Knowledge Graph Embeddings for Occupational Mobility

Exploring O*NET's graph structure with TuckER to find related occupations

The Problem

If you've ever looked into changing careers, you've probably run into a version of this question: which occupations match my current skills? One of the richest resources for answering it is O*NET, the Occupational Information Network. It describes 1,016 occupations through detailed profiles that include skills, knowledge areas, abilities, tasks, and work activities.

The standard approach to finding related occupations is straightforward: look at shared elements, and if two occupations have a lot of the same skills and knowledge areas, they're probably related. This works, as long as the data is available. However, could there be other ways to find occupations that are less obviously related?

O*NET publishes its data in RDF — a graph format — and that allows us to approach the problem from a different angle. Instead of comparing occupation profiles element by element, you can treat the entire dataset as a knowledge graph and look for patterns in its structure. This lets you identify related occupations even when they don't share many of the same skills, knowledge areas, or abilities.

Building the Graph

I extracted a subgraph from O*NET 30.3 — a network where occupations, skills, knowledge areas, abilities, tasks, and detailed work activities (DWAs) are all nodes, connected by typed edges. Here's what a single occupation looks like in that graph:

Excerpt of the Graphic Designers ego graph — 7 relation types connect occupations to skills, knowledge areas, abilities, tasks, DWAs, job zones, and other occupations. The full graph contains 21,900 entities.

Each occupation connects to its skills, knowledge areas, abilities, tasks, DWAs, job zone, and other related occupations. O*NET has a lot more information than what I've included here (work styles, interests, work context, education, wages, and more), but I kept the scope to the elements most directly tied to what a person does and knows in a role. I also filtered out low-importance elements using O*NET's own importance and level ratings, so the graph focuses on the strongest signals.

21,900

entities

relation types

923

occupations

78,282

training triples

Out of the 1,016 occupations in O*NET, 923 ended up in the graph. The 93 that didn't make it are mostly "All Other" residual categories and military specializations — occupations without detailed element profiles.

For finding related career paths, one can simply compare element profiles directly. You may have noticed in the visualization above that O*NET already catalogs related occupations, which they determine largely by doing such a comparison (although, it should be noted that they also rely on judgment from human experts). That works well, but it can only find occupations that look alike on paper or that consulting experts can identify. Two roles might be structurally related — playing similar roles in the broader occupational landscape — without sharing many of the same elements.

Since O*NET's data is already structured as a knowledge graph with typed relationships, there's another angle worth exploring: models that learn from the pattern of connections, not just which elements are shared. To that end, I used a model called TuckER, which learns a low-dimensional representation of every entity and relationship in the graph and then uses those representations to predict missing links.

What TuckER Finds

TuckER works by decomposing the full graph into a small core tensor and two embedding matrices — one for entities, one for relations — with far fewer dimensions than the original graph. This compression forces the model to learn the underlying patterns that connect occupations, skills, knowledge, and other elements, rather than memorizing individual links. It can then use those learned representations to score how likely any unobserved link is.

I trained the model on the O*NET knowledge graph and ranked all unobserved occupation pairs by predicted score, filtering out every pair that O*NET already lists as related. Here are a few examples of the top predictions that result:

Click a node to see its O*NET profile

The strongest cluster is in law enforcement and public safety. Police supervisors get paired with customs officers, fraud examiners, security managers, and fire investigators — all roles that share investigative and supervisory patterns, even though O*NET doesn't list them as related. Management analysts linked to HR specialists and forest fire inspectors linked to environmental compliance inspectors also seem like plausible fits.

The results are mixed, though. Judges, magistrate judges, and magistrates are also in the same law enforcement cluster, linked to first-line supervisors of police and detectives. Clearly, structural similarity in the graph doesn't always translate to a practical career move.

Here's a table of the top 25 predictions found by TuckER, along with their Jaccard similarity:

Pairs are ranked by TuckER's raw link-prediction score. Jaccard similarity (0–1) measures the overlap in skills, knowledge, abilities, tasks, and DWAs between the two occupations.

