Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Understanding Bias in Machine Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read Bias enters a machine learning model through the data it is trained on, and it can be made visible with interactive visualizations.

desk verdict A clear, honest tutorial on dataset bias with interactive demos, but the demos are not reproducible from the text and the paper adds no research contribution beyond the artifact. read the letter →

arxiv 1909.01866 v1 pith:TQD4O5G6 submitted 2019-09-02 cs.LG stat.ML

classification cs.LGstat.ML
keywords biasinmachinelearningVisualizationExplainableAIdatasetcovariateshiftsampleselectionimbalanceparallelcoordinateslayer-wiserelevancepropagation
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

This paper makes a conceptual point: a machine learning model can be biased because the data used to train it is biased. Using a simplified hiring scenario with four skill scores and a university attribute, the paper shows how three types of bias—covariate shift, sample selection bias, and imbalance bias—change both the model's predictions and the features it relies on. Interactive parallel-coordinates plots let a reader see the biased patterns in the data, and layer-wise relevance propagation shows which features drive individual decisions. The takeaway is that biased data can propagate into model behavior, so inspecting data and models before deployment is a necessary safeguard.

What carries the argument

The demonstration is built on two visual inspection tools. A parallel-coordinates plot (each feature is a vertical axis; each applicant is a polyline connecting their score on each axis) lets the reader brush value ranges and see which combinations of features occur together in the training and test data. Layer-wise relevance propagation (LRP) computes, by a backward pass through the network, how much each input feature contributed to the model's decision for a given applicant, displayed as a relative relevance bar chart. These two tools carry the argument: the plots show the bias in the data, and the LRP attributions show the bias in the learned model.

What would settle it

Train the same synthetic hiring task with several different network architectures and random seeds, hold the biased data fixed, and check whether the LRP relevance patterns (for example, university importance under sample selection bias) still appear in every run; if the patterns vanish in most runs, the paper's demonstration does not support its general conclusion.

Watch

Extended reading notes

Core claim

The paper's central claim is that bias enters a machine learning model through the data used to build it, and that this is not a rare corner case: any of three identifiable dataset distortions can be picked up by the model. In the synthetic hiring example, when a skill (Pytorch) was never high in the training data, the model became unreliable for applicants with high Pytorch scores; when a university was spuriously correlated with the invite label, the model treated university as a relevant feature; and when the reject class dominated the training set, the model's predictions skewed toward rejection even on a balanced test set. The paper argues that such effects can be made visible and understandable through interactive visualization of the data and explainability analysis of the model, making it possible for domain experts to recognize and counter the bias.

Load-bearing premise

The demonstrations assume that the neural network and the explanation scores shown in the plots are representative of machine learning models generally.

Editorial extensions

If this is right

  • If training data is biased, the resulting model's predictions can be biased even when the test population has changed; the Pytorch example shows a model failing on a pattern it never saw.
  • Models can pick up spurious correlations, such as university identity, and treat them as relevant indicators rather than the semantically meaningful skills.
  • Class imbalance in the training set can shift the model's decision boundary so that its rejection rate is distorted relative to the true label distribution.
  • The same inspection techniques (parallel coordinates plus LRP) can be used to detect bias in the training data and in the model before deployment.

Reading between the lines

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

  • The paper's demonstration uses one unspecified network; the three biases should appear for any model class, but the exact relevance patterns are likely architecture-dependent, so the qualitative claim is more robust than the specific numbers.
  • In practice, the three bias types often compound, so the clean separation used in the demos may not hold; detection tools would need to disentangle confounded sources.
  • A testable extension is to train a deliberately debiased model (for example, resampling to balance classes or removing the university feature) and check whether LRP attributions drop the spurious feature, turning the visual explanation into a diagnostic for mitigation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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. This short paper argues—correctly and in an accessible way—that machine learning models can inherit bias from biased training data. The authors define three types of bias (covariate shift, sample selection bias, and imbalance bias), introduce a synthetic hiring scenario in which the ground-truth label is based on a simple two-of-four-skills rule, and present interactive parallel-coordinate and LRP-relevance visualizations intended to show how each bias type distorts the predictions and feature attributions of a trained neural network. The paper closes with a discussion of countermeasures and a list of further reading.

