Mapping Partisan Fault Lines Within DAOs‗ Thomas Lloyd†, Daire Ó Broin‡, and Martin Harrigan§ South East Technological University (SETU), Republic of Ireland
arXiv:2605.10316v1 [cs.CR] 11 May 2026
Abstract Decentralised Autonomous Organisations (DAO) can fragment when partisan communities emerge within their governance structures, leading to organisational splits known as “forks.” We present a method to detect these emerging communities by analysing on-chain voting behaviour before fragmentation occurs. Our approach extracts voting events from governance smart contracts, constructs voter matrices encoding participation patterns, and applies pairwise dissimilarity analysis to quantify ideological divergence between addresses. We visualise these relationships using multidimensional scaling and identify partisan communities through k-means clustering with silhouette score optimisation. Using Nouns DAO as a case study, a protocol that has experienced multiple documented forks, we demonstrate that addresses destined to fork cluster together months before actual fragmentation events. Our analysis of 330 proposals spanning from contract deployment to the first major fork shows that 90 % of fork addresses cluster together in the final 44 proposals, compared to only 47 % in randomised data. These results indicate that partisan communities can be detected and visualised through on-chain governance analysis, offering early warnings of emerging divisions before they cause organisational fragmentation.
1
Introduction
DAOs have experienced rapid growth in both number and participation since 2019 [4], with governance decisions now affecting billions of USD in digital assets. Unlike traditional organisations with known stakeholders, DAOs are governed by pseudonymous token-holding addresses, creating challenges for understanding community dynamics and predicting organisational stability. While DAOs often present themselves as unified communities, the inherent nature of governance, involving disagreement, debate, and competing interests, can lead to the formation of distinct sub-communities that may ultimately fragment the organisation. The fragmentation of DAOs through “forking,” where dissenting members create new instances of the protocol, is an important aspect of decentralised governance. Forking divides communities and resources and undermines the legitimacy and effectiveness of the original organisation. Early detection of partisan communities that might lead to fragmentation is essential for maintaining long-term protocol stability and effective governance. This paper presents a method for identifying and visualising partisan communities within DAOs by analysing on-chain voting behaviour. Our approach transforms blockchain governance data into spatial representations that reveal ideological alignment patterns among participants. By examining the voting history of addresses across multiple proposals, we can detect the emergence of cohesive sub-communities before they manifest as organisational splits. Our method consists of several key stages: we extract voting events from on-chain governance smart contracts and construct voter matrices encoding participation and support patterns across all proposals. We then apply pairwise dissimilarity analysis to quantify ideological divergence between addresses based on their shared voting history. These dissimilarity scores are visualised using multidimensional scaling (MDS) to create two-dimensional spatial representations where proximity indicates voting alignment. Finally, we apply k-means clustering with silhouette score optimisation to identify distinct partisan communities and validate our results against randomised data. We demonstrate our approach on Nouns DAO1 as a case study, using its documented history of forks to provide ground truth validation. Nouns DAO’s governance structure, where voting power derives from NFT ownership, and its well-documented fragmentation events make it a suitable testbed for evaluating our ‗ This is a preprint of a published article [10]. † # [email protected] ‡ # [email protected] , 0000-0002-2886-5546 § # [email protected] , 0000-0002-6069-7001 1 Nouns DAO homepage:
3 https://nouns.wtf/
1
community detection method. Our analysis spans from the deployment of the governance contract through the period preceding the first major fork. Our results show that addresses destined to fork can be detected within distinct clusters months before the actual fragmentation event occurs. As proposals approach a fork, voter data shows near-unanimous clustering of future fork participants, while randomised data exhibits no such pattern. The remainder of this paper is organised as follows: Section 2 reviews related work on DAO governance behaviour, multidimensional scaling for political analysis, and decentralisation measurement frameworks. Section 3 details our multi-stage method for transforming blockchain data into community visualisations. In Section 4 we present our validation procedures and our analysis of Nouns DAO, including MDS visualisations and clustering results compared against randomised baselines. We conclude in Section 5 with implications for DAO governance.
2
Related Work
We organise related work into three areas: DAO communities and governance behaviour, multidimensional scaling for political analysis, and decentralisation measurement frameworks.
