Pith. sign in

REVIEW 3 major objections 6 minor 48 references

Unsupervised anomaly detection in large-scale estuarine acoustic telemetry data

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A neural-network autoencoder trained only on normal fish movements flags every true anomaly in three million acoustic detections, with negligible false alarms.

desk verdict Useful telemetry anomaly-detection case study, but the 100% recall is likely built into the experimental design via label-feature alignment and validation-threshold tuning, so the strong claim overstates what is shown. read the letter →

arxiv 2502.01543 v1 pith:MXZKOZ6G submitted 2025-02-03 cs.LG

Siphendulwe Zaza , Marcellin Atemkeng , Taryn S. Murray , John David Filmalter , Paul D. Cowley This is my paper · ORCID
classification cs.LG
keywords acoustictelemetryunsupervisedanomalydetectionneuralnetworkautoencoderthresholdselectiontimeseriesresamplingduskykobfishmovementreceiverarray
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that automated, unsupervised machine learning can replace manual or statistical screening of anomalous detections in large acoustic telemetry datasets. Working with over three million detections of fifty dusky kob in the Breede Estuary, the authors train a neural-network autoencoder on normal movement patterns only, then set its reconstruction-error threshold with a new algorithm that prioritises recall over precision. They report that this model flags every true anomaly in the test set, with 100% recall and zero false negatives, while misclassifying only about 0.35% of normal detections as anomalous, and that resampling the irregular time series at finer intervals lowers that fraction further. The same evaluations show that isolation forest, DBSCAN, and local outlier factor miss over 99% of true anomalies, which the authors argue makes them unsuitable for telemetry studies where undetected anomalies can distort ecological interpretations. If the result holds, ecologists would gain an automated first-pass filter that guarantees no labelled anomaly survives into downstream movement analyses.

What carries the argument

The object that carries the argument is the neural-network autoencoder, a two-part network whose encoder compresses each detection window into a two-unit latent representation and whose decoder reconstructs it; the reconstruction error is the anomaly score. The proposed threshold-finding algorithm makes the method operational: it sweeps percentile thresholds over validation reconstruction errors, keeps the thresholds that achieve maximum recall, narrows to those with maximum precision, and then to those with maximum specificity. The input features—duration at the same station, number of unique stations, consecutive missing stations, and Haversine distance travelled—encode the movement patterns the labels are designed to capture, and the resampling strategy selects a nearly regular sampling interval from the irregular daily detections.

What would settle it

Run the same autoencoder and threshold algorithm on a telemetry dataset whose anomalies are labelled by independent expert review rather than the three fixed criteria, using features that do not directly encode those criteria; if recall drops below 100% or the model misses injected anomalies that fit none of the three rules, the no-false-normal claim is an artifact of label-feature alignment rather than a general property of the method.

Watch

Extended reading notes

Core claim

The central claim is that a neural-network autoencoder, trained only on normal detections and paired with a proposed threshold-finding algorithm, detects every anomalous detection in the acoustic telemetry test set while misclassifying only a negligible fraction of normal detections. On the paper's test data the autoencoder achieves 100% recall, meaning zero false negatives, with a false-anomaly fraction of about 0.00352 without resampling and lower values after resampling at 90-second and 65-second intervals; the traditional unsupervised classifiers miss the great majority of true anomalies, with false-normal fractions above 0.9. The paper also claims this is the first automated machine-learning and deep-learning treatment of anomalous detections in acoustic telemetry data, and that the practical enabler is the threshold algorithm, which selects the percentile of reconstruction error that first maximises recall, then precision, then specificity.

Load-bearing premise

The labels come from three expert rules, and the input features are built from the same behaviours those rules describe, so the perfect recall may only show that the autoencoder relearns those rules rather than detecting genuinely new anomalies.

