Data stream classification with novel class detection: a review, comparison and challenges

Din, Salah Ud; Shao, Jun; Kumar, Jay; Mawuli, Cobbinah Bernard; Mahmud, Sakib; Zhang, Wei; Yang, Qinli

doi:10.1007/s10115-021-01582-4

Cited by 29 publications

(10 citation statements)

References 159 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The major challenge of the incremental learning task is to improve the classifier even by learning new classes, possibly characterised by few instances, without a catastrophic forgetting of the previously acquired discrimination capabilities. This capability is critical, especially when the input data originates from a continuous stream as in the case of the OBSEA cabled observatory (Delange et al, 2021;Din et al, 2021;Mai et al, 2022).…”

Section: Discussionmentioning

confidence: 99%

“…Species detection and classification in a real-world scenario requires supervised-learning methods that allow a computer system to automatically make predictions based on a series of examples (e.g., see Marini et al, 2018a;Marini et al, 2018b;Lopez-Vazquez et al, 2020;Malde et al, 2020). Unfortunately, the effectiveness of such automated approaches incurs into the "concept drift" phenomenon, consisting in a progressive decrease over time of the detection and classification performance (Hashmani et al, 2019;Jameel et al, 2020;Din et al, 2021). The concept drift is largely investigated in the community of computer vision and artificial intelligence, but very few contributions are available in the marine science context (Langenkämper et al, 2020;Kloster et al, 2020).…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Assessing the Image Concept Drift at the OBSEA Coastal Underwater Cabled Observatory

Ottaviani

Francescangeli

Gjeci³

et al. 2022

Front. Mar. Sci.

View full text Add to dashboard Cite

The marine science community is engaged in the exploration and monitoring of biodiversity dynamics, with a special interest for understanding the ecosystem functioning and for tracking the growing anthropogenic impacts. The accurate monitoring of marine ecosystems requires the development of innovative and effective technological solutions to allow a remote and continuous collection of data. Cabled fixed observatories, equipped with camera systems and multiparametric sensors, allow for a non-invasive acquisition of valuable datasets, at a high-frequency rate and for periods extended in time. When large collections of visual data are acquired, the implementation of automated intelligent services is mandatory to automatically extract the relevant biological information from the gathered data. Nevertheless, the automated detection and classification of streamed visual data suffer from the “concept drift” phenomenon, consisting of a drop of performance over the time, mainly caused by the dynamic variation of the acquisition conditions. This work quantifies the degradation of the fish detection and classification performance on an image dataset acquired at the OBSEA cabled video-observatory over a one-year period and finally discusses the methodological solutions needed to implement an effective automated classification service operating in real time.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Assessing the Image Concept Drift at the OBSEA Coastal Underwater Cabled Observatory

Ottaviani

Francescangeli

Gjeci³

et al. 2022

Front. Mar. Sci.

View full text Add to dashboard Cite

show abstract

“…A data point that lies in a neighborhood of low density is declared to be an outlier, while a data point that lies in a dense neighborhood is declared to be normal. Considering the example data set in Figure 1.4 from [21], it shows that the data set contains two outliers (𝑂 1 and 𝑂 2 ), and two clusters, one of which is much sparser than the other. Data point 𝑂 2 cannot be detected as an outlier by a distance-based algorithm unless a smaller distance threshold is used.…”

Section: The Proximity-based Outlier Detectionmentioning

confidence: 99%

“…Data point 𝑂 2 cannot be detected as an outlier by a distance-based algorithm unless a smaller distance threshold is used. LOF [21] can identify 𝑂 2 in a locally adjusted manner. A proximity-based method is naturally designed to detect both noise and anomalies, although different methods are suitable for different kinds of outliers.…”

Section: The Proximity-based Outlier Detectionmentioning

confidence: 99%

See 1 more Smart Citation

Parameter-free outlier scoring using mass ratio variance for static and streaming data

Changsakul

View full text Add to dashboard Cite

Outlier detection is a significant problem that has been studied in a variety of research and real-world applications. However, little research has been conducted on unsupervised parameter-free outlier scoring. This thesis proposes Mass ratio variance-based Outlier Factor, or MOF, which is unsupervised parameter-free outlier scoring for static data. This algorithm calculates outlier scores based on the variance of mass ratio. The data points with high outlier scores are associated with outliers while the data points with low outlier scores are associated with normal data points. This thesis also proposes Streaming Mass ratio variance-based Outlier Factor or SMOF. This algorithm calculates outlier scores based on MOF and the non-overlapping sliding window model which keeps the dense data points by weighted random sampling making highly efficient storage. Extensive experiments�have�been conducted to evaluate the performance of MOF and SMOF using synthesized and real-world data sets. The experimental results show that they have better accuracy than the state-of-the-art outlier detection techniques.

show abstract

AdaDeepStream: streaming adaptation to concept evolution in deep neural networks

Chambers,

Gaber,

Ghomeshi

2023

Appl Intell

View full text Add to dashboard Cite

Typically, Deep Neural Networks (DNNs) are not responsive to changing data. Novel classes will be incorrectly labelled as a class on which the network was previously trained to recognise. Ideally, a DNN would be able to detect changing data and adapt rapidly with minimal true-labelled samples and without catastrophically forgetting previous classes. In the Online Class Incremental (OCI) field, research focuses on remembering all previously known classes. However, real-world systems are dynamic, and it is not essential to recall all classes forever. The Concept Evolution field studies the emergence of novel classes within a data stream. This paper aims to bring together these fields by analysing OCI Convolutional Neural Network (CNN) adaptation systems in a concept evolution setting by applying novel classes in patterns. Our system, termed AdaDeepStream, offers a dynamic concept evolution detection and CNN adaptation system using minimal true-labelled samples. We apply activations from within the CNN to fast streaming machine learning techniques. We compare two activation reduction techniques. We conduct a comprehensive experimental study and compare our novel adaptation method with four other state-of-the-art CNN adaptation methods. Our entire system is also compared to two other novel class detection and CNN adaptation methods. The results of the experiments are analysed based on accuracy, speed of inference and speed of adaptation. On accuracy, AdaDeepStream outperforms the next best adaptation method by 27% and the next best combined novel class detection/CNN adaptation method by 24%. On speed, AdaDeepStream is among the fastest to process instances and adapt.

show abstract

Data stream classification with novel class detection: a review, comparison and challenges

Cited by 29 publications

References 159 publications

Assessing the Image Concept Drift at the OBSEA Coastal Underwater Cabled Observatory

Assessing the Image Concept Drift at the OBSEA Coastal Underwater Cabled Observatory

Parameter-free outlier scoring using mass ratio variance for static and streaming data

AdaDeepStream: streaming adaptation to concept evolution in deep neural networks

Contact Info

Product

Resources

About