2.1
DAO Communities and Governance Behaviour
The pseudonymous nature of DAOs creates challenges for community analysis, as participants must reach consensus on multi-billion USD decisions without traditional trust mechanisms. Research in this area has moved from basic controversy metrics to coalition detection methods that account for pseudonymity and delegation. Sun et al. [15] identified five distinct coalitions within MakerDAO [14] through clustering analysis, showing how voting blocs can be detected and their impact on protocol decisions measured. Similarly, DuPont [5] applied graph-based deep learning to detect Sybil clusters, revealing that 2 % to 5 % of voter addresses exhibit suspicious behavioural patterns. Austgen et al. [2] introduced the theoretical framework of voter-block entropy (VBE), which treats aligned addresses as unified voting entities rather than individual accounts. While their approach remains theoretical, it provides the conceptual foundation for our practical method of identifying partisan communities through historical voting behaviour. Their work on Dark DAOs, which attempt to buy votes opaquely, highlights the importance of transparency in governance analysis. Recent studies have incorporated off-chain community sentiment analysis. Quan et al. [12] analysed Discord discussions across six DAOs using natural language processing, finding generally positive sentiment in governance channels. However, this approach cannot capture the verified, immutable voting behaviour that on-chain analysis provides.
2.2
Multidimensional Scaling for Political Analysis
The application of MDS to political voting analysis has proved effective in traditional parliamentary systems, providing the methodological foundation for our DAO community detection approach. Mickevicius et al. [11] demonstrated the use of MDS for parliamentary voting analysis in Lithuania. They showed that clustering algorithms can reliably identify opposing political parties. Their work established MDS as a suitable method for visualising ideological alignment in voting systems. Amelio and Pizzuti [1] analysed Italian parliamentary votes using temporal segmentation, dividing their dataset into six-month intervals to capture evolving political dynamics. Their approach of tracking faction formation over time directly informs our method of analysing DAO voting patterns across sequential proposals. They showed that MDS visualisations can reveal the emergence of new political coalitions before they become formally recognised.
2.3
Decentralisation Measurement and DAO Assessment
Understanding DAO decentralisation has become important as regulators grapple with how to classify these novel organisational structures. Research has revealed concerning centralisation patterns across multiple DAOs. Fritsch et al. [8] found that fewer than 19 addresses control majority voting power in three major DAOs, highlighting the concentration of governance influence. This finding shows why our address-level analysis remains relevant despite apparent power concentration: coalition formation among minority holders can still influence majority decisions. Feichtinger et al. [6] performed a related analysis across 23 Ethereum DAOs, revealing widespread centralisation in voting power distribution alongside low participation rates and significant costs from governance transactions. Their work on “pointless governance transactions” shows the importance of filtering for genuine, sustained participation in our analysis.
2
The same authors’ systematisation of DAO attack vectors [7] identified behavioural factors as the primary source of governance exploits across 28 real-world incidents. Their findings show the importance of understanding community dynamics to predict and prevent governance failures. Axelsen et al. [3] developed the TIGER framework for measuring decentralisation across five dimensions, applying design-science methodology to create verifiable decentralisation metrics. Their work addresses regulatory concerns about what constitutes “sufficient decentralisation” but focuses on structural rather than behavioural measures. Historical perspectives on governance crises provide additional context. Reijers et al. [13] analysed the resolution to The DAO hack in 2016 as a “state of exception,” revealing how private interests can override established governance rules during emergencies. This analysis highlights the fragility of decentralised governance systems and the importance of detecting partisan divisions before they reach crisis points. Our work synthesises insights from all three areas: we apply MDS techniques from political analysis to DAO governance behaviour, providing a practical tool for measuring the kind of community fragmentation that decentralisation frameworks seek to prevent.
3
Method
To identify partisan communities within DAOs, we propose a multi-stage method that bridges on-chain data analysis with spatial and statistical validation. The following subsections cover each stage (Sections 3.1 to 3.7). Briefly, the method begins with blockchain data acquisition, where we gather raw transaction logs and event data from EVM-based DAOs using EVM RPC calls. The data is structured into a voter matrix, encoding voting behaviour while accounting for abstentions and non-participation. We perform a community friction assessment to quantify overall voter discord across proposals. This acts as a sieve, prioritising DAOs with significant partisan dynamics. Then we identify active voters by selecting addresses with a sufficient level of participation within a proposal window; this limits our analysis to active participants, reducing noise from sporadic voters. The core of our method lies in pairwise dissimilarity computation, where we measure ideological divergence between addresses by comparing their shared voting history. This produces a dissimilarity matrix per proposal, capturing how often two voters opposed each other relative to their co-participation. We visualise matrices using multidimensional scaling (MDS), mapping addresses into 2D space where proximity reflects voting alignment. Finally, we identify partisan communities using dynamic clustering, optimising cluster counts through silhouette scoring and testing robustness by comparing real-world data to randomised voting patterns. Our method can be applied broadly to DAOs. However, we explain it using Nouns DAO as a running example. Voting power in Nouns DAO is derived from NFT ownership and its documented history of forks provides a ground-truth with which we can compare our results.