Editorial extensions

If this is right

  • If the result is correct, all anomalies matching the three expert criteria are eliminated automatically, so downstream analyses of dusky kob movement start from a clean dataset without manual review.
  • A recall of 100% means no anomalous detection is carried into movement or habitat-use analyses, removing a source of bias the paper identifies as critical for telemetry interpretation.
  • The false-anomaly rate below 0.004 removes a negligible fraction of normal data, which the paper argues is acceptable in datasets operating at millions of detections.
  • Finer resampling, first at 90-second and then at 65-second intervals, consistently lowers the autoencoder's false-anomaly fraction, so higher temporal resolution in data collection or preprocessing improves the model's precision.
  • Isolation forest, DBSCAN, and local outlier factor miss over 99% of true anomalies, so they should not be relied on where the cost of a missed anomaly is high.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable implication the paper leaves implicit: if the engineered features are removed or decorrelated from the three labelling criteria, the autoencoder's 100% recall should persist for genuinely novel anomalies; if it collapses, the performance is rule rediscovery rather than general anomaly detection.
  • The recall-first threshold policy is a deliberate trade-off: it guarantees no missed anomalies at the price of discarding a small amount of normal data. In monitoring contexts where false alarms trigger expensive field responses, the lexicographic priority could be inverted with no change to the architecture.
  • The paper notes it struggled with gradual deviations from normal movement; a natural extension is to inject synthetic gradual anomalies, such as slow station shifts or slowly changing detection rates, into the test set to quantify the boundary of the no-false-normal guarantee.
  • Because the paper's own conclusion limits generalisability to other species and estuaries, the strongest validation would be a multi-site test where labels are produced by independent expert review without exposing the feature-construction rules.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper presents a pipeline for unsupervised anomaly detection in acoustic telemetry detections of dusky kob in the Breede Estuary. It introduces engineered features and three expert-labelling criteria, then compares Isolation Forest, LOF, DBSCAN, LSTM-AE, and a neural-network autoencoder (NN-AE) on the resulting dataset. The authors report that NN-AE with a validation-based threshold-finding algorithm achieves 100% recall (no false normals) on a held-out test set while keeping false anomalies near 0.4% (Section 3.2, Table C4), whereas traditional methods miss most true anomalies. The paper also documents resampling strategies, hyperparameters, and detailed appendices with threshold and metric tables.

Significance. If the result holds, the NN-AE pipeline would provide a practical, automated alternative to manual filtering of large telemetry datasets, with a template for pre-processing, feature engineering, and threshold selection that other telemetry studies could adopt. The manuscript is transparent in reporting confusion matrices, full threshold sweeps (Appendices A-C), and training/computation times, which is beneficial for reproducibility. However, the significance claimed in the Abstract ('high recall with no false normal') is currently contingent on the unresolved label-feature alignment and evaluation-protocol issues raised below; the value of the paper as a methodological case study remains, but the empirical claim needs to be re-baselined before it can be taken at face value.

major comments (3)
  1. [§2.4-§2.5, Abstract] Each of the three labelling criteria in §2.5 is a threshold on one of the engineered features introduced in §2.4/Table 2: criterion 1 corresponds to num_unique_stations = 1, criterion 2 to duration_in_same_station > 120 days, and criterion 3 to consecutive_missing_stations > 1. Because the NN-AE is trained only on samples that, by construction, do not trigger these rules (§2.9), any test anomaly is an extreme point in the feature space along exactly the dimensions the model was trained to reconstruct, and the threshold-finding algorithm in §3.1 is applied to a validation set that contains the same kinds of anomalies. In this design, perfect recall is a built-in consequence rather than evidence of learning beyond the rules. The manuscript should include a rule-based baseline that directly applies the three criteria to the test set and report whether NN-AE improves on it; without this, the central claim in the Abstract and §3.4 is unsupported.
  2. [§2.9, Table C4] The data split in §2.9 is performed at the detection level ('10% of normal samples ... 50% of anomalous samples') with no grouping by FishID. Many engineered features in Table 2 (e.g., num_detections, num_unique_stations, duration_in_same_station) are per-fish statistics or sequences that span the entire monitoring period, so detections of the same individual appear in both the training and test partitions. This creates information leakage and inflates the reported test metrics, including the 100% recall in Table C4. The evaluation should be repeated with a fish-exclusive split (or at least a temporal split), so that no individual contributes detections to both training and test.
  3. [§3.1, Table 4, Table C4] The threshold is selected by maximizing recall on a validation set that contains known anomalous samples (Step 2-3 in §2.9; algorithm in §3.1). This is a supervised use of the validation labels for choosing the decision boundary, and it is not mirrored for the baseline models: IF, LOF, and DBSCAN are evaluated with contamination parameters fixed at 0.001/0.01 (Table 4), which are not tuned to the actual 8.8% anomaly rate. Consequently, the gap between NN-AE and the baselines in Table C4 may reflect the availability of labelled anomaly data for threshold selection rather than the models' unsupervised detection ability. Please either select thresholds without labels (e.g., a fixed reconstruction-error percentile on normal validation data) or, if labels are used, apply the same selection procedure to the baselines and report both settings.
minor comments (6)
  1. [§3.1] The sentence 'As shown in Tables A1, A1, and A2' should read 'Tables A1, A2, and A3'.
  2. [Figure 9] The confusion matrices in Figure 9 are difficult to read because the row/column labels (Normal/Anomaly) are not clearly aligned with the numeric cells and some numbers (e.g., '329') appear without context; please redraw with explicit axis labels for each panel.
  3. [Table 2] The original feature 'time sa' is not defined; it should be 'timestamp' or 'time' to match the text and other rows.
  4. [Appendix B] Appendix B contains the typo 'resemble anomaly of typenormal/expected movements'; correct to 'type normal/expected movements'.
  5. [§3.2, Table C4] The reported confidence interval for NN-AE recall is 100±0.0000; please clarify whether these intervals are computed over multiple random splits or over different resampling implementations, because a zero width for a test set with more than 130,000 anomalies is surprising and the notation should be explained.
  6. [References] Reference [48] is a blog post ('Geek for Geeks') and does not meet the citation standard of the other references; consider replacing it with a peer-reviewed source on shallow versus deep neural networks.

