Web3, one of the arising paradigms which may rule the future Web, is also representing a source of big data stored in the underlying blockchains. Many diferent research fields are benefiting from these large collections of temporal and heterogeneous data, which capture diferent aspects of the interactions among people and between people and Web3 platforms. Specifically, since each piece of information is validated and timestamped, Web3 platforms are becoming an invaluable source for understanding the dynamics of these techno-social systems at a high temporal resolution. In this contribution, we focused on the analysis of the evolution of the networked structure of Web3 social networks through the lens of discrete choice models, and on the changes in the structure of the relationships after a shocking event has occurred on the platform - namely a hard-fork in the supporting blockchain. To support large-scale analysis, we represent Web3 platform data as temporal multigraphs manageable by modern graph database management systems. The main findings, which represent a summary of our efort in mining data from Web3 platforms, highlight some interesting aspects: i) when applied to Web3 social networks, discrete choice models allow us to decompose the evolution of social networks into diferent growing mechanisms, which are quite stable during the observation period; and ii) in a stratified context, such as Web3 platforms, interactions resulting from economic actions, such as transfers or loans of crypto-tokens, are as important as social relationships to predict how users will behave during a shocking event. These are a few examples of how Web3 social platforms may represent a challenging playground for a more in-depth understanding of the users' behaviors when social and economic interactions are strictly intertwined.
In the last years, the actual structure of Web 2.0 has been questioned by novel paradigms which are trying to reduce the over-centralization around a few big platforms and tech companies. One of the ideas gaining momentum is Web3, i.e. the design of platforms and software systems built on blockchain technologies to promote a decentralized Web. In fact, we are witnessing the birth of decentralized counterparts of Twitter or Reddit, embodied by Hive, Mind, or Steemit; but also services that are specific to the Web3 world, such as Decentralized Finance (DeFi), Decentralized Autonomous Organizations (DAOs), and non-fungible token (NFT), a financial asset, linked to data stored in the blockchain, that can be traded. Although the idea of Web3 is at the heart of a heated debate between enthusiasts and skeptics, the platforms following this paradigm ofer a great opportunity to researchers in diferent fields thanks to the huge volume of high-resolution data stored in the supporting blockchains. Indeed, a broad set of data about these techno-social systems can be easily accessible: by the nature of blockchains, data are publicly available, validated, and afordable by interfacing with the API blockchain. Moreover, data from Web3 platforms ofer two advantages: i) each piece of information is timestamped since each blockchain block has a validation timestamp; and ii) each block reported multi-faceted interactions - social, economic, financial - among people and between people and platform. So, these data sources have all the features to face tasks and issues related to modern techno-social networks and to support detailed and in-depth analysis of users’ traits. Specifically, here we focus on a few issues - which summarize our efort in the recent years to mine Web3 platforms related to the growth of Web3 social networks from the perspective of link creation mechanisms and the efects of shocking events.
Understanding and mining how the networks behind large techno-social systems grow
and react to exceptional events are fundamental elements for the comprehension of the main
processes driving the evolution of such systems and for the identification of specific patterns
of growth which are consequences of the platform design. In this regard, in the past years,
many models, mechanisms, and measures describing the network growth from a link formation
perspective have been proposed, however, most of these approaches rely on the assumption that
the growth is guided by a single parameterized mechanism. But, current techno-social networks
are the result of diferent and heterogeneous behaviors where diferent choices and mechanisms
occur [
connections with users with many common relationships are the leading factors; • it is possible to predict, with reasonably good performances, whether or not a user will migrate to a new Web3 platform born by a fork. Specifically, these outcomes have been reached even with the only information on the network structure, without including textual or external data such as the trend of cryptocurrencies. Moreover, in a stratified context with diferent types of connection, interactions resulting from economic actions, such as transfers or loans of crypto-tokens, are as important as social relationships to predict how users will behave during a shocking event, such as a hard fork. The paper is organized as follows. Section 2 provides a brief introduction of Web3 social platforms and an overview of the related studies. In Section 3 we describe how we model and manage data from Web3 social platforms. Sections 4 and 5 report the main findings of the growth of the Web3 platforms by applying discrete choice models and a discussion about predictability of user migration after a shocking event, such as a hard fork of the underlying blockchain.
Diferent services and platforms lie under the umbrella of the Web3 paradigm, but they all have the usage of blockchain technology as a common factor. Hereby we introduce the reader to a specific type of Web3 platform: Web3 social platforms or networks, broadly speaking online social networks or blogging platforms that rely on a blockchain to validate and persist all the interactions and actions established by their users.
