize=(20,20)) #Control figure size
nx.draw(G,pos,with_labels=True
        ,node_size = d * 300 # node size (degree에 비례한 노드 크기)
        ,node_color= d # data for label color (degree에 따른 컬러맵 사용)
        ,font_size= 12 # labelsize
        ,cmap=plt.cm.YlOrRd # color_map
       )
plt.title('Cocktail Recipe DB Network',size=20);
plt.savefig('Graph.png', format="PNG") # Save figure

크기가 커서 웹에서 쫌 깨져보이네요.... 이미지 크게 보기

결론: 나머지 큰 재료들은 가지고 있으니까 진 한병 사면 더 많은 칵테일을 말아 먹을 수 있겠다!!!