3.1
Blockchain Data Acquisition
We collected on-chain transaction and event logs from EVM blockchains (Ethereum and Arbitrum) through EVM RPC calls, using a self-hosted Ethereum node and Infura. We identified governance smart contracts for six DAOs: Lido, Compound, Tornado Cash, Nouns, Uniswap, and Arbitrum. In previous work, we implemented a systematic approach to identifying governance smart contracts [9]. However, for this work, our focus is a deeper analysis of a smaller number of DAOs and it sufficed to identify the contracts by inspecting the projects’ documentation and transaction history. For each DAO, we determined the relevant event signature(s) by analysing historical transactions and filtering on voting events triggered by governance contracts when DAO members executed votes. These transactions contain critical data: proposal identifiers, voter addresses, and vote direction, i.e., support, opposition, or abstention. We collected event logs across the entire operational history of each DAO, from the block in which the governance contract was deployed, up to and including block 22,575,000 (May 2025). We selected Nouns DAO as a running example for subsequent steps. Its governance contract2 was deployed in block 12,985,453 (August 2021).
3.2
Voter Matrix Construction
After parsing, each vote log entry contained the voting address, proposal ID, and support value. We transformed this data into a voter matrix, illustrated in Fig. 1, where rows correspond to voter addresses, columns represent proposal IDs, and entries denote support: 1 for “Yes”, 0 for “No,” and −1 to represent all other cases. Originally, the matrix cells included five distinct states: 1 (Yes), 0 (No), −1 (no voting power and no vote), 2 (had voting power and abstained), and 99 (had voting power but did not vote). However, we subsumed the latter three states (−1, 2 and 99) into a single −1 value to simplify our analysis. The calculations to determine voting eligibility can 2 Contract Address: 0x6f3E6272A167e8AcCb32072d08E0957F9c79223d
3
Figure 1: A sample from the voter matrix for Nouns DAO with ten voting addresses (the rows) and nine proposals (the columns). The cell values represent either “Yes” votes (1s in green), “No” votes (0s in red), or all other cases (−1s in grey).
be complex: addresses might have held tokens at the time of a vote but only gained eligibility retroactively due to delayed vesting schedules, delegation mechanics, or other temporal constraints, requiring computationally complex checks against historical token balances at specific block heights. To construct governance proposal voter lists, we extracted addresses with non-negative support values (0 or 1) for each proposal. While DAOs often use a support value of 2 to denote abstentions that contribute to quorum, we opted for a binary classification (for/against) to align with standard voting analysis frameworks.
𝑣11 𝑣21 𝑉 = . .. 𝑣 𝑛1
𝑣12 𝑣22 .. .
··· ··· .. .
𝑣𝑛2
···
𝑣 1𝑚 𝑣 2𝑚 .. .
𝑣 𝑛𝑚
For Nouns DAO, we constructed a matrix with 629 unique voter addresses (rows) and 330 proposals (columns) spanning a period from the deployment of the governance smart contract to the first Nouns DAO fork in block 18,144,239 (September 2023). The proposal IDs were numbered 1 to 362 inclusive; we excluded 32 cancelled proposals that had no voting activity, leaving 362 − 32 = 330 proposals.
3.3
Community Friction Assessment
To prioritise DAOs with meaningful partisan dynamics, we quantified community-level discord using two metrics derived directly from the voter matrix (see Section 3.2): static disagreement percentages per proposal and rolling average disagreement percentages over sequential proposals. For static disagreement analysis, we categorise proposals by the proportion of voters opposing the winning outcome. The thresholds are unanimous (0 % disagreement), low ((0, 20)%), medium ([20, 40)%), and high ([40, 50]%). These align with standard benchmarks for social consensus in collective decision-making studies. As shown in Fig. 2, Nouns DAO exhibits medium-to-high disagreement in 30 % of proposals, a stark contrast to Lido DAO’s near-unanimous voting patterns. This quantitative disparity motivates our selection of Nouns DAO as a high-friction case study, particularly given its documented fork in September 2023. The rolling average metric (see Fig. 3) reveals temporal trends in partisan behaviour. By calculating the mean disagreement percentage across a 10-proposal window (moving forward one proposal at a time), we detect persistent ideological divisions in Nouns DAO, where disagreement spikes and sustained volatility preceded its fork. This metric serves two purposes: (1) identifying DAOs with persistent rather than transient conflict, and (2) pinpointing temporal clusters of contentious governance that warrant further analysis (see Section 3.7). These friction metrics inform our subsequent dynamic clustering analysis. DAOs with medium/high disagreement exceeding 20 % (see Fig. 2) and rolling averages exceeding 15 % (see Fig. 3) are flagged for further analysis, while low-friction DAOs (e.g., Lido) are excluded. This filtering ensures our clustering experiments (see Section 3.7) focus on cases where conflict is evident.
3.4
Active Voter Identification
To address the sparsity of the voter matrix caused by fluctuating participation in DAOs we compute a voter participation list. This is a key distinction between DAO voting systems and, say, parliamentary voting systems where persistent cohorts dominate. In DAOs, voting eligibility often depends on dynamic factors like token ownership, delegation, or time-based vesting, leading to variable active voter sets across proposals. This results in many addresses having minimal or no voting activity, introducing noise into community detection efforts.
