1 of 1

Enabling Workforce Intelligence through Occupational Taxonomy Alignment

Yifan Xu¹, Isabella Loaiza², César A. Uribe³

¹Department of Computational Applied Mathematics and Operations Research, ³Department of Electrical and Computer Engineering, Rice University;�²The MIT Sloan School of Management

yx76@rice.edu

Motivation

• Labor market analysis relies on occupational taxonomies that are periodically revised, disrupting longitudinal comparisons.
• Manual crosswalks between occupational codes are coarse, opaque, and cannot capture partial similarities or many-to-many relationships.
• Alternative view: occupations are not unchanging chunks of work but rather structured collections of tasks. This raises the question: how should similarity between two occupational networks be defined?

​

We develop a network-based occupational translator using optimal transport to align occupations across taxonomies by jointly comparing task semantics and task network structure.

​

Data & Task Embeddings

Occupations (O*NET 2017: m=959, O*NET 2024: n=894)

Software

Developer

15-1252

Data

Scientist

15-2051

Financial

Analyst

13-2051

Task Descriptions (~17,000 unique tasks)

"Write code to

implement designs"

"Debug software

and fix errors"

"Analyze user

requirements"

"Develop data

models & schemas"

"Prepare financial

reports & forecasts"

"Evaluate investment

opportunities"

E(x) SentenceTransformer (all-mpnet-base-v2) → ℝ⁷⁶⁸

Embeddings → k-means Clustering (k = 750)

ℝ⁷⁶⁸ embedding space

c₁

Coding

c₂

Analysis

c₃

Reporting

c₄

Finance

⋯ 750 clusters

c₁

c₂

c₃

c₄

c₅

o₁

o₂

o₃

o₄

1

1

0

0

1

0

1

1

0

0

0

0

1

1

0

1

0

0

0

1

X =

① Co-occurrence W

How frequently do task pairs co-appear across occupations?

Dense (99.87% nonzero)

② Graphical Lasso W⁽²⁾

Direct conditional dependencies between task clusters

Sparse (6.60% nonzero)

G(o⁽²⁰¹⁷⁾)

e.g. Software Developer

c₁

coding

c₂

analysis

c₃

Data Mgmt

c₅

Testing

G(o⁽²⁰²⁴⁾)

e.g. Financial Analyst

c₂

c₂

analysis

c₃

Data Mgmt

c₄

Finance

Task Meta-Network Construction

Fused Gromov-Wasserstein Formulation

Minimization over all possible task-level transport plans

​

Penalizes by text semantic dissimilarity.

​

between cluster centroids.

​

Penalizes mismatched structure.

C: distance between the motif role profiles of task clusters a and a′ within G(o), capturing how similarly two tasks participate in local structural patterns within an occupation.

Restrict to occupation subgraph:

G(o) = G[C(o)]

Both meta-network constructions show a statistically significant positive correlation with semantic similarity (Mantel test, one-sided p = 0.001):

Co-occurrence W

Mantel r = 0.4342

Both constructions produce consistent FGW distances with clear block structure along O*NET major groups. Group 25 (Education) shows a prominent band of high dissimilarity, driven by fewer tasks per occupation.

776 / 959 occupations achieve k-score = 1 (~81% exact match)

894 / 959 achieve k-score ≤ 5 (~93%)

Result 1: Task Semantics vs Task Relatedness

Result 2: Cross-year Dissimilarity

Result 3: Translator Reliability

Conclusion & Future Work

Being able to translate occupations means we can track how jobs change over time, identify viable career transitions, and bridge fragmented labor market data across institutions. As next steps, we aim to:

1. Investigate the effects of λ (in graphical lasso) on translator performance.
2. Align heterogeneous taxonomies (e.g., ISCO ↔ SOC).

Graph Lasso W⁽²⁾

Mantel r = 0.2932