Significance. The paper addresses a topic of clear importance: understanding and detecting bias in machine learning systems. Its accessible framing and the use of interactive visualizations as a pedagogical device are commendable, and the central qualitative claim is well supported by prior literature cited in Section 5. However, the manuscript's own demonstration is not independently verifiable: it omits the data-generation details, the network architecture, the training procedure, and the exact LRP rule, and it provides no quantitative evaluation. If these gaps are filled, the paper could serve as a useful educational resource; as written, it reads more like an extended abstract for a workshop demonstration than a standalone journal article. The authors should be credited for clearly separating the three bias types and for explicitly stating that a model may or may not pick up a given bias, a nuance that is often missing in popular treatments.

major comments (3)
  1. [§3.1–3.3, Figures 1–7] The demonstration's load-bearing assumption is that the displayed behavior of the neural network and its LRP attributions faithfully represent what a typical model trained on the biased synthetic data would do. The paper does not specify the network architecture, activation functions, layer sizes, training procedure, hyperparameters, the exact LRP propagation rule, or the data-generation parameters (e.g., the number of samples, score distributions, or the university–skill correlation). Without these details, the plots in Sections 3.1–3.3 cannot be reproduced or independently assessed, and the observed relevance patterns could be idiosyncratic to one arbitrary configuration. Please provide a complete description of the model and data in an appendix, and include a stable URL or DOI for the code and interactive plots.
  2. [§3.1, side note] The side note in Section 3.1 states that the neural network 'did not pick up' the intended two-skill rule even for some in-distribution applicants. This admission is important: if the network fails to learn the ground-truth rule in the unbiased case, then the differences in LRP relevance attributed to covariate shift in Figures 2 and 3 could instead be effects of underfitting, misoptimization, or an unsuitable LRP rule. The paper should report the model's training and test accuracy on the unbiased data, a confusion matrix or quantitative comparison of model decisions against the ground-truth HR rule for each bias scenario, and a sanity check of the LRP attributions on correctly classified examples.
  3. [§4, Conclusions] The conclusion states that 'we have shown that if the data is biased, the resulting model may be biased, too.' As a scientific claim, this is not established by the preceding demonstration: the bias is injected into the synthetic data by design, and the observed model behavior is described only qualitatively. The claim is independently supported by the literature cited in Section 5, so it is not incorrect, but the paper should explicitly position the synthetic experiments as illustrative rather than as a proof or systematic study. Adding a small quantitative experiment with multiple random seeds and architectures would also strengthen the robustness claims.
minor comments (6)
  1. [Abstract] The sentence 'we develop interactive plots to visualizing the bias' should read 'to visualize the bias.'
  2. [§2, first bullet] The definition of covariate shift as 'one of the features is not covered uniformly in the dataset' is imprecise and could be confused with missing data. Please define it as a change in the marginal distribution P(X) between the training and test/deployment populations, while P(Y|X) remains the same.
  3. [§3, Figures and text] Several places use 'Lets' instead of 'Let's' (e.g., 'Lets take a simplified...', 'Lets make this more concrete'), and 'HRs decision' should be 'HR's decision.'
  4. [Figure 7 caption] The caption of Figure 7 reads 'Interactive Visualization of Relevance of each Feature,' but the figure actually shows pie charts of invite/reject ratios and correct/incorrect decisions. Please correct the caption to match the content.
  5. [§3] The interactive plots are not self-contained in a static paper: the reader is directed to a workshop homepage or to 'run our python code' without a direct URL or repository identifier. Please include a stable link and, ideally, a supplementary PDF with screenshots or static versions of the key states of the plots.
  6. [References] Reference [7] is incomplete ('In CVPR.' with no year or venue details), and some other entries lack pages or years. Please complete all references for reproducibility of the literature survey.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the paper is an illustrative demo with no derivation chain, and its single self-citation is not load-bearing.

full rationale

This paper makes no formal derivation or prediction; it is a pedagogical illustration. Section 1 defines bias as entering the model through biased training data, and Sections 3.1–3.3 construct synthetic hiring data with deliberately injected covariate shift, sample selection bias, and imbalance, then show a neural network trained on that data reflects the injected bias in its decisions and LRP attributions. Because the bias is inserted by design, the demonstration is an example of the definition rather than an independent test, which makes the illustration partly self-fulfilling; however, this is not a circular argument in the technical sense. The central claim ('if the data is biased, the resulting machine learning model may be biased, too', Section 4) is independently supported by external prior work cited in Section 5, such as references [4,5,6,7,8,9]. The only self-citation is reference [21], a co-author's ACCV paper on contrastive backpropagation, and it appears in a list of backpropagation-based attribution methods without carrying any load for the paper's argument. The side note in Section 3.1 admits the network failed to learn the intended two-skill rule for some in-distribution applicants, which is a limitation of the demo's credibility but not a circular step. Similarly, the omitted network architecture and training details are reproducibility concerns, not evidence of circularity. Overall, the paper is self-contained as a demonstration and does not reduce its conclusions to its inputs by construction.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The paper is a tutorial; its illustrative conclusions rest entirely on hand-constructed synthetic data and on trusting LRP's faithfulness. There is no derivation, so the ledger consists of data-generation choices and unstated modeling assumptions.

