Welcome

EBA addresses a long-standing issue that has hindered the ML community from developing applications for Bitcoin: a lack of ML-first data.

The Challenge

The paradox of publicly accessible blockchain data is that its cryptography-first design, a necessity for a decentralized ledger, has impeded its adoption by the ML community. Accordingly, blockchains use models such as the Unspent Transaction Output (UTxO) that do not readily provide information from which an ML model can learn the topology of financial flows for predictive tasks like market forecasting or classification tasks like anomaly detection.

Specifically, the friction for ML from this cryptography-first design is twofold. First, UTxO-like models provide only a static snapshot of financial flows (i.e., the transaction), and reconstructing the actual flow of how funds are earned and spent requires additional traversal of the blockchain. Arguably, this flow provides more information for ML applications than the snapshot. Second, unlike the common perception of a transaction where "$x$ sends $Z$ amount to $y$", the UTxO model encodes it as a complex process where "cryptographic proof of ownership (from the set $\{x\}$) unlocks a set of previously unspent atomic currency units, and this sum ($Z$) is then split and re-locked under a new cryptographic locking mechanism where only the intended recipients can unlock (the set $\{y\}$)". This latter description is what is publicly recorded on the blockchain, and it is not ML-friendly. EBA addresses both of these issues.

The Solution

EBA interfaces with the Bitcoin network and creates a graph of the full history of transactions recorded on-chain. On this graph, the nodes are Bitcoin scripts (aka addresses), and the edges between them represent transactions recorded on-chain via the UTxO model. Simply put, the graph represents a transaction between $x$ and $y$ as a time-stamped, directed edge between their corresponding nodes. Consequently, the flow of how funds are earned and spent can be traced by traversing these paths. This graph is built for ML, Graph Neural Networks (GNNs) in particular, allowing a graph-based model to aggregate information from neighbors through message passing and learn the topology of fund flows for various applications.

The Bitcoin Graph that EBA creates encompasses the complete trading details of over 8.72 billion BTC, and it consists of over 2.4 billion nodes and 39.72 billion time-stamped edges spanning more than a decade, making it a complete resource for developing models on Bitcoin and a large-scale resource for benchmarking graph neural networks.

We share the complete ETL pipeline and all the data it generates. To simplify working with the pipeline and its resources, we have split them into separate repositories. The following is the list of resources we provide for Bitcoin:

• Datasets:
  • Dataset Releases
  • Documentation
  • Bitcoin Graph in neo4j format
  • Block Statistics
• Tools:
  • Extract, Transform, and Load (ETL) pipeline
  • Community Sampling
• ML:
  • Quick-Start Example model: Trains a model on sampled communities that learns to generate node embeddings by aggregating neighborhood features through message passing, then uses these embeddings for script node clustering.
  • External and off-chain resources

The Application

This graph is built for machine learning, particularly Graph Neural Networks (GNNs), allowing a graph-based model to aggregate information from neighbors through message passing and learn the topology of fund flows for various applications across the vibrant cryptocurrency ecosystem, including:

• Exploring economic evolution and temporal behaviors.
• Analyzing network dynamics and trading patterns.
• Identifying suspicious or illicit activities.
• Benchmarking large-scale, graph-based machine learning models.