Please use this identifier to cite or link to this item:
Title: HIN embedding learning by aspect for recommendation systems
Authors: Chen, Jieyi
Keywords: Engineering::Electrical and electronic engineering::Computer hardware, software and systems
Issue Date: 2020
Publisher: Nanyang Technological University
Abstract: Network representation learning is a graph-based machine learning task, and its applications have gradually spread throughout our daily lives. Its goal is to represent nodes of the network as vectors which are low-dimensional dense. And then use low- dimensional vectors to solve subsequent machine learning tasks. The main challenge in this research area is to find a way to represent or encode the graph structure so that machine learning models can easily use it. In recent years, the heterogenous information network composed of multiple nodes or edges has been regarded as a powerful modelling algorithm for fusing complex information and has been applied to many data mining tasks successfully. In addition, due to the flexibility of heterogeneous information networks in modelling data heterogeneity, it has also been used as recommendation method to characterize data connection. The recommendation algorithm under this setting is called “Recommendation method based on heterogeneous information networks (HINs)”. Existing recommendation method based on HINs mainly use the random walk to learn the representation of the network nodes, then enter the one-hot vector of nodes into the deep learning model such as Skip-Gram to access the low-dimensional embedding vector of the nodes. However, this method can only learn the topology of the network and ignores the semantic information. Due to the limitation of a single space, the node representation (also known as node embedding) learned from deep learning method is difficult to deal with the information loss caused by topology conflict. This report introduced an aspect embedding framework (AspEm) to mitigate information loss, thus provide a more accurate recommendation method. Firstly, we calculate the incompatibility of each aspect (any three or more node types can form an aspect) based on the similarity of their adjacency matrix. And then we select proper aspects combination base on several rules. There we regard each aspect as a single space and apply deep learning method to obtain node vectors separately. In order to maintain maximum information from every aspect, we splice node vectors learned from all the selected aspect to form the final node vectors. To verify the recommendation performance of AspEm, we implement the framework on two real-worlds datasets. The experiment result demonstrates the effectiveness of AspEm framework on solving the information loss.
Fulltext Permission: restricted
Fulltext Availability: With Fulltext
Appears in Collections:EEE Theses

Files in This Item:
File Description SizeFormat 
  Restricted Access
1.64 MBAdobe PDFView/Open

Page view(s)

Updated on Jan 31, 2023


Updated on Jan 31, 2023

Google ScholarTM


Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.