Awesome Search KG

Neural Click Models

4 min read

Neural Click Models

Definition

Neural click models replace the hand-specified parametric structure of classical Click Models (PBM, Cascade, DBN, UBM) with neural networks that learn distributed representations of queries, documents, and browsing state. Instead of estimating one tabular attractiveness/examination parameter per (query, document) pair, they encode context into embeddings and predict click probability from learned functions — letting the model generalize to rare and unseen query-document pairs.

They serve the same purpose as classical click models — de-biasing click logs to recover relevance for Learning to Rank and offline evaluation — but trade interpretability for expressiveness and tail coverage.

Why Go Neural

Classical click models have two structural weaknesses neural models target:

  1. Sparsity / no generalization. PBM, DBN, and UBM store parameters per query-document pair. Tail queries and freshly indexed documents have too few impressions to estimate these reliably — the model has nothing to fall back on. Neural models share parameters through embeddings, so signal transfers across similar queries and documents.
  2. Rigid behavioral assumptions. The cascade family hard-codes a specific scan-and-stop process and strong independence assumptions. Neural sequence models learn the dependency between examinations and prior clicks from data rather than asserting it.

Common Models

NCM (Neural Click Model)

Introduced by Borisov et al. (WWW 2016). Models a browsing session as a sequence of vector states processed by an RNN: at each rank the hidden state encodes everything seen so far, and click probability is read off from that evolving state. Drops the explicit examination hypothesis — the network learns examination and relevance jointly rather than factoring them by hand.

CACM (Context-Aware Click Model)

Chen et al. (SIGIR 2020). Keeps the classical relevance × examination factorization but makes both components neural and context-aware: a relevance estimator and an examination predictor are learned over query/document/session context, then combined. More interpretable than NCM while still benefiting from learned representations.

GraphCM

Frames click modeling as inference over a query–document bipartite graph and uses a graph neural network to propagate signal between neighboring nodes. Directly attacks the data-sparsity problem: a tail document inherits behavioral signal from documents that co-occur with the same queries.

Two-Tower / Additive Position Models

Used in industrial unbiased LTR (e.g. Huawei’s PAL, and additive two-tower setups at web-search scale). One tower predicts examination from position-only features, a second predicts relevance from query-document features, and the towers are combined (multiplicatively or additively). At serving time the position tower is dropped, leaving an unbiased relevance score. A neural restatement of the examination hypothesis trained end-to-end with the ranker.

Relationship to Unbiased LTR

Neural click models are one half of the unbiased learning-to-rank picture:

  • Click-model route — fit a (neural) click model to logs, then use its relevance posteriors as soft labels for Learning to Rank.
  • Counterfactual route — estimate propensities (often via a neural examination model) and reweight clicks with IPS (Inverse Propensity Scoring) directly inside the ranking loss.

The boundary blurs: two-tower additive models are both the propensity estimator and the ranker, trained jointly.

Tradeoffs

StrengthCost
Generalizes to tail / cold-start query-doc pairsData-hungry; needs large click logs
Captures rich session and context dependenciesLower interpretability than PBM/DBN parameters
Shares signal via embeddings and graphsHarder to extract clean, reusable propensities
Trains end-to-end with the rankerMore moving parts; risk of feedback loops if not debiased

Graph View

Backlinks