Predicting AsiaYo Users’ Spending for Improving Search Results Travis Greene, Martin Hsia, Letitia She, Leo Lee
Business Goal Stakeholders: Assumptions: AsiaYo’s managers 1. Model trained only on previous bookings 2. Avg. AsiaYo customers are price sensitive Reduce time for user UX is improved by Improving the 3. All users are new users to search and decide predicting default sorting customers’ budgets (AY Sort) Challenge: Opportunity: 1. Plenty places to book Improve search results to a room for a trip increase conversion rate 2. Avg. 3 trips per year Better conversion rate
Data Mining Goal Goal Predict the amount users will spend nightly A predicted amount paid per night Outcome (numeric) Task Predictive and supervised task. We are taking past customers’ transaction data and predicting new users’ per night spending
Implementation User’s Predicted Budget $1000
INPUT OUTPUT Fri. Sat. Guests Nights Platform Days of the Week AVG. amount AVG. amount • Check-in/out day User/Accom. Accom. • Created at Country City paid/night paid/night day/month/time • Lead time
Data Description Data History order transaction data 50,546 rows 16 columns Pre-process 1. Remove internal test data, outliers, unnecessary rows 2. Bin time, Convert to day of week, keep months. Compute time differences between order creation & check-in date 3. Create new column #per_night Partition Training ~40000 data rows, Testing ~10000 data rows 80/20 split
5 x 5 cross validation Methods Ensemble RMSE 846.08 36% Improvement from NAIVE
glm Performance Evaluation residuals_glm residuals_gbm gbm
Recommendations Reducing Model Prediction Error Algorithmic Considerations Business Policy Country/City models ● Connect previous transaction history to ● Prediction intervals ● search results Booking lead times ● Speed vs. accuracy ● Use UTM Source data as predictor ● Booking time of day ● trade-off Filtering categories based on counts ● Booking behavior on key dates ● Hyper-parameter tuning ● Collect more ● Log (price) ● personal/demographic data
Recommend
More recommend
