A B C D E F G H I J K L M N O P Q R S T U V W
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
 A CRITICAL FIELD GUIDE FOR WORKING WITH MACHINE LEARNING DATASETS
20
 Written by Sarah Ciston {1}
Editors: Mike Ananny {2} and Kate Crawford {3}


Part of the Knowing Machines research project.
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
 TABLE OF CONTENTS
40
 1. Introduction to Machine Learning Datasets
41
 2. Benefits: Why Approach Datasets Critically?
42
 3. Parts of a Dataset
43
 4. Types of Datasets
44
 5. Transforming Datasets
45
 6. The Dataset Lifecycle
46
 7. Cautions & Reflections from the Field
47
 8. Conclusion
48
49
50

1
51
 INTRODUCTION TO MACHINE LEARNING DATASETS
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
 Maybe you’re an engineer creating a new machine vision system to track birds. You might be a journalist using social media data to research Costa Rican households. You could be a researcher who stumbled upon your university’s archive of handwritten census cards from 1939. Or a designer creating a chatbot that relies on large language models like GPT-3. Perhaps you’re an artist experimenting with visual style combinations using DALLE-2. Or maybe you’re an activist with an urgent story that needs telling, and you’re searching for the right dataset to tell it.
73
 WELCOME.
74
 No matter what kind of datasets you’re using or want to use, whether you’re curious but intimidated by machine learning or already comfortable, this work is complicated. Because machine learning relies on datasets, and because datasets are always tangled up in the ways they’re created and used, things can get messy. You may have questions like:
75
76
 Does this dataset tell the story of my research in the way I want?
77
 How do the dataset pre-processing methods I choose affect my outcomes?
78
 How might this dataset contribute to creating errors or causing harm?
79
80
 More than likely you will encounter at least some of these conundrums — as many of us who work with machine learning datasets do. Anyone using datasets will weigh choices and make tradeoffs. There are no universal answers and no perfect actions — just a tangle of dataset forms, formats, relationships, behaviors, histories, intentions, and contexts.
81
 When choosing and using machine learning datasets, how do you deal with the issues they bring? How can you navigate the mess thoughtfully and intentionally? Let’s jump in.
82
83
84
 INTRODUCTION TO MACHINE LEARNING DATASETS
85
 1.1
86
 WHAT IS THIS GUIDE ?
87
88
89
90
 Machine learning datasets are powerful but unwieldy. They are often far too large to check all the data manually, to look for inaccurate labels, dehumanizing images, or other widespread issues. Despite the fact that datasets commonly contain problematic material — whether from a technical, legal, or ethical perspective — datasets are also valuable resources when handled carefully and critically. This guide offers questions, suggestions, strategies, and resources to help people work with existing machine learning datasets at every phase of their lifecycle. Equipped with this understanding, researchers and developers will be more capable of avoiding the problems unique to datasets. They will also be able to construct more reliable, robust solutions, or even explore promising new ways of thinking with machine learning datasets that are more critical and conscientious. {4} , {5}
91
92
 If you aren’t sure whether this guide is for you, consider the many places you might find yourself working with machine learning datasets. This guide can be helpful if you are…
93
94
 - making a model
95
 - working with a pre-trained model
96
 - researching an existing machine learning tool
97
 - teaching with datasets
98
 - creating an index or inventory
99
 - concerned about how datasets describe you or your community
100
 - learning about datasets by exploring one
101
 - stewarding or archiving datasets
102
 - investigating as an artist, activist, or developer
103
104
105
 This list is non-exhaustive, of course. Datasets are being used widely across countless domains and industries. How else can you imagine working with machine learning datasets?
106
107
108
109
110
111
112
 The appetite for massive datasets is huge and still accelerating, fueled by the perceived promise of machine learning to convert data into meaningful, monetizable information.{6} Too often, this work is done without regard for how datasets can be partial, imperfect, and historically skewed. Take the widely publicized examples of police departments and courts selecting “future criminals” from software that relied on historical crime records, which ProPublica journalists found was grossly inaccurate and targeted Black people in its predictions [6]. More troubling still, researchers (and public and private organizations) continue to make use of such datasets despite learning of their harms — perhaps because they seem more efficient or effective, because they are already part of common practices in their communities, or simply because they are the most readily available options. This is exactly why critical care is so needed — datasets’ potential harms are subtle, localized, and complex. You will need to make conscientious decisions and compromises when working with any dataset. There is no perfect representation, no correct procedure, and no ideal dataset.
113
 This guide aims to help you navigate the complexity of working with datasets, giving you ways to approach conundrums carefully and thoughtfully. Section 1 describes how DATA and DATASETS are dynamic research materials, and Section 2 outlines the BENEFITS of working critically with datasets. Then you’ll find more on the common PARTS of datasets (Section 3), examples of the TYPES of datasets you may encounter (Section 5), and how to TRANSFORM datasets (Section 4) — all to help make critical choices easier. Then Section 6 provides a DATASET LIFECYCLE framework with important questions to ask as you engage critically at each stage of your work. Finally, Section 7 offers some CAUTIONS & REFLECTIONS for careful dataset stewardship.
114
115
 FIELD GUIDES AND DATASETS AS FORMS
116
117
118
 The field guide format frames this text because, like datasets, field guides teach their readers particular ways of looking at the world — for better and for worse. Carried in a knapsack, a birder might use their field guide to confirm a species sighting in a visual index. A hiker might read trail warnings to prepare for their trek. With these practical uses, the field guide speaks to a desire to connect deeply with dataset tools and practices, and a sense of careful responsibility that data stewardship shares with environmental stewardship. However, naturalist field guides also draw on the same problematic histories of classifying and organizing information that are foundational to many machine learning tasks today. This critical field guide aims to help bring understanding to the complexities of datasets, so that the decisions you make while using them are conscientious. It invites you to mess with these messy forms and to approach any logic of classification with a critical eye.
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
 When we say CLASSIFICATION in this guide, generally we refer to the choices, logics, and paradigms that inform sociotechnical communities of practice — how people sort and are sorted into categories, how those categories come to be seen as dominant and naturalized, and how people are differently affected by those categories. We acknowledge that the term CLASSIFICATION also refers to specific machine learning tasks that label and sort items in a dataset by discrete categories. For example, asking whether an image is a dog or a cat is handled by a classification task. These are distinguished from REGRESSION tasks, which show the relationship between features in a dataset, for example sorting dogs by their age and number of spots. In this guide, we will specify ‘tasks’ when referring to these techniques, but simply say ‘classification’ when referring to the sociotechnical phenomenon more broadly.{7}
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
 INTRODUCTION TO MACHINE LEARNING DATASETS
182
 1.2
183
 WHAT ARE DATA?
