K-means based user portrait clustering model

One 、 summary

         Use K-means The main purpose of user clustering is to realize the film recommendation system of user portrait , The recommendation consists of two parts , The first part realizes movie recommendation through collaborative filtering , The second part , Build user portrait to realize movie recommendation , The implementation is realized through static attributes ： User's gender , Age , regional , role （ Student 、 Office worker 、 unemployed ）、 regional 、 Marital status , Use clustering （K-mens） The algorithm divides the crowd , Realize the movie recommendation of the same crowd , Secondly, dynamic properties ： By recording user behavior data , Search record usage word2vec The model realizes similar search keywords and recommends movies .( I will explain the recommendation system later )

         The language used in the text is python3.6 The above implementation , The tool used is jupyter-notebook Development , Mainly used python Of sklearn、pandas、faker、matplotlib Wait for the library to realize .

        Abstract ： This article mainly describes the second part of the use of K-means Clustering builds user portrait model to realize the recommendation of movies loved by the same users .

Two 、 theory

        K-means It is an unsupervised clustering algorithm , The main idea is to randomly select K Objects as the initial clustering center . Then calculate the distance between each object and each seed cluster center , To find similar objects .

3、 ... and 、 Realization

• Data acquisition

         We need data information before clustering ： Including name, 、 Age 、 Gender 、 Address 、 Type of work 、 Marital status, etc , But our thorny problem , Where does the data come from , At present, there are two ways to obtain data ： First, you can use web crawler technology to crawl information related personal information , The second kind python Have a powerful library faker It can help us realize the automatic simulation generation of user information .faker You can refer to the blog I wrote before ~ faker Generating user information

from faker import Faker
import pandas as pd
import numpy as np
from collections import OrderedDict
fake = Faker(locale='zh_CN')
sex_dict = OrderedDict([(' male ', 0.52), (' Woman ', 0.48)])
married_dict = OrderedDict([(' unmarried ', 0.4), (' married ', 0.5), (' Divorce ', 0.1), ])
work_dict = OrderedDict([(' on-the-job ', 0.7), (' freelance ', 0.3), ])
type_dict = OrderedDict([(' white collar ', 0.45), (' Teachers' ', 0.1), (' Worker ', 0.2), (' Civil servants ', 0.1), (' sales ', 0.15)])
name = []
age = []
address = []
sex = []
work = []
married = []
type = []
#  Number of people generated 
num = 1000
#  Generate name age address 
for i in range(num):
    name.append(fake.name())
    age.append(fake.random_int(min=16,max=45))
    address.append(fake.province())
    sex.append(fake.random_element(sex_dict))
    work.append(fake.random_element(work_dict))
    type.append(fake.random_element(type_dict))
    married.append(fake.random_element(married_dict))
data = {'name':name,'age':age,'sex':sex,'address':address,'work':work,'married':married,'type':type}
users = pd.DataFrame.from_dict(data)

• Data processing

         Above, we have successfully used faker The library generates user information , But let's think about , It is impossible for computer algorithms to understand the meaning of Chinese , So how to make kmeans The model knows the meaning of these data ？ We need to vectorize the data generated above , There are many methods of text vectorization , for example ：one-hot、tfidf wait , Of course, different vectorization methods have their own advantages and disadvantages , If you are interested, you can learn about it alone , There is no detailed answer here .

         In this article, we use one-hot The idea is to transform the data of different dimensions into data vectors by dictionary mapping .

