bobye/glba

R code for Gated Latent Beta Allocation described in paper "Probabilistic Multigraph Modeling for Improving the Quality of Crowdsourced Affective Data", TAFFC

GLBA

This repository includes the R code for Gated Latent Beta Allocation (GLBA) and some other codes that implement baseline methods, which appeared in the following paper:

Jianbo Ye, Jia Li, Michelle Newman, Reginald B. Adams, Jr., James Z. Wang, "Probabilistic Multigraph Modeling for Improving the Quality of Crowdsourced Affective Data," IEEE Transactions on Affective Computing (To appear).

Dataset

This repository compiles a (ID,Score) dataset filter_four.csv for demonstration purpose, which is a derivative of dataset EmoSet described in X. Lu's dissertation work. The use of this demo set is restricted only to explore the functionality of code.

X. Lu, “Visual characteristics for computational prediction of aesthetics and evoked emotions,” Ph.D. dissertation, The Pennsylvania State Uni- versity, 2015, chapter 5. Online available

How to run

Load model

> GLBA

Load data (default: valence, the_metric = 1)

> preprocssing

where hypergraph stores the userid map hypergraph$oracles and inverse map hypergraph$inv_oracles.

Train GLBA to obtain user reliability

> result=glba_curve(hypergraph)

where the user reliability can be calculated by rowMeans(result$taus).

Remember valence is the best dimension to obtain reliability score. Later, if you want to load arousal data, you don't need to retrain. Change the_metric = 2 in preprocssing.r to load arousal data.

Obtain image confidences and scores

> experiment2