184
185
186
187
188
 DATA ARE CONTINGENT ON HOW WE USE THEM
189
190
 DATA are values assigned to any ‘thing’, and the term can be applied to almost anything. Numbers, of course, can be data; but so can emails, a collection of scanned manuscripts, the steps you walked to the train, the pose of a dancer, or the breach of a whale. How you think about the information is what makes it data. Philosopher of science Sabina Leonelli sees data as a “relational category” meaning that, “What counts as data depends on who uses them, how, and for which purposes.” She argues data are “any product of research activities [...] that is collected, stored, and disseminated in order to be used as evidence for knowledge claims” [8]. This definition reframes data as contingent on the people who create, use, and interact with them in context.
191
192
193
 DATA MUST BE MADE, AND MAKING DATA SHAPES DATA
194
195
 As a form of information,{8} data do not just exist but have to be generated, through collection by sensors or human effort. Sensing, observing, and collecting are all acts of interpretation that have contexts, which shape the data. For example, when collecting images of faces using infrared cameras, that data can provide heat signatures but not the eye color of its subjects. Studies are designed with specific equipment to achieve their goals and not others. Whether they are quantitative data captured with a sensor or qualitative data described in an interview, the context in which that data is collected has already created a limit for what it can represent and how it can be used. It is easy to think that calling information “data” makes it discrete, separate, fixed, organized, computable — static [1]. But dataset users impose these qualities on information temporarily — to organize it into familiar forms that suit machine learning tasks and other algorithmic systems.
196
197
198
 MACHINE LEARNING, DEEP LEARNING, NEURAL NET, ALGORITHM, MODEL — WHAT’S THE DIFFERENCE?
199
200
201
202
 An ALGORITHM is a set of instructions for a procedure, whether in the context of machine learning or another task. Algorithms are often written in code for machines to process, but they are also widely used in any system of step-by-step instructions (e.g. cooking recipes). Algorithms are not a modern Western invention, but predate computation by thousands of years, as technology culture researcher Ted Striphas has shown [16]. That said, algorithms stayed associated mainly with mathematical calculation until quite recently, according to historian of science Lorraine Datson, who traces their expansion into a computational catch-all in the mid-20th century [17].
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
 A MODEL is the result of a machine learning algorithm, once it includes revisions that take into account the data it was exposed to during its training. It is the saved output of the training process, ready to make predictions about new data. One way to think of a model is as a very complex mathematical formula containing millions or billions of variables (values that can change). These variables, also called model parameters, are designed to transform a numerical input into the desired outputs. The process of model training entails adjusting the variables that make up the formula until its output matches the desired output.

Much focus is put on machine learning models, but models depend directly on datasets for their predictions. While a model is not part of a dataset, it is deeply shaped by the datasets it is based upon. Traces of those datasets remain embedded within the model no matter how it is used next. (This guide won’t cover the detailed aspects of working critically with machine learning models and understanding how they learn — that’s a whole other discussion. Terms like ‘activation functions’, ‘loss functions’, ‘learning rates’, and ‘fine-tuning’ give a taste of the many human-guided processes behind model making, an active conversation and set of practices that are beyond the scope of this guide.)
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
 Artificial NEURAL NETWORKS describe some of the ways to structure machine learning models (see TYPES inSection 4), including making large language models. Named for the inspiration they take from brain neurons (very simplified), they move information through a series of nodes (steps) organized in layers or sets. Each node receives the output of the previous layers’ nodes, combines them using a mathematical formula, then passes the output to the next layer of nodes.
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
 MACHINE LEARNING is a set of tools used by computer programmers to find a formula that best describes (or models) a dataset. Whereas in other kinds of software the programmer will write explicit instructions for every part of a task, in machine learning, programmers will instruct the software to adjust its code based on the data it processes, thus “learning” from new information [ 18] . Its learning is unlike human understanding and the term is used metaphorically.

Some formulas are "deeper" than others, so called because they contain many more variables, and DEEP LEARNING refers to the use of complex, many layers in a machine learning model. Due to their increasing complexity, the outputs of machine learning models are not reliable for making decisions about people, especially in highly consequential cases. When working with datasets, include machine learning as one suite of options in a broader toolkit — rather than a generalizable multi-tool for every task.
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
 INTRODUCTION TO MACHINE LEARNING DATASETS
311
 1.3
312
 WHAT ARE DATASETS?
313
314
315
316
 A DATASET can be any kind of collected, curated, interrelated data. Often, datasets refer to large collections of data used in computation, and especially in machine learning. Information collections are transformed into datasets through a LIFECYCLE of processes (collection/selection, cleaning and analyzing, sharing and deprecating), which shape how that information is understood. (For critical questions you can ask at each phase of a dataset’s lifecycle, see Section 6.)
317
318
319
 DATASETS ARE TIED
TO THEIR MAKERS
320
321
 The many choices that go into dataset creation and use make them extremely dynamic. They always reflect the circumstances of their making — the constraints of tools, who wields them and how, even who can afford the equipment to train, store, and transmit data. For example, analysis of national datasets in the Processing Citizenship project revealed how some European nations collected information differently, with a range of specificity in categories like ‘education level’ or ‘marital status’. Software engineer Wouter Van Rossem and science and technology studies professor Annalisa Pelizza examined not only the data in that dataset, but how they were labeled, organized, and utilized to show that these reflected how nations perceived the migrants they cataloged [ 19] . When gathered by a different group, using different tools, a dataset will be quite different — even if it attempts to collect similar information.
322
323
 DATASETS ARE TIED
TO THEIR SETTINGS
324
325
 Datasets can be frustratingly limited, but this does not mean they are static; instead, the information in datasets is always wrapped up in the contexts that make and use them. Media scholar Yanni Alexander Loukissas, author of All Data Are Local, calls datasets “data settings,” arguing that “data are indexes to local knowledge.” They remain tied to the communities, individuals, organisms, and environments where they were created. Instead of treating data as independent authorities, he says we should ask, “Where do data direct us, and who might help us understand their origins as well as their sites of potential impact?” [1]. These questions extend the possibilities for exploring datasets as dynamic materials. Therefore, datasets must be used carefully, with consideration for their material connection to their origins.
