arxiv_ai 85% Match Research Paper Data Scientists,ML Engineers,Researchers in NLP and Data Mining 2 weeks ago

Merging Embedded Topics with Optimal Transport for Online Topic Modeling on Data Streams

large-language-models › training-methods

📄 Abstract

Abstract: Topic modeling is a key component in unsupervised learning, employed to identify topics within a corpus of textual data. The rapid growth of social media generates an ever-growing volume of textual data daily, making online topic modeling methods essential for managing these data streams that continuously arrive over time. This paper introduces a novel approach to online topic modeling named StreamETM. This approach builds on the Embedded Topic Model (ETM) to handle data streams by merging models learned on consecutive partial document batches using unbalanced optimal transport. Additionally, an online change point detection algorithm is employed to identify shifts in topics over time, enabling the identification of significant changes in the dynamics of text streams. Numerical experiments on simulated and real-world data show StreamETM outperforming competitors. We provide the code publicly available at https://github.com/fgranese/StreamETM.

Authors (4)

Federica Granese

Benjamin Navet

Serena Villata

Charles Bouveyron

Submitted

April 10, 2025

arXiv Category

cs.LG

Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2025, Porto, Portugal, September 15-19, 2025, Proceedings, Part VII

arXiv PDF Code

Key Contributions

Introduces StreamETM, a novel approach for online topic modeling on data streams that merges models using unbalanced optimal transport and incorporates an online change point detection algorithm to identify topic shifts. It outperforms competitors on simulated and real-world data.

Business Value

Enables businesses to gain real-time insights from evolving text data sources like social media, news feeds, or customer feedback, allowing for agile responses to market trends.

Paper Metadata

Innovation Type

Algorithmic Improvement

Deployment Feasibility

Moderate. Requires infrastructure for stream processing and integration with existing data pipelines.

Limitations Addressed

Inability of traditional topic models to handle continuously arriving data streams,Difficulty in detecting dynamic changes in topics over time

View Code on GitHub

Technical Tags

online topic modelingdata streamstopic modelingoptimal transportEmbedded Topic Model (ETM)change point detectionunsupervised learningtext analysisstreaming datadynamic topics

Research Topics

Topic ModelingOnline LearningData StreamsUnsupervised LearningNatural Language Processing

Methods & Architectures

StreamETMMerging models with unbalanced optimal transportOnline change point detection Embedded Topic Model (ETM)

Applications & Tasks

Social Media Analysis Text Data Analysis Real-time Analytics Topic modeling on continuously arriving dataDetecting shifts in topics over time Identifying topics in data streamsTracking topic evolutionUpdating topic models dynamically

Related Fields

Machine LearningData MiningNatural Language ProcessingTime Series Analysis

Keywords

online topic modelingdata streamstopic modelingoptimal transportEmbedded Topic Modelchange point detectionunsupervised learningtext analysisstreaming datadynamic topicssocial mediaNLP

Academic Context

#Topic Modeling#Online Learning#Data Streams#Unsupervised Learning#Natural Language Processing

Commercial Potential

Potential Products

Real-time trend analysis toolsDynamic content recommendation systemsCustomer sentiment monitoring platforms

Target Industries

Social MediaMarketingFinanceNews MediaCustomer Service

Use Case Examples

Monitoring public opinion on a new product launch in real-timeDetecting emerging news topicsAnalyzing evolving customer feedback trends

Competitive Edge

Offers a novel online topic modeling approach that explicitly handles data streams and topic evolution using optimal transport and change point detection.

Market Opportunity

Growing demand for real-time analytics and insights from unstructured data.

Revenue Models

SaaS for real-time analyticsintegration services.

Resource Requirements

Compute Needs

Moderate (for stream processing and model updates)

Data Requirements

Continuous stream of textual data

Deployment Constraints

Requires robust stream processing infrastructure,Computational cost of merging models and detecting changes

Scalability

Designed for scalability on data streams.

Regulatory Considerations

Data privacy and usage policies for streaming data

Production Readiness

Maturity Level

Research (with code release)

Time to Market

1-2 years

Licensing

Open Source (MIT License implied by public code)

Patent Potential

Low (algorithmic approach)

View Full Paper Back to Papers