free parameters (5)
  • Invitation threshold = score > 0.7 in at least two of four skills
    Hand-defined ground-truth rule for the synthetic HR scenario (Section 3); determines the labels the model tries to learn.
  • Training invite/reject ratio = 10% invite, 90% reject
    Slider setting used to create imbalance bias in Section 3.3; the choice is illustrative, not data-driven.
  • Covariate shift cutoff for Pytorch = no training applicants above 0.7
    Constructed so the model never sees the high-Pytorch region that appears in the test data (Section 3.1).
  • University-skills correlation = University10/9/3 concentrated among high stats and Python applicants
    Injected sample selection bias in Section 3.2; exact joint distribution is not specified.
  • Neural network architecture and training setup = unspecified
    The demo results depend on the model and training choices, which are not reported anywhere in the manuscript.
assumptions (3)
  • domain assumption A neural network with sufficient capacity can approximate the rule 'invite when at least two skills are above 0.7' when the training data covers the relevant region.
    The demo assumes the model's failures are due to data bias rather than to its inability to learn the decision boundary (Section 3).
  • domain assumption Layer-wise Relevance Propagation attributions correctly reflect the features the neural network actually uses for each decision.
    LRP is used to infer feature importance (e.g., university relevance) without validation on this task (Sections 3.1-3.2).
  • domain assumption The three-fold taxonomy of bias types (covariate shift, sample selection bias, imbalance bias) is sufficient for the tutorial's claims.
    The paper builds the entire demo around this taxonomy from Section 2, taken from the literature without derivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Understanding Bias in Machine Learning." pith.science (2026). https://pith.science/paper/TQD4O5G6

@misc{pith2026190901866,
  author       = {Pith},
  title        = {Pith review of: Understanding Bias in Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TQD4O5G6}},
  note         = {Machine review of arXiv:1909.01866}
}
read the original abstract

Bias is known to be an impediment to fair decisions in many domains such as human resources, the public sector, health care etc. Recently, hope has been expressed that the use of machine learning methods for taking such decisions would diminish or even resolve the problem. At the same time, machine learning experts warn that machine learning models can be biased as well. In this article, our goal is to explain the issue of bias in machine learning from a technical perspective and to illustrate the impact that biased data can have on a machine learning model. To reach such a goal, we develop interactive plots to visualizing the bias learned from synthetic data.

Figures

Figures reproduced from arXiv: 1909.01866 by the authors.

Figure 1
Figure 1. Interactive Visualization of Data about Previous Applicants can see in the PC plot that visualizes the training data, there are no applicants whose Pytorch score is higher than 0.7. This looks different if we visualize the scores of todays applicants in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Interactive Visualization of Data about New Coming Applicants plot are called HRs decision (either an invite or reject) and Models decision (a probability value between 0 and 1 with 0 denoting a clear reject recommendation and 1 a clear recommendation to invite the candidate). If you brush the range 0.7 to 1 on the Python and the Pytorch axis and select a range of lower values for the two other skills, you can see t… view at source ↗
Figure 3
Figure 3. Interactive Visualization of Relevance of each Feature If you play around with the values, you will recognize that features with higher score in general are considered more relevant by the model than the ones with lower scores. However, this is not true for Pytorch. Changing its score has far less effect on its relevance for the decision than for all the other skills. Interestingly, if the score of Python is the sam… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Interactive Visualization of Training Data about Previous Applicants As Data Science becomes more and more popular, more universities start to offer similar programs. Will our machine learning model be able to understand that what is important is the skills and not the…
Figure 5
Figure 5. Figure 5: Interactive Visualization of Test Data about New Coming Applicants You can continue to explore the model by making different selections. How does the model behave if in addition to the Statistical Knowledge and Python the applicant has high scores in a third skill? Is …
Figure 6
Figure 6. Figure 6: Interactive Visualization of Relevance of each Feature 3.3 Imbalance Bias An Imbalance Bias occurs if there is a significant difference in the number of samples in the training data for the different labels. In our example scenario this type of bias could occur if the …
Figure 7
Figure 7. Figure 7: Interactive Visualization of Relevance of each Feature If you drag the slider towards the middle, choosing a ratio of about 50% for the reject and invite classes in the training data, there is no imbalance bias in the training data anymore. You will then see in the pie…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. The CRAFT principles for the responsible use of large language models in policymaking

    cs.CY 2026-07 accept novelty 4.0 of 10

    Proposes the CRAFT principles (Control, Rigour, Accountability, Fairness, Transparency) as a practical framework for responsible LLM use in policymaking.