326
327
328
 For your consideration: How does framing information as “data” change your relationship to it? What other forms of information do you work with? What kinds of information should not be included in datasets?
329
330
331
332
333
334
335
336
337
338

2
339
 BENEFITS: WHY APPROACH DATASETS CRITICALLY?
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
 HERE ARE SOME EXAMPLES OF HOW DATASET STEWARDSHIP CAN BENEFIT YOUR PRACTICE, AS WELL AS BENEFIT OTHERS:
361
362
363
 MORE ROBUST DATASETS ARISE BY CONSIDERING MULTIPLE PERSPECTIVES AND WORKING TO REDUCE BIAS.
364
365
366
367
 TO-DO: Include interdisciplinary, intersectional communities in designing, developing, implementing, and evaluating your work. (See ALTERNATIVE APPROACHES TO DATASET PRACTICES)
368
369
370
371
 MORE RELIABLE RESULTS COME FROM ANTICIPATING AND ADDRESSING CONTINGENCIES LIKE DEPRECATED DATASETS AND UNINFORMED CONSENT.
372
373
374
375
 TO-DO: Apply checkpoints at each stage, asking critical questions about data provenance and reflecting on your own methodologies.
376
377
378
379
 GAIN INCREASED PROTECTION FROM LIABILITY FOR DATASETS WITH LEGAL OR ETHICAL ISSUES BY PROACTIVELY ADDRESSING POTENTIAL CONCERNS BEFORE USE.
380
381
382
383
 TO-DO: This does not constitute legal advice. However, always perform due diligence before working with existing datasets, including checking any licenses or terms of use. Simply downloading some datasets can create legal liability [20] , [21]. So try to be aware of potential consent issues, misuse, or ethical concerns beyond those outlined by the dataset creators, especially as they may have changed since creation or arise from your new usage. You can check data repositories and data journalism to see how datasets have already been used.
384
385
386
387
 CRITICAL PRACTICES ARE BECOMING FIELD-STANDARD AND REQUIRED FOR ACCESS TO TOP CONFERENCES AND JOURNALS.
388
389
390
391
 TO-DO: Help shape the future of the field by modeling and advocating for best practices. Suggest new frameworks and methods for making, using, and deprecating datasets.
392
393
394
395
 MORE CAREFUL AND CONSCIENTIOUS OUTCOMES FOR THOSE IMPACTED BY RESULTS.
396
397
398
399
 TO-DO: Engage the people and groups affected by datasets and your use of them, to learn what careful and conscientious practices mean to them.
400
401
402
403
 OPEN-SOURCE, OPEN-ACCESS, AND OPEN RESEARCH COMMUNITIES BUILD POSITIVE FEEDBACK LOOPS THROUGH DATASET STEWARDSHIP OF RELIABLE MATERIALS.
404
405
406
407
 TO-DO: Share datasets responsibly, through centralized repositories and with thorough documentation.
408
409
410
411
 NO NEUTRAL CHOICE (OR NON-CHOICE) EXISTS. “WHEN THE FIELD OF AI BELIEVES IT IS NEUTRAL,” SAYS AI RESEARCHER PRATYUSHA KALLURI, IT “BUILDS SYSTEMS THAT SANCTIFY THE STATUS QUO AND ADVANCE THE INTERESTS OF THE POWERFUL” [23].
412
413
414
415
 TO-DO: Working with datasets brings challenges that need conversations and multiple perspectives. Discuss issues with your team using the Dataset’s Lifecycle questions in Section 6, plus the wide range of critical positions shared in the “Critical Dataset Studies Reading List” compiled by the Knowing Machines research project [22].
416
 Before publishing or launching your work, ask hard questions and share your project with informal readers within your networks who can provide constructive feedback. Go slow. Pause or even stop a project if needed.
417
 Remember that taking “no position” on a dataset’s ethical questions is still taking a position. Consider the tradeoffs for choosing one dataset or technique over another.
418
419
420
421

3
422
 PARTS OF A
DATASET
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
 What actually makes up a machine learning dataset, practically speaking? Here are some of the key terms that are helpful for understanding their parts and dynamics:
447
448
449
 INSTANCE
450
451
 One data point being processed or sorted, often viewed as a row in a table. For example, in a training dataset for a classification task that will sort images of dogs from cats, one instance might include the image of a dog and the label “dog,” while another instance would be an image of a cat and the label “cat” as well as other pertinent metadata (see also LABEL, METADATA, and TRAINING DATA below in this section and SUPERVISED machine learning in Section 4).
452
453
454
 FEATURE
455
456
 One attribute being analyzed, considered, or explored across the dataset, often viewed as a column in a table. Features can be any machine-readable (i.e. numeric) form of an instance: images converted into a sequence of pixels, for example. Note: Researchers often select and “extract” the features most relevant for their purpose. Features are not given by default. They are the results of decisions made by datasets’ creators and users. (For more discussion of ENGINEERING FEATURES see Section 5.)
457
458
459
 LABEL
460
461
 The results or output assigned by a machine learning model, or a descriptor included in a training dataset meant for the model to practice on as it is built, or in a testing or benchmark dataset used for evaluation or verification. (See Section 7.2.2 for more on labels’ creation and their potentially harmful impacts.)
462
463
464
 METADATA
465
466
 Data about data, metadata is supplementary information that describes a file or accompanies other content, e.g. an image from your camera comes with the date and location it was shot, lens aperture, and shutter speed. Metadata can describe attributes of content and can also include who created it, with what tools, when, and how. Metadata may appear as TABULAR DATA (a table) and can include captions, file names and sizes, catalog index numbers, or almost anything else. Metadata are often subject- or domain-specific, reflecting how a group organizes, standardizes and represents information [23].
467
468
469
 DATASHEET
470
471
 A document describing a dataset’s characteristics and composition, motivation and collection processes, recommended usage and ethical considerations, and any other information to help people choose the best dataset for their task. Datasheets were proposed by diversity advocate and computer scientist Timnit Gebru, et al., as a field-wide practice to “encourage reflection on the process of creating, distributing, and maintaining a dataset, including any underlying assumptions, potential risks or harms, and implications for use” [24]. Datasheets are also resources to help people select and adapt datasets for new contexts.
