Hi, I'm Wei-Wei Du.

A

Applied Research Scientist at Sony specializing in recommender systems, off-policy evaluation, and LLM applications.

About

I am an Applied Scientist at Sony with three years of experience building and evaluating recommender systems for an OTT streaming platform with 30M+ monthly active users. My work spans LLM-based recommendation, off-policy evaluation, multi-objective recommendation, and agent-driven experimentation. I am the first author of a RecSys 2025 paper and a co-author of an ICLR 2026 paper. Before joining Sony, I worked as a Machine Learning Scientist Intern at Appier, developing real-time bidding models.

During my master's, my research spanned several domains, including property valuation (self-supervised learning for few-shot scenarios, graph neural networks), natural language processing (depression detection, multi-modal fact-checking), and Sports AI, with multiple publications on these topics.

With over five years of experience in the field of Machine Learning, I am passionate about turning research into real-world impact and continuously seeking new knowledge. In my free time, I enjoy photography, cycling, and practicing yoga.

ML/AI: Recommender systems, Learning-to-rank, LLM-based recommendation, Off-policy evaluation, Multi-objective recommendation, Causal discovery, Self-supervised learning
Frameworks: PyTorch, Hugging Face Transformers, scikit-learn, PySpark
Infra & Experimentation: A/B testing, SQL, Linux/Shell, AWS/GCP
Languages: English (Full Professional Proficiency), Mandarin (Native)

Experience

Sony

Applied Scientist

Applied causal discovery to identify 11 short-term KPIs that drive revenue on a 30M+ MAU OTT platform, and used them as objectives for multi-objective, multi-behavior recommenders and marketing targeting.
Led the design and build of an autonomous research loop in which LLM agents propose, evaluate, and refine recommendation strategies, cutting the offline experimentation cycle from 2 weeks to 2 days.
Built off-policy evaluation pipelines for ranking policies under deterministic logging, now used across teams to evaluate 10+ candidate ranking policies in parallel; the production constraints motivated the click-based estimators (CIPS/CDR) published at ICLR 2026.
Led research on LLM-based sequential recommendation with irregular time intervals, achieving a +4.4% average improvement over traditional and LLM-based baselines across 3 benchmarks (first-author paper accepted at RecSys 2025).
Bridged research, engineering, and business teams across Europe, India, and Japan, translating research into engineering requirements and business decisions across media and gaming.

Oct 2023 – Present | Tokyo, Japan

Appier

Machine Learning Scientist Intern

Implemented an ensemble real-time bidding model with new data-driven features from 10M+ e-commerce click-stream events, achieving 2x CTR improvement in production.
Conducted tree analysis and feature importance with SHAP to analyze model behavior.
Cooperated with 3 data scientists to build the RTB model for re-engagement campaigns.

Jun 2022 – Nov 2022 | Taipei, Taiwan

Research Papers

                
 Off-Policy Evaluation for Ranking
                  Off-Policy Evaluation for Ranking Policies under Deterministic Logging Policies (ICLR-26)
                
Accomplishments
                    Address the critical challenge of off-policy evaluation for ranking systems when the logging policy is deterministic, a common yet under-explored scenario in production systems.
                  
                    Propose a principled estimator that enables unbiased evaluation of ranking policies without requiring stochastic exploration in data collection.
                  
 LLM-based Recommender
                  Not Just What, But When: Integrating Irregular Intervals to LLM for Sequential Recommendation (RecSys-25)
                
Accomplishments
                    The first work to integrate time intervals into LLMs for sequential recommendation.
                  
                    Propose Interval-Infused Attention to capture the temporal relevance between items and intervals; +4.4% average improvement over traditional and LLM-based baselines across 3 benchmarks.
                  
                    Introduce a novel perspective on the cold-start problem by considering time intervals and reveal that existing methods suffer from significant performance drops in interval cold-start scenarios.
                  
