2015 TaoBao Challenge of User Classification

PengyangWang And I participated in a big data challenge game holded by Alibaba in our school,Beijing University of Posts and Telecommunications , in this autumn. This challenge required participants to classify customers of TaoBao,chinese Amazon, into 12 categories based on their online behaviors.Our solution had achieved 34% precision rate. In this article, we prefer to making a summarize of this challenge and our method.

code

Problem Description

Data Description

This is a Multi classification problem. We are given three files: log_train.csv, info_train.csv,log_test.csv.

Data in file info_train.csv contains Id and class of all users appeared in log_train.csv. The structure of info_train.csv is given in the table below .

Example
user_id class
254 1
456 3

Data Description
Col Name Type Description Comment
user_id int identify of user  
class int class of user number range from 1 to 12

log_train.csv and log_test.csv have same data structure .The contents of this two files is user’s online-shopping behaviors during a year. The name and meaning of all colums in this two files is given below.

Example
user_id item_id cat_id seller_id brand_id time_stamp action_type
254 1034 1206 1003 34 173 0
254 1498 1085 1243 4241 173 1

Data Description
Col Name Type Description Comment
user_id int identify of user  
item_id int identify of item  
cat_id int Category Id of The Item  
seller_id int Seller Id of The Item  
brand_id int brand Id of the Item  
time_stamp int When The Action Occured  
Action_type int Type Id of the action 0-click,1-collect,2-Buy,3-Delete

TASK

This challenge required participants to classify all the users appeared in log_test.csvinto 12 categories based on user’s online-shopping behavior.

Evaluation

The evaluation method is to calculate the precision rate of classification.The python code can be download here.

Solution

Perception of Question

Costumer’s number of all 12 categories in log_train.csv is given below

User’s number of 12 categories
class 1 2 3 4 5 6 7 8 9 10 11 12
number 12166 26830 18836 10066 8981 1999 6578 11915 8416 4163 3536 973

Class 2 has the highest numbers of users and class12 has the minimum ,only 937 ,users. The total number of users is 114459.

Major diffrences between online-behaviors of 12 classes customer’s

It’s important to understand all the key elements that can influence classification precision rate.According to our usual experience of online-shopping and the data we accessed from this competition,we concluded some potential meaningful features behavior of users:

  • Active Days of User

    We defined this term as the total number of days that user has online active record.We calculated activity days of all users in training set and plot the result to visualize the data.The R code to compute the active days of user and the figure of 12 categories activity days are given below.

     user_active_days<-sapply(train_user$V1,function(x){length(unique(train_data[train_data$user_id==x,]$time_stamp))})
    Active Days of 12 Categories

  • Action frequency of specific period

    In China,there are some essential date,such as Singles’ Day,National Day and Valentine’s Day,for online-consumer because of the discount. By analysis the log file, we found that people from different categories have different attitude towards those special date. For example, class 3 and 4 will more active in those period than ordinary days. In contrast, users in class 8 are more calm when faced the on sale.Even there is only a slightly variation between the 12 classes in this feature,but after we split those data into different time intervals,we can achieve a 13% precision rate by using this feature set(buying_frequency,clicking_frequency,action_frequency).

  • The ratio of Action Types in a time range

    From the head of this article,we know the action type is divided into 3 types.After analysis,We discovered that it’s really rare for behaviors of colloect and delete. So we ignore this two action type and just focus on the buying_clicking ratio.This ratio is refer to before a customer do a buying action how many clicking action or how many items this user viewed. The formulation of computing the buying_clicking ratio of one period is given below.

\[buying\_clicking\_ratio = \frac{clicking\_frequency}{buying\_frequency}\]

  • Brand of consumer goods

    Every record in log_train.csv have a feature named brand_idwhich represent the brand related with this action.We just consider buying_action ,we want to use the brand of items people consuming to classify users. We all know that in our life different brands are towards different group of people. So it’s also make sense that people from different categories will buy products from different brands.However,how to connect the brand with user’s categories is a problem.In our case we use K-meanson brand_id feature. K-means is a clustering algorithm in machine learning area. To use this on bran_id we require to construct a dataframe of brand_id.The dataframe is given below.Our data frame has 13 features, the first column is brand_id and the other features are the buying proportion of specific class users in 12 categories. There are 6304brand in this data.

    Brand DataFrame
    brand_id \(C_1\) \(C_2\) \(C_3\) \(C_4\) \(C_5\) \(C_6\) \(C_7\) \(C_8\) \(C_9\) \(C_{10}\) \(C_{11}\) \(C_{12}\)
    234 0.2 0.1 0 0 0 0.4 0 0 0 0 0 0.2
\[C_i\_brand_t = \frac{Numbers\_of\_Class_i\_Buying\_Brand_t}{\sum_{i=1}^12 Numbers\_of\_Class_i\_Buying\_Brand_t}\]

After we construct this dataframe ,we are using K-meanson Brand DataFrame .R code of this process is given below.


getKmeans_brand<-function(log,k)
{
brand_class<-table(log$brand_id,log$class)/rowSums(table(log$brand_id,log$class))

brand_kmeans_k<-kmeans(brand_class,centers = k)

brand_id<-rownames(brand_class)
brand_kmeans_k<-data.frame(brand_id,brand_kmeans_k$cluster)
return (brand_kmeans_k)

}

After we cluster brand into 14 calsses,then we update brand_id with their newest calss_id.

  • Category of consumers goods We also adopt same process on the categories of item as well as brands of item.The idea is as the same as we propose in previous part.

Data processing

All the R code of clean data process can be found here.

clean_data<-function(log,time=1:185,kmeans_k=6,user.all,brand_kmeans,cat_kmeans)

The argument time in clean_data function can split the whole log_train.csv into different part by time interval as your wish.We give a invoking example below.


 train_t1<-clean_data(log_train,time = 1:10,user.all = get_user(log_train),brand_kmeans = brand_kmeans,cat_kmeans = cat_kmeans)
  train_t2<-clean_data(log_train,time = 11:40,user.all = get_user(log_train),brand_kmeans = brand_kmeans,cat_kmeans = cat_kmeans)
  train_t3<-clean_data(log_train,time = 40:100,user.all = get_user(log_train),brand_kmeans = brand_kmeans,cat_kmeans = cat_kmeans)
  train_t4<-clean_data(log_train,time = 100:185,user.all = get_user(log_train),brand_kmeans = brand_kmeans,cat_kmeans = cat_kmeans)
  final_train<-merge_data(train_t1,train_t2,train_t3,train_t4)

In the above code, we split the whole 185 days into 4 parts.We also combine all the feature in 4 parts together by a merge_data function.

merge_data<-function(data1,data2,data3,data4,data5,data6,data7)
{
final_log<-merge(data1,data2,by.x="user_id",by.y="user_id")
final_log<-merge(final_log,data3,by.x="user_id",by.y="user_id")
final_log<-merge(final_log,data4,by.x="user_id",by.y="user_id")
final_log<-merge(final_log,data5,by.x="user_id",by.y="user_id")
final_log<-merge(final_log,data6,by.x="user_id",by.y="user_id")
final_log<-merge(final_log,data7,by.x="user_id",by.y="user_id")
#final_log<-merge(final_log,data7,by.x="user_id",by.y="user_id")
return(final_log)
}

we got our final dataframe final_train.Then we start to construct a classification model.

Modeling

We have tired many classification algorithm like Naive Bayes,Decision Tree,SVM ,Random forests,KNN and Adaboosting.We found that when using 5 descision trees in Adaboosting we got highest precision rate 34%.The modeling code can be found here.