Gender Differences in Collaboration Patterns in Computer Science

Abstract

The research discipline of computer science (CS) has a well-publicized gender disparity. Multiple studies estimate the ratio of women among publishing researchers to be around 15–30%. Many explanatory factors have been studied in association with this gender gap, including differences in collaboration patterns. Here, we extend this body of knowledge by looking at differences in collaboration patterns specific to various fields and subfields of CS. We curated a dataset of nearly 20,000 unique authors of some 7000 top conference papers from a single year. We manually assigned a field and subfield to each conference and a gender to most researchers. We then measured the gender gap in each subfield as well as five other collaboration metrics, which we compared to the gender gap. Our main findings are that the gender gap varies greatly by field, ranging from 6% female authors in theoretical CS to 42% in CS education; subfields with a higher gender gap also tend to exhibit lower female productivity, larger coauthor groups, and higher gender homophily. Although women published fewer single-author papers, we did not find an association between single-author papers and the ratio of female researchers in a subfield.

1. Introduction

The gender gap in science, technology, engineering, and mathematics (STEM), and in particular in computer science (CS), is a well-known and well-studied problem. It carries significant societal effects, such as inequality in economic opportunities for women and an undersupply of researchers and engineers in the rapidly growing discipline [1,2,3,4]. The gender gap among researchers is particularly noteworthy: the people who participate in research, publish about it, and have their research acknowledged for its value are predominantly men [5]. Numerous studies estimate that only about 15–30% of the CS research community are women [6,7,8,9,10]. Although some recent indications show these numbers could be growing, they remain low, and the rate of growth remains slow [11].

The gender gap, defined simply as the difference in participation between men and women, is nevertheless a complex, multifaceted societal phenomenon [12]. Numerous approaches to understand and perhaps increase the representation of women have focused on aspects such as resource availability, gender stereotypes, child care, structural barriers, gender differences, discrimination, and other factors. This article focuses on one of these factors: the collaboration patterns of paper coauthors across genders and CS fields.

Scientific collaborations are the backbone of a successful career in science [13]. For example, researchers with more collaborators have been found to publish more articles, publish in higher impact journals, and accrue citations more quickly [14]. Consequently, many studies have investigated whether women and men collaborate at different rates across disciplines, and have often found significant differences [15,16,17,18].

In CS, and in particular in its more experimental fields such as computer systems, graphics, and artificial intelligence, collaboration is crucial because the large-scale implementation efforts involved often require teams of researchers with various experience levels. In this article, we focus on gender differences in collaboration patterns across the fields and subfields of CS, both experimental and theoretical.

Our study design is descriptive and observational in nature. We did not start out with any preset hypotheses to validate. Instead, our goal was to collect and analyze up-to-date, accurate, and extensive data on CS authorship and collaboration patterns across genders. These data and this analysis provide baseline statistics for comparison across different time points and scientific disciplines. However, it also provides immediate answers and comparisons to existing work, thereby offering new insights into the current state of collaboration differences across genders and CS fields. Specifically, in this article we address the following research questions:

• RQ1: What are the ratios of women and men among CS conference authors?
• RQ2: Do women publish less than men?
• RQ3: Are productivity differences affected by collaboration size?
• RQ4: Do women collaborate with fewer people than men?
• RQ5: Do women publish fewer single-author papers?
• RQ6: Are team sizes (coauthor groups) larger in more experimental subfields?
• RQ7: Do authors exhibit gender homophily in their choice of coauthors?

To bring these questions into historical context, we next briefly survey some of the previous work in the area.

1.1. Related Work

There exists rich literature on the gender gap in the sciences in general, and in computer science research in particular. For a recent review of these works, refer to Avolio et al.’s review from 2020 [12]. Instead, we limit our focus to the relevant literature on collaboration patterns and differences.

For example, a recent study of differences in collaboration patterns across disciplines found that female scientists have a lower probability of repeating previous coauthors than males. It also found that female faculty have significantly fewer distinct coauthors over their careers than males, but that this difference can be fully accounted for by females’ lower publication rate and shorter career lengths [19].

This productivity gap, which we observed in our dataset as well, has been thoroughly explored in several other studies [17,20,21,22,23]. In the social sciences, one study has found that women generally publish fewer papers than men and that two thirds of the single-author papers were written by men [24]. In mathematics, women also publish less than men, especially early in their careers, and leave academia at a higher rate than men [25]. Women are also underrepresented in the three top-ranked journals and publish fewer single-author papers. In terms of the mean number of coauthors, women’s statistics are similar to men’s. That being said, there is even a gap in recognition, as women are also less likely to receive tenure the more that they coauthor [26].

CS researchers in particular tend to collaborate more than researchers in other fields, regardless of gender, and there are no gender-specific differences in how collaborative behavior impacts scientific success [27]. This study also found that gender homophily in CS has been increasing over the past few years.

On a related note, another study of collaboration patterns across the sciences found that CS papers average 2.84 coauthors, and electrical engineering papers average 3.04 [28], which is somewhat lower than what we have found in this study. It also found that generally, in engineering, female–female collaborations accounted for only 7% of all total pairs. In CS, the percentage was even lower (5%). Since 1990, there have been even more same-gender (gender homophily) coauthorships than expected [11]. However, this property can vary across CS fields, necessitating more nuanced analysis. For example, in the field of data visualization, women collaborated with substantially more women than men [29].

