Offline biases in online platforms: a study of diversity and homophily in Airbnb

Sharing economy platforms are new manifestations of century old phenomena. Resource circulation systems that facilitate the exchange of underutilized goods or services between consumers have long existed in the form of flea markets, garage sales, second-hand shops, just to name a few. However, what used to be small scale and local instances of collaborative consumption, have now become massive online marketplaces, where face-to-face interactions have been replaced by technology-mediated ones [1].

A fundamental question arises about the role that such decentralized, largely unregulated, online platforms play in our societies. Often thought of as level-playing fields, where all participants receive the same opportunities, sharing economy platforms might instead end up acting as online aggregators of well-known offline human dynamics and biases. Indeed, a number of studies have suggested that some of the big sharing economy players are acting as accelerators of gentrification pheonomena that are already underway in large cities. For example, Airbnb has led to the emergence of short-term rent gaps between different areas of New York City [2] and has contributed to exacerbating the affordable housing crisis in Los Angeles [3]. These phenomena, in turn, typically accelerate preexisting divides along racial lines, fostering inequalities between the Airbnb community and the communities living in the neighbourhoods where Airbnb has a significant presence (see, e.g., [4]).

Moreover, sharing economy platforms have come under multiple allegations over discrimination episodes taking place within the platforms themselves. For example, Uber drivers were found to be twice more likely to cancel trips requested by passengers with African–American sounding names compared to White-sounding names (even though Uber penalizes drivers for cancellations) [5]. Similarly, Airbnb’s hosts were found to be turning down potential guests based upon their racial background [6].

While headway has been made to tackle the unintended consequences brought about by sharing economy platforms, the debate on their socio-economic impact is still in its infancy, and only relies on a handful of studies or on anecdotal evidence. This, in turn, delays the execution of targeted interventions to expose, and possibly reduce, such consequences. The goal of this paper is to contribute to inform such a debate by performing a large scale empirical analysis aimed at detecting systematic statistical evidence of ‘offline’ biases taking place in online sharing economy platforms.

We study the Airbnb hospitality service, and focus first on the composition of its user base, with the aim of assessing its diversity both in general terms and then contextually with respect to the city hosting it. Second, we employ a network methodology to assess the statistical significance of host–guest interactions in Airbnb. In particular, we focus on homophily, i.e. the social phenomenon where people gravitate towards those like themselves [7], and on its opposite, heterophily. We also study the tendency to avoid members of a social group with different social traits, which we refer to as avoidance. While avoidance is universally deemed as unacceptable, homophily has sometimes been perceived as ‘natural’, and thus judged in a more accepting way. However, several studies have shown that the aggregation of slightly biased individual preferences can lead to unintended and collectively undesirable consequences, as evidenced by Schelling’s work on urban racial segregation [8, 9], and Neal’s work on school children’s development [10].

These results echo other findings in the literature (see next section), and provide concrete evidence about how sharing economy platforms are being appropriated in different city contexts, possibly resulting in large divides between the online communities who can enjoy the benefits of the sharing economy and the ‘offline’ urban communities who are most exposed to its expansion. They also offer an opportunity to inform the design of tailored technology interventions aimed at exposing, and possibly reducing, certain behaviours, while also providing the means to monitor their effects (Discussion section).

Upon its inception, the Internet was expected to create a global level playing field, where the inequalities of the ‘offline’ world would be overcome thanks to easy access to digital opportunities. Yet, reality has been very different. As it is invariably the case, different social groups are not equally equipped to face technological innovation in its early stages, which typically exacerbates preexisting inequalities [11].

The sharing economy, as a whole, has been no exception. Indeed, a handful of studies have shown that the ability to seize the sharing economy’s opportunities is often severely limited by geographical and socio-economic constraints. For example, Airbnb listings are usually more concentrated in wealthier, more attractive areas populated by young and tech-savvy residents [12]. Similarly, TaskRabbit users from areas with low socioeconomic status and/or low population density were found to have a harder time both when selling their services and when seeking to outsource work to potential taskers [13], while individuals living in deprived Chicago suburbs have been found to have a harder time to get an Uber ride [14].

Similarly, the Internet’s promise to circumvent physical barriers and improve communication between social groups has not always been upheld. For example, episodes of racial discrimination in online social networks have been extensively documented [15, 16]. Also, a vast amount of scholarly work has been devoted to understanding the formation of online preferential relationships between individuals. This has often been explained either in terms of interest-based homophily, e.g., showing the impact of ideological homophily in determining the opinions and content individuals are exposed to on social media [17], or in terms of homophily driven by demographics.

Studies of early social networks, e.g., MySpace [18], have identified race, gender, and age as the main demographic features driving online homophily, and such elements kept recurring in more recent studies. Indeed, evidence of racial [19] and gender [20] homophily has been reported in Facebook and Twitter, respectively, and evidence of both has been documented in the social networks underpinning location sharing applications [21]. Age homophily is somewhat less studied, but still documented in a study of the Facebook social graph [22] and in niche environments such as virtual worlds [23, 24].