4
Figure 2: A percent stacked bar chart showing disagreement across all proposals for six DAOs at the voter address level. Lido DAO voters are unanimous (0 % disagreement) or exhibit low disagreement (in the (0, 20)% range) for almost all of their proposals whereas as Nouns DAO voters exhibit medium disagreement (in the [20, 40)% range) and high disagreement (in the [40, 50]% range) for over 30 % of their proposals.
Figure 3: A plot of the rolling average of disagreement for six DAOs using a 10-proposal sliding window. The x-axis contains the proposal IDs, which acts as a proxy for time. The y-axis shows the level of disagreement. Lido DAO remains close to unanimity throughout, while Nouns DAO exhibits sustained conflict leading up to its first fork.
5
Our method calculates participation frequency using a sliding proposal window. For each proposal, we define a window containing the proposal and a fixed number of preceding proposals. For Nouns DAO, we used a window of size 10. Therefore, for Proposal 20, the window included Proposals 11 to 20 inclusive. For each voter address in each window, we compute their participation percentage: Participation % =
# of proposals in which the address voted # of proposals in the window
Addresses with a participation percentage below a fixed threshold (40 % for Nouns DAO) were dropped from the filtered voter matrix. This ensures subsequent analysis focuses on voters with sustained engagement. Early proposals (proposals 1 to 9 for Nouns DAO) used shorter windows (e.g., the window for Proposal 2 included only Proposals 1 and 2), increasing voter counts until sufficient historical data was accumulated. More formally, let 𝑤 ∈ N denote the proposal window size and 𝜏 ∈ [0, 1] the participation threshold. For each proposal 𝑝 𝑗 ∈ 𝑃, we define its sliding window 𝑊𝑗 ⊆ 𝑃 as: 𝑊𝑗 = 𝑝 max(1,𝑗−𝑤+1) , 𝑝max(1,𝑗−𝑤+2) , . . . , 𝑝 𝑗 .
For each address 𝑎 𝑖 ∈ 𝐴, we define its participation percentage 𝜋 𝑖𝑗 in 𝑊𝑗 as: 𝜋 𝑖𝑗 =
1 Õ 1{𝑣 𝑖 𝑘 ≥0} , |𝑊𝑗 | 𝑝 𝑘 ∈𝑊𝑗
where 1 is the indicator function. The filtered voter matrix 𝑉 ′ retains only rows 𝑎 𝑖 with 𝜋 𝑖𝑗 ≥ 𝜏, and the columns corresponding to the proposals in 𝑊𝑗 . This step improves the signal-to-noise ratio in sparse voter matrices. By prioritising active participants, we align the analysis with real-world governance dynamics where ideological commitment is most meaningful among addresses actively involved in decision-making. The filtered data then serves as input for the pairwise dissimilarity computation (see Section 3.5), ensuring clustering algorithms operate on representative voting patterns.
3.5
Pairwise Dissimilarity Computation
The core of our partisan community detection method relies on measuring ideological divergence between voter addresses through pairwise dissimilarity computation. This process quantifies how often two addresses vote in opposition relative to their shared participation across proposals, producing a dissimilarity matrix for each proposal that captures the voting alignment within the DAO. For each proposal 𝑝 𝑗 ∈ 𝑃, we compute pairwise dissimilarity scores between all active voter addresses within the proposal’s sliding window 𝑊𝑗 (as defined in Section 3.4). Let 𝐴 𝑗 ⊆ 𝐴 denote the set of active addresses for proposal 𝑝 𝑗 , filtered by the participation threshold 𝜏. For any two addresses 𝑎 𝑖 , 𝑎 𝑘 ∈ 𝐴 𝑗 , their dissimilarity score 𝑗
𝑑 𝑖 𝑘 is calculated as:
Í 𝑝 ∈𝑊 1{𝑣 ,𝑣 ≥0 ∧ 𝑣 𝑖ℓ ≠𝑣 𝑘ℓ } ℓ Í 𝑗 𝑖ℓ 𝑘ℓ 𝑗 𝑑𝑖 𝑘 = 𝑝ℓ ∈𝑊𝑗 1{𝑣 𝑖ℓ ,𝑣 𝑘ℓ ≥0} 1
if ∃ 𝑝ℓ ∈ 𝑊𝑗 : 𝑣 𝑖ℓ , 𝑣 𝑘ℓ ≥ 0 otherwise.
