hls__A Consumer Compensation System in Ride-hailing Service_2023_Yu

a transfer learning enhanced uplift modeling is designed to measure the elasticity
ls-type:: annotation
hl-page:: 1
hl-color:: yellow
a model predictive control based optimization is formulated to control the budget accurately
ls-type:: annotation
hl-page:: 1
hl-color:: yellow
Given a total compensation budget or an average compensation rate, find an optimal policy to subsidize queries so that the overall revenue is maximized.
ls-type:: annotation
hl-page:: 2
hl-color:: yellow
Consumer elasticity
ls-type:: annotation
hl-page:: 2
hl-color:: yellow
Consumer fairness
ls-type:: annotation
hl-page:: 2
hl-color:: yellow
Randomness in queries:
ls-type:: annotation
hl-page:: 2
hl-color:: yellow
Transfer Learning Enhanced Uplift Modeling
ls-type:: annotation
hl-page:: 2
hl-color:: yellow
abundant biased observational data and limited randomized data
ls-type:: annotation
hl-page:: 2
hl-color:: yellow
[:span]
ls-type:: annotation
hl-page:: 3
hl-color:: yellow
tabular input space and transfer learning
ls-type:: annotation
hl-page:: 2
hl-color:: yellow
Massive observational data is first fed into both inputs to pre-train the model
ls-type:: annotation
hl-page:: 3
hl-color:: red
RCT data is used to fine-tune using a different output layer
ls-type:: annotation
hl-page:: 3
hl-color:: red
Optimization Formulation
ls-type:: annotation
hl-page:: 3
hl-color:: yellow
use the mean of the historical query-wise elasticity to forecast the class-wise elasticity
ls-type:: annotation
hl-page:: 3
hl-color:: yellow
model predictive control (MPC) technology
ls-type:: annotation
hl-page:: 3
hl-color:: yellow
[:span]
ls-type:: annotation
hl-page:: 3
hl-color:: yellow
The features include query information (e.g., the origin, destination, time, weekday, and distance), spatial features (e.g., point of interest information, and order statistics in the same cells), subsidy information, and trading features (e.g., historical order placement rate). I
ls-type:: annotation
hl-page:: 4
hl-color:: yellow
The size of the common inner layers and output layer is set to 128, 64, and 32
ls-type:: annotation
hl-page:: 4
hl-color:: green
[:span]
ls-type:: annotation
hl-page:: 4
hl-color:: yellow
No Cluster Oracle
ls-type:: annotation
hl-page:: 4
hl-color:: yellow
Compared with the baseline, our system obtains a lower subsidy rate and higher revenue, for its accurate compensation, to achieve a higher ROI.
ls-type:: annotation
hl-page:: 4
hl-color:: yellow