Chen Li Email & Phone Number

New York, Ny, Us

New York, Ny, Us

New York, Ny, Us

Predictive Volumetric Video StreamingDeveloped a Temporal Graph Network to predict user attention in volumetric video, optimizing streaming with an adaptive algorithm tailored to attention patterns.Cost-Effective Edge Caching for 360 Degree Live Video StreamingEngineered a cost-effective edge caching system for 360-degree live videos, implementing a collaborative prediction model to reduce streaming costs by 35%.Predictive Edge Caching through Deep Mining of User’s Sequential PatternsDesigned a predictive model for edge caching by mining user's sequential patterns, improving hit ratio and reducing latency by 12%, with results published in Computer Networks.Temporal Sequence Prediction for Edge CacheUtilized LSTM to optimize edge caching strategies in CDNs by predicting user request patterns, published in IEEE INFOCOM.

Seattle, Wa, Us

Formulated and proposed 3 ML solutions for fully automatic Knowledge Graph Schema inference on tabular data, which has no off-the-shelf solution in current research. The final model, a novel fact-driven method, integrating semantic relations, data distribution, KG schema graph contrains achieved a 0.86 F1 score compared to expert-built scheme and outperforms zero-shot LLM and other baselines up to 21%.Attained an ’Inclined’ rating, positioning for a full-time Applied Scientist role.

Seattle, Wa, Us

Concession Abuse Detection Using Temporal Graph Networks(TGN) • Framework Development: Constructed a customer-item-attribute TGN and processed time-stamped customer purchase history events for model training.• Training Innovation: Innovated the TGN training by incorporating Amazon item metadata into node memory and enhancing link prediction for customer-item, and introduced ”daily snapshot training” to save memory.• Model Enhancement: Utilized customer and item embeddings in the risk model, increasing dollar recall and item recall by 2% and at the same time achieving 30x memory savings and production alignment.• Recognition: Attained an ’Inclined’ rating, positioning for a full-time Applied Scientist role.

Seattle, Wa, Us

Amazon Standard Identification Number(ASIN) Embedding• Embedding Design: Modelled the customer order history as sequences and employed the Word2Vec (skip-gram) model for item embedding, capturing complementarity within orders.• Self-Attention Integration: Formulated and adapted self-attention to incorporate ASIN metadata in the model.• Performance Enhancement: Elevated ASIN embedding quality, achieving a 9.3% increase in reseller detection precision over the previous top-performing model.• Cold Start Solution: Innovated a unique masking technique and integrated textual embeddings, resulting in a 6.4% improvement in AUC for cold start ASINs.Attained an ’Inclined’ rating, positioning for a returning Applied Scientist intern role.

Lehi, Ut, Us

Recommendation System for Ancestry Mobile Users• Solution Formulation: Conducted a literature review on recommendation algorithms and crafted a custom data science approach aligned with product objectives.• Data Processing: Identified the data source, processed data with SQL on AWS RedShift, performed Exploratory Data Analysis (EDA) and extracted features.• Model Development: Constructed a Behavioral Sequential Transformer neural network for content recommendations, leveraging users’ historical content rating sequences and features.• Results: Reduced rating error from 2.076 to 0.599 and the solution was deployed online and subsequently submitted for patenting.

Beijing, Cn

Multi-label Classification for JD Snapshop• Data Collection & Pre-processing: Collected and pre-processed large-scale fashion images (specifically bags) from JD e-commerce platform by categorizing and applying bounding box.• Model Customization:Utilized a pre-trained ResNet-50, adapting it with a tailored weighted loss function and balanced score to address challenges posed by highly imbalanced and noisy labels.• Performance Metrics: Achieved 76.77% exact match, nearly 70% balanced score on 55 classes, alongside more stable convergence.

Hefei, Anhui, Cn

RL Algorithm for a Intelligent Electronic Pet-Showed at Stanford University Global Alliance for Re-Design (SUGAR) in Stanford University in June, 2018

Doctor Of Philosophy - Phd, Electrical And Computer Engineering

Bachelor'S Degree, Electronic Engineering And Information Science