Paper: Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate, Ma et al., 2018.
A CVR model predicts conversion after a click, yet conversion labels are available only for clicked impressions while online ranking covers every impression. ESMM starts from this mismatch between the training and inference spaces.
Two connected problems
A conventional CVR model trained only on clicked samples faces:
- Sample-selection bias: clicks are not random, so clicked traffic differs from all exposed traffic.
- Data sparsity: conversions are much rarer than clicks, leaving little supervision for the CVR task.
Treating every unclicked impression as a negative CVR label would confuse “no opportunity to convert” with “clicked but did not convert.” ESMM instead uses the sequential structure of exposure → click → conversion.
The probability chain
Let CTR = P(click | impression) and CVR = P(conversion | click, impression). The probability of both clicking and converting is:
CTCVR = P(click & conversion | impression)
= CTR × CVR
ESMM trains CTR and CVR towers, then multiplies their outputs to obtain CTCVR. Its losses use CTR and CTCVR labels defined over the entire impression space. The CVR tower is therefore constrained through an observable full-space target instead of a standalone loss over only clicked examples.
ESMM does not invent CVR labels for unclicked impressions. Probability multiplication connects an unobserved conditional probability to a joint event that can be labeled across all impressions.
Shared representation and sparse supervision
The two towers share lower-level embeddings. The denser click task helps sparse features learn stable representations, and the CVR tower uses the same representation to capture conversion intent.
This is classic multi-task transfer: shared layers carry common information while task towers preserve objective-specific behavior. Too little sharing offers limited transfer; too much can create negative transfer. Gradient scale, conflict, and training dynamics deserve monitoring in a production system.
Implementation checks that matter
Beyond reproducing the diagram, I would verify:
- whether exposure, click, and conversion labels follow a consistent funnel definition;
- whether attribution windows and delayed conversions are handled consistently;
- whether the
pCTR × pCVRoutput remains numerically stable and calibrated; - whether evaluation covers CTR, CTCVR, CVR, and important business slices.
CVR quality should not be reduced to AUC on clicked samples. Calibration and bias across traffic positions and user groups matter, and ranking value ultimately needs online experiments.
What it solves—and what it does not
ESMM provides an elegant framework for mismatched sample spaces and sparse supervision, but entire-space training does not guarantee a perfectly unbiased estimator. Later work has examined bias introduced by probability composition, task dependencies, and the data-generating process. Real funnels may also contain cart, favorite, or other intermediate actions that a single click-conversion chain cannot express.
My takeaway
ESMM first respects the probability structure of business events and then designs the multi-task objective. Instead of manufacturing labels for missing outcomes, it finds a fully observable constraint.
When supervision is naturally missing, better modeling often begins by reconsidering the relationships among observable quantities.
DISCUSSION