PaperlayerView exact revision on arXiv

Source-linked AI summary

Data Vision: Learning to See Through Algorithmic Abstraction

Samir Passi, Steven J. Jackson

arXiv:2002.03387v1cs.HCcs.CYcs.LGstat.ML

TL;DR

The paper asks how students learn to apply algorithms while adapting to empirical contingency, since making data work with algorithms requires situated decisions as well as formal rules. Drawing on two learning environments, it argues that data analytics is rule-based rather than rule-bound and that data vision involves balancing abstraction with contingency.

  • Problem

    Learning to see the world through data requires more than mechanical rules, but how students learn to improvise around formal methods amid empirical contingency remains a central question.

    Source: INTRODUCTION · p. 1

  • Method

    The paper examines two data-learning settings: a machine-learning classroom and digital humanities workshops.

    Source: INTRODUCTION · p. 1, CASE 2: DIGITAL HUMANITIES WORKSHOPS · p. 6

  • Results

    Data analytics is rule-based rather than rule-bound, requiring analysts to align incomplete, inconsistent data with formal assumptions through situated decisions.

    Source: DISCUSSION · p. 9

  • Takeaways & Limitations

    Data vision connects methodological abstraction, empirical contingency, and situated discretion in data-analytic practice.

    Source: CONCLUSION · p. 10

    Because classrooms are only one step in professionalization, understanding data vision also requires study of industry, research centers, startups, and hackathons.

    Source: CONCLUSION · p. 11

Abstract

from arXiv · show

Learning to see through data is central to contemporary forms of algorithmic knowledge production. While often represented as a mechanical application of rules, making algorithms work with data requires a great deal of situated work. This paper examines how the often-divergent demands of mechanization and discretion manifest in data analytic learning environments. Drawing on research in CSCW and the social sciences, and ethnographic fieldwork in two data learning environments, we show how an algorithm's application is seen sometimes as a mechanical sequence of rules and at other times as an array of situated decisions. Casting data analytics as a rule-based (rather than rule-bound) practice, we show that effective data vision requires would-be analysts to straddle the competing demands of formal abstraction and empirical contingency. We conclude by discussing how the notion of data vision can help better leverage the role of human work in data analytic learning, research, and practice.

INTRODUCTION

Algorithmic data analysis makes knowledge production possible, but its apparent mechanical neutrality obscures the situated work needed to make data fit algorithms. The paper calls for understanding data vision as balancing formal methods with discretion amid empirical contingency.

INTRODUCTION

  • Algorithmic knowledge production is socially collaborative and grounded in sociocultural, economic, and political conditions.Algorithms can be selective, subjective, and biased, and embody assumptions about how the world works.Source: INTRODUCTION · p. 1
  • Data vision combines organizing the world through data and algorithms with discretion about why, how, and when to apply or adapt methods.Effective analysis requires both computational representation and mastery of how worlds and tools are combined.Source: INTRODUCTION · p. 1
  • The paper examines how novices learn to see through data and improvise around formalizations when they encounter empirical contingency.It studies these questions in classrooms and workshops, where learning and professionalization make competing demands visible.Source: INTRODUCTION · p. 1
  • Learning environments reveal how guided interactions among instructors, students, teaching assistants, and educational materials shape professional learning.Classrooms and workshops foreground participation in a discourse and community of practice, not only individual experience.Source: INTRODUCTION · pp. 1–2
  • Two sequences—a machine-learning classroom and digital humanities workshops—show that data analytics is rule-based, not rule-bound.The paper argues that analysts must balance formal abstraction and mechanical routines with discretionary action and empirical contingency.Source: INTRODUCTION · p. 2

PROFESSIONAL VISION, SITUATED KNOWLEDGE, AND DISCRETIONARY PRACTICE

Professional vision and situated-practice research show that expertise is built through socially learned ways of seeing and through adapting formal routines to particular circumstances. The paper applies these insights to data analytics, where making algorithms work requires discretion as well as abstraction.

PROFESSIONAL VISION, SITUATED KNOWLEDGE, AND DISCRETIONARY PRACTICE

  • Professional vision is learned through training, socialization, and immersion in a community of practice, not simply through mastering formal techniques.Goodwin’s cases show professionals learning to see relevant objects through practice and tools.Source: PROFESSIONAL VISION, SITUATED… · p. 2
  • Data analysts make datasets work with chosen algorithms through decisions about methods, formatting, collection, cleaning, modeling, and communicating results.These choices make meaningful combinations of data and algorithms; “raw data” is an oxymoron as a workable entity.Source: PROFESSIONAL VISION, SITUATED… · p. 2
  • Chess illustrates that formal rules can be taught, yet applying them remains contingent on the particular game.Abbott’s example distinguishes a rule-based practice from one fully bound by rules.Source: PROFESSIONAL VISION, SITUATED… · pp. 2–3
  • Formal routines encounter contingency and surprise in practice; as plans are applied amid continuous variation and uncertainty, people engage in iterative, situated problem-solving and often creative improvisation [22, 17, 33].Source: PROFESSIONAL VISION, SITUATED… · p. 3
  • Routine procedures have both structural, rule-like aspects and situated, contingent performances, whose interplay allows routines to shift while remaining recognizable.Adapting practices to new problems is part of learning by doing, and can indicate skill and membership.Source: PROFESSIONAL VISION, SITUATED… · p. 3
  • Representations do not simply reveal objects: situated interpretation and manipulation help produce phenomena as tractable objects of analysis.Research on AI and medical imaging shows that seeing professional objects involves representational devices, discussion, and action.Source: PROFESSIONAL VISION, SITUATED… · p. 3
  • Data-analytics research highlights collaborative deliberation and judgments in modeling, collection, and data sharing, challenging the idea of neutral, raw data.Erasing this work can make algorithms appear neutral, opaque, or self-efficacious.Source: PROFESSIONAL VISION, SITUATED… · p. 4
  • Data vision treats analytics as rule-based rather than rule-bound: analysts must work across formal abstraction and empirical contingency.Algorithms organize data predictably, but each dataset poses distinct challenges that require situated decisions.Source: PROFESSIONAL VISION, SITUATED… · p. 4

METHODS AND FINDINGS

Ethnographic observations in a machine-learning class and digital-humanities workshops show algorithmic learning balancing formal routines with decisions responsive to particular data.

