Relation Discoverer (RD) Overview

• Relation Discoverer (RD) is a class of modules that infer relational structures from raw data across multiple domains, outputting labels, graph edges, or scene embeddings.
• RD employs techniques like clustering, prompt-based LLM generation, and active learning to distinguish between known and novel relations under realistic test mixtures.
• Applications of RD span open-world extraction, metadata graph building, visual dialogue, and geospatial link discovery, demonstrating scalability and improved performance benchmarks.

Relation Discoverer (RD) denotes a class of modules that infer relational structure from partially observed data, but the term is not attached to a single standardized architecture. In the cited literature, it refers to open-world relation extraction systems that separate known and novel relations under realistic test mixtures, a large-language-model component that generates previously unseen relation names, a metadata-driven graph builder in distributed object storage, an object-relationship module for visual dialogue, and a geospatial topological link-discovery pipeline (Hogan et al., 2023, Tu et al., 18 Sep 2025, Li et al., 2022, Primmer et al., 2013, Wang et al., 2020, Sherif et al., 2016). The common abstraction is the conversion of raw instances into relations, edges, or labels; the major differences lie in what counts as an instance, how relation evidence is represented, and whether the output is a class label, a graph edge, a scene graph embedding, or a topological link set.

1. Scope and recurring abstractions

Across the cited work, RD appears in at least five technically distinct settings. In open-world relation extraction, it operates on sentences with marked entity pairs and must distinguish known from novel relations, often under long-tail class imbalance and the presence of no-relation instances (Hogan et al., 2023, Li et al., 2022, Tu et al., 18 Sep 2025). In storage systems, RD operates on unstructured blobs and metadata, producing a weighted graph over stored objects (Primmer et al., 2013). In visual dialogue, the analogous component preserves object-level interactions by embedding a scene graph rather than collapsing an image into a single global vector (Wang et al., 2020). In geospatial RDF, relation discovery is framed as link generation subject to exact topological predicates, with indexing and filtering used to avoid exhaustive geometry comparisons (Sherif et al., 2016).

Setting	Input representation	Output
Open-world RE, KNoRD (Hogan et al., 2023)	Labeled and unlabeled sentences with entity pairs	Known/novel relation assignments
LLM-OREF RD (Tu et al., 18 Sep 2025)	Test instances plus in-context demonstrations	Generated new relation labels
Active Relation Discovery (Li et al., 2022)	Known-relation training data and mixed test pool	Known/novel detection and label-aware novel classification
RDOS RD (Primmer et al., 2013)	Objects and metadata key–value pairs	Weighted graph edges
ORD (Wang et al., 2020)	Object regions and scene graph	Relationship-aware graph embeddings
Radon (Sherif et al., 2016)	Source and target geometries	Topological links

A recurrent misconception is to treat relation discovery as synonymous with clustering all unlabeled data into unseen classes. The OpenRE papers explicitly reject that simplification. One paper argues that existing approaches “oversimplify the problem by assuming that all unlabeled texts belong to novel classes,” while another shows that performance declines sharply when test data include seen classes (Hogan et al., 2023, Li et al., 2022). A second misconception is that RD always yields human-readable labels automatically. That is true for the LLM-based variant, which directly generates relation names, but not for clustering-only approaches that require a subsequent mapping from clusters to labels (Tu et al., 18 Sep 2025).

2. Semi-supervised generalized relation discovery in KNoRD

In "Open-world Semi-supervised Generalized Relation Discovery Aligned in a Real-world Setting" (Hogan et al., 2023), the unified RD formulation corresponds to KNoRD. The setting is transductive. There is a labeled set $\mathcal{D}_l=\{(x_i,y_i)\}$ with $y_i\in\mathcal{C}^K$ and an unlabeled set $\mathcal{D}_u=\{x_j\}$ containing instances of the same known classes $\mathcal{C}^K$, instances of novel classes $\mathcal{C}^N$, and negative instances. The class sets satisfy $\mathcal{C}^K\cap\mathcal{C}^N=\emptyset$, and the method does not assume prior knowledge of $N=|\mathcal{C}^N|$. The paper states that, in realistic data, novel classes follow a long-tail distribution.

