The article explains that premature tool calls in LLM agents stem from a data‑distribution bias in fine‑tuning, and it presents practical fixes such as adding non‑tool samples, enforcing a Thought chain, and using negative sampling to teach the model when to think before acting.
This article provides a thorough review of negative sampling techniques in recommender systems, categorizing existing methods into five groups, detailing their sub‑strategies, advantages, challenges, and future research directions to improve model accuracy and robustness.
This article reviews graph link prediction problems, categorizes existing methods from heuristics to GNN‑based approaches, introduces several self‑designed neighborhood attention networks and adversarial negative‑sampling strategies, discusses pairwise ranking objectives, reports OGB competition results, and explores practical risk‑control applications.
This article explains why cold‑start is a critical challenge for recommender systems, outlines four practical strategies—generalization, fast data collection, transfer learning, and few‑shot learning—and then details the DropoutNet model, its end‑to‑end training, loss functions, negative‑sampling techniques, and open‑source implementation.
The recall module in recommendation systems retrieves thousands of items from massive pools using parallel non-personalized and personalized paths—such as hot-item, content-based, behavior-based, and deep-model recall—prioritizing coverage and low latency while addressing challenges like hard-negative sampling, selection bias, objective alignment, and channel competition to feed downstream ranking.
This article explains how treating negative response candidates with varying difficulty through a hierarchical curriculum learning framework—combining corpus‑level and instance‑level curricula—enhances dialogue response selection models, backed by experiments on Douban, Ubuntu, and E‑Commerce datasets.
This comprehensive review by Dr. Fan Yixing surveys how pretrained language models have transformed first‑stage information retrieval, tracing the shift from traditional term‑based methods to neural sparse, dense, and hybrid approaches, and discussing key challenges such as hard‑negative mining, joint indexing‑representation learning, and generative‑discriminative training.
Researchers from Tencent Ads and Tsinghua University introduced a novel Generative Adversarial framework, the Representative Negative Instance Generator (RNIG), which creates high‑quality representative negative samples from exposure data to mitigate data imbalance and selection bias, achieving superior performance on CIKM‑2020 ad targeting benchmarks.
Meituan’s Search Ads team placed third in the KDD Cup 2020 Multimodalities Recall track by tackling training‑test distribution mismatch with diversified negative sampling and distillation learning, and improving text‑image matching via gated fully‑connected layers, bidirectional attention, and diversified fusion, then ensembling neural and tree models for strong NDCG gains later applied to their ad creative‑selection system.
This article provides a comprehensive overview of embedding techniques, explaining their role in deep learning recommendation systems, detailing Word2Vec and its Skip‑gram model with negative sampling and hierarchical softmax, and extending the discussion to Item2Vec and dual‑tower architectures for item representation.
