Summary 17 — Web Retrieval, Part II (Architecture, Ranking, Web Spam & User Interaction)

Source: 11_WebIR_Part-II.pdf

1. Table of Contents (Topics Covered)

Gist (2–3 paragraphs)

As the Web outgrew single-server engines, designs moved to massively parallel cluster-based architectures using document partitioning and replication across geographic locations (lower latency, fault tolerance). Variants include task- vs. data-parallel strategies, the distributed Harvest (gatherers + brokers), and multi-site architectures that answer local queries locally and route only externally-dependent queries elsewhere. Caching is critical: query distribution follows a power law, so caching results (and inverted lists) hugely boosts capacity; refinements include Probabilistic Driven Caching (prefetching) and knapsack-based static caching. Tiered indexes put frequent queries in a small fast index and the rest in a larger slow one.

Ranking is the hardest, most important function, drawing on three signal types: content (text/IR score, layout), structure (links, anchor text), and usage (clicks, geographic/temporal context). Link-based ranking exploits the insight that “good pages point to good pages”: PageRank models a random surfer and computes each page’s stationary probability (a Markov chain, with a damping/jump parameter q≈0.15, part of Google’s ranking); HITS computes mutually reinforcing hub and authority scores. Modern engines combine text (e.g., BM25) and link (PageRank) signals linearly (R = α·BM25 + (1−α)·PR), with α tuned (even per query type) via labeled or click data, and increasingly use Learning to Rank (pointwise/pairwise/listwise training) and neural ranking models. Evaluation uses P@5/10/20 (recall is unmeasurable on the Web) plus carefully debiased click-through data. A constant battle is Web spam (spamdexing: keyword stuffing, link farms, click spam, cloaking), distinct from legitimate SEO.

Managing Web data at scale needs special infrastructure: careful document-ID assignment (URL-sorted IDs aid compression), distributed stores (BigTable, HBase, Cassandra, Neo4j, MongoDB, ArangoDB, vector DB Milvus), Web graph compression (exploiting ~80% local links, ~3 bits/link), and duplicate detection via hashing (MD/SHA) and near-duplicate shingling. Finally, user interaction centers on the search rectangle paradigm, simple query languages with operators (+, -, site:, quotes), dynamic query suggestions (5× heavier than search; need spelling correction, diversity, freshness), the SERP layout (organic + sponsored results, title/snippet/URL, oneboxes, universal search), query assistance (“Did you mean” learned from query logs; recommendations via content-ignorant/aware/query-flow methods), faceted refinement, actionable results (translate, cached, stars), and the rise of zero-click search.

2. Key Terminologies — Meaning & Use Cases

Term Meaning Use Case
Cluster-based architecture Partitioned + replicated index across machines Web-scale serving
Document partitioning Split corpus across index servers Parallel query processing
Multi-site architecture Geographically distributed query processing Latency, scalability
Result/list caching Store frequent answers/lists in memory Huge throughput gains
Tiered index Fast small + slow large index Scalability
Ranking signals Content, structure, usage Computing relevance
PageRank Random-surfer stationary probability (Markov) Query-independent quality
Damping factor q Random-jump probability (~0.15) PageRank stability
HITS Mutually reinforcing hubs & authorities Topic-based link ranking
Anchor text Link text describing target page Structural signal
Learning to Rank ML over pointwise/pairwise/listwise data Modern ranking
P@5/10/20 Precision at top positions Web evaluation (recall unmeasurable)
Click-through data (debiased) Implicit relevance from clicks Scalable evaluation/training
Web spam / spamdexing Deceptive ranking inflation (link farms, cloaking) Adversarial IR
SEO Legitimate page-description improvement vs. malicious spam
BigTable / Milvus Distributed store / vector database Web-scale data, embeddings
Shingling Near-duplicate detection via token sequences Deduplication
Search rectangle / SERP Search box / results page User interaction
Dynamic query suggestions Prefix-based as-you-type completions Query assistance
Faceted search Narrow results along dimensions Query refinement
Zero-click search Answer shown directly on SERP Snippets, AI summaries

3. Process Workflow Diagram

flowchart TD
    A[User Query in Search Rectangle] --> B[Dynamic Suggestions / Spelling]
    B --> C[Distributed Cluster
partitioned + replicated index] C --> D[Cache check
results + inverted lists] D --> E[Candidate documents] E --> F[Ranking] F --> F1[Content signals: BM25/text] F --> F2[Structure: PageRank, HITS, anchor text] F --> F3[Usage: clicks, context] F1 --> G[Combine: R = α·BM25 + 1-α·PR
+ Learning to Rank / neural] F2 --> G F3 --> G G --> H[Spam filtering] H --> I[SERP: organic + ads
title/snippet/URL, oneboxes] I --> J[Query assistance / facets / actionable results] J --> K[Click-through data] --> F

4. ELI5 — Complex Terms Explained Simply