The numerator counts the number of proposals within 𝑊𝑗 where both addresses cast valid votes (i.e., 𝑣 𝑖ℓ , 𝑣 𝑘ℓ ≥ 0) but voted in opposition to each other. The denominator represents the total number of proposals where both addresses participated with valid votes. When two addresses have no overlapping voting history within the window, we assign the maximum dissimilarity score of 1, as pairs with no shared votes cannot show voter alignment; this occurred in less than 0.6 % of all pairs across all proposals. This formulation ensures that dissimilarity scores are normalised to the range [0, 1], where 0 indicates complete alignment (addresses always vote together when both participate) and 1 indicates complete opposition (addresses never agree when both participate). Algorithm 1 outlines the computational procedure. For each proposal 𝑝 𝑗 , we construct the sliding window 𝑊𝑗 and identify the set of active addresses 𝐴 𝑗 that meet the participation threshold. We then extract the relevant submatrix 𝑉 ′ from the voter matrix 𝑉 and compute pairwise dissimilarity scores. The resulting dissimilarity matrix 𝐷 𝑗 for proposal 𝑝 𝑗 is an 𝑛 × 𝑛 symmetric matrix where 𝑛 = |𝐴 𝑗 |. Each entry 𝑗
𝑑 𝑖 𝑘 represents the dissimilarity between addresses 𝑎 𝑖 and 𝑎 𝑘 , with the diagonal entries being zero (complete self-similarity). Figure 4 shows a sample from the dissimilarity matrix for Nouns DAO Proposal 300. For Nouns DAO, this process generated 329 dissimilarity matrices: one per proposal except the first, which has no preceding proposals from which to compute a window. Each matrix captures the voting alignment within the DAO at that point in time. These matrices serve as the foundation for subsequent multidimensional scaling visualisation (see Section 3.6) and dynamic clustering analysis (see Section 3.7).
6
Algorithm 1 ComputeDissimilarity generates pairwise dissimilarity matrices for all proposals. 1: function ComputeDissimilarity(𝑉, 𝐴, 𝑤, 𝜏) 2: 𝑅 ← {} 3: for all 𝑝 𝑗 ∈ 𝑃 do 4: 𝑊𝑗 ← {𝑝max(1,𝑗−𝑤+1) , . . . , 𝑝 𝑗 } 5: 𝐴 𝑗 ← {𝑎 𝑖 ∈ 𝐴 : 𝜋 𝑖𝑗 ≥ 𝜏}
⊲ Dictionary to store results
𝑉 ′ ← 𝑉[𝐴 𝑗 , 𝑊𝑗 ] 𝑛 ← |𝐴 𝑗 | 8: 𝐷 𝑗 ← 0𝑛×𝑛 9: for 𝑖 ← 1 to 𝑛 − 1 do 10: for 𝑘 ← 𝑖 + 1 to 𝑛 do 11: valid ← {𝑝ℓ ∈ 𝑊𝑗 : 𝑣 𝑖ℓ , 𝑣 𝑘ℓ ≥ 0} 12: shared ← |valid| 13: if shared > 0 then 14: opposing ← |{𝑝ℓ ∈ valid : 𝑣 𝑖ℓ ≠ 𝑣 𝑘ℓ }| 𝑗 15: 𝑑 𝑖 𝑘 ← opposing/shared 16: else 𝑗 17: 𝑑𝑖 𝑘 ← 1 18: end if 𝑗 𝑗 19: 𝑑 𝑘𝑖 ← 𝑑 𝑖 𝑘 20: end for 21: end for 22: 𝑅[𝑝 𝑗 ] ← {addresses : 𝐴 𝑗 , matrix : 𝐷 𝑗 } 23: end for 24: return 𝑅 25: end function 6: 7:
⊲ Sliding window ⊲ Active addresses ⊲ Filtered voter submatrix ⊲ Initialise dissimilarity matrix
⊲ No shared voting history ⊲ Symmetric assignment
Figure 4: A sample from the dissimilarity matrix for Nouns DAO Proposal 300. The rows are the addresses and the columns are the same addresses (abbreviated). A score of 1 represents complete dissimilarity; a score of 0 represents complete similarity.
7
Figure 5: MDS visualisation for Nouns DAO Proposal 334 with ground-truth colouring. Each node represents a voter address, with positions determined by pairwise dissimilarity scores. Red nodes indicate addresses that participated in a subsequent fork, while blue nodes represent addresses that remained in the original DAO.
