A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Efficient Link Prediction with Clusterized Batch Training in Graph Convolution Network
https://orcid.org/http://orcid.org/0000-0003-1862-3671
In the world of information, data are represented in different forms. Much real world information can be represented in the form of a network where an entity is linked or related with another. These relationships depend on different factors like entity types, strength of relationship, dynamic or static context and thus, are complex in nature. Many researches are conducted for the analysis of the network through machine learning approach. Generally, the prime motive of such research is to build an efficient model to learn entity classification, link prediction and community detection. This experiment focuses on the link prediction problem in the homogeneous citation networks; Cora, CiteSeer and PubMed. This research anchors the link completion section within the link prediction problem. The methodology in this experiment, to be termed as CLGCN (Cluster Link Graph Convolution Network), attempts to build an efficient method using Graph Convolution Networks for link predictions. The approach here is preparation of batches from clusters generated from an efficient clustering algorithm (METIS) rather than selecting the random instances from the data to form the batch for training. We explore how the preparation of batches this way affects the performance of the model and how different this approach is compared with other approaches in terms of accuracy, time complexity and memory usage. It is observed that the methodology results drastic reduction in memory usage and training time as compared to SAGEL (GraphSAGE Link Prediction), and also, memory usage is lower than that in original FGCN (Full Batch Graph Convolution Network). This leads to conclude that the methodology undertaken in this experiment improves the training time and memory usage maintaining the performance of the model in terms of accuracy and loss.
Efficient Link Prediction with Clusterized Batch Training in Graph Convolution Network
https://orcid.org/http://orcid.org/0000-0003-1862-3671
In the world of information, data are represented in different forms. Much real world information can be represented in the form of a network where an entity is linked or related with another. These relationships depend on different factors like entity types, strength of relationship, dynamic or static context and thus, are complex in nature. Many researches are conducted for the analysis of the network through machine learning approach. Generally, the prime motive of such research is to build an efficient model to learn entity classification, link prediction and community detection. This experiment focuses on the link prediction problem in the homogeneous citation networks; Cora, CiteSeer and PubMed. This research anchors the link completion section within the link prediction problem. The methodology in this experiment, to be termed as CLGCN (Cluster Link Graph Convolution Network), attempts to build an efficient method using Graph Convolution Networks for link predictions. The approach here is preparation of batches from clusters generated from an efficient clustering algorithm (METIS) rather than selecting the random instances from the data to form the batch for training. We explore how the preparation of batches this way affects the performance of the model and how different this approach is compared with other approaches in terms of accuracy, time complexity and memory usage. It is observed that the methodology results drastic reduction in memory usage and training time as compared to SAGEL (GraphSAGE Link Prediction), and also, memory usage is lower than that in original FGCN (Full Batch Graph Convolution Network). This leads to conclude that the methodology undertaken in this experiment improves the training time and memory usage maintaining the performance of the model in terms of accuracy and loss.
Efficient Link Prediction with Clusterized Batch Training in Graph Convolution Network
https://orcid.org/http://orcid.org/0000-0003-1862-3671
In the world of information, data are represented in different forms. Much real world information can be represented in the form of a network where an entity is linked or related with another. These relationships depend on different factors like entity types, strength of relationship, dynamic or static context and thus, are complex in nature. Many researches are conducted for the analysis of the network through machine learning approach. Generally, the prime motive of such research is to build an efficient model to learn entity classification, link prediction and community detection. This experiment focuses on the link prediction problem in the homogeneous citation networks; Cora, CiteSeer and PubMed. This research anchors the link completion section within the link prediction problem. The methodology in this experiment, to be termed as CLGCN (Cluster Link Graph Convolution Network), attempts to build an efficient method using Graph Convolution Networks for link predictions. The approach here is preparation of batches from clusters generated from an efficient clustering algorithm (METIS) rather than selecting the random instances from the data to form the batch for training. We explore how the preparation of batches this way affects the performance of the model and how different this approach is compared with other approaches in terms of accuracy, time complexity and memory usage. It is observed that the methodology results drastic reduction in memory usage and training time as compared to SAGEL (GraphSAGE Link Prediction), and also, memory usage is lower than that in original FGCN (Full Batch Graph Convolution Network). This leads to conclude that the methodology undertaken in this experiment improves the training time and memory usage maintaining the performance of the model in terms of accuracy and loss.
Efficient Link Prediction with Clusterized Batch Training in Graph Convolution Network
J. Inst. Eng. India Ser. B
Khadka, Ekesh (author) / Adhikari, Nanda Bikram (author)
Journal of The Institution of Engineers (India): Series B ; 104 ; 693-701
2023-06-01
9 pages
Article (Journal)
Electronic Resource
English
Batch , Links , Nodes , Edges , Network , Graph , Memory , Time , Prediction , Clusters Engineering , Communications Engineering, Networks
IWQP4Net: An Efficient Convolution Neural Network for Irrigation Water Quality Prediction
| DOAJ | 2023
| DOAJ | 2022
A deep learning traffic flow prediction framework based on multi-channel graph convolution
| Taylor & Francis Verlag | 2021