472
473
474
 SAMPLE
475
476
 A selection of the total dataset, whether chosen at random or using a particular feature or property; samples can be used to analyze a dataset, perform testing, or train a model. For more on practices like sampling that transform datasets, see Section 5.
477
478
479
 TRAINING DATA
480
481
 A portion of the full dataset used to create a machine learning model, which will be kept out of later testing phases. Imagining a model like a student studying for exams, you could liken the training data to their study guide which they use to practice the material. For example, in supervised machine learning (see Section 4), training data includes results like those the model will be asked to generate, e.g. labeled images. Training datasets can never be neutral, and they commonly “inherit learned logic from earlier examples and then give rise to subsequent ones,” says critical AI scholar Kate Crawford [25] .
482
483
484
 VALIDATION DATA
485
486
 A portion of the full dataset that is separated from training data and testing data, validation data is held back and used to compare the performance of different design details. Validation data is separate from testing data, because validation data is used during the training process to optimize the model while adjustments are being made; therefore, the resulting model will be familiar with its data. That means separate testing data is still needed to confirm how the final model performs. Imagine validation data as practice tests that programmers can administer to check on the model’s progress so far.
487
488
489
 TESTING DATA
490
491
 A portion of the full dataset that is separated from the training data and validation data, and that is not involved in creation of a machine learning model. Testing data is then run through the completed model in order to assess how well it functions. Testing data for the model would be similar to the student’s final exam.
492
493
494
 TENSORS: SCALARS, VECTORS, MATRICES (oh my!)
495
496
 Software for working with machine learning datasets organizes information in numerical relationships, in grids called TENSORS. Understanding tensors can help you understand how data are viewed, compared, and manipulated in computational models. Their grids can have many dimensions, not only two-dimensional X-and-Y graphs [26] . A SCALAR describes a single number. A VECTOR is a list (aka an array), like a line of numbers. A MATRIX is a 2D tensor, like a rectangle of numbers. And a grid of three (or more) dimensions is a TENSOR, like a cube of numbers, or a many-dimensional cube of numbers.
497
498
499
 DATA SUBJECTS
500
501
 The people and other beings whose data are gathered into a dataset. Even if identifying information has been removed, datasets are still connected to the subjects they claim to represent.
502
503
504
 DATA SUBJECTEES
505
506
 This new and somewhat unwieldy term is used here to describe people impacted directly or indirectly by datasets, distinct from data subjects. Data subjectees include anyone affected by predictions made with machine learning models, for example someone forced to use a facial detection system to board a flight or eye-tracking software to take a test at school. Similarly, Batya Friedman and David G. Hendry of the Value Sensitive Design Lab distinguish between “direct” and “indirect stakeholders” to describe the different types of entanglement with technologies [ 27] .
507
508
509
510
 For your consideration: What other parts of a dataset are not included here but could be? How do you see dataset parts differently when you consider them within “data settings,” or contexts, tied to data subjects and data subjectees? [1] What kinds of contexts are impossible to include in datasets?
511
512
513
514
515
516
517
518
519
520
521
522

4
523
 TYPES OF DATASETS
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
 TYPES OF DATASETS
547
 4.1
548
 WHAT DISTINGUISHES
TYPES OF DATASETS?
549
550
551
552
553
 You may choose a dataset based on what it contains, how it is formatted, or other needs. For example, computer vision datasets include thousands of IMAGE or VIDEO files, while natural language processing datasets contain millions of bytes of TEXT. You may work with waveforms as SOUND files or time series data, or network GRAPH stored in structured text formats like JSON. In tables you might find PLACE data as geographic coordinates or X-Y-Z coordinates, TIME as historical date sequences or milliseconds. Likely, you’ll work with other types, too, or with combinations of MULTIMODAL data. Each dataset may include corresponding METADATA, documentation, and (hopefully) a complete DATASHEET.
554
555
556
 You can also consider datasets based on whether the information is STRUCTURED, such as tabular data formatted in a table with labeled columns, or UNSTRUCTURED, such as plain text files or unannotated images. Annotating or coding a dataset prepares it for analysis, including supervised machine learning; and annotation raises important questions about labor, classification, and power. (See Section 6.1 for more on annotation and labeling.)
557
558
559
 Datasets for SUPERVISED machine learning need to include labels for at least a portion of the data that the system is designed to “learn.” This means, for example, that a dataset for object recognition would contain images as well as a table to describe the manually located object(s) they contain. It might have columns for the object name or label, as well as coordinates for the object position or outline, and the corresponding image’s file name or index number.
560
561
562
 In contrast, UNSUPERVISED machine learning looks for patterns that are not yet labeled in the dataset. It uses different kinds of machine learning algorithms, such as clustering groups of data together using features they share. However, it would be a misnomer to think that conclusions drawn from unsupervised machine learning are somehow more pure or rational. Much human judgment goes into developing an unsupervised machine learning model — from adjusting weights and parameters to comparing models’ performance. Often supervised and unsupervised approaches are used in combination to ask different kinds of questions about the dataset. Other kinds of machine learning approaches (like reinforcement learning) don’t fall neatly into these high-level categories.

(For a discussion of deprecated datasets, see Section 7.2.3, and for critical questions at every stage of working with datasets, see Section 6.)
563
564
565
566
 TYPES OF DATASETS
567
 4.2
568
 EXAMPLES: HOW HAVE RESEARCHERS, ENGINEERS, JOURNALISTS, AND ARTISTS PUT DATASETS TO USE?
569
570
571
572
 When starting a project, you may not know what kind of dataset you need. You might work with a particular kind of media or file type most often, so you start there — or maybe you want to try a new form. You may start with a curiosity, and you’re open to datasets in any format, from any source. To spark your imagination, here are four projects that used pre-existing datasets in novel and creative ways:
573
574
 JOURNALISTS UNCOVER RAINFOREST EXPLOITATION WITH GEOSPATIAL DATA
575
 Brazilian investigative journalists at Armando.info, collaborating with El País and Earthrise Media, used field reports and satellite images to find deforestation, hidden runways, and illegal mining in the Venezuelan and Brazilian Amazon. Through computer vision analysis developed from analog maps and used on imagery from a European Space Agency satellite, the journalists compared this analysis with existing information, including complaints from Indigenous communities. “It's not that this was a technology-only job,” says Joseph Poliszuck, Armando.info’s co-founder. “The technology allowed us to go into the field without being blindfolded” [ 28] . Using similar methods, Pulitzer fellow Hyury Potter detected approximately 1,300 illegal runways, more than the number of legally registered ones in the Brazilian Amazon. Combining data work with fieldwork in international collaborations helped these journalists connect local stories to larger scale climate crises and to support communities’ efforts to create change.
