Paper: http://ceur-ws.org/Vol-2283/MediaEval_18_paper_54.pdf Youtube: https://youtu.be/BZvHNP_S92w Sergio Oramas, Dmitry Bogdanov, Alastair Porter, MediaEval 2018 AcousticBrainz Genre Task: A baseline combining deep feature embeddings across datasets. Proc. of MediaEval 2018, 29-31 October 2018, Sophia Antipolis, France. Abstract: In this paper we present a baseline approach for the MediaEval 2018 AcousticBrainz Genre Task that takes advantage of stacking multiple feature embeddings learned on individual genre datasets by simple deep learning architectures. Although we employ basic neural networks, the combination of their deep feature embeddings provides a significant gain in performance compared to each individual network. Presented by Dmitry Bogdanov

1. MediaEval 2018 AcousticBrainz Genre Task:
A baseline combining deep feature embeddings
across datasets
Sergio Oramas (Pandora Media Inc.)
Dmitry Bogdanov, Alastair Porter (Universitat Pompeu Fabra)

2. Features
● Use all features
● One-hot encode categorical features (key, scale)
● Standardize (zero mean, unit variance)
● In total 2669 input features

3. Single Dataset Model Architecture (Task 1)
● Basic DNN, only one hidden layer
● One model per dataset
● Model as feature extractor once trained
● Intermediate layer as a new 256-dim
feature embedding of each track

4. Fusion Architecture (Task 2)
● Obtain 4 feature embeddings per track
● Apply l2-norm to each one
● Concatenate embeddings
● Train fusion network

5. Validation Results (ROC AUC)
Discogs
Discogs + AllMusic
Discogs + AllMusic + Lastfm
Discogs + AllMusic + Lastfm + Tagtraum
0.7592
0.8713
0.8814
0.8863
● Macro ROC AUC improves combining feature embeddings

6. Maximizing macro F-score
● Network output: within [0,1]
● Plug-in rule approach
● Individual threshold for each genre/subgenre label
● The problem (Task 1): some labels are always
predicted
○ Infrequent labels + Uninformative classifier
○ We decided to keep those for simplicity
○ ~30 subgenres always predicted for Allmusic
○ This hurts the per-recording micro F-score

7. Test Results (Precision, Recall, F-score)

8. Conclusions
● Making use of all four datasets improves performance
● Our approach: Combine feature embeddings learnt by weak neural networks
○ More deep feature embeddings = Better predictions
○ Early fusion = No need to work on genre mapping
● What to optimize?
○ Task 1: Optimizing for per-label (macro) F-score worsens per-track micro F-score
● We ignored genre-subgenre hierarchy
○ Inheriting parent genres lead to worse performance scores

9. Future work
● ISMIR 2019
● Better individual NN models
○ More complex architectures (e.g., melbaseline 2018 or JKU 2017)
● How useful are different datasets for stacking?
○ What matters? (Size, taxonomy, source of annotations?)
○ Run an experiment on the intersection of all four datasets