My general areas of research are data mining, recommender systems, and business analytics,
with applications in urban computing and supply chain management. My primary research
focuses on predictive analytics, constrained clustering, large-scale optimization and their combinations for solving emerging supply chain management problems. I started my research on Bike
Sharing System and conducted numerous researches on large-scale station inventory rebalancing and facility site selection problems. Furthermore, I expanded my research to multi-product
supply chain analytics and developed intelligent ordering, replenishment, and transshipment systems for perishable products and fast fashion.
Big Data Analytics in Supply Chain Management
Rebalancing Bike Sharing Systems: A Data-Driven Hierarchical Optimization.
Facility Site Selection and In-service Area Expansion.
Due to the geographical and temporal unbalance of bike usage demand, a number of bikes need to be
reallocated among stations so as to maintain a high service level of the bike sharing system. To
optimize such bike rebalancing operations, I have addressed two challenges: (1) to accurately
predict the station-level bike pick-up and drop-off demand, so as to determine the rebalancing
target for each station, and (2) to efficiently solve the large-scale multiple capacitated vehicle
routing problems by developing a Capacity Constrained K-center Clustering (CCKC) algorithm
to decompose the multi-vehicle routing problem into smaller and tractable single-vehicle routing
Intelligent Ordering, Replenishment, and Transshipment Systems.
Another key to the success of a bike sharing system is the in-service area expansion and the bike demand prediction for expansion areas. There are two major challenges in this demand prediction
problem: (1) the bike transition records are not available for the expansion area and (2) sta-
tion level bike demands have big variances across the urban city. To address these challenges,
I have developed a hierarchical station bike demand predictor which analyzed bike demands
from functional zone level to station level.
The bike station site selection was studied by integrating a station bike demand & operation
cost prediction model (based on Artificial Neural Network Predictor) and a site optimization model (based on Genetic Algorithm).
Industrial Collaboration: LineZone Data Technology 2018
Spatioal and Temporal Data Analytics
Mobile Users' Activity Discovery under Encrypted Internet Traffic
The system is built to keep the inventory level of retailers at its target level to improve the service level while reducing
the inventory cost. I built three modules to support the system: retailer aggregation and sales prediction, dynamic inventory model, and product
replenishment and transshipment optimization. The sales of clusters of retailers were predicted by our sequence to sequence time series prediction model.
The dynamic inventory model determined the factory production quantity based on the prediction value, prediction error distribution, and
current inventory level. Finally, before distributing the products to retailers, I implemented a Mixed Integer Linear Programming
model to determine the replenishment quantity from factory to retailers and the transshipment
quantity from oversupplied retailers to the retailers with shortages.
Route Recommender System for Taxi Drivers
In-App service usage analytics has recently become critical for mobile companies
to enhance user experiences and for Internet providers to provide intelligent network resource distributions.
A key task of In-App service usage analytics is to effectively classify mobile Internet traffic into different usage categories in
a real-time manner. We develop an online iterative mobile app traffic analyzer that comprises a recursive time continuity constrained
KMeans clustering (rCKC) algorithm for traffic flow segmentation and a Random Forest classifier for segmented traffic classification
based on time window representation with adaptable features.
The design goal of the recommender system for taxi drivers is to maximize
their net profits by following the recommended routes for searching passages. In particular, the system can provide an entire driving route
and the drivers are able to find a customer with the largest potential profit using our efficent recursive searching strategy.
The following problems have been investigated:
1) Dynamic taxi pick-up possibility prediction model;
2) Taxi driver Maximum Net Profit (MNP) route recommendation system;
3) Road rebalancing strategy for taxi crowdedness.