Circularity Check

2 steps flagged · score 7.0 of 10

NN-AE's perfect recall is a fitted property: labels are thresholds on engineered features, and the threshold is chosen to maximize recall on validation anomalies.

  1. self definitional [Section 2.4 (Engineered features) and Section 2.5 (Labelling process)]
    "These included "duration at the same station", "number of detections", "number of days detected", "number of unique stations", and "consecutive missing stations" ... If an individual fish was recorded at only one station throughout the study period, then it was flagged as an anomaly: If an individual moves as per normal, but then remains in the vicinity of the same station for more than 120 days: If an individual fish is not detected by consecutive stations, but is missed by more than one consecutive station:"

    The three anomaly classes are literally threshold functions of the engineered features: only-one-station is num_unique_stations == 1; same-station > 120 days is duration_in_same_station > 120; missed > 1 consecutive station is consecutive_missing_stations > 1. Since the autoencoder receives these exact features, every label-triggering input is extreme in feature space and lies off the normal-only manifold, so high reconstruction error is structurally expected. The reported detection performance is therefore a rediscovery of the rule-based labels rather than evidence of discovering novel anomalous patterns.

  2. fitted input called prediction [Section 2.9 (Experimental design), Section 3.1 (Finding optimal threshold), Section 3.2 (Performance)]
    "For the NN-AE, hyperparameter tuning was conducted using the validation set, which contained both normal and anomalous samples, in order to determine the optimal threshold for the autoencoder. ... Among the metrics of precision, recall, and specificity we aim to maximize the recall ... The goal is to achieve a recall of 1, which ensures that no FN are present in the observations. ... The results indicate that NN-AE achieves a 100% TA detection rate (i.e. no FN)."

    The threshold-finding algorithm explicitly selects the reconstruction-error percentile that maximizes recall on the validation split, which contains 131985 anomalous samples. Reporting 100% recall after optimizing for recall is presenting the optimization objective as an empirical result. The held-out test set is drawn from the same labelling rules and the same feature encodings, so the perfect test recall is statistically forced rather than an independent prediction of the model's ability to find unknown anomalies.

full rationale

This paper does not rely on self-citations or imported uniqueness theorems, so the citation-based circularity patterns do not apply. The circularity is instead in the construction of the experiment. In Section 2.4 the engineered features are 'duration in same station', 'number of unique stations', and 'consecutive missing stations'; in Section 2.5 each anomaly label is defined by a threshold on exactly those quantities (one station, >120 days, >1 consecutive missing station). The NN-AE is trained only on normal samples, so any input satisfying a labelling rule is a point mass or extreme excursion in the feature space and yields large reconstruction error. The threshold-selection step in Section 3.1 then searches percentiles to maximize recall on a validation set that contains 131,985 anomalous samples (Section 2.9), and Section 3.2 reports the resulting 100% recall as the model's capability. Thus the headline result is a fitted property plus a label-feature alignment, not an independent discovery of novel anomalies. The comparison with IF, LOF, and DBSCAN on the same features retains some empirical content, which justifies a partial rather than total circularity score.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central result depends on labels constructed from the same features the model uses, a threshold fitted to maximize recall on validation, and an ad hoc resampling choice. These are the main inputs that the reader did not pay for upstream, and they limit the independence of the reported performance.

