Unsupervised Latent Type Discovery in Relational Graphs

This direction is split into other directions: unsupervised latent type discovery. Its last version stays readable and citable here.

Version 1 of 172 papers90 conceptsfrozen 2026-10-04

In traditional knowledge representation, we define the world from the top down: first, we create an ontology (e.g., "City," "Athlete," "Chemical Compound"), and then we assign entities to those categories. However, real-world data is often messy, incomplete, or evolving too fast for manual labeling.

A growing body of research suggests a more bottom-up approach: latent type discovery. The core intuition is "duck typing" for data. If an entity consistently participates in relations like capital_of, has_population, and located_in, it "walks and quacks" like a geographic region. We don't need an explicit label to treat it as one. By observing the relational signatures of entities, models can spontaneously induce semantic hierarchies and functional categories.

The Relational Signature as a Proxy for Type

The fundamental mechanism behind this research is the idea that an entity’s "type" is defined by the set of relations it can plausibly enter. Sameer Singh and Sebastian Riedel demonstrated in Towards Combined Matrix and Tensor Factorization for Universal Schema Relation Extraction that by embedding entities and relations into a shared space, models implicitly learn fine-grained categories. Entities that play similar roles across different relations naturally cluster together, forming a latent taxonomy without any human-provided labels.

This concept is formalized as "duck typing" by Mattia Atzeni and Nora Kassner in Polar Ducks and Where to Find Them. They argue that defining categories through relational patterns provides robustness against incomplete knowledge bases. If a model knows that a specific role (e.g., the subject of a produced_by relation) typically accepts a certain "kind" of entity, it can infer the category of a new, unseen entity simply by observing its connections.

Geometric Constraints and Emergent Clustering

Why does this clustering happen? Much of the intuition comes from the geometric constraints we place on the learning process. In the seminal work Learning Hierarchical Structures with Linear Relational Embedding, Alberto Paccanaro and Geoffrey Hinton showed that if you force relations to be linear transformations (matrices) that map entity vectors to other entity vectors, the system is under immense pressure to organize the vector space logically. To make the math work across thousands of facts, the model must group "similar" entities (like nationalities or family generations) into specific regions of the space.

Modern iterations of this idea, such as the AutoETER framework by Guanglin Niu and Bo Li, treat relations as translations between latent type-level embeddings. In Automated Entity Type Representation with Relation-Aware Attention for Knowledge Graph Embedding, they show that these learned types act as a "soft" schema. This allows the model to handle "many-to-many" relations more effectively—for instance, knowing that the relation plays_instrument should only connect entities of the latent type "Musician" to entities of the latent type "Instrument."

From Raw Data to Formal Ontologies

The ultimate goal of this direction is to bridge the gap between raw, unclassified data and structured knowledge. Mark Derthick explored this as a data compression problem in A Minimal Encoding Approach to Feature Discovery. By using the Minimum Description Length (MDL) principle, the system "invents" categories that make the overall relational data easier to encode. If grouping a set of entities into a category called "Sex" or "Nationality" makes their relationships more predictable, the system adopts that category as a latent feature.

This transition from instances to abstract concepts is further refined in the JOIE framework by Junheng Hao and Muhao Chen. In Universal Representation Learning of Knowledge Bases by Jointly Embedding Instances and Ontological Concepts, they demonstrate that modeling specific entities and abstract concepts simultaneously allows the model to discover "meta-relations"—high-level patterns between categories that aren't explicitly stated but are implied by the behavior of the instances.

Why It Works: The Power of Co-occurrence

At its simplest level, this research direction works because it exploits the high-order co-occurrence of entities and relations. As Xu Wang and Jiale Han point out in their work on Global Dynamic Graphs, latent schema discovery allows information to flow between distant parts of a knowledge base. If the model recognizes that two nodes in different subgraphs share the same "latent type" because they have similar relational distributions, it can apply reasoning patterns learned in one area to the other.

By moving away from rigid, human-defined labels and toward fluid, learned representations, these models become more than just databases of facts; they become systems capable of inducing the underlying structure of the world they describe.

Papers in this version

Showing 60 of 72 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_unsupervised_latent_type_discovery_in_re_bf49afcc,
  title = {Unsupervised Latent Type Discovery in Relational Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unsupervised-latent-type-discovery-in-relational-graphs/rdr_bf49afcc2bea4a1f80248b753073ab9e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_unsupervised_latent_type_discovery_in_re_bf49afcc_v1,
  title = {Unsupervised Latent Type Discovery in Relational Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unsupervised-latent-type-discovery-in-relational-graphs/rdr_bf49afcc2bea4a1f80248b753073ab9e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}