Simple, Fresh, and Fair: Revisiting Linear Collaborative Filtering for Large-Scale Job Recommendation

Abstract

Job-board recommendation is shaped by two pressures that are rarely studied together: which model family wins under a single controlled protocol, and how quickly recommendation quality decays as the job catalog churns. We study both on a large real-world job-application log (about 40.4 million applications, August to November 2025). We benchmark nine recommenders, spanning linear (EASE, Temporal EASE), neighborhood (ItemKNN), latent-factor (PureSVD), a production word2vec model (Prod2Vec/ATTR), deep sequential (SASRec, BERT4Rec), and graph (LightGCN) families, under a single catalog, item-identifier space, candidate set, and metric function, on both a leave-one-out and an out-of-time split. A shallow linear item-item model (Temporal EASE, with EASE close behind) is the strongest overall and by far the cheapest, training in seconds rather than up to ninety minutes. In-sample it clearly outranks the deep and production models; strengthened deep models become competitive out-of-time, yet the shallow models keep the lead in aggregate, the diversity advantage, and a 100- to 3000-fold lower training cost. Second, we quantify catalog freshness as the dominant production bottleneck: a frozen model loses roughly an order of magnitude of its real-world reach within three months, whereas cheap rolling-window retraining of the same model keeps reach essentially constant. The shallow-model ranking also replicates on a public job-classifieds dataset. For high-churn job recommendation, retraining cadence matters more than model family, and a fast linear model is a strong, economical default.