free parameters (4)
  • threshold percentile = 65th, 67th, and 69th percentiles for no resampling, 90s resampling, and 65s resampling respectively
    Chosen by lexicographic recall-precision-specificity optimization on the validation set, so the recall=1 result is a fitted property of the algorithm.
  • resampling interval = 90 seconds and 65 seconds
    Selected by iterative trade-off between sampling regularity and computational feasibility, not derived from a formal criterion.
  • autoencoder hyperparameters = 128 units per layer, 2 compression units, learning rate 0.001, batch size 512, 50 epochs
    Tuned via GridSearch on the validation set; standard model selection but still fitted to the data.
  • baseline model hyperparameters = IF contamination 0.001, 100 estimators; LOF 5 neighbors, contamination 0.01; DBSCAN epsilon 0.5, min samples 10
    Tuned via GridSearch; these affect the reported baseline comparisons.
assumptions (4)
  • ad hoc to paper The three labelling criteria define all true anomalies in the dataset.
    Section 2.5 states that any detection meeting at least one criterion is labelled anomalous. The model's performance is measured against this constructed ground truth.
  • domain assumption Reconstruction error from an autoencoder trained on normal data is a valid anomaly score.
    This is a standard assumption in autoencoder anomaly detection, but it is not proven for this dataset and the paper does not validate it against independent ground truth.
  • ad hoc to paper Resampling at 90s or 65s preserves the signal while making training feasible.
    Section 2.6 chooses these intervals by an iterative trade-off without a formal justification for why these particular rates preserve signal integrity.
  • domain assumption The test set is independent of the training set despite splitting by detection rather than by fish.
    The paper does not state whether fish are split across train and test; if the same fish appears in both, correlations between detections of the same individual could inflate performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unsupervised anomaly detection in large-scale estuarine acoustic telemetry data." pith.science (2026). https://pith.science/paper/MXZKOZ6G

@misc{pith2026250201543,
  author       = {Pith},
  title        = {Pith review of: Unsupervised anomaly detection in large-scale estuarine acoustic telemetry data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MXZKOZ6G}},
  note         = {Machine review of arXiv:2502.01543}
}
read the original abstract

Acoustic telemetry data plays a vital role in understanding the behaviour and movement of aquatic animals. However, these datasets, which often consist of millions of individual data points, frequently contain anomalous movements that pose significant challenges. Traditionally, anomalous movements are identified either manually or through basic statistical methods, approaches that are time-consuming and prone to high rates of unidentified anomalies in large datasets. This study focuses on the development of automated classifiers for a large telemetry dataset comprising detections from fifty acoustically tagged dusky kob monitored in the Breede Estuary, South Africa. Using an array of 16 acoustic receivers deployed throughout the estuary between 2016 and 2021, we collected over three million individual data points. We present detailed guidelines for data pre-processing, resampling strategies, labelling process, feature engineering, data splitting methodologies, and the selection and interpretation of machine learning and deep learning models for anomaly detection. Among the evaluated models, neural networks autoencoder (NN-AE) demonstrated superior performance, aided by our proposed threshold-finding algorithm. NN-AE achieved a high recall with no false normal (i.e., no misclassifications of anomalous movements as normal patterns), a critical factor in ensuring that no true anomalies are overlooked. In contrast, other models exhibited false normal fractions exceeding 0.9, indicating they failed to detect the majority of true anomalies; a significant limitation for telemetry studies where undetected anomalies can distort interpretations of movement patterns. While the NN-AE's performance highlights its reliability and robustness in detecting anomalies, it faced challenges in accurately learning normal movement patterns when these patterns gradually deviated from anomalous ones.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 45 canonical work pages

  1. [1]

    Ecological Applications 27(4), 1031–1049 (2017)

