Aligning vision-language for graph inference in visual dialog

Jiang, Tianling; Shao, Hailin; Tian, Xin; Liu, Chunping

doi:10.1016/j.imavis.2021.104316

Cited by 7 publications

(1 citation statement)

References 6 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, ML techniques without dimensionality reduction perform better with low-dimensional datasets Pei et al 31 2022 This work proposes scene graph semantic inference for cross-modal image and text matching. Using visual and textual scene graphs and graph convolutional networks, the technique analyses the local semantic correlations of inter-modal object relationships Jiang et al 32 2021 This paper proposes visual dialogue for establishing semantic relationships between visual and written content. Aligning visual and textual knowledge reduce the gap between modalities.…”

Section: Introductionmentioning

confidence: 99%

Application of wrapper based hybrid system for classification of risk tolerance in the Indian mining industry

Kumar

Bhattacharjee

2023

Sci Rep

View full text Add to dashboard Cite

The degree to which an individual is willing to take risks i.e., risk tolerance is often cited as a significant causal element in the majority of workplace accidents. It is essential to determine the risk tolerance level of miners and utilise their risk profiles to design improved training modules, safety, recruitment, and deployment policies. This paper aims to identify the most critical factors (or features) influencing miners’ risk tolerance in the Indian coal industry and develop a robust prediction model to learn their risk tolerance levels. To do end, we first conducted a questionnaire survey representing the complete feature set (with 36 features) among 360 miners and divided their responses into five classes of risk tolerance. Next, we propose a wrapper based hybrid system that combines particle swarm optimization (PSO) and random forest (RF) to train a multi-class classifier with a subset of features. In general, the proposed system selects the best feature subset by iteratively generating different feature combinations using the PSO and training an RF classifier model to assess the effectiveness of the generated feature subsets for the F1-score. At last, we compared the PSO-RF with four traditional classification methods to evaluate its effectiveness in terms of precision, recall, F1-score, accuracy, goodness-of-fit, and area under the curve.

show abstract