Web3 social platforms. By Web3 social platform, we denote a web application which i) ofers a set of “social actions” - following, commenting and voting - facilitating online interactions among accounts; and ii) whose core functions are ground in an underlying blockchain that guarantees the persistence and the validity of the operations. One of the most interesting consequences of this architecture is a strong link between economical aspects and online social behaviors, in fact, most of the current Web3 social platforms implement: i) a token ecosystem based on blockchain technology for promoting high-quality content and users, and validating social and economic operations; and ii) a rewarding system for distributing the wealth of the platform. In particular, the rewarding system defines the set of rules and mechanisms regulating the distribution of tokens among the users who actively participate in the platform activities. In most of these platforms, rewards are assigned to accounts that publish content - posts or comments - and accounts that promote content through upvoting, downvoting or sharing. Specifically, content promotion is based on a stake-based voting system, where the voter decides how much of its economic power - the amount of gained crypto-tokens - to put behind a vote.
In the landscape of Web3 social platforms, most of the research studies have been focused on
Steemit. Launched in 2016, Steemit has been one of the most widespread Web3 social platforms
and it is considered a pioneer for the Web3 ecosystem since it has introduced the seminal
concepts of rewarding system and proof-of-stake consensus algorithm for block validation. In
detail, the platform is hosted on a blockchain called Steem, and implements three diferent
tokens: STEEM, Steem Dollar - (SBD), and Steem Power - SP, where the last is on the basis of the
internal rewarding system and the first two tokens are tradable on exchange markets. Steemit
has gathered the interest of researchers for its characteristics and has been dissected in many
aspects. For example, a few studies have focused on the features of diferent types of social
networks resulting from diverse interactions or specific subsets of accounts ([
Hard fork and user migration. Web3 social platforms may ofer data about a
phenomenon which is peculiar to blockchain-based systems. Indeed, in these systems situations
where miners/validators change the consensus protocol may happen, leading to what is known
as blockchain fork. Specifically, two types of fork may occur: i) soft forks, where changes retro
compatible with the previous consensus protocol are introduced so that new blocks are added
to the same chain; and ii) hard forks, where miners do not consider as valid the blocks validated
with the new protocol so that two diferent branches are created if validators do not reach a
consensus on which protocol to use. In the latter case, the members of the original branch may
opt to migrate to the platform based on the new branch, leading to a phenomenon denoted as
user migration. User migration is a “universal” process spanning both centralized and
decentralized online social media but is not fully understood yet, especially in the Web3 world. Most
of the studies are based on centralized social platforms. For instance, Kumar et al. [
Web3 social platforms make available to their users a rich set of operations to support diferent kinds of interaction, namely interaction actions. From a network perspective, interaction actions - comments, likes, reacting and following - result in diferent types of relationships connecting accounts/users. Interaction actions in Web3 social platforms have two further important features inherited from blockchain: i) each interaction action has a timestamp corresponding to the time the block containing the action has been validated and recorded into the chain, and ii) interaction actions are not merely social, rather also related to economic or financial operations, such as borrowing or transferring tokens or assets between accounts. The latter aspect makes Web3 social platforms complex techno-social systems with diferent intertwined layers of actions.
Formally, interaction actions are represented as a set of tuples = {(, , , )} where and
are accounts, who interact through an action of type , validated by the blockchain at time
. Given the temporal information associated to each tuple in , we build two diferent kinds
of sequence of directed multigraphs [
Finally, it is worth noting that the above definition does not take into account the meaning of operations, especially complementary operations, such as “follow” and “unfollow” where the latter determines a removal of the link created by the former operation. This way actions can only increase the state of a multidigraph, while semantic constraints can be introduced in a successive phase of the analysis.
Handling data produced by Web3 social platforms requires an efort not only in modeling and representing temporal and heterogeneous data but also in designing and developing scalable analytics platforms which have to gather data from diferent blockchains and manage largescale multidigraphs. To this aim and to support our analysis of the temporal aspects of Web3 social platforms, we have developed FlowChains, an analytics platform for Web3 social and trading data, whose architecture is depicted in Figure 1. The core of the architecture is the event streaming platform Apache Kafka which collects data through a set of Kafka Connectors. Each Kafka Connector handles a specific stream of blocks gathered through the APIs released by Web3 platforms and applications. Then, syntactically diferent operations coming from diferent blockchains which express the same interaction action are grouped in the same Kafka topic. This aggregation step facilitates the alignment among the data produced by the rich set of blockchains. Finally, the interaction actions belonging to a specific topic are transformed into a temporal annotated link in a graph database management system. In FlowChains the graph database is Neo4J, the leading solution for handling and managing large-scale multigraph. The choice of Neo4J has been dictated by the fact that it natively supports multiedges between two vertices, complex attributes can be assigned both to vertices and links, and it follows the schema-free paradigm so allowing a certain level of flexibility in the data model.