    Crossin, G.T., Heupel, M.R., Holbrook, C.M., Hussey, N.E., Lowerre- Barbieri, S.K., Nguyen, V .M., Raby, G.D., Cooke, S.J.: Acoustic telemetry and fisheries management. Ecological Applications 27(4), 1031–1049 (2017)

    work page 2017

  2. [2]

    ICES Journal of Marine Science 76(2), 477–488 (2019)

    Lowerre-Barbieri, S.K., Kays, R., Thorson, J.T., Wikelski, M.: The ocean’s movescape: fisheries management in the bio-logging decade (2018- 2028). ICES Journal of Marine Science 76(2), 477–488 (2019)

    work page 2019

  3. [3]

    Fisheries Research 247, 106179 (2022)

    Hughes, J.M., Meadows, N.M., Stewart, J., Booth, D.J., Fowler, A.M.: Movement patterns of an iconic recreational fish species, mulloway (ar- gyrosomus japonicus), revealed by cooperative citizen-science tagging programs in coastal eastern australia. Fisheries Research 247, 106179 (2022)

    work page 2022

  4. [4]

    (a) An overview from 2016 to 2021, with blue lines indicating latitudinal movements, red dots highlight- ing anomalies, and magenta crosses marking TA

    Thorstad, E.B., Rikardsen, A.H., Alp, A., okland, F.: The use of electronic Machine Learning and Deep Learning for Anomaly Detection 39 Figure D6 This figure shows the movement patterns of dusky kob in the Breede Estuary from 2016 to 2021, with a focus on highlighting anomalies and TA detected by the DBSCAN. (a) An overview from 2016 to 2021, with blue li...

    work page 2013

  5. [5]

    Science 348, 1255642 (2015)

    Hussey, N.E., Kessel, S.T., Aaerustrup, K., Cooke, S.J., Cowley, P .D., Fisk, A.T., Harcourt, R.G., Holland, K.N., Iverson, S.J., Kocik, J.F., Mills Flem- ming, J.E., Whoriskey, F.G.: Aquatic animal telemetry: a panoramic window into the underwater world. Science 348, 1255642 (2015)

    work page 2015

  6. [6]

    Trends in Ecology and Evolution 37(1), 79–94 (2022)

    Matley, J.K., Kilnard, N.V ., Barbosa Martins, A.P ., Aarestrup, K., As- pillaga, E., Cooke, S.J., Cowley, P .D., Heupel, M.R., Lowe, C.G., Lowerre- Barbieri, S.K., Mitamura, H., Moore, J.-S., Simpfendorfer, C.A., Stokes- bury, M.J.W., Taylor, M.D., Thorstad, E.B., vandergoot, C.S., Fisk, A.T.: Global trends in aquatic animal tracking with acoustic teleme...

    work page 2022

  7. [7]

    Dhellemmes, F., Aspillaga, E., Monk, C.T.: Atfiltr: A solution for man- aging and filtering detections from passive acoustic telemetry data. MethodsX 10, 102222 (2023) 40 Machine Learning and Deep Learning for Anomaly Detection (a) (b) (c) (d) Figure D7 This figure shows the movement patterns of dusky kob in the Breede Estuary from 2016 to 2021, highlight...

    work page 2023

  8. [8]

    Charlie; Steckenreuter, Tattersll, X

    Simpfendorfer, C.A., Huveneers, A. Charlie; Steckenreuter, Tattersll, X. Katherine; Hoenner, Harcourt, R., Heupel, M.R.: Ghosts in the data: false detections in vemco pulse position modulation acoustic teleme- try monitoring equipment. Animal Biotelemetry 3(55), 40317–0150094 (2015)

    work page 2015

Show all 48 references
  1. [9]

    Ecology 96, 332–339 (2015)

    Chambert, T., Miller, D.A.W., Nichols, J.D.: Modeling false positive detections in species occurence data under different study designs. Ecology 96, 332–339 (2015)

    2015

  2. [10]

    R Foundation for Statistical Computing, Vienna, Austria (2022)

    R Core Team: R: A Language and Environment for Statistical Comput- ing. R Foundation for Statistical Computing, Vienna, Austria (2022). R Foundation for Statistical Computing. https://www.R-project.org/

    2022

  3. [11]

    Methods in Ecology & Evolution 12(1) (2021)