Corroborating this result for biotech patent networks, women have been found to be more likely to collaborate with women, and benefit from it, but both genders mostly collaborate with men [13]. There are also fewer women “stars“, which we also found to hold specifically for the subfield of high-performance computing [30].

A surprising result came from a survey of 1714 scientists in 2011, finding that when accounting for various confounding factors, women actually have more collaborators than men [31]. The paper also reported that regression models that take into account different collaboration strategies are better at predicting a researcher’s number of collaborators.

Some studies analyzed the researcher gender gap by aggregating the coauthors of each paper into one “gender”. One study analyzed different aggregations based on the proportion of female authors, gender of most senior authors, and single-author papers [32]. Looking at author position for aggregation, another study found that there are fewer women in first and last author positions in science overall, as well as in single-author papers [33]. Other ways to aggregate genders include counting all papers that have at least one female author, and those where at least half the authors are female. In this study, we do not aggregate papers by gender, except for the trivial cases when they have a single author.

Research in computer science, and in particular in its more applied and experimental fields, can sometimes require expensive resources. Several studies have found that the gender gap in research tends to be higher in disciplines with expensive barriers to entry [22,34,35,36], which appears to agree with our findings for CS fields. A related question to that of research expense is that of travel expense, since we are focusing on conference publications, which require additional funds and could interfere with family life for people in care roles. Both factors may lead us to speculate that women would be more likely to publish in conferences in their home countries, but that is not what we found in our dataset [37]. In the same study, we also attempted to look at conference attendance and gender, for which we found almost no literature or published data. However, for one large conference that has been publishing attendance demographics since 2016, SC, we found that women’s attendance rates remained near constant around 13–14%, suggesting little variation in time.

1.2. Organization

The rest of this paper is organized as follows. In the next section (Section 2), we describe in detail our data collection methodology, including the manual assignment of genders to authors to avoid the well-known issues of name-based gender inference. In the results section (Section 3), we enumerate our findings, organized by research question, and then summarize an answer to each of the questions. The discussion (Section 4) that follows then elaborates on these answers in an attempt to synthesize insights. Finally, we conclude in Section 5 and suggest directions for future research.

2. Materials and Methods

To answer our research questions, we needed to collect expansive data on CS publications and their authors. Such data collection involves many choices, such as which publications to collect and how to assign gender to authors. The following list enumerates our main data decisions. Each choice necessarily involves trade-offs, and we attempt to justify our choices by explaining which aspects we prioritized.

Conference data instead of journal data. In CS, original scientific results are typically first published in peer-reviewed conferences [38,39], and then possibly in archival journals, sometimes years later [40]. To increase the coverage and relevance of our dataset, we only looked at conference publications. The complete list of selected conferences can be found in

Table 1. All conferences, ordered by subfield and acceptance rate.

Conference	Subfield	Papers	Authors	Acceptance
TACAS	Algorithms	52	189	0.31
SODA	Algorithms	181	442	0.34
AAAI	Artificial Intelligence	599	1760	0.23
ACL	Computational Linguistics	195	610	0.26
SIGCSE	Computer Science Education	87	259	0.23
ITICSE	Computer Science Education	64	186	0.27
CVPR	Computer Vision	532	1702	0.2
ICDM	Data Science and Mining	72	269	0.09
KDD	Data Science and Mining	64	237	0.09
WSDM	Data Science and Mining	80	285	0.16
SIGGRAPH	Graphics	166	560	0.38
CHI	Human–Computer Interaction	599	2020	0.25
ICML	Machine Learning	434	1314	0.26
MM	Multimedia	225	821	0.33
NIPS	Neural Networks	679	2038	0.21
POPL	Programming Languages	64	190	0.23
ICSE	Software Engineering	96	360	0.23
FSE	Software Engineering	94	354	0.32
STOC	Theoretical Computer Science	103	263	0.24
FOCS	Theoretical Computer Science	90	219	0.28
WWW	World-Wide Web	170	614	0.18
ISCA	Architecture	54	295	0.17
ASPLOS	Architecture	56	247	0.18
MICRO	Architecture	61	306	0.19
HPCA	Architecture	50	215	0.22
PACT	Architecture	25	89	0.23
SIGMETRICS	Benchmarking	27	101	0.13
IMC	Benchmarking	28	124	0.16
MASCOTS	Benchmarking	20	75	0.24
ISPASS	Benchmarking	24	98	0.3
ICPE	Benchmarking	29	102	0.35
IISWC	Benchmarking	31	121	0.37
CCGrid	Cloud Computing	72	296	0.25
CLOUD	Cloud Computing	29	110	0.26
HotCloud	Cloud Computing	19	64	0.33
SOCC	Cloud Computing	45	195	Unknown
PLDI	Compilers	47	173	0.15
OOPSLA	Compilers	66	232	0.3
SLE	Compilers	24	68	0.42
PPoPP	Concurrency	29	122	0.22
EuroSys	Concurrency	41	169	0.22
HiPC	Concurrency	41	168	0.22
IPDPS	Concurrency	116	447	0.23
SPAA	Concurrency	31	84	0.24
PODC	Concurrency	38	101	0.25
EuroPar	Concurrency	50	179	0.28
ICPP	Concurrency	60	234	0.29
SIGMOD	Databases	96	335	0.2
PODS	Databases	29	91	0.29
CIDR	Databases	32	213	0.41
IGSC	Energy	23	83	Unknown
HCW	Heterogeneous Computing	7	27	0.47
HPDC	High-Performance Computing	19	76	0.19
SC	High-Performance Computing	61	325	0.19
Cluster	High-Performance Computing	65	273	0.3
ISC	High-Performance Computing	22	99	0.33
HPCC	High-Performance Computing	77	287	0.44
SIGIR	Information Retrieval	78	264	0.22
SIGCOMM	Network	36	216	0.14
NSDI	Network	42	203	0.16
MobiCom	Network	35	164	0.19
CoNEXT	Network	32	145	0.19
HotI	Network	13	44	0.33
SOSP	Operating Systems	39	217	0.17
ICAC	Operating Systems	14	46	0.19
ATC	Operating Systems	60	279	0.22
Middleware	Operating Systems	20	91	0.26
HotOS	Operating Systems	29	112	0.31
SP	Security	60	287	0.14
NDSS	Security	68	327	0.16
CCS	Security	151	589	0.18
FAST	Storage	27	119	0.23
SYSTOR	Storage	16	64	0.34
HotStorage	Storage	21	94	0.36
VEE	Virtualization	18	85	0.42

