Happy mood detection in CAL500

Today, I am testing the discrimination ability of the sparse representation classifier (SRC) using auditory temporal modulations (ATMs) using the CAL500 dataset. The first application to autotagging of such features and such machine learning is: Y. Panagakis, C. Kotropoulos, and G. R. Arce, “Sparse multi-label linear embedding nonnegative tensor factorization for automatic music tagging,” in Proc. EUSIPCO, (Aalborg, Denmark), pp. 492-496, Aug. 2010. Here, we take SRC and apply ATMs (derivation described in B. L. Sturm, “Two systems for automatic music genre recognition: What are they really recognizing?,” in Proc. ACM MIRUM Workshop, (Nara, Japan), Nov. 2012. We also normalize the features such that each dimension is mapped to [0,1].)
To decide on the emotional label of a song, which has several ATMs, we pick the class giving the lowest sum of the errors over all frames.

When I create classes based on the tags:

tagsinclass1 = {'Emotion-Happy'};
tagsinclass2 = {'Emotion-Sad'};

I get the following statistics:

meansongaccuracy =
0.7254
normsongaccuracy =
0.6562
SONG TRUE
H 	 S
H 	 0.83 	 0.52
S 	 0.17 	 0.48
--- CLASS ---
Recall	 0.83	 0.48
Prec.	 0.79	 0.55
F   	 0.81	 0.51

When I create classes based on the tags:

tagsinclass1 = {'Emotion-Happy','NOT-Emotion-Sad'};
tagsinclass2 = {'Emotion-Sad','NOT-Emotion-Happy'};

and select only those excerpts that have 1 in both, I get the following statistics:

meansongaccuracy =
0.7091
normsongaccuracy =
0.6517
SONG TRUE
H 	 S
H 	 0.79 	 0.49
S 	 0.21 	 0.51
--- CLASS ---
Recall	 0.79	 0.51
Prec.	 0.79	 0.51
F   	 0.79	 0.51

Looks like we have an ok “happy” tagger, but its ability to choose “sad” is random.

Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s