    Fl ´avio, H., Baktoft, H., Codling, E.: actel: Standardised analysis of acoustic telemetry data from animals moving through receiver arrays. Methods in Ecology & Evolution 12(1) (2021)

    2021

  4. [12]

    Methods in Ecology and Evolution 11(12), 1733–1742 (2020)

    Niella, Y., Fl ´avio, H., Smoothey, A.F., Aarestrup, K., Taylor, M.D., Ped- demors, V .M., Harcourt, R.: Refined shortest paths (rsp): Incorporation of topography in space use estimation from node-based telemetry data. Methods in Ecology and Evolution 11(12), 1733–1742 (2020)

    2020

  5. [13]

    Spaulding, T.: telemetR: Filter and Analyze Generalised Telemetry Data from Organisms. (2024). R package version 1.0. https://github.com/ tspaulding-esa/telemetr

    2024

  6. [14]

    Nature Communications 14, 3072 (2023)

    Wu, Z., Zhang, C., Gu, X., Duporge, I., Hughey, L.F., Stabach, J.A., Skidmore, S.K., Hopcraft, J.G.C., Lee, S.J., Atkinson, P .M., McCauley, D.J., Lamprey, R., Ngene, S., Wang, T.: Deep learing enables satellite- based monitoring of large populations of terrestrial mammals acr...

    2023

  7. [15]

    IEEE access 8, 224372–224382 (2020)

    Wang, J.-H., Lee, S.-K., Lai, Y.-C., Lin, C.-C., Wang, T.-Y., Lin, Y.-R., Hsu, T.-H., Huang, C.-W., Chiang, C.-P .: Anomalous behaviors detection for underwater fish using ai techniques. IEEE access 8, 224372–224382 (2020)

    2020

  8. [16]

    Oecologia 198, 605–618 (2022)

    Notte, D.V ., Lennox, R.J., Hardie, D.C., Crossin, G.T.: Application of ma- chine learning and acoustic predation tags to classify migration fate of atlantic salmon smolts. Oecologia 198, 605–618 (2022)

    2022

  9. [17]

    Klinard, V . Natalie, Matley, J.K., Ivanova, S.V ., Larocque, S.M., Fisk, A.T., Johnson, T.B.: Application of machine learning to identify predators of stocked fish in lake ontario: using acoustic telemetry predation tags to inform management. Journal of Fish Biology 98(1), 23...

    2020

  10. [18]

    Frontiers in Marine Science 8, 631262 (2021)

    Griffin, L.P ., Casselberry, G.A., Hart, K.M., Jordaan, A., Becker, S.L., No- vak, A.J., DeAngelis, B.M., Pollock, C.G., Lundgren, I., Hillis-Starr, Z., Danylchuk, A.J., Skomal, G.B.: A novel framework to predict relative habitat selection in aquatic systems: applying machine ...

    2021

  11. [19]

    Estuar- ine, Coastal and Shelf Science 238, 106713 (2020)

    Gu ´enard, G., Morin, J., Matte, P ., Secretan, Y., Valiquette, E., Mingel- bier, M.: Deep learning habitat modeling for moving organisms in rapidly changing estuarine environments: a case of two fishes. Estuar- ine, Coastal and Shelf Science 238, 106713 (2020)

    2020

  12. [20]

    Computers 12(12), 253 (2023)

    Schindler, T.F., Schlicht, S., Thoben, K.-D.: Towards benchmarking for evaluating machine learning methods in detecting outliers in process datasets. Computers 12(12), 253 (2023)

    2023

  13. [21]

    Renewable Energy 172, 829–840 (2021)

    Chen, H., Liu, H., Chu, X., Liu, Q., Xue, D.: Anomaly detection and crit- ical scada parameters identification for wind turbines based on lstm-ae neural network. Renewable Energy 172, 829–840 (2021)

    2021

  14. [22]

    Sensors 24(3), 746 (2024)

    Benova, L., Hudec, L.: Comprehensive analysis and evaluation of anomalous user activity in web server logs. Sensors 24(3), 746 (2024)

    2024

  15. [23]

    African Journal of Marine Science 45(3), 201–213 (2023)

    Ziko, B., Murray, T., Næsje, T., Filmalter, J., Cowley, P .: Acoustic teleme- try reveals the drivers behind estuary–sea connectivity of an important estuarine-dependent fishery species, pomadasys commersonnii, in the breede estuary, south africa. African Journal of Marine Sci...