.

Choice of conferences. Our dataset evolved from our previous study of conferences related to one major field, computer systems [41]. The conferences we selected include some of the most prestigious systems conferences (based on indirect measurements such as Google Scholar’s metrics), as well as several smaller or less-competitive conferences for contrast. For this specific study, we decided to expand the analysis to include some of the most influential conferences in most subfields of CS, based on the same measures, for a total of 6949 papers across CS. Obviously, not all subfields have equal numbers of participants or conferences, and we had no set quota for either to be included in our dataset. Instead, we tried to ensure that each subfield is represented by at least a few hundred authors for statistical power.

Limit data to a single year. Many fields and researchers shift characteristics over time, complicating collaboration analyses. To control for these effects, all of the conferences in our dataset are from a single year, 2017.

Focus on manual gender assignment. Most studies of author gender at scale use automated approaches to assign gender to authors, typically inferred from given names [21,42]. These statistical approaches can be reasonably accurate for names of Western origin, and especially for male names [6,43,44], but can fall short when inferring from Indian and East Asian names. We opted instead to rely primarily on a manual approach that can overcome the limitations of name-based inference. Using web lookup, we assigned the gender of 14,001 of the unique researchers for whom we could identify an unambiguous web page with a recognizable gendered pronoun, or absent that, a photo. (For example, many Linkedin profiles may lack a photo, but include a gendered pronoun in the recommendations section.) For 8939 others, we used genderize.io’s automated gender designations if it was at least 90% confident about them [44]. The remaining 576 persons were assigned “NA” instead of a gender and were excluded from most analyses. This method provided more gender data and higher accuracy than automated approaches based on forename and country, especially for women [11,22,42,43,45]. Consequently, we have very few NA genders relative to comparable studies. We believe that high coverage is critical when analyzing coauthorship networks, because omitting a large number of connected sub-networks (such as people from Asia) may distort our results.

Assignment of field and subfield. We could find no standard definition and delineation of fields and subfields of CS, so we had to come up with our own, which was necessarily subjective (

Table 2. All CS subfields analyzed, arranged by fields.

Subfield	Field
Artificial Intelligence	Artificial Intelligence (AI)
Computational Linguistics	Artificial Intelligence (AI)
Computer Vision	Artificial Intelligence (AI)
Machine Learning	Artificial Intelligence (AI)
Neural Networks	Artificial Intelligence (AI)
Graphics	Human–Computer Interaction (HCI)
Human–Computer Interaction	Human–Computer Interaction (HCI)
Multimedia	Human–Computer Interaction (HCI)
Programming Languages	Software Engineering and Programming Languages
Software Engineering	Software Engineering and Programming Languages
Data Science and Mining	Knowledge
Information Retrieval	Knowledge
World-Wide Web	Knowledge
Computer Science Education	CS Education
Computer Architecture	Systems
Communications	Systems
Operating Systems	Systems
Benchmarking	Systems
Cloud Computing	Systems
Compilers	Systems
Concurrency	Systems
Databases	Systems
High-Performance Computing	Systems
Security	Systems
Storage	Systems
Algorithms	Theory
Theoretical Computer Science	Theory

). Moreover, conferences do not always fall neatly into a single subfield, and some papers may stray from the primary focus of the conference. We note, however, that in most of our analyses, papers in subfields assigned to the same field often exhibited similar characteristics to each other and were distinct from other subfields. This affinity provides some evidence that these assignments are not entirely arbitrary. That said, other researchers may choose different assignments of papers or conferences to subfields and fields. Since our dataset and code are both open and available, we encourage such reevaluations of the data.