sex_map = {' male ':0, ' Woman ':1}
married_map = {' unmarried ':0,' married ':1,' Divorce ':2}
work_map = {' on-the-job ':0, ' freelance ':1}
type_map = {' white collar ':0,' Teachers' ':1,' Worker ':2,' Civil servants ':3,' sales ':4}
addr_map = {' Shanghai ':0,' Yunnan Province ':1,' Inner Mongolia Autonomous Region ':2,' The Beijing municipal ':3,' Taiwan Province ':4,' Jilin Province ':5,' Sichuan Province ':6,' tianjin ':7,' Ningxia Hui Autonomous Region ':8
            ,' Anhui Province ':9,' Shandong Province ':10,' Shanxi Province ':11,' Guangdong province, ':12,
            ' Guangxi Zhuang Autonomous Region ':13,' Xinjiang Uygur Autonomous Region ':14,' Jiangsu Province ':15,' Jiangxi Province ':16,' Hebei Province ':17,' Henan province ':18,' Zhejiang Province ':19,' Hainan ':20,' Hubei province ':21,' Hunan province ':22,' Macao SAR ':23,' Gansu Province ':24,
            ' Fujian Province ':25,' Tibet Autonomous Region ': 26,' Guizhou Province ':27,' Liaoning Province ':28,' Chongqing City ':29,' Shaanxi Province ':30,' Qinghai Province ':31,' Hong Kong Special Administrative Region ':32,' Heilongjiang Province ':33} 

         And then use pandas Medium map Method vectorize each dimension data according to the above dictionary type , After conversion, as shown in the figure below ：

users['married'] = users['married'].map(married_map)

•   Model implementation

         We use sklearn Machine learning library creation kmeans Model , What we need to pay attention to is that we are using kmeans Before classification , The data is normalized once .

         The purpose of data standardization is to readjust the value of each dimension of data , Make the final data vector fall in [0,1] Within the interval , After standardized treatment , The original data is transformed into dimensionless index evaluation value , Each index value is at the same quantity level , Comprehensive evaluation and analysis can be carried out . If the original index value is directly used for analysis , It will highlight the role of indicators with higher values in comprehensive analysis , Relatively weaken the role of indicators with low numerical level .

         I will standardize the steps and kmeans The model makes a connection, so they are all put together , The implementation code is as follows ：

model = Pipeline([
    ('BN',preprocessing.StandardScaler()),('KMS',KMeans())
])
model.fit(X)
#  Save the model 
joblib.dump(model,  "./model/model.pkl")
model = joblib.load("./model/model.pkl")
y_pred = model.predict(X)
#  Store the prediction results in the vectorized total data 
users['lable'] = y_pred

•   The model is saved and tested

model = joblib.load("./model/model.pkl")
y_pred = model.predict([X[0]])
y_pred

• Visualization of clustering results

         Use adviz The principle of visualization is to map a series of points in multi-dimensional space to two-dimensional space through nonlinear method .

import matplotlib.pyplot as plt
from pandas.plotting import radviz
"""
adviz The visualization principle is a visualization technology that maps a series of points in multi-dimensional space to two-dimensional space through nonlinear method , It is a multi-dimensional visualization method based on the design idea of circular parallel coordinate system . circular m Bar radius representation m Dimensional space , Use a point in the coordinate system to represent many information objects , Its realization principle refers to the force balance theorem of objects in physics .
       mm Points in dimensional space {Ai1,Ai2,...,AimAi1,Ai2,...,Aim} The position mapped to the two-dimensional visual space is determined by the spring gravity analysis model .
       First, divide a circle into  mm Share , The number of bisectors is determined by the data dimension , Record the bisection points as {R1,R2,...,RnR1,R2,...,Rn}, A spring is fixed on each bisection point , The elastic coefficient of each spring is different , Spring  jj The coefficient of elasticity is AijAij, That is the first.  jj Value of dimension , all mm The other end of a spring is fixed on a small circle , If the small circle finally reaches an equilibrium position  Bi={Xi,Yi}TBi={Xi,Yi}T, Then the elastic balance point  {Xi,Yi}{Xi,Yi} Namely m Dimensional space points { Ai1,Ai2,...,AimAi1,Ai2,...,Aim} Projection point in two-dimensional space .
       Blog ：https://blog.csdn.net/Haiyang_Duan/article/details/78985225
"""
plt.figure(' User portrait clustering ',figsize=(15, 7.5), dpi=80)
plt.title('radviz')
radviz(users.loc[:,'age':], class_column='lable')
plt.show()