 SSL for NSTD
                  A Survey on Self-Supervised Learning for Non-Sequential Tabular Data (ACML-24)
                
Accomplishments
                    The first comprehensive survey of recent advancements in SSL4NS-TD, consisting of problem definitions, taxonomy, application issues, NS-TD datasets, and evaluation protocols
                  
                    Evaluate representative SSL4NS-TD approaches of each learning category on the most recent and large-scale benchmark, TabZilla.
                  
                    Highlight the addressed challenges and future directions in SSL for the existing NS-TD methods.
                  
Graph-Based Learning
                  Look Around! A Neighbor Relation Graph Learning Framework for Real Estate Appraisal (PAKDD-24; AAAI-23 MUFin Workshop, Best Paper Award)
                
AccomplishmentsIncorporate the relationship between the target transaction and neighbors with an attention mechanism
Utilize the neighbors’ price information to predict a preliminary value
Introduce dynamic predictor to model the price of target transactions with different characteristics

SSL for Few-Shot Learning
                  DoRA: Domain-Based Self-Supervised Learning Framework for Low-Resource Real Estate Appraisal (CIKM-23)
                
AccomplishmentsThe first work focusing on low-resource real estate appraisal, which meets the needs of real-world scenarios.
Introduced with novel and effective intra- and inter-sample SSL objectives to learn robust geographical knowledge from unlabeled records.
Illustrate a developed system of DoRA and the real-world industrial scenarios for cities and towns with extremely limited transactions.

AI for Sport
                  Benchmarking Stroke Forecasting with Stroke-Level Badminton Dataset (IJCAI-24 Demo)
                
Accomplishments
                    Introduce ShuttleSet22, a stroke-level badminton singles dataset collected from real-world high-ranking matches in 2022
                  
                    Initiated a challenge within CoachAI Badminton Challenge 2023 (https://sites.google.com/view/coachai-challenge-2023/) in conjunction with IJCAI 2023.
                  
Depression Detection
                  Ensemble Models with VADER and Contrastive Learning for Detecting Signs of Depression from Social Media (ACL-22 Workshop)
                
Accomplishments
                    Developed an ensemble model with VADER and contrastive learning for detecting depression.
                  
                    Won second place in 30+ teams without any auxiliary information.
                  
Multi-Modal Fact Checking
                  Parameter-Efficient Large Foundation Models with Feature Representations for Multi-Modal Fact Verification (AAAI-23 Workshop)
                
Accomplishments
                    Introduced a parameter-efficient large foundation model by utilizing adapters and additional features.
                  
                    Incorporated co-attention modules for different modalities (image and text) and different types (claim and
                    document).
                  
                    Surpassed 25.9% compared with the official baseline.

Personal Projects - Agentic Development

Where To Go

LLM-Powered Travel Destination Recommender

Highlights

End-to-end retrieval-and-ranking pipeline modeled on industrial recommender systems: LLM-based query understanding with schema-constrained extraction, candidate filtering, embedding retrieval, and LLM listwise reranking with explanations grounded in user intent.
Reduced latency by caching hash-invalidated embedding indexes and prefetching external API enrichment concurrently with LLM reranking.

Home screen of the First and Last multiplayer word game

First & Last

Real-Time Multiplayer Word Game

Highlights

Server-authoritative architecture with a phase state machine and server-driven timers, synchronizing clients via personalized WebSocket state snapshots.
External API validation with offline fallback for resilient real-time gameplay.

Honors & Services

Invited Speaker

RecSys 2025 Workshop on Evaluating and Applying Recommender Systems with LLMs (EARL) — LLMs for Next-Generation Recommender Systems: From Understanding User Behavior to Deployment, Sep 2025 [link]

Program Committee

The Web Conference (WWW) 2026–2027
CIKM 2023–2026

Organizer

Workshop on Intelligent Technologies for Precision Sports Science (IT4PSS) @ IJCAI 2024 and IJCAI 2023 [site]
CoachAI Badminton Challenge @ IJCAI 2023 [site]

Competitions