user_order_relationship_feature_engineering.R

source('load_data.R')
d = read_and_preprocess_data_file('data/BADS_WS1718_known.csv')
library(dplyr)

# 1st idea (the image I previously shared on WhatsApp)
# relationship between total orders (item_count, not transaction count) and return count (again the item_count)
# Are users who order more "wiser" in terms of picking the right items than small-volume users?
orders_and_returns_per_user = d %>%
  group_by(user_id) %>%
  summarize(nreturn=sum(return), norder=n())
orders_and_returns_per_user$return_ratio = orders_and_returns_per_user$nreturn/orders_and_returns_per_user$norder
orders_and_returns_per_user = arrange(orders_and_returns_per_user, norder)
plot(x=orders_and_returns_per_user$norder, y=orders_and_returns_per_user$return_ratio, type="p")

# 2nd idea
# do users return less with each returned item? In relationship to time, that is.
# I.e. do they "learn" over time that some things fit, and others do not?
d = arrange(d, order_date)
userids = sort(unique(d$user_id))
cumulated_returns = unlist(sapply(userids, function(x) cumsum(d$return[d$user_id == x]))) 
d$cumulated_returns[with(d, order(user_id))] = cumulated_returns
cumulated_itemcount = unlist(sapply(userids, function(x) cumsum(d$return[d$user_id == x] != 10))) 
d$cumulated_itemcount[with(d, order(user_id))] = cumulated_itemcount
d$volatile_return_ratio = d$cumulated_returns / d$cumulated_returns
plot(x=d$cumulated_itemcount, y=d$cumulated_returns, type="p")