The scientific product of my PhD Research.

• My response to topic X is A
• but I have incentives to make it recorded that my response is B.
• B is someway different than A.
• Possibly, converging into B will avoid that consensus on X will be C.

This is formally identical to tactical voting in political theory, where people would desire A to be elected, but they choose B because they do not believe that A would be elected, and they fear that C would, instead.

Sock-puppets are online fake identities, or just ‘fake accounts’.

• I am Aleph and everybody think that I believe C regarding X.
• I will make the fake online identity of Beth;
• everybody will think that she believes C, too.
• Conclusion: now the contribution of my opinion on X will count for double.

Here one can notice that Fake identities could induce a shift of response (not opinion, just response or “voting”) in tactical voters, inducing a false, disinformed, perception of future scenarios.

Social Botnetsare a extreme case of disinformation through sock-puppets.

The case for disinformation as a conceal of information about ourself - not about the outside world - is well discussed in Bergstrom & West’s Calling Bullshit. This theory, originally proposed by philosopher Harry Frankfurt, was further developed by authors as:

• Ian Bogost
• Gordon Pennycook
• David Graeber in the marvelous book Bullshit Job.

The best way to describe Agent Based Networks is:

…they are just randomly generated dynamic Networks where there is at least one class of nodes that are Agents.

Hyperparameters of the nodes and Agent behaviour are coded following a Model. By the actions of agents, the network will evolve into peculiar topologies that can be summarised with statistical indicators of outcome.

I think that a good ABNM accounts for agents observing past statistical summarisation of the whole systems, and acting accordingly in the present. Agent Based Network Models are developed at Indiana University - CNetS:NaN

Why is the word Network so important to define a subclass of Agent Based Models (ABM)? Because networks are both flexible and precise for model identification. I think that the epistemology of networks is still developing, so I will just provide an example of why networks are so interesting to model (social) action:

Imagine a recommender systems that observes actions from the users; these are very heterogeneous agents that interact with items in the catalogue of the systems. The items have no direct agency, we will call this structure a bipartite directed network. However, the recommender system acts as a architecture, shaping the topology of the connections between users and items. Even if not explicit, this mechanism can be found in one of the foundational paper on Consensus/ORPs systems.

My personal insight on this is that for facilitate and possibly enrich statistical inference, networks can always be converted into a tabular format, for example with softwares like tidygraph. Statistically speaking, that format would be a structured dataset - that means that one can apply Multilevel Statistics on (bipartite) networks.

(…Even if probably not exactly the classical estimators, since the assumptions are different…)

In the other panel it is said that agents doing stuff in a network lead into stable topologies that can be represented through statistical indicators.

There is an issue: while the class of the topology could be classified as unique, that means that all of randomly generated observation from the ANBM will share the same unique and known hyper-parameterisation, the statistical outcomes are still random variables with a unknow a priori variance.

Monte Carlo (MC), that is random resampling of the model would help identification of location of statistical indicators of outcome of the model. I personally find MC helpful for identification of the condition of neutrality of the model (null hypothesis), too.

1. A context for consensus disinformation is identified
2. Data is collected (possibly, scraped)
3. Data is analysed to understand basic agent mechanics.
   • Possibly, disinformation is disentangled from genuine, sincere, opinions from the agents. So, on pair of mechanics to characterise honest agents, different mechanics would characterise disinformation agents.
   • If all the agents have potential for disinformation, then disinformation is in the act, not in the agent.

• Models run through instances (parallel spaces, also called universes or worlds) and iteration (time).
  1. Fixed hyperparameters of the simulation, parameters for the instance are randomly generated. For example, the hyperparameter could be “The prior probability for a user to be a tactical voter is 30%”. But since the istance is random, the final parameter for that statistics would be: “There are 32 tactical voters for each 100 users”.
  2. Then, first iteration starts.
  3. All agents execute their program.
  4. Summary statistics of the state of the system are collected and recorded as the current state of the system.
  5. Second iteration starts. It is possible that the behaviour of the agents will be influenced by the summary statistics at the end of the first instance.
  6. The iterations repeat until the final iteration, which could be pre-fixed (as a hyperparameters) or until the state of the system is steady.
  7. The it starts the second universe. Universes could even run in parallel since they are statistically independent.
  8. For each universe, final summary statistics at the last iteration are collected and recorded as outcomes in the table for results.
• Through hyperparameterisation of the model, it is possible to explore many different setups. For example, the % of imputed disinformation could range uniformly from 0% to 30%, but never exceed 30%.
  • Given one of more target parameters per universe, do a raise in disinformation induce a stochastic raise in the bias for the statistics of a
  • For example: imagine that agents are programmed to infer the number of objects in a jar (which is a parameter for that universe). In the end, the average estimate across agents, that is a statistic, would be equal to that parameter plus one error. This error is tied to that universe in particular.
    • So… is the square of that error correlated with the % of disinformation in the system? This is a causal question that the ANBM - MC allows to answer to.
• Validation of the ANBM - MC model: do the distribution of outcomes fit, or at least is it coherent in its projection over empirical data? And if there are strong differences, are these admissible or at least useful for the research? If not, the model is not valid to go further into step III.

There are two kind of counter-measures:

• Endogenous interventions: which is likely a change of architecture within the consensus system.
  • Here is important that the intervention does not change the way that agents are programmed or parameterised. Only the condition to access about summary information about the sate of the systems. For example, this information could be totally negated to the agents, or presented in a different way, or only to some agents and not for others.
  • In this case the summary statistics for the table of results are exactly the same. The difference would be that in half of the universes, the intervention is applied, and in the other half is not. If the bias, that is the difference between parameter and consensus-reached estimate is lower in the sample with intervention, then the intervention for reduction of impact of disinformation has been estimed successful!
• Exogenous interventions: this is trickier but it can be summarised in a very simple sentence: are there alternative statistics? i.e. are there alternative inferential models that out-perform the null statistics for consensus estimation in a disinformed system? Here the estimation method is just a bit more complicated, and for once, equations make it simplier:

$$ DIR = \tau(\theta_1(\hat{\alpha}),\alpha) - \tau(\theta_0(\hat{\alpha}),\alpha) $$

• α is a vector of parameters.

  • For example: the latent quality of items in a ORP

• θ are the estimators.

  • For example: the average of the rating received (theta zero) vs. the average of the rating received by the 100 most active user nodes (theta one)

• τ is a coefficient of rank-correlation.

  In other words, for a exogenous change of estimator it only matters that it is robust to disinformation in order to be more desirable.