Clustering & Segmentation

Recency-Frequency-Monetary framework still widely used in retail and e-commerce.

Probability models (BG/NBD) for CLV estimation enabling value-based segmentation.

Clustering taste profiles to power personalized playlists at scale.

Framework for predicting individual-level CLV using past transaction data—foundational for value-based marketing.

Pareto/NBD model for customer counting and transaction prediction—still used at scale in industry.

Linking customer acquisition, retention, and expansion to CLV—connects marketing spend to lifetime value.

Definitive k-means formulation with convergence guarantees—still the default.

Ward's method for agglomerative clustering minimizing within-cluster variance.

Parallel k-means initialization achieving logarithmic rounds—default in Spark MLlib.

Density-based clustering finding arbitrarily shaped clusters and handling noise—foundational for spatial clustering.

Hierarchical clustering using CF-trees for single-scan scalability—enables clustering of millions of records.

Density-based ordering producing cluster hierarchy without fixed epsilon—extends DBSCAN for variable densities.

Non-parametric mode-seeking algorithm for clustering without specifying number of clusters.

Mini-batch k-means enabling streaming updates—Google's approach for web-scale clustering.

Comprehensive treatment of Gaussian and non-Gaussian mixture estimation.

Foundational discrete mixture model for categorical response patterns.

Scalable non-parametric Bayesian clustering without specifying K.

Generative probabilistic model for topic modeling—foundational for discovering latent topics in text collections.

Non-parametric Bayesian approach for sharing clusters across grouped data—automatic topic number selection.

Gaussian mixture model framework with automatic model selection via BIC—implemented in R's mclust package.

Joint representation learning and clustering via autoencoders.

Theoretical foundation for spectral methods in network/embedding clustering.

Self-supervised contrastive objectives for cluster-friendly representations.

Foundational spectral clustering using Laplacian eigenvectors—NeurIPS best paper, widely implemented.

Self-supervised visual learning via online clustering—state-of-the-art for unsupervised image representations.

Two-step unsupervised classification via representation learning then clustering—strong ImageNet results.

VAE-based clustering combining variational autoencoders with GMM priors for end-to-end learning.

Iterative clustering and CNN training for unsupervised feature learning—Facebook AI's breakthrough.

Causal forests for estimating HTEs and deriving optimal targeting policies.

Foundational uplift/treatment-effect modeling enabling segment-specific interventions.

Large-scale user segmentation powering real-time recommendations.

Segment-based image selection improving engagement through visual personalization.

Linear UCB algorithm for personalization at Yahoo—foundational contextual bandit for segment-based recommendations.

Production system for personalization via contextual bandits—deployed across Microsoft products.

Dynamic user clustering for bandits—learns segment structure while optimizing recommendations.

Benchmark dataset for music analysis enabling audio feature clustering research at scale.

Survey of audio feature extraction for music similarity and clustering—foundational for MIR.

Deep generative model for audio—enables learned audio embeddings for clustering.

Standard Python library for audio feature extraction—MFCCs, spectrograms for clustering.

CNN for music auto-tagging enabling tag-based clustering and organization.

Audio analysis API powering playlist generation and music clustering at scale.

Benchmark for genre classification—evaluates clustering approaches on real music data.

Two-tower architecture for video clustering and candidate generation at YouTube scale.

Overview of Netflix's recommendation system including movie clustering and personalization.

8 million videos with labels for video understanding—benchmark for video clustering research.

Simple but effective collaborative filtering for movie recommendations—Netflix competition winner approach.

Netflix Prize winning approach using latent factors—foundational for content clustering.

VAE-based collaborative filtering—Netflix research on implicit feedback clustering.

Survey of player modeling techniques including behavior clustering and segmentation.

Combining survival models with player clusters for churn prediction in games.

Foundational study of player motivations and behavioral clustering in online games.

Survey of ML for game content—includes UGC clustering and content organization.

Comprehensive guide to mining game data including player segmentation techniques.

Collapsed Gibbs sampling for short text clustering—handles sparse, short documents like tweets.

Efficient sentence embeddings for semantic similarity—enables fast document clustering.

State-of-the-art short text clustering combining contrastive learning with self-training.

Comprehensive survey of document clustering methods from TF-IDF to neural approaches.

Multi-task visual embeddings for image clustering and similarity search at Pinterest scale.

ResNet architecture enabling powerful visual features for image clustering.

Vision-language model enabling zero-shot image clustering via text descriptions.

Evolution of visual search and image clustering at Pinterest—practical lessons at scale.