    2023

  16. [24]

    Marine Ecology Progress Series 537, 233–245 (2015)

    Childs, A.-R., Cowley, P .D., Næsje, T.F., Bennett, R.H.: Habitat connec- tivity and intra-population structure of an estuary-dependent fishery species. Marine Ecology Progress Series 537, 233–245 (2015)

    2015

  17. [25]

    NINA rapport (2006)

    Cowley, P .D., Kerwath, S., Næsje, T., Childs, A.-R., Gillespie, L.D., Leg- gitt, C., Thorstad, E.B., Økland, F.: Space use patterns and movements of juvenile dusky kob argyrosomus japonicus in the great fish estuary (south africa): implications for management. NINA rapport (2006)

    2006

  18. [26]

    Ichthyologica Bulletin JLB Smith Institute of Ichthyology 65, 1–40 (1995)

    Griffiths, M.H., Heemstra, P .C.: A contribution to the taxonomy of the marine fish genus argyrosomus (perciformes: Sciaenidae), with descrip- tions of two new species from southern africa. Ichthyologica Bulletin JLB Smith Institute of Ichthyology 65, 1–40 (1995)

    1995

  19. [27]

    South African Journal of Marine Science 17, 135–154 (1996) Machine Learning and Deep Learning for Anomaly Detection 45

    Griffiths, M.H.: Life history of the dusky kob argyrosomus japonicus (sciaenidae) off the east coast of south africa. South African Journal of Marine Science 17, 135–154 (1996) Machine Learning and Deep Learning for Anomaly Detection 45

    1996

  20. [28]

    South African Journal of Aquatic Sciences 22, 52–61 (1996)

    Ter Morhuizen, L.D., Whitfield, A.K., Paterson, A.W.: Influence of fresh- water flow regime on fish assemblages in the great fish river and estuary. South African Journal of Aquatic Sciences 22, 52–61 (1996)

    1996

  21. [29]

    African Journal of Marine Science 30(2), 247–253 (2008)

    Cowley, P ., Kerwath, S., Childs, A., Thorstad, E., Økland, F., Næsje, T.: Estuarine habitat use by juvenile dusky kob argyrosomus japonicus (sci- aenidae), with implications for management. African Journal of Marine Science 30(2), 247–253 (2008)

    2008

  22. [30]

    Marine Ecology Progress Series 460, 221–232 (2012)

    Næsje, T.F., Cowley, P .D., Diserud, O.H., Childs, A.-R., Kerwath, S.E., Thorstad, E.B.: Riding the tide: estuarine movements of a sciaenid fish, argyrosomus japonicus. Marine Ecology Progress Series 460, 221–232 (2012)

    2012

  23. [31]

    PhD thesis, Rhodes University (2013)

    Childs, A.-R.: Estuarine-dependency and multiple habitat use by dusky kob argyrosomus japonicus (pisces: Sciaenidae). PhD thesis, Rhodes University (2013)

    2013

  24. [32]

    BUILDING BRIDGES–INNOVATION, ENTREPRENEURSHIP AND SOCIETY, 107 (2021)

    MARCIUC, A.M.: Data preprocessing techniques. BUILDING BRIDGES–INNOVATION, ENTREPRENEURSHIP AND SOCIETY, 107 (2021)

    2021

  25. [33]

    ” O’Reilly Media, Inc.”, ??? (2018)

    Zheng, A., Casari, A.: Feature Engineering for Machine Learning: Prin- ciples and Techniques for Data Scientists. ” O’Reilly Media, Inc.”, ??? (2018)

    2018

  26. [34]

    In: 2017 International Conference of Electronics, Communication and Aerospace Technology (ICECA), vol

    Santhosh, A., Idicula, S.M.: Home range estimation of migrating organ- isms considering the spatiotemporal aspects of gps tracked data. In: 2017 International Conference of Electronics, Communication and Aerospace Technology (ICECA), vol. 1, pp. 162–167 (2017). IEEE

    2017

  27. [35]

    In: 2008 Eighth Ieee International Conference on Data Mining, pp

    Liu, F.T., Ting, K.M., Zhou, Z.-H.: Isolation forest. In: 2008 Eighth Ieee International Conference on Data Mining, pp. 413–422 (2008). IEEE