576
577
578
 HISTORIANS ASSEMBLE FRAGMENTS OF ANCIENT TEXTS
579
 Researchers from Google’s DeepMind used a neural net on an existing scholarly dataset to complete, date, and find attributions for fragments of ancient texts. They drew on 178,551 ancient inscriptions written on stone, pottery, metal, and more. that had been transcribed in the text archive Packard Humanities Institute’s Searchable Greek Inscriptions [29] . They said that the “process required rendering the text machine-actionable [in plain text formats], normalizing epigraphic notations, reducing noise and efficiently handling all irregularities” [ 30] . They collaborated with historians and students to corroborate the machine learning outputs, calling it a “cooperative research aid” showing how machine learning research can include humans in the training process. They also created an open-source interface: ithaca.deepmind.com
580
581
582
 SOUND ARTIST EXPERIMENTS WITH REFUGEE ACCENT DETECTION TOOLS
583
 Pedro Oliveira’s work [31] explores the accent recognition software used since 2017 by the German Federal Office for Migration and Refugees (BAMF). Though BAMF does not disclose the software’s datasets, Oliveira traced the probable source to two annotated sound databases from the University of Pennsylvania — unscripted Arabic telephone conversations named “CALL FRIEND” [32] and “CALL HOME” [33] . In 2019 the software had an error rate of 20 percent despite its deployment 9,883 times in asylum seekers’ cases [34] . Oliveira utilizes sounds removed from the datasets and reverse engineers the algorithm (as musical transformations rather than for classification tasks), in order to show how politically charged it is to define and detect accents. “How can you say it’s an accurate depiction of an accent?” he says. “Arabic is such a mutating language. That’s the beauty of it actually” [35] . He presents this through live performance and the online sound essay “On the Apparently Meaningless Texture of Noise” [36] .
584
585
586
 HUMAN RIGHTS ACTIVISTS ACCOUNT FOR WAR CRIMES WITH SYNTHETIC DATA
587
 Sometimes important training data is missing from a dataset, because not enough of it exists, and these absences can amplify narrow assumptions about a diverse community. In other cases that don’t involve human subjects, synthetic data can fill gaps in creative ways. When human rights activists from Mnemonic, who were investigating Syrian war crimes using machine learning, struggled to find enough images of cluster munitions to train their model, computer vision group VFRAME created synthetic data — 10,000 computer-generated 3D images of the specialized weapon and its blast sites — which researchers then used to sift through the Syrian Archive’s 350,000 hours of video, searching for evidence of war crimes [37] , [38] . Such systems can reduce the number of videos people need to comb through manually, while still keeping humans involved with pattern review and confirmation.
588
589
 THERE ARE MANY MORE EXAMPLES LIKE THESE OF HOW TO SOURCE, USE, AND COMBINE DATASETS THAT ALREADY EXIST. THE CRITICAL AND CREATIVE POSSIBILITIES ARE NEARLY ENDLESS.
590
591
592
 For your consideration: What kind of dataset(s) will you use, and how can you approach it more critically? How will you apply what you’ve learned here to your next machine learning project?
593
594
595
596
597
598
599
600
601

5
602
 TRANSFORMING DATASETS
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
 Just as there is no such thing as neutral data, no dataset is ready to use off the shelf. From preprocessing (sometimes confusingly called ‘cleaning’) to model creation, transformations reflect the perspectives of the dataset creators and users. This overview covers some of the technical details of getting a dataset ready for your tasks [18], [23], [26], [39], [40], [41], {9}, while asking critical questions along the way.

As artist and researcher Kit Kuksenok argues, “Data cleaning is unavoidable! Each round of repeated care of data and code is an opportunity to invite new perspectives to code/data technical objects” [42]. Preprocessing is a key part of building any system with a dataset, so it is crucial to document and reflect upon preprocessing transformations.
624
625
626
 CAUTION: Be on the lookout for dataset transformations that result in lost meanings, new misconceptions, or skewed information.
627
628
629
630
631
632
633
634
635
 STORING DATA
636
637
 A dataset must live somewhere, and once it grows beyond a single manageable file, it usually lives in a DATABASE. While ‘dataset’ describes what the data are, ‘database’ describes how data are stored — whether as a set of tables with columns and rows (e.g. a relational database like `SQL`, or a collection of documents with keys and values like `MongoDB`). Database structures should suit what they hold, but they will also shape what they hold and reflect how database designers see data. “They also contain the legacies of the world in which they were designed,” says media studies scholar Tara McPherson [43].
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
 ACCOUNTING FOR MISSING DATA
661
662
 You may have entries in your dataset that read `NaN` (not a number) or `NULL`, which may or may not cause errors, depending on what kinds of calculations you do. You may also have manual entries like, ‘?’, ‘huh’, or blanks that lack context. Should you remove the missing information? If you replace it, how will you know what goes in its place? Is data missing in uniform ways such that whole categories can be eliminated, or is it only missing for subgroups in ways that could skew results? How will you know what impacts your edits may have? Consider what missing data might mean. “Unavailable [is] semantically different from data that was simply never collected in the first place,” says data scientist David Mertz [41]. Filtering out data and filling in data have very different implications. Could you consult data subjects to get more context on missing data or the implications of removal or substitutions? How have others handled similar challenges? Can you run tests that treat missing data differently and compare the results? Mimi Ọnụọha’s “The Library of Missing Datasets” reflects on how missing data imply what will not or cannot be collected, or what has been considered not worthy of collection. The project creates a physical archive of empty files, covering topics that are excluded despite our data-hungry culture. She says, “That which we ignore reveals more than what we give our attention to” [44].