#	Source	Target	Score	Jaccard
1	First-Line Supervisors of Police and Detectives	Customs and Border Protection Officers	11.20	0.391
2	First-Line Supervisors of Police and Detectives	Fraud Examiners, Investigators and Analysts	10.80	0.333
3	First-Line Supervisors of Police and Detectives	Judges, Magistrate Judges, and Magistrates	10.72	0.347
4	First-Line Supervisors of Police and Detectives	Airfield Operations Specialists	10.65	0.314
5	First-Line Supervisors of Police and Detectives	Security Managers	10.59	0.364
6	First-Line Supervisors of Police and Detectives	First-Line Supervisors of Personal Service Workers	10.37	0.301
7	Compensation, Benefits, and Job Analysis Specialists	Treasurers and Controllers	10.30	0.321
8	First-Line Supervisors of Police and Detectives	Management Analysts	10.23	0.397
9	Management Analysts	Human Resources Specialists	10.22	0.338
10	Concierges	Cashiers	10.22	0.050
11	Railroad Conductors and Yardmasters	Commercial Pilots	10.19	0.244
12	Police and Sheriff's Patrol Officers	Arbitrators, Mediators, and Conciliators	10.18	0.215
13	Management Analysts	Clinical Data Managers	10.17	0.384
14	Legal Secretaries and Administrative Assistants	Title Examiners, Abstractors, and Searchers	10.17	0.286
15	First-Line Supervisors of Security Workers	Airfield Operations Specialists	10.16	0.333
16	Forest Fire Inspectors and Prevention Specialists	Environmental Compliance Inspectors	10.15	0.277
17	Dispatchers, Except Police, Fire, and Ambulance	Air Traffic Controllers	10.07	0.263
18	Bailiffs	Compliance Officers	10.03	0.220
19	First-Line Supervisors of Police and Detectives	Government Property Inspectors and Investigators	10.03	0.368
20	Dispatchers, Except Police, Fire, and Ambulance	Freight Forwarders	10.02	0.273
21	Dispatchers, Except Police, Fire, and Ambulance	Airfield Operations Specialists	10.02	0.306
22	Legal Secretaries and Administrative Assistants	Statistical Assistants	10.02	0.263
23	First-Line Supervisors of Police and Detectives	Eligibility Interviewers, Government Programs	10.01	0.297
24	First-Line Supervisors of Police and Detectives	Fire Inspectors and Investigators	10.00	0.302
25	Management Analysts	Securities, Commodities, and Financial Services Sales Agents	9.99	0.346

Why This Matters

Element-based similarity methods find occupations that look alike on paper — shared skills, shared knowledge, shared abilities. TuckER tries a different angle: learning from the pattern of connections rather than counting shared elements. When it works, it surfaces connections like fraud examiners and police supervisors that share investigative reasoning patterns a skills checklist might not capture. When it doesn't, you can end up with things like railroad conductors paired with commercial pilots — superficially similar in the graph but not in practice.

These predictions are hypotheses, not recommendations, and the mixed results suggest this approach works better as a starting point. The natural next step would be validation against real-world data. The Bureau of Labor Statistics publishes occupational transition data that could tell us whether these predicted paths correspond to career moves people actually make.

TuckER is also just one approach. Other knowledge graph embedding models — RotatE, TransR, ConvE — model relational patterns differently and might surface different connections. Graph neural network methods like R-GCN could go further by incorporating node features and multi-hop reasoning.

Technical Details

Model Architecture

TuckER uses Tucker decomposition to model relationships in a knowledge graph. For a given subject entity $s$ and relation $r$ , it scores all possible object entities:

\text{score}(s, r, \cdot) = \left( \mathcal{W} \times_2 \mathbf{r} \times_1 \mathbf{e}_s \right) \cdot \mathbf{E}^\top

where $\mathcal{W}$ is a learnable core tensor, $\mathbf{e}_s$ is the subject entity embedding, $\mathbf{r}$ is the relation embedding, and $\mathbf{E}$ is the full entity embedding matrix.

Hyperparameters

Parameter	Value
Entity embedding dim	200
Relation embedding dim	30
Batch size	128
Learning rate (initial)	0.005
LR decay (per epoch)	0.995
Label smoothing	0.1
Dropout (input / hidden₁ / hidden₂)	0.2 / 0.2 / 0.3
Loss function	Softmax cross-entropy
Optimizer	Adam
Max epochs	500
Actual epochs (early stopping)	237
Patience	50

Training Curves

Training Loss

Validation MRR

Test Set Evaluation

Filtered link prediction metrics with 95% bootstrap confidence intervals (n = 1,000 resamples).

37.1%

MRR

[35.8–38.4%]

19.5%

Hits@1

[17.9–20.9%]

45.9%

Hits@3

[44.0–47.8%]

71.9%

Hits@10

[70.1–73.6%]

13.7 mean rank out of 21,900 entities

A Note on the Loss Function

The original TuckER paper uses binary cross-entropy (BCE) loss. That didn't work here. With O*NET's extreme class imbalance — roughly 3 positive entities out of 21,900 per training example (0.014%) — BCE lets the model get away with predicting all-negative scores and still achieving near-zero loss. The model basically learns to say "nothing is related to anything," which isn't very useful.

Softmax cross-entropy fixes this by normalizing scores across all entities before computing the loss, which forces the model to actually rank positive entities above negatives. There's no trivial all-negative solution. Label smoothing ( $\varepsilon = 0.1$ ) keeps the model from getting overconfident about its predictions.