Spam has become the norm in our day and age. Whether it is on social media platforms or in our email inbox, it's absolutely everywhere.

A common experience when recruiting: In an attempt to fill an empty position, you're reaching out to a great number of candidates and only a small percentage gets back to you. Those who do respond often end up lacking important qualities. It's a hit-or-miss scenario. 

But how did we end up in this conflicting situation and how can we resolve it?

The problem of spam exists because for everybody individually, spamming people is the most effective solution. The marginal costs of distributing spam-mails are practically zero whereas building a healthy, honest relationship between recruiters and candidates is tough work. 

Furthermore, free access to spamming without any regulations collectively leads to a worse experience for everyone. It's a typical tragedy of the commons: Through the pollution of the information environment, the relevant messages can't be filtered out anymore because of the excessiveness of spam.

This is a vicious cycle. The more people are getting spammed, the more important messages are lost, resulting in declining response rates. Declining response rates lead to further spamming because even “honest” senders now have to start sending more messages to get a response. To improve this situation there needs to be a system in place where the sending behaviour has to be tracked over time i.e. spamming starts to be costly.

For this reason, we're building a reputation system for Primeflow where users who spam the system will bear the cost. The lower the reputation of the spamming agent, the fewer of his messages get prioritised in the future. With this method, the aforementioned vicious cycle can finally be broken. 

Characteristics of agent types:

In our research to find an adequate reputation system that could effectively resolve the problem of spamming without discouraging relevant messages, we created an agent-based simulation. In this simulation, four different groups of agents with different propensities to spam, hoard, be passive or take initiative (see above) compete against each other to generate matches within a typical social graph structure

Simulation performance report:

The image above shows how the "greedy” agent, who is characterized by his "spamminess" - his ruthless drive to claim cheap deals at every occasion - is actively depleting his reputation. The agent has no reputation control and ruins his networking ability because of it whereas its peers fared much better. 

Some main take-aways from conducting this simulation are as follows:

There is no free lunch

For the simulation to be fully resolved, messages must be constantly propagated through the graph: this comes with a cost, in reputation. 

Location is crucial (but effects depend on the agent)

Greedy and passive agents demonstrated distinctive behaviours when isolated at the extremities of the graph: whilst greedy agents were able to profit from the lack of neighbours, passive agents suffered the position tremendously. 

Passivity beats hyperactivity, if you are already well connected

Passive agents have risked less and won more (reputation). Default agents, however, come close, with even a better equilibrium between score/reputation.

Network-based matching can further improve

C.A.S. is a measure of which fraction of all the resources in the network find their match - the larger the C.A.S., the more efficient the collaboration within the network. The highest C.A.S.  we recorded was 0.84. There is still room to improve the matching since even the most aggressive strategies didn’t get all their resources matched in deals. An external “helper” that suggests matches based on the data and behaviours observed within the system could significantly improve the matching.  

We are using these observations on the path to the Primeflow prototype version 5 - reach out to us below if you want to be a part of our initial pilot user group.

Follow us on Twitter and LinkedIn for future updates on our reputation model and let us know if you want to learn more about how we conducted the simulation above. Get in touch at if you’d like to see the rest of the slides.

Read next