The architecture combines a sentence encoder, prompt-based explicit/implicit relation representations, and a prototype/clustering module. A transformer encoder such as BERT or DeBERTa maps a sentence to contextual token states, and the entity-pair representation is obtained by wrapping the head and tail with special tokens, mean-pooling over each span, and concatenating the resulting vectors into $\mathbf{r}\in\mathbb{R}^{2d}$ (Hogan et al., 2023). In parallel, the model trains a prompt-based masked LLM on labeled data in two modes. Constrained prediction restricts the $[MASK]$ token to words occurring in the sentence and is intended to capture explicit triggers such as “born” and “founded.” Unconstrained prediction allows the full vocabulary and is intended to capture implicit cues such as dates, patterns, or world-knowledge words. The top-3 predicted tokens from each mode are embedded through a fixed phrase-embedding model and concatenated into a joint vector $\mathbf{p}\in\mathbb{R}^{2D}$ for clustering.

All prompt-based representations from $y_i\in\mathcal{C}^K$0 and $y_i\in\mathcal{C}^K$1 are fed into a Gaussian Mixture Model with $y_i\in\mathcal{C}^K$2 components. Clustering is augmented by entity-type consistency through reassignment of low-confidence points to the nearest cluster with the same entity type. The supervised objective on labeled data is standard cross-entropy over the $y_i\in\mathcal{C}^K$3 known classes, later extended to include discovered novel classes. The formulation also allows an optional InfoNCE loss on the prompt embeddings to pull same-cluster points together and push different clusters apart. After clustering, each unlabeled instance is assigned to its highest-probability GMM component, and within each cluster the top-$y_i\in\mathcal{C}^K$4 members by posterior probability are selected as high-quality pseudo-labels. These pseudo-labels are incorporated through a cross-entropy term, together with entropy minimization on the entire unlabeled set.

The discovery loop proceeds by encoding all instances, fitting the GMM, marking clusters as known or novel by majority vote over labeled points, adding unlabeled instances from novel clusters to the pseudo-label pool, selecting the top-$y_i\in\mathcal{C}^K$5 in each novel cluster, retraining the classifier on $y_i\in\mathcal{C}^K$6 together with the selected pseudo-labeled set, and optionally repeating the procedure (Hogan et al., 2023). The paper also states that steps (1–3) can be viewed as prototype initialization and steps (4–6) as prototype refinement and classifier retraining.

Experimentally, KNoRD is evaluated on TACRED, ReTACRED, and FewRel. For TACRED and ReTACRED, the original 42 relations are ranked by frequency, with the top 21 treated as known and the bottom 21 as novel; FewRel uses a similar 50/50 split augmented with negative examples from ReTACRED. The reported metric is micro-$y_i\in\mathcal{C}^K$7 on all, on known, and on novel classes. The baselines include RoCORE, MatchPrompt, TABs with and without pre-confidence splitting, ORCA adapted to text, and GPT-3.5 with in-context prompts. KNoRD achieves the best overall micro-$y_i\in\mathcal{C}^K$8 on all three datasets and improves especially on long-tail novel classes by 10–30 points over the nearest baseline. Ablations show that removing the cross-entropy fine-tuning collapses performance, and using only constrained or only unconstrained prompt embeddings reduces novel-class $y_i\in\mathcal{C}^K$9 by 5–10 points. The best pseudo-label trade-off is obtained by selecting the top 15% of cluster members by GMM posterior. Constrained predictions excel on rare relations, unconstrained predictions on common ones, and the combined representation is strongest (Hogan et al., 2023).

3. LLM-based Relation Discoverer in LLM-OREF

In "LLM-OREF: An Open Relation Extraction Framework Based on LLMs" (Tu et al., 18 Sep 2025), RD is a front-end component that predicts previously unseen relations directly, without human intervention. The input consists of a test set $\mathcal{D}_u=\{x_j\}$0, where each $\mathcal{D}_u=\{x_j\}$1, and a training set $\mathcal{D}_u=\{x_j\}$2 whose labels come from a known relation set $\mathcal{D}_u=\{x_j\}$3. The new test relations $\mathcal{D}_u=\{x_j\}$4 are disjoint from $\mathcal{D}_u=\{x_j\}$5. RD aims to produce

$\mathcal{D}_u=\{x_j\}$6

where $\mathcal{D}_u=\{x_j\}$7 is a small in-context demonstration set sampled from $\mathcal{D}_u=\{x_j\}$8 and guaranteed not to contain the target relation of $\mathcal{D}_u=\{x_j\}$9.