METHODS AND FINDINGS

  • Fieldwork comprised four months in a graduate machine-learning class and three digital-humanities workshops, each lasting two hours.The class had about 80 students; workshop attendance ranged from nine to thirteen.Source: METHODS AND FINDINGS · p. 4

CASE 1: MACHINE-LEARNING CLASSROOM

  • Students initially saw three clusters, but the instructor showed that the answer depended on the similarity measure and that other clusterings were possible.The examples included different arrangements of two or three clusters.Source: CASE 1: MACHINE-LEARNING CLASSROOM · p. 5
  • The clustering exercise taught a mechanical sequence of rules, with students visually comparing points and clusters and repeatedly merging the closest members.The exercise began with each dot as a cluster and ended with all eight dots in one cluster.Source: CASE 1: MACHINE-LEARNING CLASSROOM · pp. 5–6
  • Label-less dots and visual diagrams presented data abstractly, as points in two-dimensional space that could be organized into clusters.The authors note that the algorithm appears to work when data is represented as dots in n dimensions.Source: CASE 1: MACHINE-LEARNING CLASSROOM · p. 6

CASE 2: DIGITAL HUMANITIES WORKSHOPS

  • Workshop participants chose Gothic novels partly because a complete digital set was readily available, while data access, file formats, and cleaning complicated preparation.Students and conveners also had to negotiate differing research interests and ways to make algorithms do the work of reading.Source: CASE 2: DIGITAL HUMANITIES WORKSHOPS · pp. 6–7
  • Quote-extraction work required adapting to inconsistent text: OCR errors, hyphens, and other conventions made quotation marks unreliable proxies for character speech.Students opted to use quotation marks alone to keep the exercise relatively simple, though the workshop also surfaced hyphens and vertical-bar artifacts.Source: CASE 2: DIGITAL HUMANITIES WORKSHOPS · pp. 6–7
  • Students first mistook section counts for quote counts, then found further OCR, formatting, and stylistic variation; the workshop ended without reliable automated extraction.The script created one more section than the number of quotation marks, so an even quote-mark count implied an odd section count.Source: CASE 2: DIGITAL HUMANITIES WORKSHOPS · pp. 7–8

DISCUSSION

The two cases show data analysis as both rule-following and situated work: analysts adapt formal methods to incomplete, inconsistent, and contingent data. A rule-based view preserves the role of abstraction while making discretion, collaboration, and the limits of algorithmic knowledge visible.

DISCUSSION

  • Machine-learning students learn to organize data through abstract forms and fixed algorithmic steps, making the world computationally actionable.The demonstration uses label-less dots on a nameless graph and treats algorithm application as a methodical sequence of rules.Source: DISCUSSION · pp. 8–9
  • Digital-humanities work requires iterative decisions and workarounds to make scripts fit datasets that may be incomplete, unavailable, or inconsistent.Analysts adjust prior knowledge, empirical contingencies, and formal methods to one another.Source: DISCUSSION · p. 9
  • A rule-bound view makes formal methods and their assumptions visible, but understates the creative work needed to fit methods to empirical contingencies.It can teach predictable manipulation while leaving the craftsmanship involved in making data usable under-described.Source: DISCUSSION · p. 9
  • A rule-based view treats algorithms as structured but not fully determined by mechanical rules, foregrounding situated decisions and practical issues.It frames algorithmic performance and formal routines as mutually informing aspects of practice.Source: DISCUSSION · p. 9
  • Data vision describes the ongoing negotiation between abstract algorithmic routines and situated, reflexive applications of those routines.Analysts align incomplete and inconsistent data with formal assumptions through discretionary and improvisational work.Source: DISCUSSION · p. 9
  • Making situated decisions explicit can contextualize algorithmic knowledge, reduce opacity, and help teach the nontechnical work behind analytic claims.The paper also argues that sharing contingent research practices can counter overly neat accounts and support acknowledgment of often-invisible collaborative work.Source: DISCUSSION · p. 10

CONCLUSION

The conclusion presents data vision as a way to understand the connection between methodological abstraction, empirical contingency, and situated discretion in data analysis. The authors note that their learning-environment study is an early step and call for research in other contexts of algorithmic knowledge production.

CONCLUSION

  • Data vision connects methodological abstraction, empirical contingency, and situated discretion in data analytic practice.The paper argues for a rule-based rather than rule-bound understanding, based on empirical examples from classrooms and workshops.Source: CONCLUSION · p. 10
  • Studying classrooms and workshops reveals formative aspects of data analysts’ training and professionalization.The paper presents these settings as early steps toward defining and understanding data vision.Source: CONCLUSION · p. 10
  • Because classrooms are only one step in professionalization, the authors call for study of data analysis in industry, research centers, startups, and hackathons.They note that professional expectations and organizational imperatives shape practice across contexts.Source: CONCLUSION · p. 11
Loading 2002.03387v1…