2.1. Limitations

The decisions listed above also represent some compromises that limit the generalization or applicability of our analysis. One such limitation is that the data reflects a snapshot in time to avoid the complexities of gender differences in retention rates. However, this choice precludes analyses of changes and trends in collaboration patterns over time.

Another limitation is our choice of which conferences to include out of the hundreds or thousands of annual CS conferences. Moreover, CS is a diverse discipline with significant multidisciplinary collaborations, and in some fields and subfields, journals are preferred to conferences. Consequently, our choice of conferences may not be not representative of all of CS or even a proportional representation of subfields with CS. That said, the relative metrics we measured comparing different subfields are nevertheless meaningful, but metrics over the entire dataset should be taken with a grain of salt. We believe that the large number of authors we included in our analysis provides some statistical robustness and therefore does not significantly deviate from a representative sample of the field of CS as a whole.

For this study, the most critical piece of information on these researchers is their perceived gender. Gender is a complex, multifaceted identity, but most bibliometric studies still rely on binary genders—either collected by the journal, or inferred from the first name—because that is the only designator available to them [6,7,8,9,10,11,46]. In the absence of self-identified gender information for our authors, we also necessarily compromised on using binary gender designations. We therefore use the gender terms “women“ and “men“ interchangeably with the sex terms “female“ and “male“. The conferences in our dataset did not collect or share specific gender information, so we had to collect this information from other public sources.

This labor-intensive approach does introduce the prospect of human bias and error. For example, a gender assigned by an outdated biography paragraph with pronouns may no longer agree with the self-identification of the researcher. To verify the validity of our approach, we compared our manually assigned genders to self-assigned binary genders in a separate survey we conducted among 918 of the authors [41]. We found no disagreements for these authors, which suggests that the likelihood of disagreements among the remaining authors is low. However, the main limitation that arises from this manual process of data collection and gender assignment, is that it does not scale well to a larger number of conferences or years.

Finally, the nature of the current analysis is more descriptive than prescriptive. Rather than presenting preconceived hypotheses and testing them with the data, we ask and answer open-ended research questions that fit in the scope of this paper. The answers to these questions will surely instigate further hypotheses and questions requiring deeper analysis, such as social network analysis, which is also important to understanding many collaboration patterns [13]. The open dataset we provide with this article should enable any interested researcher to perform such analyses.

2.2. Statistics

For statistical testing, group means were compared pairwise using Welch’s two-sample t-test and group medians using the Wilcoxon signed-rank test; differences between distributions of two categorical variables were tested with the ${\chi }^2$ test; correlations between two numerical variables were evaluated with Pearson’s product-moment correlation coefficient. All statistical tests are reported with their p-values.

3. Results

For each research question, we start with descriptive statistics across the entire sample population, and then break the statistics down by field and subfield.

3.1. RQ1: What Are the Ratios of Women and Men among CS Conference Authors?

Before we can look at collaboration patterns, we need to establish a baseline for authorship numbers across genders. For example, the question of how many women or men an author collaborates with makes little sense without the context of how many women and men are available to collaborate with overall. Consequently, the first question we ask is what is the female author ratio (FAR) in our dataset.

Summarizing across all 27,743 authors (with repeats for multiple papers) and omitting the 645 repeated authorships for which we could establish no gender, we find a total of 3833 women, which represents an overall FAR of 14.1% across authors. This result is on the low end of previously reported statistics in the range of 15–30% [6,7,8,9,10]. It is quite possible that our results are on the low end because our choice of conferences, with its emphasis on computer systems, overrepresents fields with lower representation of women. Keep in mind, however, the differences between previous studies and the current one, both in data and in methodology. Our data includes only conferences and only from one year and is by no means exhaustive or necessarily representative of the entire discipline. On the other hand, the smaller sample size allowed us to apply a primarily manual approach to gender assignment, which provides higher accuracy and coverage of researchers. In contrast, most comparable studies use a gender inference approach based on given names, which can fail for names with unclear or no gender association at all, as are many east Asian names, and tend to misidentify women in particular [6,43,44].

Breaking down FAR values for each field and subfield (Figure 1), we find large differences across CS. The highest FAR was exhibited in CS education conferences (42.2%) and the lowest in theoretical CS (6.4%). Most conferences in the field of computer systems hovered around 10% FAR, while the average across the entire field of AI was a little bit larger at (11.8%).

3.2. RQ2: Do Women Publish Less Than Men?

Many papers across disciplines discuss the existence and potential reasons for a productivity gap, that is, the observation that men generally publish more scholarly articles than women. Here, we continue our exploration of the data by looking at the productivity rates across genders and subfields of CS.

Figure 2 shows the overall distributions of paper productivity in CS across genders. Aside from the now-obvious observation that men far outnumber women authors, we can also observe a longer tail for the men’s distribution overall. The interpretation is that the most prolific authors are especially skewed male. On the opposite tail, we find that 62.4% of female authors published only one paper in our dataset, compared to 53% of men.

