Foundations of Recommender Systems: Paradigms, Taxonomies, and Core Evaluation Metrics

The history of recommender systems (RecSys) traces back to 1992, when Xerox PARC launched *Tapestry*, the first system to apply collaborative filtering (CF) to filter emails and documents for enterprise users. Over the past three decades, RecSys has evolved from heuristic rule-based systems to deep learning-powered architectures that power 80% of Netflix’s viewing volume, 35% of Amazon’s total sales, and 70% of TikTok’s video plays as of 2024. For advanced engineers, mastering the foundational paradigms, taxonomies, and evaluation frameworks is not just an academic exercise—it is the prerequisite for

is not just an academic exercise—it is the prerequisite for designing systems that avoid the cold start trap, mitigate selection bias, and align offline metrics with real-world business outcomes. To build this foundational intuition, we first trace the three distinct evolutionary waves of recommender systems (RecSys) over the past 30 years, each defined by a paradigm shift in how user preferences are modeled and matched to items.

1.1 Evolutionary Waves of Recommender Systems

The first wave (1992–2005) was dominated by heuristic and memory-based collaborative filtering (CF) approaches. The field was launched in 1992 by Xerox PARC’s *Tapestry* system, the first to apply CF to filter enterprise emails and documents by leveraging user-generated annotations of “relevant” content. This was quickly followed by the GroupLens project at the University of Minnesota, which extended CF to filter Usenet news articles for end users, formalizing the core intuition that “users with similar past preferences will have similar future preferences.” The most impactful industrial deployment of this era was Amazon’s 1998 launch of item-based CF for its “Customers Who Bought This Also Bought” feature, which drove a documented 35% lift in cross-sell revenue by recommending complementary products to shoppers. The core algorithms of this era—user-based CF (UserCF) and item-based CF (ItemCF)—relied on explicit similarity calculations between users or items, with prediction scores computed via weighted averages of neighbor ratings. For UserCF, the predicted rating for user $u$ on item $i$ is formalized as: $$\hat{r}_{u,i} = \frac{\sum_{v \in N(u)} \text{sim}(u,v) \cdot r_{v,i}}{\sum_{v \in N(u)} |\text{sim}(u,v)|}$$ where $N(u)$ is the set of $k$ nearest neighbors of user $u$, $\text{sim}(u,v)$ is the similarity score between users $u$ and $v$ (typically cosine similarity or Pearson correlation), and $r_{v,i}$ is the rating user $v$ gave to item $i$. For ItemCF, the equivalent formula swaps user and item roles: $$\hat{r}_{u,i} = \frac{\