nextminingIn-Batch Negatives Can Silently Cripple LLM-Encoded Sequential Recommenders
How training objectives and negative sampling affect LLM-encoded sequential recommendation, and why in-batch negatives can quietly degrade it.
AI for recommendation & personalization
Helping people find what matters.
We research how large language models understand intent and evolving behavior, and build recommendation and personalization systems for content, learning, and products.
Research
Recommender systems and personalization, with LLMs.
Recommendation systems still struggle to capture nuanced intent and to keep up as behavior and catalogs change. Our work looks for better ways to represent and reason about that behavior.
View Google ScholarLLM-encoded sequential recommendation
Training techniques, such as better contrastive objectives and negative sampling, that make LLM-encoded recommenders more accurate and robust.
Multimodal recommendation
Combining text, visual, and behavioral signals into unified representations that capture relevance more completely.
Evolving preferences & behavioral memory
Remembering how interests shift over time, so recommendations keep up as people change.
Lightweight, transferable personalization
Models that carry across domains without large-scale data, and evaluation that measures real-world relevance beyond offline benchmarks.
Publications
MEMOIR: Temporal Behavioral Memory for Recommendation Across the Preference-Drift Spectrum
LLM-based behavioral memory for recommendation as preferences change over time.
Applications
One foundation, from content to products.
The same understanding of intent and behavior powers content discovery and learning today, and is the basis of our research into product discovery.
Content & learning · Live
Scan NewsYour interests. Your words.
Scan NewsYour interests. Your words.
Your English.
Personalized content and learning: read English news you care about, tap unfamiliar words, and build your own vocabulary.
Product discovery · In research
Product discovery & recommendation that keeps up with changing preferences.
Research into how LLM-based behavioral understanding can make product recommendation more accurate and adaptive over time, with lightweight models that don't depend on large-scale data.
- Preference driftRecommendations that adapt as tastes change, not just to the last click.
- Content & visual similarityRelating products by what they are and how they look.
- Behavior-aware recommendationLearning from sequences of behavior with LLM-encoded representations.
About
Fifteen years of understanding people through text and behavior.
NextMining is an independent AI research and product studio founded by Younggue Bae, Ph.D. The work has followed one thread: from understanding opinions and intent in text, to recommendation and personalization systems that put that understanding to use.
Publications on Google Scholar- Doctoral research
Opinion mining, sentiment analysis, and text mining: understanding intent in unstructured text.
- Search & discovery at scale
Document quality, shopping keyword extraction, and related-search recommendation for web search and content discovery.
- LLMs for customer insight
Aspect-based summarization of customer reviews with fine-tuned LLMs.
- Today
Applying LLMs to retrieval, discovery, and personalization at scale, alongside independent research on LLM-encoded recommenders.