Click model

As explained in Experimentation for Engineers

A click model is a predictive model that estimates the likelihood that a user will click a particular advertisement.

A click model estimates the probability that a user will click each advertisement. An ad selector applies those estimates across the available inventory and serves the ad with the highest estimated probability. section 3.1

Figure 3.1 An ad-serving system. The ad selector pulls ads from the ad inventory, then asks the click model to estimate the probability that the user will click on each ad. The ad selector sends the ad with the highest estimated click probability to the user. The page tells the server whether the user clicked on the ad, shown here as a variable clicked taking a value of either 0 (not clicked) or 1 (clicked). The rest of the web page is delivered to the user by a web server, not shown here.

Trace the production path from ad inventory, through click-probability estimates, to ad selection and the logged clicked outcome. section 3.1

The model serves a business metric: click-through rate (CTR). With a $1 cost per click, daily revenue is CPC × CTR × homepage visitors, so improving click predictions can improve revenue. Offline quality is not sufficient when competing models have similar estimates; production evaluation determines which model produces the highest CTR. section 3.1 section 3.2

Comparing two click models

Suppose model A produces a CTR of 0.005 and model B produces 0.007. A simple A/B test randomly serves either model with probability 0.5, logs each clicked value as 0 or 1, and compares the two measured CTRs. A simulated run uses 91,561 measurements and returns z = 2.95. The overall CTR is about 0.006, halfway between the two models because traffic is split evenly. Epsilon-greedy instead increasingly serves the model with the higher observed CTR while retaining some exploration; with four arms whose CTRs range from 0.005 to 0.011, it learns that the best arm is the fourth and achieves about 0.0108. section 3.1.1 section 3.2

Evaluate click models with these practical checks:

• Account for exposure: measuring a model exposes real users to its ads and can increase or reduce revenue during the experiment. section 3.1.1
• Test production behavior when offline CTR estimates tie, because model quality can change after interaction with the ad selector and users. section 3.2
• Tune related ad-selector parameters, such as the probability threshold below which no advertisement is shown. section 3.2
• Test interactions when changing both the click model and business logic; the combined system may perform differently from either change alone. section 3.2

How Experimentation for Engineers covers it

How deep it goes

Explains serving flow, binary click measurement, A/B evaluation, and epsilon-greedy experimentation, but not model-training internals.

Examples

Uses simulated models with CTRs 0.005 and 0.007, a 91,561-measurement A/B test, and four epsilon-greedy arms.

What it assumes

Assumes familiarity with advertisements, probability, CTR, and basic A/B-test reasoning while defining click outcomes from zero.

What you can do after

Evaluate click models in production and configure experiments that balance CTR gains against exploratory exposure.

How it connects

Connects click models with ad selection, click-through rate, A/B testing, epsilon-greedy, business logic, and interaction effects.

liveBookmanning.com

Read Experimentation for Engineers on liveBook Get the book at manning.com

Also explained in

• 

  liveBookmanning.com
  AI-Powered SearchA click model is a model that represents or predicts user clicks on search results to infer their relevance or examination behavior.

Compare how 2 books explain click model →

Related concepts

• probabilistic judgmentproduces · A probabilistic judgment is a relevance assessment represented as a probability that a user considers a result relevant.
• human-in-the-loopproduces · Human-in-the-loop is a system pattern in which people provide input, judgment, review, or intervention within an automated process.
• judgment listproduces · A judgment list is a collection of labeled query-document pairs containing human or system relevance judgments.
• click-through rateuses · Click-through rate is the proportion of displayed or offered items that receive a click.
• simplified dynamic Bayesian networkincludes · A simplified dynamic Bayesian network is a probabilistic click model that estimates search-result relevance from user examination and click behavior.
• implicit feedbackproduces · Implicit feedback is indirect evidence about user preferences or judgments inferred from observed behavior rather than explicitly stated ratings or evaluations.

This explainer was generated by Manning's knowledge graph from the text of Experimentation for Engineers. The section links point to the parts of the book it draws on.