BHANU TEJA MALINENI← All projects
Portfolio / Project 01 / Causal Graph-Augmented Multimodal LLM

PROJECT 02  ·  Master’s thesis · Machine learning research

Causal Graph-Augmented Multimodal LLM
Making AI decisions more explainable

CG-MM-LLM combines multimodal representations with causal structure to explore more informed decision-making. The thesis brought together language, image, graph learning, causal discovery, and effect estimation in one research pipeline.

CG-MM-LLM / DECISION INTELLIGENCE
TEXT
+ IMAGE
↘ CAUSE
Bhanu Teja Malineni02 / 04
RoleThesis researcher
PeriodMay—Aug 2026
ContextUniversity of Europe for Applied Sciences
FocusMaster’s thesis

The idea

The thesis asks how a system can move beyond correlations in mixed data and produce decision support with an explicit causal view. I developed CG-MM-LLM, a causal graph-augmented multimodal LLM pipeline, as my master’s thesis research at the University of Europe for Applied Sciences in Berlin.

How it works

The pipeline brings together Sentence-BERT for text representation, ResNet-50 for image features, heterogeneous graph neural networks, NOTEARS causal discovery, and DoWhy causal effect estimation. The evaluation used Amazon Reviews 2023, RetailRocket, and Instacart datasets.

01RepresentEncode text and image signals with Sentence-BERT and ResNet-50.
02ConnectModel heterogeneous relationships with graph neural networks.
03EstimateDiscover causal structure with NOTEARS and estimate effects with DoWhy.

Reported outcomes

+27.8%Composite Decision Intelligence score improvement over the strongest baseline.
0.707Composite score, compared with 0.553 for the strongest baseline.
0.019 → 0.519Demand-forecast R² across the reported evaluation.
3 datasetsAmazon Reviews 2023, RetailRocket, and Instacart.
98/100Master’s thesis grade.
These are results reported in the thesis source material. The framework was described as intended for future submission; this page does not claim it has been submitted or published.

Tools & methods

Sentence-BERTResNet-50Heterogeneous GNNNOTEARSDoWhyAmazon Reviews 2023RetailRocketInstacart
More about the project

This project is a research prototype rather than a deployed product. Its contribution is the integration of multimodal representations, graph learning, causal discovery, and causal effect estimation into a single decision intelligence workflow. The results above are the reported thesis metrics; further independent validation would be needed before generalizing them beyond the evaluated datasets.

What I took from it

The work sits at the intersection of machine learning and decision support: model quality is useful, but understanding relationships and estimating effects can make outputs more actionable. The thesis was supervised by Prof. Dr. Raja Hashim Ali.