ESMM Entire Space Multi-Task Model

ESMM (Entire Space Multi-Task Model, proposed by Alibaba Alimama at SIGIR 2018) is a multi-task learning architecture specifically engineered to resolve the twin industry dilemmas in Conversion Rate (CVR) estimation: Sample Selection Bias (SSB) and Data Sparsity (DS); using probability chain rules: $p\text{CTCVR} = p\text{CTR} \times p\text{CVR}$, ESMM trains CTR and CTCVR simultaneously across the entire un-gated impression space, implicitly deriving $p\text{CVR} = \frac{p\text{CTCVR}}{p\text{CTR}}$ without conditioning training exclusively on clicked samples.

E-commerce purchase conversion rate estimation, performance advertising bidding (pCVR), and multi-step conversion funnel modeling.

Conventional CVR models train solely on clicked items but serve across the full impression pool, triggering severe distribution shifts (SSB) and over-fitting on sparse positive data; ESMM's full-space joint training completely eliminates selection bias and data sparsity.

Derive the joint loss function of ESMM over full impression space combining CTR and CTCVR targets?

Why does online serving extract the raw output of the CVR Tower directly as the predicted $p\text{CVR}$ without dividing by CTR?

How does embedding weight sharing between CTR and CVR towers leverage abundant CTR impressions to alleviate CVR data sparsity?

How does ESM2 extend the ESMM framework across multi-stage conversion DAG funnels (Impression -> Click -> Cart -> Purchase)?

Why does optimizing CVR via explicit division cause numerical instability, and how does multiplicative CTCVR loss avoid gradient explosions?