3.6
Spatial Visualisation via MDS
To visualise the ideological alignment captured by our dissimilarity matrices, we use multidimensional scaling (MDS) to project high-dimensional voting relationships into 2D space. MDS preserves pairwise distances between data points while reducing dimensionality, making it suitable for revealing spatial patterns in voter alignment. For each proposal 𝑝 𝑗 with dissimilarity matrix 𝐷 𝑗 , we apply classical MDS to obtain a 2D embedding 𝑗
X 𝑗 ∈ R𝑛×2 where each row x𝑖 represents the coordinates of address 𝑎 𝑖 in the visualisation space. The MDS algorithm minimises the stress function:
v u tÍ Stress =
𝑗 𝑗 𝑗 2 𝑖<𝑘 (𝑑 𝑖𝑘 − ∥x 𝑖 − x 𝑘 ∥2 ) Í 𝑗 2 𝑖<𝑘 (𝑑 𝑖𝑘 )
where ∥·∥2 denotes the Euclidean distance in the 2D embedding space. This ensures that addresses with low dissimilarity scores appear close together in the visualisation, while highly dissimilar addresses are positioned far apart. To maintain temporal continuity across proposals, we use the embedding coordinates from proposal 𝑝 𝑗−1 as initialisation for the MDS algorithm at proposal 𝑝 𝑗 . This prevents arbitrary rotations and reflections between consecutive visualisations, enabling coherent tracking of voter movement over time. We performed MDS using the scikit-learn library with the following parameters: metric MDS with Euclidean distance, maximum of 300 iterations, and convergence tolerance of 10−6 . The embedding positions X 𝑗 were stored for subsequent clustering analysis (see Section 3.7). Nouns DAO provides a natural ground truth for validating our visualisation approach. The DAO has undergone three forks since deployment, each initiated when a subset of NFT holders (Noun owners) signalled their intent to establish a new DAO instance via on-chain transactions. We identified fork participants by
8
Figure 6: MDS visualisation for Nouns DAO Proposal 334 with k-means clustering. Note the overlap with the ground-truth clusters in Fig. 5.
analysing NFT transfers to the new fork smart contracts, creating a binary classification of addresses as either “forkers” or “stayers”. Figure 5 shows the MDS visualisation for Proposal 334, where red nodes represent addresses that participated in a subsequent fork and blue nodes represent those that remained. The spatial clustering of the nodes shows that our dissimilarity-based approach captures ideological divisions that manifest in real-world governance actions. This ground truth shows that MDS translates abstract voting patterns into interpretable spatial relationships that can be used for automated partisan community detection.
3.7
Dynamic Cluster Analysis
The final stage of our method applies clustering algorithms to the MDS-projected voter positions to identify distinct partisan communities within the DAO. This transforms the spatial patterns visible in our visualisations into quantifiable community structures that can be analysed and validated. For each proposal 𝑝 𝑗 , we perform k-means clustering on the 2D coordinates X 𝑗 obtained using MDS (see Section 3.6). Since the number of active voters varies across proposals and partisan structures evolve over time, we dynamically determine the optimal number of clusters rather than use a fixed value. We use the silhouette score to select the optimal cluster count 𝑘 ∗ for each proposal. The silhouette score 𝑠(𝑖) for voter 𝑖 is calculated as: 𝑠(𝑖) =
𝑏(𝑖) − 𝑎(𝑖) max{𝑎(𝑖), 𝑏(𝑖)}
where 𝑎(𝑖) is the average distance from voter 𝑖 to all other voters in the same cluster, and 𝑏(𝑖) is the average distance from voter 𝑖 to all voters in the nearest neighbouring cluster. The overall silhouette score for a clustering configuration is the mean of individual silhouette scores across all voters. We evaluate cluster counts from 𝑘 = 2 to 𝑘 = 5 inclusive and select the configuration with the highest silhouette score, as shown in Fig. 7. This range reflects our belief that partisan structures in DAOs manifest as a small number of ideological groups rather than highly fragmented communities. In supplementary tests up to
9
Figure 7: Silhouette coefficient scores for cluster counts 𝑘 = 2 to 𝑘 = 5 on Nouns DAO’s Proposal 334. 𝑘 = 10, 22 % of clusterings exceeded five groups overall, but fewer than 3 % did so after Proposal 319. Hence, restricting to 𝑘 ≤ 5 has a negligible impact on detecting partisan structures. As a preliminary validation, we can use the ground truth provided by Nouns DAO’s documented forks. Figure 6 shows the clustering results for Proposal 334, where 𝑘 ∗ = 2. The clusters closely match the ground truth division shown in Fig. 5, with forking addresses predominantly clustered together and distinct from addresses that remained in the original DAO.
4
Results
Our analysis of Nouns DAO governance data spans from the deployment of its governance contract at block 12,985,453 (8th August 2021) through Proposal 362 at block 18,050,498 (2nd September 2023). This time-frame captures the complete pre-fork governance history, providing an insight into how partisan communities emerged and solidified before the DAO’s fragmentation.
4.1
Visualisation of Community Structure
The MDS visualisation in Fig. 5 shows a spatial representation for Proposal 334, which we selected for detailed analysis due to its high concentration of participating fork addresses. The graph maps 2D positions of governing addresses based on their voting alignment, where proximity indicates similar voting behaviour and distance reflects ideological divergence. A clear bimodal structure emerges, with two distinct clusters separated along the horizontal axis. The rightmost cluster contains addresses that subsequently participated in one of Nouns DAO’s three forks (represented in red), while the leftmost cluster contains addresses that remained in the original DAO (shown in blue). This spatial segregation shows that our dissimilarity-based approach captures the underlying ideological divisions.