Overall, men average 2.15 papers per author, compared to women’s 1.77 ($t=-15.41$, $p<{10}^{-9}$). Looking at medians—to attempt to attenuate the large effect of the long tail on means—does not help much. Both medians are naturally 1 but a Wilcoxon signed-rank test still shows a significant difference ($W=\mathrm{39,773,088}$, $p<{10}^{-9}$).

Figure 3 shows the breakdown of productivity distribution by gender, field, and subfield. The field of AI and its subfields show the highest average productivity (and highest outliers, for men), while software engineering, programming languages, and CS education exhibit the lowest average papers per author. In terms of gender, the largest differences in productivity exhibit in the subfield of theoretical CS (difference in means $\Delta$ = 0.8; $t=4.72$, $p<{10}^{-4}$), followed by machine learning ($\Delta$ = 0.62; $t=4.22$, $p<{10}^{-4}$), benchmarking ($\Delta$ = 0.59; $t=3.93$, $p<{10}^{-3}$), HPC ($\Delta$ = 0.53; $t=4.51$, $p<{10}^{-4}$), AI ($\Delta$ = 0.48; $t=4.35$, $p<{10}^{-4}$), and security ($\Delta$ = 0.42; $t=4.12$, $p<{10}^{-4}$). A few subfields show a small productivity advantage for women, such as programming languages ($\Delta$ = −0.04; $t=-0.11$, $p=0.91$), but none are statistically significant.

3.3. RQ3: Are Productivity Differences Affected by Collaboration Size?

Comparing raw productivity across subfields in this way can be misleading, because the typical collaboration size is also related to productivity, but varies by subfield. For example, if in some subfield the typical number of authors per papers is larger than another field, we might also expect that each author’s name would show up in more papers under otherwise equal assumptions. We therefore also look at the “fractional count“ of publications instead, dividing each authorship event by the number of authors on the paper [14].

Overall, men average a total of 0.54 fractional papers per author, compared to women’s 0.44, which is statistically significant ($t=-14.95$, $p<{10}^{-9}$).

This normalized productivity metric exhibits on the one hand smaller differences across fields, but on the other, larger differences across genders (Figure 4). The medians for men’s fractional productivity appear noticeably higher in most subfields, and men exhibit much larger and more numerous outliers on the prolific tail of the spectrum. However, fractional productivity does not appear to affect the relative rankings of subfields very much. Of the fields that are more affected by this metric, the most notable perhaps is theoretical CS, which now exhibits the highest mean fractional productivity.

The modified productivity metric also segues into the next three research questions that examine in depth the differences in collaboration sizes in CS.

3.4. RQ4: Do Women Collaborate with Fewer People Than Men?

This question could be addressed by two distinct measures: the mean number of coauthors per paper and the size of the total coauthor network for each author.

Women in our dataset average 4.03 coauthors per paper, while men average 4.11 ($t=-1.22$, $p=0.22$). This metric appears to show no significant differences in the aggregate. Breaking it down by field (Figure 5) shows that the gender differences still remain minimal throughout almost all of CS. The largest differences appear in the subfields of computer architecture ($t=-0.54$, $p=0.59$) and operating systems ($t=-1.36$, $p=0.18$). In general, the field of computer systems stands out with an average gender gap of 0.09 fewer coauthors for women than for men ($t=-0.41$, $p=0.68$).

The second measure is how many distinct authors each person collaborates with across all of their papers—in other words, the size of the network of all collaborators of a person. This time, the difference is more pronounced, with women averaging 4.99 total cohort size, while men average 5.49 ($t=-4.82$, $p<{10}^{-5}$). Unfortunately, this metric cannot be neatly broken down by fields, because coauthor networks often include authors that span more than one field.

Overall, men exhibit slightly higher collaboration metrics than women (more coauthors per paper and more coauthors overall), but not dramatically so.

3.5. RQ5: Do Women Publish Fewer Single-Author Papers?

Next, we turn our attention to single-author papers (

Table 3. Number of papers by subfield and gender, sorted by the overall percentage of single-author papers.

Subfield	Number of Papers	Percent Single	All Single Author	Women	Men	Unknown
Computer Science Education	151	12.58%	19	3	14	2
Theoretical Computer Science	193	12.44%	24	0	24	0
Programming Languages	64	9.38%	6	0	6	0
Algorithms	233	9.01%	21	0	20	1
World-Wide Web	170	8.82%	15	4	10	1
Artificial Intelligence	599	4.34%	26	1	24	1
Multimedia	225	4%	9	2	7	0
Machine Learning	434	3.92%	17	1	16	0
Compilers	137	3.65%	5	0	5	0
Concurrency	406	3.45%	14	1	13	0
Neural Networks	679	3.24%	22	0	22	0
Benchmarking	159	3.14%	5	0	5	0
Graphics	166	2.41%	4	0	4	0
Data Science and Mining	216	2.31%	5	1	4	0
Operating Systems	180	2.22%	4	0	4	0
Human–Computer Interaction	599	2.17%	13	4	9	0
Communications	158	1.9%	3	0	3	0
High-Performance Computing	244	1.64%	4	0	4	0
Storage	64	1.56%	1	1	0	0
Computer Architecture	276	1.45%	4	0	4	0
Information Retrieval	78	1.28%	1	0	1	0
Computational Linguistics	195	1.03%	2	0	2	0
Computer Vision	532	0.75%	4	0	4	0
Security	279	0.72%	2	0	2	0
Databases	157	0.64%	1	0	1	0
Software Engineering	190	0.53%	1	0	1	0
Cloud Computing	165	0%	0	0	0	0
Total	6949	3.34%	232	18	209	5

). In our dataset, there is a total of 232 such papers, of which 18 were written by a woman, 209 by a man, and the rest unknown. The ratio of women among single authors with known gender is 7.9%, which is significantly lower than that of the overall 14.1% FAR ($\chi {}^2=6.78$, $p<0.01$).