Our work follows this stream of literature and investigates whether well known ‘offline’ biases also take place in sharing economy platforms. A handful of recent studies have started to look at such platforms from this perspective. Indeed, recently published work [5] found evidence of both gender and racial discrimination in Uber and Lyft, as female passengers were disproportionally taken on longer and more expensive routes, while passengers with African American-sounding names were twice as likely to receive trip cancellations from Uber drivers compared to passengers with White-sounding ones (even though Uber penalizes drivers for cancellations). Similarly, another study [25] found gender and race to have an impact on worker evaluations in online freelance marketplaces.

Evidence of biased behaviour was also found in Airbnb by means of a field experiment [6]. In particular, guests were found to be 16% less likely to have their booking accepted if they had a distinctly African American-sounding name when compared to identical guests with White-sounding names instead. Similarly, in [26] it was found that non-Black hosts charged on average 12% more for an equivalent rental compared to Black hosts, and similar results were replicated in a subsequent study on Airbnb [27], where Asian and Hispanic hosts were found to rent at prices 9.3% and 9.6% lower, respectively, than their White counterparts.

While the above works investigate some specificities of user demographics and interactions in sharing economy platforms, a systematic analysis of these dimensions across the fundamental features of gender, age, and race is still lacking. This work aims at filling this void, by providing (i) an overview of the composition and diversity of Airbnb’s community, and (ii) a quantitative method to dissect the anatomy of user-user interactions in sharing economy platforms (and Airbnb in particular), providing statistical evidence of homophily and avoidance between certain user groups.

In order to perform this study, we needed two types of data: demographic characteristics of hosts and guests (i.e., gender, age, race); and their pairing dynamics (i.e., who stayed with whom). Since we hypothesise that peers’ behaviours might vary in different geographic (and cultural) contexts, we chose to perform this study on a per city level, rather than treating the whole of Airbnb as a single analytical context.

To begin with, we accessed city snapshots that the website InsideAirbnb¹ already makes available. We chose five cities (Amsterdam, Chicago, Dublin, Hong Kong, Nashville) so to have high geographic diversity (these cities span three different continents), as well as high diversity in terms of population composition and cost of living. Records of Airbnb hosts, guests and stays go from 2008 to 2016 for all cities except Nashville, whose Airbnb records start in 2009. For each city, InsideAirbnb makes available a full list of host IDs (from their ‘listings’ file). We used these IDs to query the Airbnb website and further acquire a host profile picture, the type(s) of property they were renting out (i.e., full property, private room in a shared property, or shared room), and the full list of IDs of all guests that ever left a review to such host (and for what property). We then further queried the Airbnb website with the guest IDs to acquire their profile pictures. Since Airbnb does not explicitly make available a peer’s gender, age and race as attribute-value pairs in the peers’ profile, we used image processing software on the collected profile pictures to automatically extract this information. In particular, we first used face localisation software to detect whether the profile picture contained a human face, and if so, to identify the portion in the picture containing it. We tested both FaceReact² and Indico³ on a manually curated sample of 50 Airbnb images, so to contain a mix of pictures with and without human faces, and with and without background clutter. We found Indico to be significantly more accurate, especially for human images taken at an angle rather than straight-facing the camera. We thus continued only with the latter. Having extracted the bounding box containing a human face, we then used face recognition software to extract attributes. We tested Betaface,⁴ Sightcorp F.A.C.E,⁵ and Face++⁶ on a subset of 250 Airbnb images. We found all three to be equally accurate when detecting gender. Sightcorp was found to be significantly more reliable in recognising age groups, and Betaface in extracting race (note that our analyses will focus exclusively on race, not on ethnicity; in particular, we will focus on three main race categories, i.e., White, Black, and Asian). We thus worked with Sightcorp and Betaface in parallel. We manually verified their accuracy on all 250 test images, and found the confidence levels reported by both products to be \(0.3 \in[0,1]\) or higher on images annotated correctly. Hence, we kept such value as a threshold for the ensuing automatic annotation; furthermore, we only retained pictures for which both face recognition software products agreed on both gender and ethnicity. To understand how robust our results are when varying facial annotation accuracy, we repeated all our analyses after (i) increasing the above threshold to 0.5, and (ii) manipulating the data by changing the race annotation on a random sub-sample of the images. The results obtained from such analyses are reported in the Additional file 1.

In terms of pairing dynamics, Airbnb does not make visible who stays with whom, nor whether a stay request has been refused or cancelled. However, what it does make visible are reviews that hosts and guests leave to one another. We use these as proxies for the actual pairing dynamics. Studies show that over 65% of stays result in a guest review and 72% result in a host review [28], so most stays are indeed captured by reviews. At present, it is not known whether those who do not leave reviews in Airbnb belong to specific users’ groups; a past survey study of Tripadvisor reviews [29] did find that certain age and gender groups were more vocal than others, and this might also be the case in this context. Although the method we present next is still applicable, the validity of some of our findings might be impacted, and we will come back to this when we discuss limitations and future work (Conclusion section).

We model Airbnb hosts and guests as nodes in a bi-partite graph, with a directed \(g \rightarrow h\) edge with weight \(w_{gh}\) representing the number of times guest g stayed at host h. Since we hypothesise that pairing dynamics may vary across cities, as well as across type of rented property (full property vs. shared—the latter comprising both private and shared rooms), we create and analyse a total of 10 (5-cities ×2-property types) bipartite networks.