4.2
Quantitative Clustering Analysis
Applying k-means clustering with silhouette score optimisation to the MDS coordinates confirms our visual observations. For Proposal 334, the algorithm selected 𝑘 ∗ = 2 as the optimal number of clusters, aligning with the
10
Figure 8: A stacked area chart showing the assignment of forked addresses in the Nouns DAO to k-means clusters across Proposals 257 to 362. The addresses gradually coalesce into a single cluster as the fork approached. binary fork/no-fork ground truth. Of the 15 fork addresses, 14 fell into the same cluster; one was misclassified into the other cluster. The clustering algorithm, operating solely on geometric proximity without knowledge of future fork participation, reconstructed the partisan divide that would later lead to the DAO’s fragmentation.
4.3
Temporal Evolution of Partisan Behaviour
Our analysis shows a shift in fork addresses’ governance engagement over time. During the early period (Proposals 1 to 256), fork addresses participated sporadically, with an average of only 1.86 addresses voting per proposal. Beginning with Proposal 257 (22nd March 2023), their participation increased. In the final 91 non-cancelled proposals (Proposals 257 to 362, of which 15 were cancelled), at least four fork addresses participated in every vote, with average participation jumping to 8.27 addresses per proposal. This suggests that addresses destined to fork the DAO increased their governance involvement as ideological tensions mounted. While non-forking addresses also increased their participation (from 28.86 to 39.93 average participants), the proportional increase was much smaller, indicating that the impending fork motivated greater engagement among dissenting community members.
4.4
Consolidation into Partisan Communities
Figure 8 shows the temporal consolidation of fork addresses into a unified partisan community. The stacked area chart tracks cluster assignments across Proposals 257 to 362, revealing how initially fragmented fork addresses gradually coalesced into a single, dominant cluster as the DAO approached its breaking point. This process demonstrates that partisan communities do not emerge suddenly but rather form through sustained governance participation. The visual evidence shows fork addresses becoming increasingly unified in their cluster assignment, suggesting that ideological alignment strengthened over time through repeated interaction and voting.
4.5
Validation Against Randomised Data
Table 1 compares genuine voting data with 100 randomised iterations in which participants and outcomes were preserved but vote assignments shuffled (seeds 0 to 99), isolating the effect of genuine alignment from random behaviour. The results are as follows. Firstly, genuine voter data produces fewer clusters per proposal (2.77 average) compared to randomised data (3.36 average), indicating that genuine voting behaviour exhibits more coherent community structure than would occur by chance. This gap widens as we focus on periods closer to the fork, with the final 14 proposals showing just 2.08 clusters on average versus 3.87 in randomised data. Secondly, the percentage of fork addresses concentrated within a single cluster shows differences between genuine and randomised data. Across all proposals, 72.20 % of fork addresses cluster together in genuine data, compared to only 62.64 % in randomised data. This differential becomes even more pronounced in the pre-fork
11
Table 1: Average number of clusters and percentage of fork addresses in a single cluster, comparing genuine voting data against 100 randomised iterations across four proposal ranges. Range
Metric
Value
Rand. Min.
Rand. Max.
Rand. Avg.
Proposals 1 to 362
Avg # Clusters % in Forked Cluster
2.77 72.20
3.14 57.70
3.57 67.64
3.36 62.64
Proposals 257 to 362
Avg # Clusters % in Forked Cluster
2.48 79.36
3.23 44.60
4.05 59.31
3.66 50.40
Proposals 319 to 362
Avg # Clusters % in Forked Cluster
2.14 90.96
3.06 41.66
4.14 59.69
3.62 47.36
Proposals 349 to 362
Avg # Clusters % in Forked Cluster
2.08 90.28
2.50 34.82
4.67 63.24
3.87 46.99
period: in the final 44 proposals, 90.96 % of fork addresses cluster together in genuine data versus just 47.36 % in randomised data. Our results demonstrate that partisan communities destined to fragment a DAO can be detected before an actual fork occurs. The clustering analysis shows that addresses ultimately participating in forks were already exhibiting coordinated voting behaviour months before their decision to leave was enacted. DAO governance systems could therefore implement early warning mechanisms to identify emerging partisan divisions before they result in organisational fragmentation.