Another way to look at the same data is from the perspective of papers instead of authors. As the data in Table 3 shows, in most CS subfields fewer than 5% of papers were written by a single author, adding up to only 232 papers. (Contrast this, for example, with astronomy, physics, and biology, where the rates of single-author papers average over 10% [47]).

Breaking down the data by gender offers little information because the numbers of single-author papers per subfield are too small for statistical significance. In fact, the numbers of single-author papers are so low, as is the number of women authors overall, that their intersection is actually empty for most subfields. That said, women published relatively fewer single-author papers than men (by percentage) in all but three of the subfields: WWW, data science, and storage. In all three, the numbers are simply too small to draw any conclusive inferences.

It is also worth noting that in much of the systems field the percentage of single-author papers is extremely low. This observation may be another indication that the systems field in particular depends on larger collaboration teams for published research. We suspect that the field’s emphasis on complex implementations and experimental platforms requires larger teams to pull off, which we address in the next research question.

3.6. RQ6: Are Team Sizes Larger in More Experimental Subfields?

The data we collected on coauthorship size and single-authorship shows that authorship norms vary significantly by field, if not by gender. The largest coauthorship groups appear in computer systems papers, averaging 6.09 coauthors per paper, followed by HCI (5.06), knowledge systems (4.89), software engineering and programming languages (4.65), AI (4.42), CS education (4.12), and finally theory (3.59).

These data do appear to confirm the hypothesis that experimental fields generally require larger teams to design, engineer, implement, and measure research results. For example, research in computer architecture, the most collaborative of our subfields, often requires large investments in effort (and often, in capital as well). This characterization extends to most computer systems subfields that occupy the top spots in terms of collaboration sizes. It appears indeed that the larger effort and resource requirement is associated with larger collaborations, as expressed in mean number of coauthors.

On the opposite end, research in computer theory requires virtually no equipment and is often carried out by individuals, as we have previously observed. The characteristics of theory research are naturally very similar to those of mathematicians as a whole, so it is perhaps not surprising that the mean number of coauthors we found for theory is nearly identical to the one found for mathematics [25].

3.7. RQ7: Do Authors Exhibit Gender Homphily?

For our last research question, we follow the approach of Wang et al. to estimate whether authors collaborate with coauthors of the same gender at rates higher than expected [11]. For this computation, we look at every pairing of coauthors as one coauthoring event (omitting single-author papers), and ask whether same-gender pairings occur at a higher frequency than we would observe from a random pairing. A random pairing is expected to follow the same overall statistics for gender distribution, i.e., the expected probability of any (co)author to be a woman should be the same as the overall FAR.

As Wang’s study has also found, our data suggest that for CS, authors—especially women—are actually more likely to collaborate with coauthors of the same gender. Overall, the probability of a woman’s coauthor to be a woman in our dataset is 21.2%, nearly ten percentage points above the overall FAR. For men, the probability to collaborate with a woman is 12.3%, slightly below the overall FAR.

We can also break down these probabilities by subfield (Figure 6). As a generalization, most subfields exhibit gender homophily, especially in the two subfields with the highest FAR (CS education and HCI). A few subfields exhibit gender heterophily, but typically very little. A curious exception is the subfield of programming languages, where women only have a probability of 3.8% to collaborate with a woman, less than half the overall FAR for the field. Together with the subfield of software engineering, this field appears to show consistent heterophily among its female authors.

All other fields show fairly consistent gender homophily to varying degrees. For example, in the large field of computer systems, women have a probability of 12.3% to collaborate with a woman, above the $\approx 10\%$ FAR. In all cases, the more pronounced deviations from the expected probability are for women, suggesting that perhaps collaborating with same-gender authors is more important to women. Note, however, that since most authors are men and therefore overall FAR is mostly determined by men, we would expect men’s deviations from FAR to be smaller than women’s.

4. Discussion

In this section, we dive deeper into the data by exploring the relationships between the different measurements and metrics across subfields, summarized in

Table 4. Comparison of subfields by different gender metrics, ordered by FAR. Metrics include mean productivity (papers per author), fractional paper total, mean total coauthors (with repeats), ratio of papers written by a single author, and the ratio between a woman’s probability to coauthor with a woman and FAR.