The base model is a single open-source LLM, such as LLaMA-2 or Qwen2.5, fine-tuned under LoRA. The paper’s design motivation is that clustering methods yield unaligned clusters and require human labeling, whereas an LLM can directly generate a natural-language relation name if shown relation examples in context. It also reports a preliminary contrast: zero-shot $\mathcal{C}^K$0 on FewRel is approximately 40%, but rises to approximately 80% when demonstrations include examples of the target relation. RD is trained with a fine-tuning objective that augments standard cross-entropy with a distillation term from a stronger Relation Predictor teacher, intended to bridge the gap between discovery without target examples and prediction with target examples (Tu et al., 18 Sep 2025).

At inference time, RD samples $\mathcal{C}^K$1 distinct known relations from $\mathcal{C}^K$2, with $\mathcal{C}^K$3 in the experiments, and selects one example for each relation to construct $\mathcal{C}^K$4. It then forms an input $\mathcal{C}^K$5 and autoregressively decodes a relation string. To boost recall, this procedure is repeated $\mathcal{C}^K$6 times with different random demonstrations, with $\mathcal{C}^K$7 by default, yielding a bag of preliminary predictions $\mathcal{C}^K$8 (Tu et al., 18 Sep 2025). The paper notes that one can view each generation as assigning a log-score to each possible relation label and taking a one-best decision, but in practice it collects the top-1 output from each run.

RD is only the first stage of a self-correcting inference pipeline. In the denoising stage, the union of discovered relation strings becomes a candidate set $\mathcal{C}^K$9. For each instance and candidate relation, the Relation Predictor constructs $\mathcal{C}^N$0 demonstrations that contain one example of the candidate relation and $\mathcal{C}^N$1 examples of other members of $\mathcal{C}^N$2, then checks whether its prediction agrees across all $\mathcal{C}^N$3 runs. Pairs that survive this cross-validation are marked high-reliability. This denoising stage is repeated for $\mathcal{C}^N$4 iterations to collect reliable test instances for each new relation. In the final prediction stage, RP treats each candidate relation as a pseudo-class, forms new demonstrations from reliable examples, and chooses the best relation for every test instance by comparing scores over candidate sets (Tu et al., 18 Sep 2025).

The paper reports RD’s stand-alone contribution explicitly. On FewRel, RD with $\mathcal{C}^N$5 attains Precision $\mathcal{C}^N$6, Recall $\mathcal{C}^N$7, Macro-$\mathcal{C}^N$8, and Accuracy $\mathcal{C}^N$9. RD(Pass@3) reaches 78.8%, meaning that the correct relation appears among the top-3 RD outputs. Removing RD entirely collapses performance to $\mathcal{C}^K\cap\mathcal{C}^N=\emptyset$0-$\mathcal{C}^K\cap\mathcal{C}^N=\emptyset$1 on FewRel, compared with 0.672 for the full system, and to classification $\mathcal{C}^K\cap\mathcal{C}^N=\emptyset$2, compared with 0.718 for the full system. The full LLM-OREF system improves final macro-$\mathcal{C}^K\cap\mathcal{C}^N=\emptyset$3 by approximately 10–20 points over RD($\mathcal{C}^K\cap\mathcal{C}^N=\emptyset$4) on all datasets, indicating that denoising and prediction contribute substantially beyond raw discovery (Tu et al., 18 Sep 2025).

4. Active Relation Discovery and label-aware OpenRE

"Active Relation Discovery: Towards General and Label-aware Open Relation Extraction" (Li et al., 2022) treats relation discovery as a two-part problem: first detect whether a test instance belongs to a known or novel relation, then assign human-readable labels to the novel relations with minimal annotation. The supervised corpus is

$\mathcal{C}^K\cap\mathcal{C}^N=\emptyset$5

while the realistic test set is a mixture of known and novel relations:

$\mathcal{C}^K\cap\mathcal{C}^N=\emptyset$6

Under General OpenRE, an instance is labeled novel when the maximum known-relation score falls below a threshold $\mathcal{C}^K\cap\mathcal{C}^N=\emptyset$7. Supervised pre-training minimizes the usual cross-entropy over known classes (Li et al., 2022).

