How to Build a Customer Churn Warning Model with R and Discover

Churn warning is crucial for enterprises to predict and retain valuable customers early.

Data Loading

In a Discover Notebook, select R as the interpreter and load the built‑in dataset losingWarn, which contains three columns: trade_time (transaction timestamp), user_id (customer ID), and consume_amt (transaction amount).

library(discoverR)
library(ggplot2)
discover.init()

data("losingWarn")
df_losingWarn <- createDataFrame(losingWarn)

Data Processing and Analysis

Convert the transaction time to month and year, filter for the year 2014, and compute monthly transaction counts and amounts.

df_losingWarn <- df_losingWarn %>%
  mutate(trade_month = month(df_losingWarn$trade_time)) %>%
  mutate(trade_year = year(df_losingWarn$trade_time))

df_losingWarn <- df_losingWarn %>% filter("trade_year == 2014")

Summarize per‑user monthly statistics:

colName <- c("user_id", "trade_month", "consume_amt")
df_user_summary <- select(df_losingWarn, as.list(colName)) %>%
  groupBy("user_id", "trade_month") %>%
  summarize(trade_num = n(df_losingWarn$user_id),
            trade_amt = sum(df_losingWarn$consume_amt))

Summarize overall monthly totals:

colName <- c("trade_month", "trade_num", "trade_amt")
df_trade_summary <- select(df_user_summary, as.list(colName)) %>%
  groupBy("trade_month") %>%
  summarize(sum_trade_num = sum(df_user_summary$trade_num),
            sum_trade_amt = sum(df_user_summary$trade_amt))

Visualize transaction amount distribution (log‑scaled) and customer count trends.

Feature Extraction

Define a GetFeature function that extracts, for a given time window, the last trade month, maximum amount, maximum transaction count, total amount, total count, and their averages.

GetFeature <- function(data, startMonth, endMonth) {
  result <- data %>%
    filter(paste("trade_month >=", startMonth, "and trade_month <=", endMonth))
  month <- as.integer(endMonth) - as.integer(startMonth) + 1
  out <- result %>% groupBy("user_id") %>% summarize(
    lastTradeMonth = max(result$trade_month),
    maxAMT = max(result$trade_amt),
    maxNum = max(result$trade_num),
    sumAMT = sum(result$trade_amt),
    sumNum = sum(result$trade_num)
  )
  out %>% mutate(
    averageAMT = out$sumAMT / month,
    averageNum = out$sumNum / month
  )
}

Labeling

Define a GetLabel function that marks a user as "losing" if the total transaction count in the observation window is zero, otherwise as "remain".

GetLabel <- function(data, startMonth, endMonth) {
  out <- data %>% filter(paste("trade_month >=", startMonth, "and trade_month <=", endMonth)) %>%
    group_by("user_id") %>%
    summarize(sumNum = sum(data$trade_num))
  out <- out %>% mutate(label = ifelse(out$sumNum > 0, "remain", "losing")) %>%
    select(as.list(c("user_id", "label")))
  names(out) <- c("id", "label")
  out
}

Model Building

Combine features and labels for the training window (June‑September) and train a random‑forest classifier.

GetModelData <- function(data, startMonth, endMonth, observeSize) {
  feature <- GetFeature(data, startMonth, endMonth)
  labelStartMonth <- endMonth + 1
  labelEndMonth <- endMonth + observeSize
  label <- GetLabel(data, labelStartMonth, labelEndMonth)
  out <- join(feature, label, feature$user_id == label$id, "left_outer") %>% drop("id")
  out %>% mutate(label = ifelse(out$label == "remain", "remain", "losing"))
}

training_data <- GetModelData(df_user_summary, 6, 9, 2)
rf_model <- txRandomForest(data = training_data,
                           formula = label ~ lastTradeMonth + maxAMT + maxNum + sumAMT + sumNum + averageAMT + averageNum,
                           type = "classification")

Prediction

Apply the model to the test window (July‑October) and obtain predictions.

test_data <- GetModelData(df_user_summary, 7, 10, 2)
result <- predict(rf_model, test_data)
showDF(result)

Model Evaluation

Evaluate using ROC AUC and inspect feature importance.

areaUnderROC <- txBinaryClassificationEvaluator(result, metricName = "areaUnderROC",
                                                probabilityCol = "probability",
                                                labelCol = "label",
                                                labels = summary(rf_model)$labels)
areaUnderROC

variable_importance <- importance(rf_model)
variable_importance

The evaluation shows that averageNum and sumNum are the most influential features, confirming the earlier hypothesis that a user's average and total transaction counts strongly affect retention.

Conclusion

This case study outlines the complete analytical pipeline for building a churn warning model in Discover, from data ingestion to model evaluation, illustrating how Discover’s built‑in functions simplify predictive analytics and help enterprises implement timely marketing actions to improve customer value.