Subfield	FAR	Mean Productivity	Fractional Productivity	Mean Coauthors	Single-Author Ratio	Women’s Homophily
CS Education	0.422	1.228	0.393	3.125	0.114	0.511
HCI	0.336	1.721	0.405	4.131	0.022	0.418
World-Wide Web	0.218	1.696	0.419	4.292	0.083	0.217
Multimedia	0.188	2.436	0.563	3.966	0.040	0.198
CL	0.182	2.090	0.552	3.244	0.010	0.194
SE	0.148	1.650	0.402	3.751	0.005	0.130
AI	0.139	2.420	0.657	3.068	0.042	0.155
DS and Mining	0.136	2.130	0.523	3.618	0.023	0.199
Benchmarking	0.127	2.194	0.498	3.960	0.031	0.156
Cloud Computing	0.124	1.817	0.430	3.932	0.000	0.180
Computer Vision	0.122	2.366	0.582	3.654	0.008	0.141
Security	0.117	1.976	0.424	4.639	0.007	0.159
Databases	0.116	1.684	0.411	5.019	0.006	0.143
IR	0.106	2.772	0.671	3.632	0.013	0.121
Neural Networks	0.104	2.428	0.649	3.513	0.032	0.128
Architecture	0.098	2.101	0.459	8.585	0.014	0.083
Algorithms	0.098	1.847	0.596	2.767	0.086	0.177
HPC	0.096	2.033	0.439	5.204	0.017	0.132
Communications	0.095	1.901	0.398	5.870	0.019	0.112
Compilers	0.095	1.755	0.485	3.437	0.036	0.124
Graphics	0.095	1.861	0.441	3.928	0.024	0.096
PL	0.095	1.593	0.466	2.975	0.094	0.038
Concurrency	0.092	2.058	0.518	3.634	0.034	0.119
Storage	0.091	1.846	0.408	4.696	0.016	0.055
OS	0.085	2.068	0.444	4.972	0.022	0.075
Machine Learning	0.081	2.599	0.705	3.471	0.039	0.098
Theoretical CS	0.064	2.303	0.812	2.375	0.124	0.068
Overall mean	0.141	2.078	0.519	5.109	0.036	0.156
FAR correlation	1.000	−0.463	−0.336	−0.139	0.263	0.956

. Our goal is to see whether the variations in any of the separate metrics we collected can help explain higher—or lower—values of FAR in the field, possibly providing hints to causal relationships.

The first relationship we investigate is between productivity and FAR. In other words, can the higher observed productivity of men explain why we observe so many more male authors than female authors?

The answer appears to be “mostly not.“ Obviously, the observation that men publish more than women implies that we would find more names of men on papers than we would of women, leading to lower FAR. When aggregating the data on a subfield basis we do indeed find a moderate negative correlation between a subfield’s FAR and the mean productivity of its practitioners ($r=-0.46$, $p=0.02$). This correlation weakens when using fractional paper counts, ($r=-0.34$, $p=0.09$), since the slightly larger team sizes for men attenuate their productivity advantage. Some subfields show no apparent relationship between FAR and productivity, like storage and machine learning, which have very similar FARs but very dissimilar productivity gaps. Other subfields exhibit a stronger opposite relationship, like HCI and compilers, with similar productivity metrics but very dissimilar FARs.

Overall, the relatively modest advantage in productivity for men (21.81% more papers per author) does not translate directly to the nearly 9:1 ratio of male-to-female authorship. In fact, if we ignore repeated publications altogether and look simply at the ratio of unique women among all unique authors, we still observe a ratio of 15.3% women overall. While this ratio represents a slight improvement of one percentage point over the non-unique FAR, it is still far from parity, suggesting that higher productivity alone cannot fully explain the gender gap.

We can extend this analysis of correlation with FAR to three other research questions we asked: collaboration size, single-author counts, and gender homophily.

A hypothetical relationship between FAR and the typical collaboration size in a subfield can also be easily refuted with counterexamples. Consider the subfields of algorithms and architecture. Although their FAR values are nearly identical, they are on extreme ends of the average team sizes. The overall correlation between the two metrics is indeed negative, but too close to zero for significance. ($r=-0.14$, $p=0.49$).

From the related perspective of single-author papers, a similar hypothesis would be that subfields with more single-author papers would have lower FAR because single-author papers have an even lower FAR than the overall sample population. However, the small number of single-author papers and their minuscule weight in computing FAR present a statistical obstacle to testing this hypothesis. As a matter of fact, our data shows a positive but nonsignificant correlation between the two ($r=0.26$, $p=0.19$), so we must reject this hypothesis as well.

Our last metric for comparison is gender homophily. The method we previously used to measure homophily, deviation from the expectation (FAR), produces two measures per field, one for men and one for women. We focus on the latter because the deviation from FAR for men is nearly negligible, owing to the high ratio of men in the data. We therefore look at “women’s homophily”, defined as the ratio between a subfield’s probability for a woman to coauthor with a woman and its FAR. This variable too appears uncorrelated with FAR ($r=0.07$, $p=0.73$). This finding is explained by the already low statistical probability of a woman (or anyone) to collaborate with a woman.