    2008

  28. [36]

    IEEE transactions on knowledge and data engineering 33(4), 1479–1489 (2019)

    Hariri, S., Kind, M.C., Brunner, R.J.: Extended isolation forest. IEEE transactions on knowledge and data engineering 33(4), 1479–1489 (2019)

    2019

  29. [37]

    In: Proceedings of the 2000 ACM SIGMOD International Conference on Management of Data, pp

    Breunig, M.M., Kriegel, H.-P ., Ng, R.T., Sander, J.: Lof: identifying density-based local outliers. In: Proceedings of the 2000 ACM SIGMOD International Conference on Management of Data, pp. 93–104 (2000)

    2000

  30. [38]

    Big Data and Cognitive Computing 5(1), 1 (2020)

    Alghushairy, O., Alsini, R., Soule, T., Ma, X.: A review of local outlier factor algorithms for outlier detection in big data streams. Big Data and Cognitive Computing 5(1), 1 (2020)

    2020

  31. [39]

    In: Proceedings of the Conference on Research in Adaptive and Convergent Systems, pp

    Cheng, Z., Zou, C., Dong, J.: Outlier detection using isolation forest and 46 Machine Learning and Deep Learning for Anomaly Detection local outlier factor. In: Proceedings of the Conference on Research in Adaptive and Convergent Systems, pp. 161–168 (2019)

    2019

  32. [40]

    In: Kdd, vol

    Ester, M., Kriegel, H.-P ., Sander, J., Xu, X., et al.: A density-based algo- rithm for discovering clusters in large spatial databases with noise. In: Kdd, vol. 96, pp. 226–231 (1996)

    1996

  33. [41]

    In: The Fifth International Conference on the Appli- cations of Digital Information and Web Technologies (ICADIWT 2014), pp

    Khan, K., Rehman, S.U., Aziz, K., Fong, S., Sarasvady, S.: Dbscan: Past, present and future. In: The Fifth International Conference on the Appli- cations of Digital Information and Web Technologies (ICADIWT 2014), pp. 232–238 (2014). IEEE

    2014

  34. [42]

    ACM Transactions on Database Systems (TODS) 42(3), 1–21 (2017)

    Schubert, E., Sander, J., Ester, M., Kriegel, H.P ., Xu, X.: Dbscan revisited, revisited: why and how you should (still) use dbscan. ACM Transactions on Database Systems (TODS) 42(3), 1–21 (2017)

    2017

  35. [43]

    arXiv preprint arXiv:2101.08635 (2020)

    Nwadiugwu, M.C.: Neural networks, artificial intelligence and the com- putational brain. arXiv preprint arXiv:2101.08635 (2020)

    2020 arXiv

  36. [44]

    science 313(5786), 504–507 (2006)

    Hinton, G.E., Salakhutdinov, R.R.: Reducing the dimensionality of data with neural networks. science 313(5786), 504–507 (2006)

    2006

  37. [45]

    In: 2018 Wireless Telecommunications Symposium (WTS), pp

    Chen, Z., Yeo, C.K., Lee, B.S., Lau, C.T.: Autoencoder-based network anomaly detection. In: 2018 Wireless Telecommunications Symposium (WTS), pp. 1–5 (2018). IEEE

    2018

  38. [46]

    In: Proceedings of the MLSDA 2014 2nd Workshop on Machine Learning for Sensory Data Analysis, pp

    Sakurada, M., Yairi, T.: Anomaly detection using autoencoders with nonlinear dimensionality reduction. In: Proceedings of the MLSDA 2014 2nd Workshop on Machine Learning for Sensory Data Analysis, pp. 4–11 (2014)

    2014

  39. [47]

    arXiv preprint arXiv:2104.14677 (2021)

    Zahedi, L., Mohammadi, F.G., Rezapour, S., Ohland, M.W., Amini, M.H.: Search algorithms for automated hyper-parameter tuning. arXiv preprint arXiv:2104.14677 (2021)

    2021 arXiv

  40. [48]

    Geek for Geeks, (2024)

    ramlakhan: Difference Between Shallow and Deep Neural Networks. Geek for Geeks, (2024). Geek for Geeks. https://www.geeksforgeeks. org/difference-between-shallow-and-deep-neural-networks/

    2024

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.