ESMMEntire Space Multi-task Model
Learn conversion rate indirectly through pCTR × pCVR, so both losses are defined over all impressions.
Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate§1The key contribution
Fix sample selection bias in conversion-rate prediction by never training CVR on clicked samples alone: supervise pCTR and pCTCVR = pCTR × pCVR over the entire impression space.
CVR models are usually trained only on clicked impressions (where a conversion can be observed) but used on all impressions at serving time: sample selection bias. Clicks are also rare, so CVR training data is very sparse.
Use two towers sharing embeddings: one predicts pCTR, the other pCVR. Their product is pCTCVR = p(click & convert | impression). Losses are put only on pCTR (label: click) and pCTCVR (label: click & convert), both defined over every impression. pCVR is learned as a latent quantity, with no loss of its own.
Figure 1. ESMM architecture. Arrows show data flow; annotations show tensor shapes or symbols. Click a block for details, or use the walkthrough to step through the forward pass.
§2Breaking it down
The contribution, piece by piece. Select a card to highlight the blocks it refers to in Figure 1.
§3How it works
ESMM has a shared embedding layer feeding two towers: a CVR tower and a CTR tower. Their outputs are multiplied to give pCTCVR. Training uses two losses over all impressions: cross-entropy of pCTR against click labels, and of pCTCVR against 'clicked and converted' labels. pCVR has no direct loss; it is learned as the factor that makes the product correct. This removes the sample selection bias of training CVR on clicks only and lets the sparse CVR task share embeddings with the dense CTR task. On Taobao data it clearly beat CVR models trained on the click space.
Lineage. It became the template for 'entire-space' and sequential-funnel multi-task models (ESM², AITM) and is often combined with MMoE/PLE towers.
§4Key equations
§5Why it works
- Modeling the funnel structure (impression → click → conversion) directly in the architecture removes the bias at its source.
- It is a simple, general trick that combines with any tower design.
§6Limitations & trade-offs
- pCVR is only identifiable through the product, so when pCTR is tiny, pCVR receives weak gradients.
- It assumes a strict sequential funnel; behaviors that skip steps need extensions (ESM²).