Enabling Privacy-Preserving Concept Stock Recommendation with Federated Learning

MPhil Thesis Defence


Title: "Enabling Privacy-Preserving Concept Stock Recommendation with 
Federated Learning"

By

Mr. Zhuoyi PENG


Abstract

The prosperity of the stock market has witnessed the rapid booming of the stock 
tradings, especially the tradings from individual investors. In all major stock 
exchanges, individual investors are of vital importance, and even dominate the 
stock market by pumping up the stock prices. Among stock markets and individual 
investors, there is one important stock phenomenon named concept stock, where 
stock tradings are associated with a given topic or theme. In this phenomenon, 
topics/themes with special connotations are so-called concepts, and investors 
usually trade concept-related stocks in large volumes. Many social media 
platforms would recommend concept stocks to aid concept stock tradings and thus 
build an automated system to label concept stock for emerging unseen concepts. 
However, the existing automated recommendation system only uses public social 
media (e.g., news and firm reports), which are inadequate to model concept 
stocks accurately and timely. Flourishing in social media platforms, individual 
traders usually use private messages to share their trading ideas like the 
potential concept stocks. These private messages are valuable resources to 
capture the potential concepts, stocks, and the associations between concepts 
and stocks. However, the key challenge is how to build a concept stock 
recommendation system from these messages without compromising user privacy. In 
this thesis, we propose to use the Federated Learning framework to enable 
privacy-preserving concept stock recommendation using private messages to 
capture concept stock information with accuracy and timeliness We first propose 
a federated concept stock recommendation baseline named Federated Meta 
Embedding (FedME), which jointly learns embedding from private messages and 
public social media. To further improve the performance, we also propose 
Federated Domain Meta Embedding (FedDME) to enhance embedding with domain 
expert knowledge in labeling concept stocks. Extensive experiments on two 
concept stock datasets show our proposed two methods improve the current 
concept stock recommendation system substantially. For example, our FedME 
outperforms baseline with Mean Average Precision (MAP) improvement on 
Jinrongjie dataset from 0.2624 to 0.3670, and our FedDME yields 0.0911 higher 
MAP than FedME.


Date:  			Wednesday, 18 August 2021

Time:			4:00pm - 6:00pm

Zoom meeting: 
https://hkust.zoom.us/j/95963211676?pwd=eE5XTUVwaU9KNDdsQ21JSW12ZS9ydz09

Committee Members:	Dr. Kai Chen (Supervisor)
 			Dr. Qifeng Chen (Chairperson)
 			Dr. Yi Yang (ISOM)


**** ALL are Welcome ****