The representation module inserts entity markers around the head and tail and uses BERT or another PLM to encode the sentence. The relation representation is the concatenation of the contextual embeddings at the start markers of the head and tail entities, $\mathcal{C}^K\cap\mathcal{C}^N=\emptyset$8. Pre-training adds a supervised contrastive loss with a one-layer projection to $\mathcal{C}^K\cap\mathcal{C}^N=\emptyset$9, encouraging tighter within-class clustering. The paper’s central assumption is that known relations form dense clusters in feature space and novel relations tend to lie outside them.

Novelty detection is implemented by the Local Outlier Factor. Using Euclidean distance, the method defines the $N=|\mathcal{C}^N|$0-nearest-neighbor distance, reachability distance, local density, and then the LOF score $N=|\mathcal{C}^N|$1. An instance is declared novel if $N=|\mathcal{C}^N|$2 (Li et al., 2022). This module directly addresses the failure mode of prior clustering-based methods under mixed known/novel test sets.

Once the model has separated $N=|\mathcal{C}^N|$3 into known and novel subsets, ARD enters a relational active learning loop. A labeled seed set $N=|\mathcal{C}^N|$4 and an unlabeled remainder $N=|\mathcal{C}^N|$5 are maintained. An adversarial encoder–discriminator pair is trained so that the discriminator distinguishes labeled from unlabeled novel instances, while the encoder and discriminator are jointly updated through $N=|\mathcal{C}^N|$6. At each iteration, the system selects the top-$N=|\mathcal{C}^N|$7 unlabeled instances with the highest discriminator score, asks a human annotator whether each shares a relation type with any instance in $N=|\mathcal{C}^N|$8, assigns either an existing novel label or a new one, adds the queried instances to $N=|\mathcal{C}^N|$9, and retrains a standard MLP classifier with cross-entropy on the updated labeled set (Li et al., 2022).

The evaluation distinguishes conventional OpenRE, where test data contain only unseen relations, from General OpenRE settings labeled Noi and Imb, where known and novel relations are mixed and, in Imb, novel-class frequencies are rebalanced. On FewRel, ARD reaches instance-level $\mathbf{r}\in\mathbb{R}^{2d}$0 values of 73.6 in Ori, 80.8 in Noi, and 76.5 in Imb, outperforming RSN-CV, SelfORE, and OHRE in all three settings. The paper reports that removing LOF and treating all test data as novel reduces FewRel (Noi) $\mathbf{r}\in\mathbb{R}^{2d}$1 from 80.8 to 73.4 after 5 query epochs. It also states that with only 288 human-labeled queries out of approximately 11,200 novel instances, ARD matches or exceeds random labeling of 30% of the novel pool, while alternative active-learning strategies underperform the discriminator-based selection by 4–15 $\mathbf{r}\in\mathbb{R}^{2d}$2 points (Li et al., 2022). This suggests that, in label-aware OpenRE, RD can be decomposed into an anomaly detector plus an annotation-efficient labeling mechanism rather than a single monolithic classifier.

5. Relation Discoverer in relational distributed object storage

In "Creating a Relational Distributed Object Store" (Primmer et al., 2013), RD is a storage-system component rather than a text classifier. The object universe is $\mathbf{r}\in\mathbb{R}^{2d}$3, each object $\mathbf{r}\in\mathbb{R}^{2d}$4 has metadata $\mathbf{r}\in\mathbb{R}^{2d}$5, and discovered relations are represented as a weighted graph $\mathbf{r}\in\mathbb{R}^{2d}$6 with $\mathbf{r}\in\mathbb{R}^{2d}$7. The edge set is

$\mathbf{r}\in\mathbb{R}^{2d}$8

where $\mathbf{r}\in\mathbb{R}^{2d}$9 is a relation-scoring function and $[MASK]$0 is a tunable threshold. Explicitly declared relations receive weight $[MASK]$1. Implicit relations are derived from metadata similarity using

$[MASK]$2

RD depends on a Metadata Generation Module pipeline. Each namespace defines an ordered pipeline of $[MASK]$3 MGM modules, each acting as a data dictionary or script. On ingest or retroactive pass, each MGM examines the object and its accumulated metadata and produces an updated metadata set. The complexity per object is $[MASK]$4 (Primmer et al., 2013). After metadata population, RD discovers two types of relations: explicit edges from metadata tags of the form $[MASK]$5, and implicit edges via metadata overlap. Pairwise similarity search is implemented through a distributed inverted metadata index mapping each key–value pair to a posting list of URIs. For each metadata key, RD scans the posting list, increments pair counts, computes intersection and union sizes, derives the score, and upserts edges into the graph database whenever the score exceeds the threshold.