This high skew towards men also means we cannot use standardized metrics for homophily such as Krackhardt’s Index [48], defined simply as $EI=\frac{External-Internal}{External+Internal}$ (where $Internal$ represents all the same-gender pairings in our dataset and $External$ all other pairings). In our data, this metric is strongly correlated with a subfield’s FAR ($r=0.95$, $p<{10}^{-9}$), to the point of adding no valuable information. This finding also makes sense: if there are very few women in a field, most coauthor pairings will be internal male–male, all other things being equal. The upshot here is that homophily measures appear to be much more the result of a skewed FAR than its cause. In other words, gender homophily appears to have little role in explaining variations in FAR, at least when it deviates significantly from 50% as it does in our dataset.

In summary, none of the collaboration metrics we collected in the various research questions can produce a satisfactory explanation for FAR in a given subfield. For productivity (RQ2, RQ3), the correlation we found can only explain a small fraction of the large gender gap. For team sizes (RQ4, RQ5, RQ6) we found no correlation with FAR. For gender homophily (RQ7), the strong correlation suggests, if anything, an opposite-direction causal relationship, i.e., low FAR causing high homophily, and not the other way around.

Despite these negative results, we can still draw some interesting general observations on specific CS fields. One such example is theoretical computer science, which shows extreme values in many of the metrics we collected. Not only does it exhibit the lowest FAR, but also the lowest average team size, the highest single-author ratio, and the highest fractional productivity. In these metrics it matches the observations of Mihaljevic for mathematicians, which is not surprising, given the thematic similarity of the two fields [25]. Theoretical CS is an extreme point by almost all metrics and resembles closely what has been found for mathematics. It is a broad and foundational subfield of CS but also exhibits different characteristics from the rest of CS in some of its aspects, including the collaboration aspects discussed in this paper.

Another example is the field of systems, which also generally exhibits very low FAR. Systems is a large and influential field, with many industrial and technological applications [41]. It is therefore particularly of interest to try to explain and reduce the gender gap, as this could have far-reaching societal impact [2]. As discussed in RQ6, one possible explanation for the magnitude of the gender gap is the high cost in participation in experimental fields such as systems, which has been associated with higher gender disparity [34]. It is also possible that the combination of larger team sizes, fewer single-author papers, and stronger gender homophily makes systems a particularly unwelcoming field because of self-reinforcing network effects [49]. Teams are large but mostly male, which discourages women seeking female collaborators from participating, which in turn could drive more women away from the field, further exacerbating the challenge of finding female collaborators for the next woman to join the field.

It has also been suggested that women are more likely to work in human-centered fields [50,51,52]. We can certainly observe corroboration of this hypothesis in our data, with CS education and HCI as the two fields with the highest FAR, followed by WWW and multimedia, which also put the human in the center of the research. Our FAR figure for CS education is in fact remarkably similar to the one found by West et al. for the field of education as a whole [33].

5. Conclusions and Future Work

Computer science is a collaborative discipline. Many papers exhibit larger team sizes than is typical in most other scientific disciplines, and very few papers are written by single authors. Understanding collaboration patterns in CS authorship is therefore imperative to understanding related bibliometric and societal phenomena, such as the gender gap in CS. However, research in CS is also highly heterogeneous across fields and subfields. Although a number of papers have studied collaboration patterns in CS and the sciences, and even more have studied the gender gap, this study may be the first to examine the two at a finer resolution of CS subfields, and with highly accurate gender data.

Because subfields vary so much by their representation of female researchers—varying from 6% to 42%—we can examine how these variations relate to different collaboration metrics, especially in the extremes. Our findings do confirm that some collaboration patterns appear indeed to be associated with the gender gap in each subfield. Fields that exhibit lower FAR than average also tend to exhibit larger team sizes, smaller cohort sizes for women, higher gender homophily, and higher author productivity. Additionally, although we found a significant productivity gap across genders, as has been measured in many other fields, it is too small to explain most of the overall gender gap across fields.

All of these associations have exceptions, and perhaps the most notable is the subfield of theoretical CS, which stands alone in the extreme of most metrics, perhaps because of the foundational, mathematical nature of theoretical CS, in contrast to many applied subfields of CS. On the opposite end, CS subfields that have increased focus on people rather than computers, such as CS education and HCI, show better overall representation of women and less extreme gaps across most metrics. On the other hand, many of the subfields of computer systems, which focuses on building, measuring, and improving the concrete implementation of computers and their tools, show larger gender gaps across most metrics.

This last field of computer systems is of particular interest to us and the focus of our future work, because of its large size, its impact on technology and the economy, and its very low representation of women, at about 10%. We will therefore turn our attention next not just to statistical association with low FAR, but to causal association. We plan to collect and analyze additional data to try to address the question: why is the representation of women in computer systems lower than in most other CS fields? We have already presented a few hypotheses for this question in this article: the low representation could be partially explained by the higher cost of research in the field and the network effects that perpetuate the lack of female peers and mentors. These hypotheses require additional data before we can accept or reject them.

This dataset currently exposes facts and factors from a single year, 2017. It could also be instructive to follow up on this study with data collected from later years to observe any trends and changes in collaboration patterns and representation of women. Additional future work can leverage the dataset we collected to dive deeper into the analysis of collaboration patterns, using tools such as social network analysis, degree centrality, citation networks, and information diffusion.