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
 HANDLING EXTRA DATA
708
709
 You’ll probably encounter dataset anomalies, outliers, and duplicates and then need ways to identify, adjust, or remove them. In text datasets for unsupervised learning, you’ll likely remove punctuation and “stop words” (commonly used conjunctions or articles like ‘an’ or ‘the’, for example). But, as software engineer Francois Chollet says, ”even perfectly clean and neatly labeled data can be noisy when the problem involves uncertainty and ambiguity” [18]. Outliers can also be accurate and contain meaningful information. As Crawford emphasizes, such acts of data cleaning and categorization create their own concepts of outside and otherness that can restrict “how people are understood and can represent themselves” [25]. Defining outliers, anomalies, or extra data means deciding what is ‘normal’, unexpected, or distracting — what is signal and what is noise.
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
 DISCRETIZING DATA:
740
741
 AKA “binning” or grouping instances together may be useful when you don’t need the original level of detail provided (see ‘dimensionality reduction’ in ENGINEERING FEATURES, below). For example, you might switch continuous data like temperature readings into bins grouped by every five or ten degrees. There are built-in functions for doing so, but remember that creating data ranges can skew results and may not be appropriate for all cases. Make sure to document any changes and provide the original dataset as well as the modified version.
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
 TOKENIZING OR CHUNKING DATA:
764
765
 Breaking up data into smaller units. Tokens are often individual words or sentences. Other text-related operations might include removing punctuation and stop words that are commonly used (see HANDLING EXTRA DATA). Datasets should use tokenization strategies that account for linguistic differences, since a ‘word’ as a unit of meaning can vary significantly among languages. Other sequence-like data types like audio and video are also broken up into chunks, sometimes in order to be processed, compressed, or streamed.
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
 NORMALIZING DATA
787
788
 Altering numerical data to bring them all within the same range and using the same units, e.g. between 0–1, is called normalization, or scaling [18]. In text datasets, this can also mean converting all text to lowercase and standardizing each word by reducing it to its root form (stemming or lemmatizing). In image data, this could also mean cropping images to the same dimensions or around the same subject, changing the color profile to grayscale, and so on. Remember that, while useful for many tasks, normalizing data potentially removes important context or adds ambiguity to the data (e.g. cropped images ignore any information outside the frame, acronyms may be read as their homonyms, scaled numbers may then be rounded and lose specificity).
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
 SEPARATING TESTING & VALIDATION DATA
816
817
 If it has not already been separated, mark off a portion of your data that will not be exposed to your model or used to train it in any way. Keep these testing and validation portions separate from your training data, so that they can be used later to test your model’s performance on new information once it has been trained on the remaining training data. (See Section 3, TRAINING DATA and VALIDATION DATA for more context.)
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
 ENGINEERING FEATURES
837
838
 You may need to create features (e.g., add columns to your table) to show data from new perspectives. This can impact how the dataset can be analyzed going forward, how the model can be designed, and how the data subjects and subjectees might be affected. For example, the unsupervised machine learning approach called ‘dimensionality reduction’ analyzes a dataset for its most relevant features so that the rest can be ignored. It can also involve combining or altering existing features to simplify the whole. However, this runs directly counter to what legal scholar Kimberlé Crenshaw calls intersectional analysis — an approach that rejects grouping people together in categories without attending to the unique experiences (and the data) of people for whom those categories intersect, who are most negatively impacted by systems with the power to categorize people [45] (For more on intersectionality, see ALTERNATIVE APPROACHES TO DATASET PRACTICES).
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
 EXPLORING DATA
872
873
 Sorting, sampling, combining, pivoting, and visualizing data are other transformations you will likely use during dataset preprocessing. These will differ greatly depending on the type of dataset and your project’s objectives, but they all require asking critical questions about how such explorations influence the meaning of the dataset, the model developed, and the system deployed.
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
 For your consideration: How have your data transformations shaped your dataset so far? Which transformations are most thought-provoking and worth exploring? Who else could offer perspective on your preprocessing decisions?
891
892
893
894
895
896
897
898
899
900
901
902

6
903
 THE DATASET'S LIFECYCLE
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
 DATASETS OFTEN HAVE SURPRISING HISTORIES, USES, AND AFTERLIVES. FROM LOCATING THE BEST DATASET FOR THE JOB, TO WORKING WITH IMPERFECT EXISTING DATA, TO SHARING RESULTS AND ACCOUNTING FOR IMPACTS, HERE ARE SOME CRITICAL QUESTIONS TO ASK AT EACH STAGE OF A PROJECT THAT USES MACHINE LEARNING DATASETS:
922
923
924
 THE DATASET'S LIFECYCLE
925
 6.1
926
 ORIGINS: WHAT IS YOUR DATASET’S STORY?
927
928
929
930
 “Datasets are the results of their means of collection,” says artist and technology researcher Mimi Ọnụọha [46]. They are influenced by the many people who contribute to them, who participate in their creation (knowingly or unknowingly), who collect data, who annotate or label data, or who are affected by a machine learning system that uses the dataset.

These questions will help you select a dataset to work with, and to understand how its creation could inform your project if you use it. Often the answers to these questions can be found in a dataset’s DATASHEET or a related research paper. Although including datasheets is becoming a standardized practice, not all datasets have datasheets, and many datasheets are incomplete. Look for datasets with complete datasheets, updated documentation, and current contact information for its creators.
931
932
933
 WHO CREATED THIS DATASET? WHO FUNDED IT? WHAT WERE THEIR MOTIVATIONS OR AIMS, AND HOW DO THEY COMPARE TO YOURS? [24]
934
935
936
937
938
939
940
941
 If they differ in significant ways, consider how using the dataset for other purposes will impact both the original data subjects, as well as data subjectees, and the outcomes of your project. Would an alternative dataset be more appropriate? Document the rationale for the dataset you choose.
942
943
944
945
946
947
948
949
950
951
952
953
954
 HOW WAS THE DATA COLLECTED? HOW WAS IT ANNOTATED AND BY WHOM? ARE THE ORIGINAL ANNOTATION INSTRUCTIONS AVAILABLE? HOW ARE THE LIMITATIONS OF THOSE METHODS ACCOUNTED FOR? [47] WERE DATA SUBJECTS PART OF THE DATASET’S DESIGN AND CREATION? WAS THE RESULTING DATA VALIDATED BY ITS SUBJECTS? [24], [48]
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
 If information about collection and annotation is missing, or if the collection and annotation methods were inappropriate or misaligned with your objectives, you may want to consider an alternative dataset. Also consider the contexts of labeling and annotation, as crowd-sourced data can lack the nuance of individual annotators from diverse perspectives who had the opportunity to collaborate [49].