The paper gives the dominant complexity of pair counting as

$[MASK]$6

with total index entries approximately $[MASK]$7 when $[MASK]$8 and each object has on average $[MASK]$9 tags. It states that in practice $\mathbf{p}\in\mathbb{R}^{2D}$0 and each key’s list is small, so the overall cost is near-linear in $\mathbf{p}\in\mathbb{R}^{2D}$1 (Primmer et al., 2013). Incremental maintenance updates only the affected posting lists and local graph neighborhoods when objects are added or modified, and edges are garbage-collected when metadata expire or objects are deleted.

The distributed architecture includes a public front-end for protocols such as HTTP, S3, CIFS, and NFS, and a private back-end using gossip for membership and consistent hashing for placement. Metadata indexes and the object-reference database are partitioned and replicated synchronously within a cluster and asynchronously across clusters. RD itself runs as a distributed batch or stream job whose workers process posting-list partitions or update queues and write to graph shards (Primmer et al., 2013).

The reported performance targets petabyte-scale deployments with $\mathbf{p}\in\mathbb{R}^{2D}$2–$\mathbf{p}\in\mathbb{R}^{2D}$3 objects. Typical figures include per-node object capacity greater than $\mathbf{p}\in\mathbb{R}^{2D}$4 objects, per-cluster capacity greater than $\mathbf{p}\in\mathbb{R}^{2D}$5 objects, MGM throughput of 5–20 K objects/sec per node, relation-discovery update of approximately 1.5 M edge-computations/sec per worker node, graph-database write latency below 5 ms per edge insert or update, single-hop read latency below 2 ms, and end-to-end ingest plus metadata plus relation latency of 50–200 ms (Primmer et al., 2013). In this usage, RD is a persistent graph-construction and maintenance service rather than a predictive model over labels.

6. Object relationship discovery for visual dialogue

In "ORD: Object Relationship Discovery for Visual Dialogue Generation" (Wang et al., 2020), the discovery target is a scene graph rather than a symbolic relation label. At dialogue round $\mathbf{p}\in\mathbb{R}^{2D}$6, the inputs are an image $\mathbf{p}\in\mathbb{R}^{2D}$7, dialogue history $\mathbf{p}\in\mathbb{R}^{2D}$8, and a new question $\mathbf{p}\in\mathbb{R}^{2D}$9; the output is a ranked list over 100 candidate answers. Object regions are detected by Faster-R-CNN, producing features $y_i\in\mathcal{C}^K$00, and a scene graph $y_i\in\mathcal{C}^K$01 is built with object nodes and directed predicate-typed edges of the form $y_i\in\mathcal{C}^K$02 (Wang et al., 2020). The paper argues that plain CNN encodings and even co-attention over grid or region features do not explicitly model predicate-typed edges and longer-range interactions among objects.

ORD addresses this limitation with a Hierarchical Graph Convolutional Network. The first stage is a local GCN that converts each directed edge into two undirected links through explicit predicate nodes, forming a graph with object nodes and predicate nodes. With initial embeddings in $y_i\in\mathcal{C}^K$03 and adjacency $y_i\in\mathcal{C}^K$04, local propagation is

$y_i\in\mathcal{C}^K$05

After discarding the predicate nodes, the refined object embeddings are passed through a global GCN over object-only adjacency $y_i\in\mathcal{C}^K$06:

$y_i\in\mathcal{C}^K$07

The result is a graph embedding matrix whose columns encode each object’s own visual region, local predicate context, and global interactions (Wang et al., 2020).

A graph-attention mechanism then conditions the object-node weights on the current question and attended dialogue history. For each node,

$y_i\in\mathcal{C}^K$08

followed by softmax normalization over nodes and formation of the attended graph context

$y_i\in\mathcal{C}^K$09

This context is fused with question, history, and region-attention features by concatenation and projection into a joint representation $y_i\in\mathcal{C}^K$10, which is used by either a discriminative decoder trained with N-pair loss or a generative decoder trained by maximum likelihood (Wang et al., 2020).