5
Conclusions
DAOs enable open, pseudonymous participation in significant governance decisions, yet this openness can produce partisan divisions that are often difficult to identify but can culminate in disruptive outcomes such as forks. We proposed a method to map such divisions by analysing on-chain governance data. It starts with a voter matrix representing a DAO’s governance activity across its proposals and voter addresses. We measure the overall friction within the community based on the outcomes of governance proposals. We explore this friction at the address level, filtering the data to include active voting participants within the DAO, conducting a pairwise dissimilarity analysis on voters within a proposal window, and visualising the alignment using multidimensional scaling. Once projected, we identify clusters using 𝑘-means clustering and select the optimal configuration using silhouette scoring. This framework reveals the dynamics of community formation within DAOs. We use Nouns DAO as a case study, as its documented forks provide a ground truth for validating our method. We show that addresses that later participated in a fork are increasingly grouped into a single cluster, with over 90 % aligned just before the split. Moreover, the average number of clusters falls to two distinct groups by this stage, indicating two partisan communities. We validate these results by randomising voter assignments while preserving participation and outcomes; in 100 iterations of randomised voter assignments 47 % of forked addresses clustered together and the number of average clusters rose above four. Our findings indicate that partisan communities can be detected before forks occur. If left undetected and unresolved, such divisions can fragment a DAO and negatively impact its stability. Our method provides a means to monitor these dynamics, offering a foundation for early-warning systems. In future work, we will extend our analysis to other DAOs that have undergone a fork.
References [1] A. Amelio and C. Pizzuti. Mining and analyzing the Italian parliament: Party structure and evolution. In Recommendation and Search in Social Networks, pages 249–279. Springer, 2015. ISBN 9783319364803. doi: 10.1007/978-3-319-14379-8_12. [2] J. Austgen, A. Fábrega, S. Allen, K. Babel, M. Kelkar, and A. Juels. DAO decentralization: Voting-bloc entropy, bribery, and Dark DAOs. 3 https://arxiv.org/abs/2311.03530 , 2023. [3] H. Axelsen, J. R. Jensen, and O. Ross. When is a DAO decentralized? Complex Systems Informatics and Modeling Quarterly, 176(31):51–75, 2022. doi: 10.7250/csimq.2022-31.04.
12
[4] C. Bellavitis, C. Fisch, and P. P. Momtaz. The rise of decentralized autonomous organizations (DAOs): A first empirical glimpse. Venture Capital: An International Journal of Entrepreneurial Finance, 25(2):187–203, 2023. [5] Q. DuPont. New online communities: Graph deep learning on anonymous voting networks to identify Sybils in polycentric governance. 3 https://arxiv.org/abs/2311.17929 , 2023. [6] R. Feichtinger, R. Fritsch, Y. Vonlanthen, and R. Wattenhofer. The hidden shortcomings of (D)AOs —- An empirical study of on-chain governance. In Financial Cryptography and Data Security (FC), 2023. [7] R. Feichtinger, R. Fritsch, L. Heimbach, Y. Vonlanthen, and R. Wattenhofer. SoK: Attacks on DAOs. 3 https://arxiv.org/abs/2406.15071 , 2024. [8] R. Fritsch, M. Müller, and R. Wattenhofer. Analyzing voting power in decentralized governance: Who controls DAOs? Blockchain: Research and Applications, 5(3):100208, 2024. [9] T. Lloyd, D. Ó Broin, and M. Harrigan. The on-chain and off-chain mechanisms of DAO-to-DAO voting. In The IEEE International Conference on Blockchain (BLOCKCHAIN), pages 649–655. IEEE, 2024. doi: 10.1109/BLOCKCHAIN62396.2024.00095. [10] T. Lloyd, D. Ó Broin, and M. Harrigan. Mapping partisan fault lines within DAOs. In The IEEE International Conference on Blockchain (BLOCKCHAIN), pages 38–45. IEEE, 2025. doi: 10.1109/BLOCKCHAIN67634.2025. 00015. [11] V. Mickevicius, T. Krilavicius, and V. Morkevicius. Analysing voting behavior of the Lithuanian parliament using cluster analysis and multidimensional scaling: Technical aspects. In The International Conference on Electrical and Control Technologies (ECT), pages 84–89, 2014. [12] Y. Quan, X. Wu, W. Deng, and L. Zhang. Decoding social sentiment in DAOs: A comparative analysis of blockchain governance communities. In The IEEE International Conference on Software Quality, Reliability, and Security Companion (QRS-C), pages 216–224, 2024. [13] W. Reijers, I. Wuisman, M. Mannan, P. D. Filippi, C. Wray, V. Rae-Looi, A. C. Vélez, and L. Orgad. Now the code runs itself: On-chain and off-chain governance of blockchain technologies. Topoi: An International Review of Philosophy, 2018. doi: 10.1007/s11245-018-9626-5. [14] Sky (formerly MakerDAO). 3 https://sky.money/ , n.d. [15] X. Sun, X. Chen, C. Stasinakis, and G. Sermpinis. Voter coalitions and democracy in decentralized finance: Evidence from MakerDAO. 3 https://arxiv.org/abs/2210.11203 , 2022.
13