[Paper Review] Multi-Label Classifier Chains for Bird Sound

Forrest Briggs, Xiaoli Z. Fern+1 more|arXiv (Cornell University)|Apr 22, 2013

Animal Vocal Communication and Behavior14 references6 citations

TL;DR

This paper proposes an ensemble of classifier chains (ECC) with a histogram-of-segments representation for multi-label bird sound classification, outperforming binary relevance in most cases and showing competitive performance against MIML methods despite lacking parameter tuning. The approach effectively models label correlations in multi-species bird recordings, achieving state-of-the-art results on two real-world datasets with complex acoustic environments.

ABSTRACT

Bird sound data collected with unattended microphones for automatic surveys, or mobile devices for citizen science, typically contain multiple simultaneously vocalizing birds of different species. However, few works have considered the multi-label structure in birdsong. We propose to use an ensemble of classifier chains combined with a histogram-of-segments representation for multi-label classification of birdsong. The proposed method is compared with binary relevance and three multi-instance multi-label learning (MIML) algorithms from prior work (which focus more on structure in the sound, and less on structure in the label sets). Experiments are conducted on two real-world birdsong datasets, and show that the proposed method usually outperforms binary relevance (using the same features and base-classifier), and is better in some cases and worse in others compared to the MIML algorithms.

Motivation & Objective

• To address the challenge of classifying multiple simultaneously vocalizing bird species in real-world audio recordings with overlapping sounds and background noise.
• To explore whether multi-label classification, particularly using ensemble of classifier chains (ECC), can better exploit label set correlations than binary relevance in bird sound recognition.
• To evaluate ECC against existing MIML methods on two real-world datasets, including a novel iPhone-recorded dataset with high background noise.
• To compare the performance of ECC with binary relevance and MIML algorithms using standard multi-label metrics, while accounting for differences in parameter tuning.

Proposed method

• Represent each audio recording as a fixed-length histogram-of-segments using a 2D time-frequency supervised segmentation algorithm and a clustered codebook.
• Apply a histogram-of-segments representation to encode variable-length bird sound recordings into fixed-length feature vectors suitable for multi-label classification.
• Use random forest-based ensemble of classifier chains (ECC-RF) to model label correlations, where each chain learns a sequence of binary classifiers over class labels.
• Train and evaluate the ECC model using 5- or 10-fold cross-validation, with all models using the same base classifier (random forest) and feature representation.
• Compare ECC-RF against binary relevance (BR) and three MIML algorithms (MIMLSVM, MIMLRBF, MIML-kNN), with the latter using post-hoc parameter tuning for optimal performance.
• Evaluate all methods using standard multi-label metrics: Hamming loss, set 0/1 loss, rank loss, 1-error, and coverage.

Experimental results

Research questions

• RQ1 Can ensemble of classifier chains (ECC) outperform binary relevance in multi-label bird sound classification when using the same feature representation and base classifier?
• RQ2 Does modeling label set correlations through ECC improve performance on multi-species bird sound recordings compared to methods that ignore such structure?
• RQ3 How does ECC compare to state-of-the-art MIML algorithms on real-world bird sound datasets with overlapping vocalizations and high background noise?
• RQ4 To what extent do performance differences between ECC and MIML methods stem from differences in modeling multi-instance structure versus label correlation?
• RQ5 Is the histogram-of-segments representation sufficient to capture the necessary acoustic and label structure for effective multi-label bird species classification?

Key findings

• ECC-RF outperforms binary relevance (BR) in 7 out of 10 performance comparisons across two datasets, with a win-loss count of 7-3, indicating a consistent advantage in modeling label correlations.
• On the iPhone birdsong dataset, ECC-RF achieved a Hamming loss of 0.1168 and a set 0/1 loss of 0.8121, showing strong performance despite higher background noise.
• On the HJA Birdsong dataset, ECC-RF achieved a Hamming loss of 0.0485 and a coverage of 1.6555, outperforming BR and showing competitive results against MIML methods.
• MIML- $k$ NN achieved the best overall performance on the HJA dataset with a Hamming loss of 0.039, but this advantage stems from post-hoc parameter tuning, giving it an unfair edge.
• MIMLRBF and MIML- $k$ NN outperformed ECC-RF on the HJA dataset (1-4 and 0-5 win-loss counts), suggesting that better modeling of multi-instance structure may be more critical than label correlation in some cases.