Muhammad Umer Anwaar, PhD

Product Owner – Analytics CUE (Content Understanding & Enhancement) Driving state-of-the-art AI innovation at Unite. We are building cutting-edge AI solutions—including Embeddings, Specialized Language Models (SLMs), and Large Language Models (LLMs)—to unlock the full potential of content. Our mission is to understand and enhance product content at scale, empowering all Unitees to make better, faster decisions through high-quality, enriched data. By transforming raw data into structured, meaningful knowledge, we bridge the gap between content complexity and business clarity.

Part of project team who worked on the ~1 Million EUR research grant from the Free State of Bavaria, Germany Project tile: Goods & Knowledge - intelligent networking platform for business relationships (Waren & Wissen - intelligente Vernetzungsplattform für Geschäftsbeziehungen) Publications: ● Compositional Learning of Image-Text Query for Image Retrieval (​https://arxiv.org/pdf/2006.11149.pdf​) - Accepted at WACV 2021 ● Epitomic Variational Graph Autoencoder - Social Networks, Knowledge Graphs (​​https://arxiv.org/pdf/2004.01468.pdf​) - Accepted at ICPR 2020 ● Prior Knowledge Enhanced Explainable Recommender System via Path-aware Graph Attention (​https://github.com/ecom-research/graph_recsys_benchmark and https://arxiv.org/pdf/2010.11793.pdf​) ● Learning fused representations for large-scale multimodal classification (​https://ieeexplore.ieee.org/iel7/7782634/8913656/08962114.pdf​) ● Mend The Learning Approach, Not the Data: Insights for Ranking E-Commerce Products (​​https://arxiv.org/pdf/1907.10409.pdf​) - Accepted at ECML 2020

Assistant Professorship of Geometric Optimization and Machine Learning, TUM and IT- Analytics Team, Mercateo AG ● Learning to Rank E-Commerce Products with multiple noisy feedback signals extracted from user interactions ● Analysis of effectiveness of handcrafted features and features learnt through Deep Learning methods in Ranking E-Commerce Products

Online Learning to Rank through Reinforcement Learning ● Part of the team that developed and implemented Online Learning to Rank (LTR) model in the Live platform (www.mercateo.com) ● LTR Algorithm is based on Duelling Bandit Gradient Descent (DBGD) and its derivatives

Machine Learning Course for Masters Students at TUM “Information Retrieval in High Dimensional Data” Lectures/Tutorials (Recorded due to Covid-19): https://youtu.be/xdeelKgbA4s?si=NGp4OvWkYLmtgoNQ

Probabilistic Constellation Shaping for Bandwidth Efficient and Rate-Matched LDPC Coded Modulation for Next Generation Satellite Broadcasting Systems

Project “Estimation of (Generalized) Mutual Information for Real-time Applications”. Description : "Evaluating the post-FEC BER of a coded-modulation system using the uncoded (pre-FEC) BER is a commonly used method in fiber-optics. However, this approach can be inaccurate. Instead, mutual information (MI) is a natural figure of merit to consider as it is an achievable rate for a decoder that operates on symbol metrics. For the practical bit-interleaved coded modulation systems, the generalised mutual information (GMI) has been shown to accurately predict the post-FEC performance. In this project, we're investigating GMI estimation under typical constraints for real-time applications. "

Institute of Data Processing, Department of Electrical and Computer Engineering, Technische Universität München (TUM) Tutored a Machine Learning course “Information Retrieval in High Dimensional Data” in TUM in WS 2018/19, WS 2019/20, WS 2020/21, WS 2021/22, WS 2022/23 with 100+ students.