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
 HOW HAS THE DATASET BEEN PROCESSED ALREADY? IS IT A SMALL SAMPLE OF A LARGER COLLECTION? IS IT A COMPILATION OF OTHER PRE-EXISTING DATASETS? HAS IT BEEN STANDARDIZED OR TRANSFORMED IN ANY WAY (SEE SECTION 5)?
987
988
989
990
991
992
993
994
995
996
997
998
999
 If the dataset is a sample from a larger dataset or a compilation of smaller datasets, investigate those original sources to see if its data matches the sample or is more appropriate for your work. Does information in its datasheet impact your decision to use this dataset? If it has been transformed, see if the documentation also includes the original version or a description of its methods.
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
 WHAT DOES THE DATASET CONTAIN? DOES IT INCLUDE A CODEBOOK DESCRIBING ITS PARTS? WHICH PERSPECTIVES ARE INCLUDED, AND WHICH ARE MISSING? WHICH OUTLIERS ARE DISMISSED, AND WHAT DATA IS UNACCOUNTED FOR? CAN YOU AUDIT THE DATASET, OR HAS IT ALREADY BEEN AUDITED? WHAT DOES THE AUDIT SHOW AND HOW CAN YOU ACCOUNT FOR ITS FINDINGS?
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
 If the dataset has gaps that neglect important considerations or that might affect your project, would a different dataset be more appropriate?
1032
1033
1034
1035
1036
1037
1038
 Compare with other datasets in this area to see how they account for similar issues. Consult with your project group or community of practice to see how they interpret these issues and their importance. Get their help to spot issues you might have missed, and consider any other datasets that might better fit the project. Offer in-kind support.
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
 Regardless of whether you use the dataset, be sure to document your questions and concerns. Describe the limitations you see in the dataset, discuss their relevance to your project, and share what you have done to mitigate their impacts. Proceed with caution, if at all.
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
 WHEN WAS THE DATASET MADE? IS THIS ITS LATEST VERSION? IF IT HAS BEEN DEPRECATED (OR REMOVED FROM PUBLIC CIRCULATION), WHY? DOES IT CONTAIN INFORMATION THAT IS INACCURATE OR OFFENSIVE? FOR MORE DISCUSSION OF DEPRECATED DATASETS, SEE SECTION 7.2.3.
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
 If the dataset is no longer valid, you will need to find another dataset. Maybe an updated version of the dataset addresses the issues that led to its removal, or perhaps you can make these revisions yourself. However, resolving ethics or accuracy issues is not as simple as updating a table, since the underlying structure of the dataset may remain problematic. Proceed with caution, if at all.
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
 Just because a dataset is *not* deprecated does not mean it is safe to use. There is not yet any standardized way to audit, update, or remove existing datasets, or even a central repository where issues can be documented and addressed. Much remains at the creators’ discretion [50].
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
 WHO IS FEATURED IN THE DATASET? WHO IS LEFT OUT? HOW DOES THE DATASET ACCOUNT FOR WHAT IS MISSING? WHAT ASSUMPTIONS, INTUITIONS, THEORIES, STEREOTYPES, OR INEQUITIES ARE CONTAINED IN THE DATA OR BUILT INTO THE DATASET’S STRUCTURE? HOW MIGHT THESE FRAMEWORKS HAVE BEEN PERPETUATED THROUGH FORMATTING AND TRANSFORMATION PROCESSES AS THE DATASET WAS MADE MACHINE-LEGIBLE?
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
 If you are unsure how the dataset and your use of it may impact a community, consult with its members to understand their perspectives. “Build with, not for,” says the Design Justice Network, which emphasizes that “community members already know what they need and are working towards solutions that work for them” [51].
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
 Consider the consequences of inclusion. In an unjust system (e.g. racist criminal sentencing) more representation is not the right answer. Completeness may do more harm.
1133
1134
1135
1136
1137
1138
1139
 Document your processes. Noticing what information is left out can be just as important as what is shown [48]. If too much information is missing, or omissions are too problematic, you may need to find another dataset.
1140
1141
1142
1143
1144
1145
1146
1147
 Maybe just don’t build it. The best solution to a problem is not always more technology or a new system. Explaining why you did not build a model or did not use a dataset can still be a valuable contribution – showing the need for caution, skepticism, and alternative approaches.
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
 HOW WAS CONSENT GIVEN FOR INCLUSION IN THE DATASET? WAS CONSENT FULLY INFORMED, VOLUNTARY, AND REVOCABLE? HOW ARE SUBJECTS’ ANONYMITY PROTECTED?
1160
1161
1162
1163
1164
1165
1166
1167
1168
 If the dataset does not adequately address how consent was provided, or if this consent does not extend to the uses of your project, you will need to find a different dataset. If your use potentially risks subjects’ anonymity as established in the original dataset design, you will need to find a different dataset. Privacy cannot be guaranteed; be wary of the potential for re-identification of so-called anonymous data when they are combined with other datasets [52].
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
 If full consent was not given — especially if consent was not possible to obtain — “just don’t build it” is always a valid and responsible option.
1184
1185
1186
1187
1188
1189
1190
1191
 DO YOU HAVE LEGAL, ETHICAL ACCESS TO THIS DATASET? DOES YOUR USE OF IT ALIGN WITH ITS LICENSING AND TERMS OF USE, AND WITH YOUR OWN CODES OF CONDUCT? ARE ANY OF THE DATA INTENDED FOR RESTRICTED USE BY SPECIFIC COMMUNITIES? [53] – [55]
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
 We are not lawyers and this is not legal advice, of course. Check at the beginning, middle, and toward the end of your project — before publication or launch — that your use is authorized and appropriate to its creators and its subjects. It should be in keeping with both the letter and the spirit of terms of use and codes of conduct. Talk with others doing similar work to see what pitfalls they wish they had avoided upfront.
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
 OVERALL, IS THIS THE BEST DATASET FOR YOUR PROJECT? WHAT ARE THE TRADEOFFS OF USING THIS DATASET VERSUS ANOTHER ONE?
