Overcoming Signal Sparsity in Conversion Rate Prediction

32 concepts28 papers1 proposal

In digital advertising and e-commerce, the ultimate goal is often to predict whether a user will perform a high-value action, such as purchasing an item or subscribing to a service. This metric, known as the Conversion Rate (CVR), directly dictates the return on investment for advertisers and the revenue of platforms.

However, predicting CVR is notoriously difficult due to a fundamental structural challenge: conversion signal sparsity. While modern deep learning models thrive on massive datasets, conversion events sit at the very bottom of a deep user-interaction funnel. Billions of users see ads (impressions) and millions click them, but only a tiny fraction—often fewer than 0.1%—ultimately convert.

This extreme data imbalance creates a severe mathematical hurdle. When a model attempts to learn complex feature embeddings for millions of items and users from a handful of positive conversion labels, it suffers from severe overfitting, poorly calibrated probabilities, or gradient vanishing. To make matters worse, traditional models often evaluate CVR only on the subset of items that users clicked, introducing a severe mismatch between training and real-world deployment (sample selection bias).

To build robust CVR models, researchers have shifted away from treating conversion as an isolated binary classification problem. Instead, modern approaches leverage the entire user journey to densify the learning signal.

Entire Space Multi-Task Learning

The most intuitive way to combat the scarcity of conversion labels is to borrow strength from the abundant data available at earlier stages of the funnel, such as click-through rates (CTR).

Traditionally, CVR models were trained exclusively on data where a click had already occurred. As identified by Minxiao Chen et al., this introduces a dual crisis: extreme data sparsity (since clicks are already a small fraction of impressions) and sample selection bias (the model never learns how to evaluate unclicked items, which it must do during real-time ranking).

To solve this, the Entire Space Multi-Task Model (ESMM) framework decomposes the prediction mathematically. Instead of estimating the post-click conversion rate $P(\text{Conversion} \mid \text{Click})$ in isolation, the system models two tasks simultaneously over the entire space of all impressions:

  1. The probability of a click: $P(\text{Click})$
  2. The probability of both a click and a conversion: $P(\text{Click} \times \text{Conversion})$
[ All Impressions ]
       │
       ├─► Task 1: P(Click) ──┐
       │                      ▼
       └─► Task 2: P(Click ∩ Conversion) = P(Click) × P(Conversion | Click)
                              ▲
                              └─ (Shared Embedding Space)

Because $P(\text{Click} \times \text{Conversion}) = P(\text{Click}) \times P(\text{Conversion} \mid \text{Click})$, the actual CVR is derived implicitly. The beauty of this architecture lies in parameter sharing. The embedding layers—which map sparse features like user IDs and ad keywords into continuous vectors—are shared between the dense click task and the sparse conversion task. The abundant gradients from the click task effectively "drag" the shared embeddings into a meaningful representation space, ensuring that the conversion predictor has access to robust features even for rare items.

Building on this foundation, recent frameworks like the one proposed by Fei Ke et al. utilize variational information exploitation and teacher-student structures to maximize the mutual information between these funnel layers, ensuring that the knowledge transferred from the click space does not distort the downstream conversion objective.

Decomposing the Post-Click Journey

While splitting the funnel into clicks and conversions helps, the jump from a click to a purchase is still massive. A user rarely clicks an ad and immediately buys the product; instead, they navigate a sequence of intermediate actions. They might view reviews, add the item to a wishlist, or place it in a shopping cart.

These intermediate steps are known as micro-conversions. They are significantly denser than final purchases, yet they carry a much stronger signal of purchase intent than a simple click.

Hong Wen et al. introduced ESM$^2$, which explicitly models these behavioral pathways. Instead of a linear funnel, the post-click space is treated as a graph of potential user states:

               ┌──► Add to Cart ──┐
               │                  ▼
Click ─────────┼──────────────────┴─► Purchase
               │                  ▲
               └──► Wishlist ─────┘

By formulating the loss function to account for these intermediate states across the entire impression space, the model receives a multi-grained gradient signal. If a user adds an item to a cart but does not buy it, the model doesn't just treat it as a flat negative conversion. It recognizes the high intent, allowing the network to capture subtle behavioral patterns that distinguish casual browsers from high-intent shoppers.

Taking this sequential view further, Menghan Wang et al. leveraged differentiable sorting mechanisms to model user behaviors as an ordered sequence rather than independent multi-task targets. By learning to rank the intensity of these behaviors, the model preserves the natural chronological progression of user intent, stabilizing predictions in low-volume scenarios.

Representation Learning and Auxiliary Supervision

When downstream labels are exceptionally scarce—such as in predicting long-term customer lifetime value (Ruize Wang et al.) or modeling niche advertising markets (Kan Ren et al.)—architectural tweaks alone may not suffice. The model requires a foundational understanding of user and item relevance before it even looks at conversion data.

One powerful strategy is self-supervised pre-training. As demonstrated by Alex Shtoff et al., models can be pre-trained on user interaction sequences using self-supervised objectives (like predicting the next clicked item) before being fine-tuned on sparse conversion labels. This ensures that the underlying neural network weights are already well-optimized, preventing the catastrophic embedding collapse that occurs when training large models on sparse labels from scratch.

Similarly, multi-modal and multi-stage embedding frameworks help bridge this gap. For instance, the ItemSage framework developed by Paul Baltescu et al. combines visual and textual features with multi-task learning. By anchoring item representations in rich, dense content features (like product images and descriptions), the system mitigates the risk of overfitting to frequent but less valuable engagement metrics like clicks.

Countering Optimization Pathologies

Even with multi-task learning and rich representations, standard optimization objectives like Binary Cross Entropy (BCE) can fail under extreme sparsity. Junwei Pan et al. demonstrated that when positive feedback is rare, standard classification losses suffer from severe gradient vanishing on negative samples. The massive volume of negative instances overwhelms the rare positive signals, causing the model's weights to stagnate. Incorporating auxiliary ranking losses provides a crucial "gradient boost," forcing the model to distinguish relative intent boundaries even when absolute positive signals are scarce.

Furthermore, the binary assumption of "clicked but not converted equals a negative sample" is often flawed. Users might have converted if they had seen the item, or they may intend to convert in the future but face a delayed feedback loop. To address this, frameworks like IdeFN (Weiyi Zhong et al.) use optimal transport mechanisms to match unclicked items with known conversion distributions, generating reliable pseudo-labels that densify the training space and correct for unobserved positive intent.

Key Foundational Work

Robust classification of rare queries using web knowledge (2007) — contributed 1 concept elements to this direction.
Predicting bounce rates in sponsored search advertisements (2009) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_data_sparsity_conversion_data_sparsity_4e13f718,
  title = {data sparsity (conversion data sparsity)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/data-sparsity-conversion-data-sparsity/rdr_4e13f718df784250ba03c8ce733e3a21},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_data_sparsity_conversion_data_sparsity_4e13f718_v1,
  title = {data sparsity (conversion data sparsity)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/data-sparsity-conversion-data-sparsity/rdr_4e13f718df784250ba03c8ce733e3a21/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake