A graph-based feature is a predictive input derived from relationships, node neighborhoods, network distance, or other measures of node closeness and similarity. In link prediction, these features help estimate which node pairs are likely to form future connections. section 10.2 section 10.3
Common design choices include:
- Use manually defined network metrics or learned representations such as node2vec embeddings, then compare embedding pairs with cosine similarity. section 10.3
- Calculate features only from a designated feature set of relationships; reusing train or test relationships can reveal the target and create data leakage. section 10.2 section 10.3
- Prefer a time-based split when relationship timestamps exist: generate features from earlier links, then train and test on later links. section 10.2.1
- Start with straightforward features and evaluate them before adopting more complex methods; transductive embeddings may require retraining when new nodes appear. section 10.3
Predicting future Twitch connections
For shared-audience relationships between Twitch streamers, calculate network features from an earlier relationship set and use them to recommend likely future connections. Because the dataset lacks relationship timestamps, another splitting method is required. section 10.2.1 section 10.3
