A lot of existing projects aiming to create an electoral system for ICOs revolve around the idea of power hierarchies based on the possession of community tokens.
While being an intuitive and easily manageable indicator of consensus inside the holder group, they usually constitute the only basis for the business model of the company issuing them, thus becoming a strict requirement for participation.
As such, they both represent an unnecessary dependence and an asset carrying the emotional burden of financial speculators, allowing for the creation of separative cults, as consistently demonstrated by the recent trends.
This creates a significant adoption barrier for early-stage projects and results in a relatively exclusive and biased electoral base, eventually discouraging ICO projects with a diverse target audience from adopting the escrow platform.
A large-scale analysis of the wealth of information provided by the existing social networks would allow for the creation of an accurate and independently verifiable predictive model, employed for classifying and ranking user profiles on the basis of well-defined parameters. A mathematical representation of the current crypto community would, therefore, make it possible to separate token holders from users, and ultimately allow the escrow platform to have a more extensive reach with no discrimination of sorts.
Two different classes of approach to the ranking of social profiles can be defined in the current state of research: profile-specific and network-wide.
Models belonging to the first category, while being computationally inexpensive and scalable, are much more eludible and inaccurate since they do not account directly for the community structure and the relationships between user accounts.
On the other hand, models of the second type require a more advanced computational infrastructure and do not allow for the possibility of being executed on-chain, but provide a highly realistic representation of the community of choice.
Thanks to the research advances in the very latest years, we propose a solution which lies somewhere in between the two, by mapping the whole network of interactions of the most representative social network (i.e. Twitter) to an oriented graph a and generating node attributes from profiles on secondary, yet indicative, services (i.e. GitHub, Steemit).
The creation/update steps of the model should unfold as follows:
- A wide, initial node pool is extracted from the most representative social network of choice by means of snowball sampling, starting from a selected group of influential seed nodes (i.e. @vitalikbuterin, @NickSzabo4). During the process, each node undergoes the examination of a classifier which cuts off the sampling tree on the basis of different characterizing parameters such as a public list of keywords contained on user posts² ³.
- The resulting graph is then processed by state-of-the-art Sybil detection algorithms⁴ ⁵, approaching each node from both global and specific perspectives. Bot accounts are thus eliminated, providing an accurate representation of the state of the community, with an estimated size of around 1 million nodes.
- Various network-wide studies are then applied to the graph obtaining relative scores of influence and connectedness for each node thanks to well-known centrality metrics (i.e. PageRank).
- Providing the oracle invoking functions with additional social profile information instructs the database for a scheduled insertion of new node attributes in a non-penalizing fashion, and without altering the base structure of the network.
The values of those attributes are processed by simpler scoring functions, computing, for example, the number of contributions in proportion to the number of starred GitHub repositories. - The resulting model is then both hosted on a public Neo4j database and shared on IPFS with an estimated size around 1 GB, available for both online querying and offline data validation. Accordingly to financial availability, the model will be periodically rebuilt on a monthly basis as a minimum, updating both structural and parametric properties.
On the contract side, the resulting scores, once queried, are respectfully weighted and linearly combined to obtain the KYC score as indicated on the whitepaper⁷.
On the client side, the same two different domains of computation are required to be processed in reverse, to ensure the authenticity of the model:
- The client makes sure of cooperatively validating node attributes with the other clients, by querying the public APIs of their sources and broadcasting any found incongruency.
- As for structural validation, it historically required a prohibitive amount of computational power to process large graphs on single machines, but with the very latest advances in both research and engineering, we’ve reached satisfying performance and usability even on consumer-grade personal computers¹. Furthermore, it’s been demonstrated the possibility to reduce the size of the networks in question to about 20% of the original (about 150k nodes in our case) if needed, without significantly compromising expensive graph metrics like betweenness centrality⁶.
An incensurable incongruency reporting service will be made available on a decentralized protocol of choice, which will be employed for broadcasting any possible, although unincentivized, misbehavior by the centralized model generator. In this case, differently from blockchains, the maximum possible number of adversaries is one, making it trivial for the network to distrust it. Specifically, with a network size as illustrated above, an estimated minimum of 2000 coordinated users should be able of automatically validating the whole model in a matter of minutes.
Finally, should the platform gain enough traction, the development of a model based on a multidimensional graph will be considered in the medium term, replacing node attributes from secondary social networks with more accurate attributes based on the structural properties of their relative communities.
[1] Christian Staudt, Aleksejs Sazonovs, Henning Meyerhenke, NetworKit: A Tool Suite for Large-scale Complex Network Analysis, 2014.
[2] M. E. J. Newman, Aaron Clauset, Structure and Inference in Annotated Networks, 2015
[3] Jaewon Yang, Julian McAuley, Jure Leskovec, Community Detection in Networks with Node Attributes, 2014
[4] Sneha Kudugunta, Emilio Ferrara, Deep Neural Networks for Bot Detection, 2018
[5] G. Wang, T. Konolige, C. Wilson, X. Wang, H. Zheng, Ben Y. Zhao, You are How You Click: Clickstream Analysis for Sybil Detection, 2013
[6] Gerd Lindner, Christian L. Staudt, Michael Hamann, Henning Meyerhenke, Dorothea Wagner, Structure-Preserving Sparsification of Social Networks, 2015
[7] JUST: Crowdfunding the 21st Century
