Giacomo Livan, University College London
Tomaso Aste, Fabio Caccioli
The digital economy is increasingly self-organizing into platforms where users exchange knowledge, goods, and resources on a peer-to-peer (P2P) basis. Such platforms are epitomized by the digital marketplaces of the Sharing Economy, which have already replaced substantial portions of well-established business-to-consumer sectors, such as the hotel and taxi industries.
P2P platforms foster interactions between users who have never interacted before, and therefore must guarantee that trust between the users themselves is established. This is typically achieved by requiring users to develop a reputation through a digital peer-review process. Given the projected growth of the P2P business paradigm, in the next few years digital reputation will acquire greater importance, as it will grant or prevent access to substantial economic opportunities. It is therefore crucial to ensure the fairness of digital peer-review systems, and to ensure that individual reputation scores reflect the behavior of users in an accurate and unbiased way.
Being decentralized, P2P systems are often supposed to promote more economic freedom and more democratization. Yet, their current lack of regulation exposes them to new forms of fraud, information asymmetry, and malicious behavior which can distort their outcomes. In particular, users are often incentivized to reciprocate their ratings, i.e. to exchange positive ratings in order to mutually boost their reputation, and to retaliate after receiving negative ones.
P2P platforms naturally give rise to networked structures, whose degree of complexity depends on the patterns of interactions between users. In this paper we focus on three platforms whose interactions are binary, i.e. users rate their peers either positively or negatively. As such, they can be conveniently represented in terms of signed networks, i.e. directed networks whose links carry a #1 weight. Clearly, those structures represent an oversimplified counterpart of the most popular platforms where users build a peer-review based reputation, such as Uber and Airbnb. Yet, they retain the full complexity of those richer environments, both in terms of interaction patterns and user heterogeneity. Because of this ‘stylized yet complex’nature, we identify signed networks as suitable environments where to analyze online reputation and its reciprocity-induced biases.
We analyze data from Slashdot, a technology news website whose users can label each other as ‘friend’or ‘foe’based on their public comments, Epinions, a platform for crowdsourced consumer reviews, and Wikipedia, where co-edits or antagonistic edits are interpreted as positive or negative interactions. We introduce a simple proxy for user reputation in the network, defined as the normalized difference between the number of positive and negative ratings received.
In all three datasets we find a clear overexpression of reciprocated relationships (both in the positive and negative case) with respect to a null hypothesis of random link rewiring constrained to preserve each user’s reputation. This finding highlights how the same distribution of user reputations could be obtained through markedly different patterns of network interactions, which in turn shows that neither positive nor negative reciprocity represent a strategic tool necessary to achieve the observed reputations.
We further investigate reciprocity in P2P environments by means of an extended class of null models which we obtain through random link rewiring while, again, preserving each node’s reputation and simultaneously forcing the network to produce a predefined target value of positive or negative reciprocity. We find that all three systems reach a saturation, i.e. a reciprocity threshold above which the networks run out of links that can be used to produce reciprocated relationships (compatibly with the constraints). We find that the more polarizing the interactions, the more the corresponding platforms tend to exist close to their saturation threshold. In fact, we find that Slashdot can only accommodate an additional 15% of positive reciprocity and an additional 60% of negative reciprocity before reaching saturation. On the other hand, the Wikipedia network can sustain more than twice the positive reciprocity and more than five times the negative reciprocity it displays ‘in real life’.
At the macro level, we find that, on average, reciprocated ratings contribute more to the formation of user reputation than non-reciprocated ones. We prove that this is not a required ‘structural’feature of the networks we analyze by showing that small random perturbations of the link organization are enough to completely overturn the relative importance of the contributions of reciprocated and non-reciprocated activity to reputation, making the latter predominant.
On the other hand, at the ‘microscopic’level we detect overexpressed patterns of interactions between certain groups of users. Namely, we find that high reputation users preferentially interact with each other. In more general terms, we find that the relationship between a user’s reputation and that of the users she has interacted with is not compatible with a wide array of null hypotheses.
In summary, in this work we perform an extensive empirical analysis of real-world P2P systems whose interactions can be mapped onto stylized network models. We provide evidence that rating reciprocity in such systems is heavily overexpressed, and actively exploited by users in order to inflate their own reputation or damage those of others. In our conclusions, we shall argue that the null models we employ to support our claims can be used as baseline toy models to experiment with different policy options for P2P systems.