Each such network is analysed using a statistical rewiring approach designed to assess the significance of pairing patterns in each of the cities studied. More precisely, the method starts from a null hypothesis that a given guest–host pairing occurred randomly. It then proceeds to verify whether this hypothesis holds by creating ensembles of null network models through the rewiring of the original networks’ edges, and by comparing the properties of such null network model against those observed in the actual, empirical networks. Crucially, the procedure is designed to preserve the heterogeneity of the original networks, as it produces null network configurations where the number of stays that each guest and host have had are both kept intact, therefore preserving correlations between demographic features and activity on Airbnb. In the following, we describe the details of this methodology, and the rationale for adopting it.

Starting from an empirical bipartite network, we create a randomised version of it by iteratively performing xSwap operations [30]. These amount to selecting two guest nodes (\(g_{1}\) and \(g_{2}\)) and one of their corresponding host nodes (\(h_{1}\) and \(h_{2}\), respectively) at random, erasing the existing edges (\(g_{1} \rightarrow h_{1}\) and \(g_{2} \rightarrow h_{2}\)) between both pairs, then reassigning them to each other (\(g_{1} \rightarrow h _{2}\) and \(g_{2} \rightarrow h_{1}\)), as shown in Fig. 1 (left). Should either of the selected edges have a weight larger than one, the strength of the link is reduced by one and a unit weight is redistributed to the new host node (see Fig. 1-right). These operations preserve both the outgoing weight of guest nodes and the incoming weight of host nodes. Therefore, repeated xSwap operations yield configurations which are exactly equivalent to the original networks in terms of their heterogeneity, but are instead fully randomized in terms of the relationships between their nodes [31]. In order to determine how many xSwap operations were needed before the rewired network configurations could be considered distinct enough (i.e., sufficiently uncorrelated) from their original counterparts, we computed the Kendall’s Correlation Coefficient as suggested in [32].

For each of the 10 empirical networks under consideration, we generated an ensemble of 1000 randomized network configurations with the above rewiring procedure, thus preserving the original networks’ degree sequences. Since such a rewiring procedure generates null configurations that do not deviate substantially from the original networks themselves, we can resort to a fairly parsimonious numerical investigation based on such relatively small number of null configurations. On such configuration ensemble, we computed the frequency of interaction between different groups of hosts and guests based on gender, age, and race (e.g., female guests to male hosts, White guests to White hosts, etc.). We then took the 95% confidence interval to define lower and upper bound probabilities for the occurrence of each pairing combination. This range represents the expected probabilities of pairing combinations taking place by chance. We then computed the pairing probabilities among the same groups in the original empirical network: should the actual values fall below or above the ‘by-chance’ range, we take them to be statistically significant under/over-expressions of certain group interactions with respect to the null hypothesis. Note that the over-expression of interactions between two groups (or within a group) does not necessarily translate into the under-expression of interactions between other groups; we will use the former to detect a tendency of certain groups to associate, and the latter to separately detect a tendency to avoid certain groups instead.

In this section, we report the results obtained when applying the above method to analyse Airbnb users’ pairing dynamics. Before doing so, we provide an overview of the demographic characteristics of Airbnb hosts and guests, whose profile pictures we scraped and annotated using the process described in the Airbnb Data section. We will discuss and elaborate on these results in the Discussion section.

In this section, we discuss the results presented above, trying to offer an interpretation based on their societal context, and proposing recommendations concerning the design of sharing economy platforms. We break down the discussion into two parts: we first consider the Airbnb’s user base in each city under study, and reflect upon it relative to the city’s demographic, economic, and historical context; we then move our discussion to our findings on pairing dynamics.

In this paper, we have gathered and analyzed data to assess the diversity of the Airbnb community, especially with respect to the cities where it is embedded, and we have presented a method based on the statistical analysis of networks to detect homophily, heterophily and avoidance between different groups in the Airbnb community. To the best of our knowledge, network rewiring techniques, and, more generally, null network ensembles have never been employed as a tool to detect bias, and this application represents an element of novelty of our work.

Our findings suggest that, in all cities under study, certain user groups (e.g., young, White, female) are substantially over-represented compared to the local population; furthermore, statistically significant signals of gender and racial homophily were detected, across all cities and regardless of the property rental type.

Taken together, our findings echo sentiment that perhaps, contrary to all branding, the sharing economy community might not be that diverse. Rather, platforms such as Airbnb might be acting as accelerators for gentrification processes that are already well underway in major cities. While policy and legislation interventions are needed to regulate who benefits from sharing economy platforms [12], technological considerations also deserve attention: for example, Airbnb, like most sharing economy platforms, requires hosts to have a traditional bank account at the time of registration, to which money will then be deposited when guests visit. This might hinder the ability for many hosts from socio-economic deprived backgrounds to join the platform in the first place (which might be reflected in the very low representation from certain racial backgrounds in our studies); alternative solutions, such as on-demand payment services provided by platforms like BitPesa,¹² could lower the barrier to entry.