Previous work overlooked sparse networks in which the majority of experts participated only in a few teams. These networks only accounted for a small portion of the history of collaboration and were computationally expensive. These shortcomings yield in poor future team prediction and high training costs.
Proposed method
The author proposes a variational neural network model namely the variational Bayesian neural network (vBnn) which leverages the Bayes theorem to assign uncertainty weights to the parameters.
My Summary
The proposed method delivers better performance than the existing state-of-the-art neural network-based model by Sapienza et al. (which is one of the runner-ups in this paper's test results aside from RRN by Wu et al.). However, I find it's a model which requires some deeper mathematical knowledge to fully understand the workings of the architecture as it uses the Bayes theorem to compute the uncertainty weights of the parameters. There's also a mention of needing to double the number of parameters for this computation to work. So it's definitely a complex architecture compared to something like the vanilla Transformer model, in my opinion. However, the paper only tested with one dataset, DBLP, with about 33,000 teams. I'd like to see how it performs in other datasets (i.e., a dataset with 150,000 teams or more).
Datasets
DBLP with 33,002 teams, 2,000 skills, and 2,470 experts
Link: https://dl.acm.org/doi/10.1145/3340531.3412140
Main problem
Previous work overlooked sparse networks in which the majority of experts participated only in a few teams. These networks only accounted for a small portion of the history of collaboration and were computationally expensive. These shortcomings yield in poor future team prediction and high training costs.
Proposed method
The author proposes a variational neural network model namely the variational Bayesian neural network (vBnn) which leverages the Bayes theorem to assign uncertainty weights to the parameters.
My Summary
The proposed method delivers better performance than the existing state-of-the-art neural network-based model by Sapienza et al. (which is one of the runner-ups in this paper's test results aside from RRN by Wu et al.). However, I find it's a model which requires some deeper mathematical knowledge to fully understand the workings of the architecture as it uses the Bayes theorem to compute the uncertainty weights of the parameters. There's also a mention of needing to double the number of parameters for this computation to work. So it's definitely a complex architecture compared to something like the vanilla Transformer model, in my opinion. However, the paper only tested with one dataset, DBLP, with about 33,000 teams. I'd like to see how it performs in other datasets (i.e., a dataset with 150,000 teams or more).
Datasets