TAFFC2018 - "Novel audio features for music emotion recognition"

This repository contains (WIP) scripts to reproduce the classification results obtained in our TAFFC paper [1].

At the moment only one script is available (best_result.R), used to obtain the F1-Score of 0.7651 with the best 100 features (standard and novel).

Data

The features extracted, annotations, feature ranking and metadata are all available inside the data/ folder. For more details check our website and the original paper [1].

Installation

The classification script is written in R language. Thus, R is needed and RStudio is recommended too.

For reproducibility purposes this project uses the renv package/dependency manager. When the project is loaded the exact versions of the packages used (and R) are installed locally. It should also set the working directory to the current folder.

Usage

The script can be run with Rstudio or using R command line directly. Both should open the .Rprofile file and configure the environment. In both cases you should get the following output

* Project '<PATH_TO_PROJECT>/TAFFC2018' loaded. [renv 0.9.2]

RStudio

Open the TAFFC2018.Rproj file using RStudio and wait for the initial environment setup. After that run:

source('best_result.R')

R Command Line

Open an R command line in the project folder, e.g., by executing:

X:\TAFFC2018>"C:\Program Files\R\R-3.6.1\bin\R.exe"

After the environment configuration just run:

source('best_result.R')

Expected Output

Seed Value =  1 (replicability purposes)
FEATURES USED ( 100 ): SET = PANDA TAFFC2018 ( 900 ) 10 fold cv x 20 reps / svm type = C-classification / kernel = radial 
quadrant_annotations
 Q1  Q2  Q3  Q4 
225 225 225 225 
SVM params optimized: cost = 8 / gamma = 0.001953125 
Accuracy = 0.7619892 (std = 0.04071989 )
Precision: macro weighted = 0.7683199 (std = 0.03988924 )
Recall: macro weighted = 0.7619892 (std = 0.04071989 )
F1-Score: macro weighted = 0.7651222 (std = 0.04014196 )
Q1: Precision = 0.7477606 / Recall = 0.8162222  / F1 Score = 0.780493 
Q2: Precision = 0.8888889 / Recall = 0.848  / F1 Score = 0.8679632 
Q3: Precision = 0.7184797 / Recall = 0.7015556  / F1 Score = 0.7099168 
Q4: Precision = 0.697796 / Recall = 0.6824444  / F1 Score = 0.6900348 
Confusion Matrix: 
      y_pred
y_true     Q1     Q2     Q3     Q4
    Q1 183.65  14.10   9.10  18.15
    Q2  23.55 190.80   6.80   3.85
    Q3  14.30   8.35 157.85  44.50
    Q4  24.10   1.40  45.95 153.55

References

[1] Panda, R., Malheiro, R., & Paiva, R. P. (2018). Novel audio features for music emotion recognition. IEEE Transactions on Affective Computing – TAFFC, 1–1. http://doi.org/10.1109/TAFFC.2018.2820691