Introduction
The Societal Aspects dimension includes four indicators addressing Diversity in Research. They measure different aspects related to diversity among participants (i.e., keynote speakers, conference organisers and authors of papers) at a set of international AI conferences. All the indicators are presented in Freire et al. (2020) and are derived from Shannon and Pielou indexes. Maximum heterogeneity or diversity corresponds to an index value of 1, which means equal distribution of the r population categories – assessed (gender, geography, business types). For each year, diversity indicators are calculated for each conference. We present here an average of the indicators in a set of selected AI conferences: AAAI, NeurIPS, IJCAI, ICML, RecSys and ECAI.
The S1 – Gender Diversity Index (GDI) measures the average representation of researchers from different genders (male, female, other) at AI conferences, thus possibly revealing the impact of gender equality policies on AI research, and is useful for raising awareness about the need for more diverse research communities (Freire et al., 2020), such as the Affective Computing community addressed in Hupont at al., 2021.
The S2 – Geographic Diversity Index (GeoDI) tracks the average representation at AI conferences of participants from different geographical locations, representing the location of the institution to which they are affiliated.
The S3 – Business Diversity Index (BDI) assesses the participation of researchers from academia, research centres and industry in the AI research field.
The S4 – Conference Diversity Index (CDI), is an average of the previous three indices (Freire et al., 2020) and provides a single and overall indication on the different observed trends regarding diversity in the AI research field.
The European Union
We can see that all indicators have a tendency to increase over time. This shows how the AI research community has recently tried to incorporate a diversity of profiles. This might drive towards a potential reduction of bias in research which would be derived from a very homogeneous composition of researchers.
The indicator with the smallest increase is the one on gender (S1 – GDI), which shifts from 0.65 in 2016 to 0.69 in 2020. After a sharp decrease in 2017, the series follows an upward trend, as gender heterogeneity increases almost 4 points per year.
The Geographic diversity index (S2 – GeoDI) is the one showing the most considerable improvement (plus 0.28 points from 2016 to 2020, or an increase of 70%). However, at the same time it is also the one presenting the most alternating dynamic. Indeed, after an increase of 0.26 points from 2016 to 2017, for two years in a row (2018 and 2019) the observed diversity falls, and then it significantly increases again in 2020 (by 0.13 points). These fluctuations might reflect changes in the geographical location of conferences and researchers’ affiliation. Likewise, but only with regard to 2020, this could be due to increases in attendance in online mode due to the COVID-19 pandemic, which would have facilitated participation in conferences from a greater variety of countries.
The Business diversity index (S3 – BDI) shows the largest value of all four diversity indices in 2020, with a peak in 2019 (0.85). This might be explained by the greater presence of industry researchers at scientific conferences, which confirms that AI research raises considerable interest from the private sector too. The fact that the trend is overall positive but slightly decreases in 2020 could also reflect a shift in the dynamics between academy and industry in the AI field (from academia to industry).
Given that the conference diversity index (S4 – CDI) is the result of the simultaneous consideration of all the examined aspects, it is relevant to observe that diversity in conferences seems to have overall increased consistently over time since 2016. This enables one to have an optimistic view of possible advancements in diversity and inclusion for attendees of major AI conferences in the future. Nevertheless, the studied period is still too short to draw firm conclusions, it being too soon to assess the long-term impact of current diversity initiatives being carried out in the AI field (such as mentoring programs, visibility efforts, travel grants, committee diversity chairs and special workshops).