Through the FlowChains platform, we were able to collect a large dataset covering the lifespan of two important blockchains Steem and Hive, at the basis of the two Web3 social platforms: Steemit and Hive blog. Specifically, the blockchain Hive has originated by a hard fork of Steem on the 20th of March 2020, after a 51% attack. Although they are two diferent blockchains both Steem and Hive have released more than 50 operations in common. Among them, we are interested only in operations generating interaction actions, of which a subset has been reported in Figure 1b. These actions have been further grouped into two main categories: i) ifnancial and ii) social operations. Financial operations are for rewards and token management, and asset and share transfer; whereas social operations correspond to posting, rating, voting, sharing and following. From the modeling perspective, this aggregation corresponds to defining as {, } .
The details about blocks and operations for both platforms have been gathered through oficial public APIs. Moreover, due to the implementation of the hard fork, data between the two blockchains are identical up to the fork event. From there, Hive and Steem have recorded diferent data, as they have become two separated entities. So, we collected operations from the very first block on Steem blockchain, produced on 24th March 2016, up to January 2021. For Hive, we start from the first block after the fork (20/03/2020), and up to January 2021. Overall, from Steem, we extracted 993,641,075 operations related to social interaction actions and 72,370,926 operations related to financial actions; from Hive we have a total of 206,224,132 social actions and 4,041,060 financial actions. It is worth noting that both blockchains record the operations with a three-second granularity.
The understanding of how online social networks grow and evolve is a central issue in temporal
network analysis, however, traditional approaches have failed to catch the complexity of
such phenomenon, especially when online social networks are the result of a mix of growth
mechanisms based on diferent aspects, from social to economic or financial. Recently, a few
approaches based on discrete choice models [
as a choice made by to connect with , where the set of possible choices for is the
set of all other nodes. For further details about discrete choice models applied to social network
formation we refer the reader to Overgoor et al.’s seminal work [
Experimental setting.
function parameters, first we identify a set of dates - 2017-03-15, 2017-04-15, 2017-05-15,
201706-15, 2017-07-15, 2017-08-15, 2017-09-15 - which correspond to specific stages of the growth of
Steemit, as shown in [
Finally, from the set of “follow” links created in [ , ], we extract elements which corresponds to the choices (, ) . On the other hand, alternative choices have to be selected. Indeed we have to estimate the utility assigned by node to other alternatives, to approximate the likelihood function. To speed up the computation, a negative sampling of the alternatives is applied, i.e. for each positive choice (, ) , we randomly select nodes not connected to in [ , ]. Once, the choice data have been selected, we compute for each pair (, ) basis of the information provided by and infer the parameters .
the feature vector on the
Parameters inferred for the conditional logit model. Each column represents a stage of the evolution of 3. Hops (HS): length of the shortest path between and in the social network; 4. Reciprocity (RS): boolean flag indicating whether ( , ) ∈ ; 5. Common neighbors - social(CNS): number of in-neighbors in common between and in the social network; 6. Friends of friends (FoFS): boolean flag indicating if CNS is diferent from 0; 7. In financial (IF) : boolean flag indicating whether is in the financial network;
Results.
In Table 1 we report the parameters of the conditional logit choice model which has achieved the best performance in terms of accuracy, i.e. the choice model whose random utility function is , = 1 log() + 2 + 3 + 4 + 5 ( ) + 6 + , . Each column reports the weights - - users have assigned to a particular feature when they chose to connect to other nodes during a specific stage of the growth of the Steemit social network. From the analysis of the parameters and their trends we observe that: • the weights of the diferent features are quite stable in each of the time windows. This might be a possible indicator of the stability of the users’ behavior when choosing to establish new relationships. It is also interesting to observe that the preferential attachment mechanism, captured by the feature () , has lost its importance as the network evolved; • reciprocating links ( ) and making connections with users with at least one common relationship (
) have got the highest impact on the growth of the Steemit social network. Specifically, while the choice of reciprocating links has been strong and stable throughout the observation period, the mechanism based on common neighbors has strengthened its importance as the network evolved; • in the choice of link formation the financial features are marginal and almost irrelevant.
In fact, most of the financial features are not even present in the formulation of the utility function, and the only financial feature got values close to zero during the whole observation period.
To sum up, through discrete choice models we are able to disentangle the complexity behind the evolution of a Web3 social platform as Steemit, and identify which are the main growth mechanisms that are leading the evolution of Web3 social platforms.
