# Gitcoin Grants Research - ML Working Group

TODO (presentation project):
- Raw contributions data (per round) as individual NetworkX objects, visualized over time
- node2vec embedding
- visualize in TSNE

<!-- - inject attack vectors
- compare with a similarity measure -->

## :memo: Research questions

## Q1: How can we visualize the trade-off of collusion detection sensitivity?

Identified problem = Collusion algorithm sensitivity (collusion prevention versus credible neutrality)
Identified solution = Classify collusion nodes with different sensitivity parameters, observe false positive / false negative rate

Methodology:
1. Create a Gitcoin Grants bipartite graph with NetworkX and inject (flagged) archetypal attack vectors
2. Create node2vec embeddings of the graph data
3. Calculate a similarity measure comparing injected attack vector nodes to empirical contributors/nodes to surface potential signals of empirical attack vectors
4. Investigate potential malicious activity via interactive analysis

Notebooks:
- [Gitcoin Grants Research - Graph machine learning](https://colab.research.google.com/drive/12JzmgAhBqzOAVEED_e2oRSnH3YWat6zF?usp=sharing)
- [Gitcoin Grants Research - Round data overview](https://colab.research.google.com/drive/15EbzxNHqjjrDEvrNoFgsB7JRzlbm2c15?usp=sharing)
- [Gitcoin Attack Vector Research (Jiajjia, Blockscience)](https://github.com/jiajia20/GitCoin_attack/blob/main/attack_vector.ipynb)

## Q2: How can we quantify the saturation effect?

Track the before/after saturation point and observe the change in behavior (how strong is the assumption that grant contributors are aware of the saturation dynamic?)

Draw-back of the research question: What percentage of contributors are aware of the saturation point? Hypothesis is that not many. Thus there is not much to observe in the temporal patterns before/after.

## Q3: How can we model the Matthew effect?

Better understand Matthew effect -> Better handle its negative consequences (rich getting richer).

Potential starting point methodology:
Comparing theoretical (Barabasi-Albert model) and empirical (Gitcoin data) dynamics.

Draw-back of the research question: There is inherent difference in quality of projects (= propensity of people to contribute) that has nothing to do with collusion/malicious activity. To get any kind of pattern, we'd need to get a high false negative rate which provides little value to the Gitcoin team (low F-score).

**Archetypes (Heuristics):**
Types of behavior based on observations, behavior does not indicate collusion
Indicators a starting point to look for behavioral patterns, that might be in scope/out of scope for Gitcoin Grants
Grants:
- "Mature Community": grant that already gathered a large community and is able (or has even established a process) to activate the community to donate, indicators: large number of supporters, grant starts with a significant number of contributors already, contributors come back round after round (loyal contributors)
- "Seeds": grant starts with small number of contributors, grows round by round, continously active on Gitcoin Grants, loyal contributors
- "Grants Marketing": majority of contributors donate on the same 1-2 days, many contributors donate exact same amount
- "Isolated Grants": contributors donate to only this particular grant, no overlap with other grants
- "Mayfly": grant only active in 1-2 rounds, or able to attract contributors in only 1-2 rounds, highly volatile number of contributors across rounds
- "Distributed Grants": various grants published by the same team/owner

Contributors:
- "Web3": contributor active round after round, loyal to most grants, variety in contribution amount and set of grants
- "Follower": contributor active round after round, loyal to a particular grant/set of grants
- "Optimizer": interface provides matching, contribution close to optimum of contribution/matching ratio at that point in time
- "Trial": contributes in only 1 round

Value:
- define impact, how much value is extracted by a particular behavior, is it problematic
- similar to Trust Bonus add bonus for certain behavior
- implement penalty for behavior

## Applications

We should notice that contributor or grant behavior does not necessarily indicate collusion but could be derived from game theory, that is to maximize agent's utility using the rules of the game (system). We might come up with different incentives or algoritms ("applications") to mitigate bad effects, then we first have to address what kind of effects are considered bad from Gitcoin's perspective.
Gitcoin values: Self Reliance, Intellectual Honesty, Collaboration, Empathy, Stress Reducers, Inclusivity, Giving First

Let's look at the Behavioral patterns (if derived from the research efforts) from an application perspective.

Grants:
- "Mature Community": bad would be the Matthew effect, a community already well funded and doesn't need Gitcoin grants anymore. Scaling up would be done in other platforms or VC-like. **Application**: algoritms to check for maturity, #recurring contributors, #semi whale contributors, nudging assistants to consider other funding, recommendation engines to split up funding to other grants
- "Seeds": loyal contributors suggest a foundation layer of investors already aboard, same as above. "Giving first" value already addressed.
- "Grants Marketing": bad would be the "no idea what this means, I just know the guy" - effect. **Application**: educate on grants content, to align with personal incentives to split up or direct funding otherwise
- "Isolated Grants": conflicts with "Inclusivity" value. **Application**: educate on Gitcoin values and align with corresponding personal values to re-assess funding possibilities
- "Mayfly": conflicts with "Collaboration". Is this grant aligned with the open ecosystem perspective or just a wild shot? **Application**: exclude grant from funding by using a contributer's treshold or use other means to grow more traction. Curation - however sensitive - might be wise
- "Distributed Grants": possibly conflicts with "Self Reliance" value. **Application**: curate grant addresses by using SSI concepts, what kind of people are behind grants? Sensitive though..

Contributors:
- "Web3": contributor active round after round, loyal to most grants, variety in contribution amount and set of grants. This would be considered "good effect", aligned with Gitcoin values?
- "Follower": contributor active round after round, loyal to a particular grant/set of grants. **Application**: try to nudge towards other grants, possibly better aligned with personal values. Amazon's "Buyers of this thing also bought...."
- "Optimizer": interface provides matching, contribution close to optimum of contribution/matching ratio at that point in time. Possible Attack vector, useful to serach for optimality gaps. **Application**: mitigate "attacking" contributors to block certain funding strategies or nudge towards other grants.
- "Trial": contributes in only 1 round. **Application**: marketing (spamming?) effort to get recurring behavior

Overall:
Bonus for certain behavior aligned with Gitcoin values, penalties are weak offerings.