1218
1219
1220
 THE DATASET'S LIFECYCLE
1221
 6.2
1222
 USAGE: WHAT IS THE STORY YOU WILL TELL WITH YOUR DATASET?
1223
1224
1225
1226
 Your project aims will steer your dataset selection, the features you choose to interpret, the model you pick, and the adjustments you make. Across all of these small and large technical choices, you can use critical lenses to achieve your aims and minimize harms.
1227
1228
1229
 WHAT IS YOUR PROJECT’S GOAL? HOW DOES THE DATASET HELP YOU ACHIEVE IT?
1230
1231
1232
1233
1234
 Prioritize your own purpose and approach over the popularity of the dataset or your familiarity with it. What is the best dataset for this task, to answer these questions?
1235
1236
1237
1238
1239
1240
1241
1242
 WHAT ASSUMPTIONS ARE YOU PRIORITIZING OR EXCLUDING BY HOW YOU’VE FRAMED YOUR PROJECT? HOW MIGHT IT BE REFRAMED TO GAIN MORE INSIGHT? HOW MIGHT YOU COLLABORATE WITH PEOPLE FROM DIFFERENT DISCIPLINES OR BACKGROUNDS — E.G. ARTISTS, PRACTITIONERS, OR COMMUNITY STAKEHOLDERS – TO DEVELOP THE PROJECT’S AIMS MORE RICHLY?
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
 Gather and prioritize the perspectives of data subjects and data subjectees. Connect with experts from other domains for fresh eyes and constructive feedback. Seek out information that you would not normally encounter; don’t automatically rule out information in an unfamiliar vocabulary or format.
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
 WHAT ASSUMPTIONS ARE YOU MAKING AS YOU PROCESS AND CLEAN THE DATASET, AND AS YOU SELECT FEATURES FOR ANALYSIS? COULD CATEGORIES (FEATURES) IN THIS DATASET BE DISAGGREGATED TO TELL A DIFFERENT STORY THROUGH THE DATA? SOME COMMON ASSUMPTIONS:
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
 How have you treated the collected data as neutral in any obvious or more subtle ways?
1282
1283
1284
1285
 Consider how transforming data [40] during the cleaning process may misrepresent information or remove important detail from the dataset. e.g. `NaN` (Not a Number) may conceal data that was never collected. (See Section 5 for more.)
1286
1287
1288
1289
1290
1291
1292
1293
 A dataset is frequently treated as a generalizable, multi-purpose tool, applicable across many tasks and disciplines, when more often it is best suited or only suited to its original purpose.
1294
1295
1296
1297
1298
1299
1300
 Data cleaning is not one-and-done, but an iterative, integral process [42]. Though often undervalued labor, data cleaning is part of the ‘real’ work of making datasets.
1301
1302
1303
1304
1305
1306
1307
1308
 WHAT ASPECTS OF THE DATASET WILL YOU INCLUDE OR EXCLUDE IN YOUR PROJECT, AND HOW DO THEY CONVEY THE INFORMATION? HOW WILL YOU ENSURE THESE CHOICES DO NOT OVERLY SKEW THE RESULTS?
1309
1310
1311
1312
1313
1314
1315
1316
1317
 FEATURE SELECTION & ENGINEERING. Distilling a dataset into pertinent columns is an essential part of dataset work because it determines what information categories will be important for later analysis. This process is descriptive and creative, not self-evident [26].
(See Section 3: FEATURE and Section 5: ENGINEERING FEATURES for more.)
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
 HOW FAR DOES YOUR PROJECT DEVIATE FROM THE DATASET’S ORIGINAL PURPOSE? IF IT IS SIGNIFICANTLY DIFFERENT, HOW WILL YOU TRACK ANY NEW IMPACTS? [54]
1330
1331
1332
1333
1334
1335
1336
1337
 When diverging from the original objectives of a dataset, ensure that your dataset transformation processes align with both your own goals and any guidelines put in place by the dataset’s creators. Consider consent and licensing restrictions, as well as other potential legal and ethical issues. Furthermore, what kinds of impacts would the creators not have foreseen in your use of their dataset? How might communities be newly impacted by your use of this dataset, and how can you engage them in this process?
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
 COULD YOUR USE OF THE DATASET CAUSE HARM? HOW WILL YOU MEASURE AND ADDRESS ADVERSE IMPACTS? WHAT STEPS WILL YOU TAKE TO MINIMIZE HARM? [54]
1355
1356
1357
1358
1359
1360
1361
 Work closely with data subjects and potential data subjectees who may be impacted by your use of the dataset, in order to discover what mitigation strategies would be best for them. How you address potential impacts will depend greatly on the types of risks your project presents; listening across diverse communities and disciplines will help you uncover new issues, create checkpoints, and take useful action. Encourage in your project team a culture of openness and learning from mistakes [48].
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
 HOW WILL YOU MAINTAIN THE CONSENT AND ANONYMITY OF ANY DATA SUBJECTS? HAVE THEY BEEN TOLD ABOUT THE RISKS SPECIFIC TO YOUR USE CASE? CAN THEY CHECK BACK ON HOW THEIR DATA HAS BEEN USED?
1378
1379
1380
1381
1382
1383
1384
1385
1386
 Consent requires that subjects understand both the implications of data’s use and impact, as well as their digital rights [48]. This goes beyond terms of service disclosures and should include the purpose of the project, any risks, and instructions on how to revoke consent if necessary.
1387
1388
1389
1390
1391
1392
1393
1394
1395
 There is no such thing as “anonymizing” identifying data because data can easily be combined with other sources to re-identify people [48].
1396
1397
1398
1399
1400
1401
1402
1403
 DOES YOUR WORK WITH THIS DATASET RESULT IN A NEW, DERIVATIVE DATASET? HOW WILL YOU ACCOUNT FOR NEW ETHICAL CONCERNS ARISING FROM THE DERIVATIVE DATASET WHILE STILL ADDRESSING ISSUES RAISED BY THE ORIGINAL?
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
 Refer to Section 6.1 ORIGINS, as well as to resources for creating datasets like Gebru, et al.’s, “Datasheets for Datasets” to ensure that you are considering the questions that arise from creating a new or derivative dataset [24]. Your documentation, including a new datasheet, will help others who may want to use your new dataset.
1414
1415
1416
1417
1418