On VisDial v0.9, the full discriminative model ORD-D with scene graph, relation modeling, and graph attention improves MRR from 0.6222 to 0.6447 and recall@1 from 48.48% to 51.22% relative to HCIAE-D-NP-ATT. In the generative setting, ORD-G improves MRR from 0.5382 to 0.5502 and recall@1 from 44.07% to 45.63%. Ablation results show a cumulative trend from region features alone (MRR 0.6266), to undirected scene graph (0.6340), to directed edge features with local GCN (0.6383), to the full graph-attention model (0.6447) (Wang et al., 2020). Here, relation discovery functions as structural preservation and graph embedding for downstream multi-round reasoning.

7. Rapid discovery of topological relations in geospatial RDF

"Annex: Radon - Rapid Discovery of Topological Relations" (Sherif et al., 2016) addresses relation discovery for geospatial resources represented according to Linked Data principles. The task is to find all pairs $y_i\in\mathcal{C}^K$11 in two datasets $y_i\in\mathcal{C}^K$12 and $y_i\in\mathcal{C}^K$13 such that a spatial relation $y_i\in\mathcal{C}^K$14 holds. A naive solution requires $y_i\in\mathcal{C}^K$15 geometry comparisons, which the paper characterizes as intractable for millions of polygons. Radon replaces exhaustive comparison with a three-part pipeline: a swapping heuristic that indexes the smaller or “lighter” dataset first, a sparse tiling index built on minimum bounding boxes, and an MBB-based filter that prunes impossible relation checks before exact evaluation through a DE-9IM predicate (Sherif et al., 2016).

The mathematical substrate is the Dimensionally Extended nine-Intersection Model, whose $y_i\in\mathcal{C}^K$16 matrix records intersections among the interior, boundary, and exterior of two geometries. Common relations such as intersects, within, and covers are expressed through the matrix entries. For a geometry $y_i\in\mathcal{C}^K$17, its minimum bounding box $y_i\in\mathcal{C}^K$18 is the axis-aligned rectangle with minimal area containing it, with corners determined by the coordinate-wise minima and maxima. Radon exploits the implication that if $y_i\in\mathcal{C}^K$19, then $y_i\in\mathcal{C}^K$20 (Sherif et al., 2016).

The sparse tiling index divides the plane into a uniform grid of square tiles of side length $y_i\in\mathcal{C}^K$21, equivalently with granularity $y_i\in\mathcal{C}^K$22. A geometry’s MBB is mapped to all tiles whose integer coordinates fall between the floor and ceiling of the scaled bounding coordinates. Radon first builds an index for the source dataset and, as an optimization, inserts target geometries only into tiles already present in the source index. Candidate links are generated only from tiles shared by source and target geometries, and a cache prevents retesting the same pair. For relations such as within, covers, and equals, an auxiliary TestMBB predicate uses MBB inclusion to eliminate pairs that cannot satisfy the full topological predicate (Sherif et al., 2016).

The paper analyzes Radon’s runtime as

$y_i\in\mathcal{C}^K$23

where $y_i\in\mathcal{C}^K$24 and $y_i\in\mathcal{C}^K$25 are the average numbers of tiles per geometry and $y_i\in\mathcal{C}^K$26 is the number of candidate pairs. Space is dominated by the sparse indexes, with complexity $y_i\in\mathcal{C}^K$27 for $y_i\in\mathcal{C}^K$28 nonempty tiles. If the grid is chosen so that tile counts and candidate counts grow sublinearly, the behavior approaches linear time in practice (Sherif et al., 2016).

Empirically, on 308 small-scale experiments involving CLC subsets and NUTS over seven relations, Radon outperforms Silk by 65.6× on average, with peaks of 450× for within and a minimum of 4.7× for intersects. Against Strabon, it is 12× faster on average on small sets, although Strabon has a slight edge on pure intersects queries. On large-scale deduplication of $y_i\in\mathcal{C}^K$29 against itself with 2,209,538 geometries, Radon completes all tasks in under 95 minutes single-threaded and under 13 minutes with 8 threads, while Silk and Strabon do not finish within 2 hours. The paper reports near-linear speedup from 1 to 8 threads and gives the equals relation as an example, dropping from 24 minutes on 1 thread to 3.8 minutes on 8 threads (Sherif et al., 2016). In this line of work, RD is not a classifier at all, but an indexing-and-pruning strategy for exact topological link discovery.