AOBTM - Adaptive Online Biterm Topic Model

Problem Statement

App reviews have dynamic nature
Discussed topics in the app-reviews change over time for different versions of app.
The changes in the topics should be analyzed to reveal important issues in the app update

Conventional (Inadequate) Approaches

Conventional topic models, such as LDA, PLSA, BTM
Online Topic Modeling algorithms, such as OLDA, OBTM

Proposed Approach & Contributions

Adaptive Online Biterm Topic Model (AOBTM)
Parallel algorithms to automatically determine the optimal number of topics
Parallel algorithms to automatically determine the best number of previous version to consider
Open sourced Code: github.com/Mohammad-Abdul-Hadi/AOBTM-Adaptive-Online-Biterm-Topic-Modeling

Research Questions

Can AOBTM achieve better performance compared to baselines?
How different parameter settings impact the performance of AOBTM?
How discriminative and coherent are the topics discovered when parameters are set by proposed parallel algorithms?

Observations

AOBTM delivers the highest PMI_Scores in every dataset
AOBTM delivers the highest Dis_Scores in every dataset except for Tweets2020
AOBTM also delivers the highest scores in every dataset for Precision, Recall, and F_hybrid

We acknowledge that AOBTM is time-expensive; but the run-time is still comparable to adaptive online algorithms when the dataset is small.
AOBTM outperforms AOLDA in runtime for NOAA Radar dataset, which has the lowest number of average short texts per version

Conclusions

Proposed a novel Adaptive topic modeling algorithm, AOBTM
AOBTM discovers coherent and discriminative topics from short texts
AOBTM addresses the problems with conventional topic models by adopting a version sensitive strategy

Proposed two parallel algorithms to determine the value of the two most important parameters of our model automatically
The results of several experiments on different datasets conform the performance of AOBTM compared to the state-of-the-art algorithms.

Questions?