Book Overview and Technology Map

📝 Before You Continue: Please read the three perspectives in 1.1 first. This chapter unfolds that "three-dimensional understanding" into a navigable technology map, helping you locate every chapter that follows.

Once you understand the core logic of recommender systems, the real challenge is turning that understanding into actionable technical solutions. Researchers and engineers have proposed hundreds of algorithms, and beginners often feel lost: how do these models relate to each other? Which one for which scenario? How do you assemble a complete system?

This chapter gives you a map. The book is organized along two main storylines — discriminative and generative — while also threading together the capability evolution from "memorization and generalization" to "understanding and reasoning."

After reading this chapter, you will be able to:

• State the logic behind the book's two halves and where each lands
• Map every chapter of the fundamentals half (Ch0–Ch4) onto the three-stage pipeline or frontier trends
• Explain why "retrieval → ranking → re-ranking" is the organizing thread of the discriminative storyline
• Anticipate how the generative storyline will unfold in later versions (Ch5–Ch10)
• Complete 4 leveled practice problems to verify your grasp of the technology map

1.2.0 One Map, Two Storylines

The book's organization can be summarized in one diagram: the horizontal axis is time/capability evolution (from pure-ID memorization to LLM reasoning), and the vertical axis is the two paradigm storylines (discriminative vs generative).

• First half: industrial practice of discriminative recommendation — progressively deepening along the "retrieval → ranking → re-ranking" pipeline; the main battlefield of memorization and generalization capabilities.
• Second half: the technology landscape of generative recommendation — starting from the foundational paradigm, through Scaling Laws, end-to-end modeling, and reasoning capability to diffusion models; the exploration ground of understanding and reasoning capabilities.

💡 Key Insight: The two storylines are not isolated. Generative architectures often "end-to-end-ify" mature discriminative modules; meanwhile, discriminative techniques such as semantic IDs and feature crossing provide the representational foundation for generative approaches.

1.2.1 First Half: Industrial Practice of Discriminative Recommendation (Covered in This Edition)

This edition (the fundamentals half) fully covers the first half plus the bridging trends chapter — 5 parts in total:

Retrieval: From 100 Million to Thousands (Part 2)

Retrieval is the pipeline's starting point: it must filter from hundreds of millions of items down to thousands of candidates within milliseconds. Its technical evolution unfolds across five sections:

• Collaborative filtering — the classic starting point: from ItemCF's item similarity, through Swing's industrial optimization and UserCF's user perspective, to matrix factorization mapping users/items into latent vectors, pioneering vectorization.
• I2I vector retrieval — transplanting Word2Vec's sequence-modeling ideas into recommendation: from Item2Vec's direct transfer, to EGES fusing side attributes, to Airbnb embedding business objectives into sequence construction.
• Two-tower models (U2I) — encoding users and items separately as vectors, represented by FM, DSSM, and YoutubeDNN, enabling efficient vector retrieval.
• Sequence retrieval — attending to the temporal information earlier methods ignored: MIND uses multiple vectors to represent diverse interests; SDM separates long- and short-term preferences and fuses them dynamically with gating.
• Streaming index retrieval — stepping outside model-internal compression: Trinity uses clustering statistics to preserve full historical interests; Streaming VQ lets the index structure adapt to data distribution in real time.

Ranking: From Thousands to Hundreds (Part 3)

Ranking scores thousands of candidates precisely — the main battlefield of deep generalization:

• Wide & Deep — jointly training a linear model and a deep network, establishing the foundational "memorization + generalization" framework.
• Feature crossing — from FM's second-order crossing, through DeepFM and xDeepFM, toward automatic high-order crossing.
• Sequence modeling — DIN uses attention to dynamically activate history based on the candidate; DIEN explicitly models the temporal evolution of interests.
• Multi-objective / multi-scenario — MMoE and ESMM balance multiple objectives; multi-tower and dynamic weights adapt to cross-scenario differences.

Re-ranking: From Hundreds to One Screen (Part 4)

Ranking output is often highly homogeneous; re-ranking optimizes the whole-list experience while preserving relevance:

• Greedy re-ranking — MMR linearly combines relevance and diversity; DPP uses a determinant framework to control diversity more precisely.
• Personalized re-ranking — PRM uses a Transformer to model mutual influence among items, achieving end-to-end personalized list generation.

1.2.2 Preview of the Second Half: Generative Recommendation (Covered in Later Versions, Ch5–Ch10)

To help you build a complete picture, here is a brief preview of the second half, so you know where the road leads after this edition:

📝 Note: This edition focuses on the fundamentals half (Ch0–Ch4). The second half will be continued in later editions under the same book-writer conventions.

⚠️ Common Mistakes in 1.2

Chapter Summary

📌 Key Takeaways

❓ FAQ

Q1: Why does the fundamentals half stop at Ch4 instead of including Ch5's generative content?

A: Ch5 is the cornerstone of the generative storyline, and its conceptual density rises sharply (semantic IDs, Transformers, diffusion models). First solidifying the industrial practice of discriminative recommendation, then moving into generative, makes for steadier learning.

Q2: Must retrieval, ranking, and re-ranking all use deep learning?

A: No. Retrieval often uses lightweight methods (collaborative filtering, two-tower); only ranking deploys the most complex deep models; re-ranking can use rules or lightweight models. Complexity increases as candidates shrink.

Q3: What order should I read in?

A: Strictly in Part 1→5 order. Later chapters often presuppose earlier ones (e.g., sequence modeling builds on feature crossing).

🔗 Connections to Later Chapters

• Part 2 Retrieval (Ch2.1–2.5) expands this chapter's "Retrieval: from 100 million to thousands" into five algorithm families.
• Part 3 Ranking (Ch3.1–3.5) dives into the industrial implementation of the discriminative scoring function $f$.
• Part 4 Re-ranking (Ch4.1–4.2) wraps up the three-stage pipeline and connects to the trends chapter.
• Part 5 Trends (Ch5.1–5.3) bridges with "debiasing / cold start / generative," pointing toward the second half in later editions.

Practice Problems

Problem 1.2.1 — Locating Chapters 🟢 Easy

Place each of the following algorithms into its correct position in the pipeline (write the corresponding Part/chapter): (a) DPP re-ranking　(b) YoutubeDNN two-tower　(c) DeepFM feature crossing　(d) Swing collaborative filtering

Problem 1.2.2 — Classifying Capability Stages 🟢 Easy

Classify the following techniques into one of the four capability stages — "memorization / generalization / understanding / reasoning": (a) Collaborative-filtering co-occurrence　(b) Deep feature crossing　(c) Semantic IDs (RQ-VAE)　(d) OneRec-Think explicit reasoning

Problem 1.2.3 — Probing Motivations 🟡 Medium

Why did "feature crossing" evolve from FM's second order to xDeepFM's automatic high order? Use a concrete example to illustrate: what kind of pattern would second-order crossing alone miss?

🏆 Challenge: Designing Your Reading Route

Suppose a colleague "only knows SQL and basic logistic regression, has never touched deep learning," but urgently needs to get up to speed on your company's recommendation ranking module within two weeks. Based on this chapter's technology map, design a two-week learning route for them (listing which chapters to read per day/phase, what to skip, and why), and explain your reasoning (within 150 words).