In the previous section we dealt with a certain level of stability of the mechanisms leading the evolution of Web3 platforms. In this section, we drastically change our setting by focusing on the efects of a shocking event such as a hard fork of the blockchain supporting a Web3 social platform - Steem in our case. Specifically, we are interested in the user migration consequent to the hard fork. In the light of the data representation described in Section 3, modeling user migration is quite straightforward: both the original blockchain - Steem - and the new branch Hive - are described by two distinct evolution multidigraphs: and , respectively, with a common ancestor representing the multidigraph at fork time . Given these definitions, we assign to each account a state. Specifically, a user migrates - migrant - from platform to after a fork, if after
s/he does at least one action on ; while a user remains on the original platform - resident - if s/he keeps performing actions on the platform and no actions on after the fork event. We also introduce a third category - inactive users, i.e. people who are inactive or have abandoned both platforms. So, for each user, we are interested in the predictability of the choice to adopt a new platform given some user’s characteristics or activities. In particular, we ask whether early signals indicating that s/he will move to a new platform exist. We cope with these research questions by casting this issue into a machine learning task, specifically a node classification task.
Definition: User migration prediction task. successive timestamps ′, where ′ > , we define the user migration prediction task as the prediction of a node migration in one of the successive time steps.
Given the graph and considering the
Indeed, the main goal is classifying a user as migrant or resident as a function of several features related to the structure of the temporal multidigraph describing the platform; whereas the assumption is that user features, at the network structure level, could be predictive of a future user migration. It is worth noting that features are extracted by exploiting both financial and social links.
in many network-based prediction tasks. Such features encode information about a node and its neighborhood. Specifically, for each user in , we compute in-degree and out-degree, weighted in-degree, Pagerank, neighborhood average degree, and local clustering coeficient. Alongside the structural information, we also include information on the status of nodes in the neighborhood: i) the percentage of inactive neighbors, i.e. the number of neighbors whose status is inactive at time , divided by the total number of neighbors, and ii) the percentage of resident neighbors, i.e. the number of neighbors whose status is resident at time divided by the total number of neighbors. In addition to the structural features, we compute a set of features related to the activity of the users before the hard fork, mostly inspired by the similar task of churn prediction: i) the number of active/inactive days in the three-month period before the fork, ii) the average number of daily actions in the three-month period before the fork, iii) the lifetime in the original blockchain, iv) the distance - in days - between the first and the last action in the three-month period before the fork, and v) the average length of the activity sessions2. These features have been extracted from the financial and social layers, i.e. subgraphs of a multidigraph where links have a specific label, social or financial actions in this case.
Experimental setting.
The experimental context is given by the migration from Steem to Hive. Hence, we rely on the Steem evolving multidigraph , and its financial and social layers. More precisely, we select the snapshot at fork time, = 2020/03/20, at 2:00 PM. Then, we obtain the labels migrant and resident by inspecting the sequences of multidigraph and after the fork time. The two classes are imbalanced. For instance, social layer, where there is a more severe imbalance, residents are 3/4x more than migrants (66.9 %, 33.1%); vice-versa in the financial layer there are more migrants (56.1%) than residents (43.9%). In this case we deal with the sample imbalance by oversampling through the SMOTE method.
We perform the performance evaluation in a 5-fold cross-validation setting. For each fold,
we apply oversampling on the training portion of the fold. Then, we train a model, compute a
set of evaluation metrics and we average the performances over the five folds. To compare the
performances of the diferent learning algorithms, we compute the main evaluation metrics
for classification tasks: weighted F1, accuracy, precision, recall and AUC. For the classification
task, we rely on standard machine learning methods: Logistic Regression, Random Forest,
2An activity session corresponds to a bursty train of actions as described in [
Support Vector Machine with linear kernel, a Gradient Boosting classifier, and two multilayer perceptrons - MLP - with 100-unit hidden layer and three hidden layers (128 units - 64 units 32 units), respectively.
Results. In Table 2 we report the results for the user migration prediction task. In this task, we are combining structural and activity-related features from both the social and financial layers, fully leveraging both the evolving graphs. This latter aspect is resulted crucial for improving the performances of all the models w.r.t. settings where features are taken from a single layer only - financial or social - have been considered. By comparing the model, as expected, the ensemble methods - Random Forest and Gradient Boosting - have obtained the best performances, which in general are quite good for the task. In addition, we observe that SVM and Logistic Regression have benefited from the addition of features coming from both layers. In short, the results suggest that in a stratified context with diferent types of links, interactions and features resulting from financial actions should be used together with social-based features to enhance the predictability of users in the case of a user migration.
Finally, we also performed a feature importance analysis to highlight the most predictive features to identify the early signals of willingness to migrate. The features ordered by their importance are depicted Figure 2. We observe that the most important features have been extracted from both social and financial layers and concern both structural and activity-related aspects. In fact, the average length of the session in the financial layer, and the neighbor degree in both social and financial ones are among the most important features. To sum up, this analysis on feature importance confirms the importance of taking into account information derived from both types of interaction actions when we tackle the user migration prediction task in Web3 social platforms.