Node Representation Learning in Bipartite Graphs
Published:
Node Representation Learning in Bipartite Graphs by Sajjad Athar (2021)
The Network Embedding research has been proved to be extremely useful for transforming large network data on to low dimensional embedding space and enables a large spectrum of vector based algorithms (link prediction, recommendation and classification etc.) to be applicable on the graph data. Most of the researcher’s focus have been on the homogeneous networks. Much less attention was dedicated towards the embedding methods for bipartite networks. However, there can be found extensive set of application which follow the bipartite network mining pattern e.g. e-commerce, recommendation systems etc. The network embedding related published work, we find in the literature, mostly deals with network topology while these networks also contain with rich textual attribute information. Considering these textual attributes along with the topological information would significantly improve the quality of the embedding vectors. In this thesis, it is proposed a novel embedding method Attributed and Structural Bipartite Network Embedding (ASBiNE). The ASBiNE incorporates both the network topological information in the context of inter partition and intra partition links and attribute information by generating proximities between nodes having attribute similarities. Both types of links are modelled separately and combined to produce final embeddings through a joint optimization framework. The novelty of the proposed method lies in dynamically controlling the attribute and topological information contributions in the embeddings. Extensive experiments have been performed on real life datasets and results are quite convincing against the state of the art baseline embedding methods.