Reference graph

Works this paper leans on

21 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [1]

    Parallel coordinates: a tool for visualizing multi-dimensional geometry

    Alfred Inselberg and Bernard Dimsdale. Parallel coordinates: a tool for visualizing multi-dimensional geometry. In Proceedings of the 1st conference on Visualiza- tion ’90, pages 361–378. IEEE Computer Society Press, 1990

  2. [2]

    http://heatmapping.org/

    Layerwise relevance propagation(lrp). http://heatmapping.org/. Accessed: 2019- 08-30

  3. [3]

    On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation

    Sebastian Bach, Alexander Binder, Gr´ egoire Montavon, Frederick Klauschen, Klaus-Robert M¨ uller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one , 10(7):e0130140, 2015. 12 Jindong Gu 1,2, Daniela Oelke 2

  4. [4]

    Gender shades: Intersectional accuracy dispar- ities in commercial gender classification

    Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy dispar- ities in commercial gender classification. In Conference on fairness, accountability and transparency, pages 77–91, 2018

  5. [5]

    Man is to computer programmer as woman is to homemaker? debiasing word embeddings

    Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in neural information processing systems , pages 4349–4357, 2016

  6. [6]

    Men also like shopping: Reducing gender bias amplification using corpus-level con- straints

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Men also like shopping: Reducing gender bias amplification using corpus-level con- straints. arXiv preprint arXiv:1707.09457 , 2017

  7. [7]

    Unbiased look at dataset bias

    Antonio Torralba, Alexei A Efros, et al. Unbiased look at dataset bias. In CVPR

  8. [8]

    Undoing the damage of dataset bias

    Aditya Khosla, Tinghui Zhou, Tomasz Malisiewicz, Alexei A Efros, and Antonio Torralba. Undoing the damage of dataset bias. In European Conference on Com- puter Vision , pages 158–171. Springer, 2012

Show all 21 references
  1. [9]

    A deeper look at dataset bias

    Tatiana Tommasi, Novi Patricia, Barbara Caputo, and Tinne Tuytelaars. A deeper look at dataset bias. In Domain adaptation in computer vision applications , pages 37–55. Springer, 2017

  2. [10]

    Why should i trust you?: Explaining the predictions of any classifier

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Why should i trust you?: Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining , pages 1135–1144. ACM, 2016

  3. [11]

    Causal interpretations of black-box models

    Qingyuan Zhao and Trevor Hastie. Causal interpretations of black-box models. Journal of Business & Economic Statistics , (just-accepted):1–19, 2019

  4. [12]

    Permutation importance: a corrected feature importance measure

    Andr´ e Altmann, Laura Tolo¸ si, Oliver Sander, and Thomas Lengauer. Permutation importance: a corrected feature importance measure. Bioinformatics, 26(10):1340– 1347, 2010

  5. [13]

    Visualizing and understanding convolutional networks

    Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, pages 818–833. Springer, 2014

  6. [14]

    Visualiz- ing deep neural network decisions: Prediction difference analysis

    Luisa M Zintgraf, Taco S Cohen, Tameem Adel, and Max Welling. Visualiz- ing deep neural network decisions: Prediction difference analysis. arXiv preprint arXiv:1702.04595, 2017

  7. [15]

    Deep inside convo- lutional networks: Visualising image classification models and saliency maps

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convo- lutional networks: Visualising image classification models and saliency maps. In ICLR, 2013

  8. [16]

    Ried- miller

    Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin A. Ried- miller. Striving for simplicity: The all convolutional net. In ICLR, 2014

  9. [17]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In ICML, 2017

  10. [18]

    Smoothgrad: removing noise by adding noise

    Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Vi´ egas, and Martin Wat- tenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017

  11. [19]

    Grad-cam: Visual explanations from deep networks via gradient-based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedan- tam, Devi Parikh, Dhruv Batra, et al. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, pages 618–626, 2017

  12. [20]

    Learning important features through propagating activation differences

    Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In ICML, 2017

  13. [21]

    Understanding individual decisions of cnns via contrastive backpropagation

    Jindong Gu, Yinchong Yang, and Volker Tresp. Understanding individual decisions of cnns via contrastive backpropagation. In ACCV, 2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.