In search engines and recommendation systems, user clicks are often treated as "votes" for quality. However, a click is a noisy signal. A user might click the first result simply because it is at the top (position bias), because it has an attractive thumbnail (appearance bias), or because they trust the system's brand (trust bias).
The research direction of Unbiased Learning to Rank (ULTR) seeks to recover the "true" relevance of an item by mathematically decoupling these presentation biases from the underlying content quality.
The Generative Model of a Click
To understand why these methods work, we must first view a click not as a single event, but as the outcome of a two-step generative process. As framed by Jun Wang and Marc Sloan, a user clicks on a document only if:
- Examination: They actually look at the document (influenced by its position).
- Relevance: They find the document useful (influenced by its content).
Mathematically, this is often expressed as $P(\text{Click}) = P(\text{Examination}) \times P(\text{Relevance})$. The challenge is that we only observe the click; both examination and relevance are latent variables—hidden factors we must infer.
Decoupling Bias with Expectation-Maximization
The core mechanism used across this field to solve for these hidden variables is the Expectation-Maximization (EM) algorithm.
In the context of community question-answering, Heyan Huang and Xiaochi Wei demonstrated that by iteratively estimating the hidden quality parameters of an answer and the bias of its position, a system can "un-bias" voting data. The EM process works like a feedback loop: it assumes a certain level of bias to estimate relevance, then uses that relevance to refine its estimate of the bias, repeating until the values converge.
This approach allows systems to learn from biased historical logs without requiring expensive, randomized "interleaving" experiments that might degrade the user experience, as noted by Michael Bendersky and Donald Metzler.
From Simple EM to Regression-Based Estimation
While traditional EM works well for items with many interactions, it struggles with "sparse" data—items that have only been seen a few times. To solve this, researchers like Xuanhui Wang and Marc Najork introduced regression-based EM.
Instead of treating the relevance of every single document as an independent parameter to be learned, they use a regression function (like a neural network) to predict relevance based on document features. This allows the model to generalize: if it learns that "technical depth" makes a document relevant in one context, it can estimate the relevance of a brand-new document with similar features, even if that document has never been clicked before.
Recent reproducibility studies by Jiaxin Mao and Zechun Niu have confirmed that this regression-based approach is significantly more robust than traditional EM, particularly when dealing with the complexities of real-world search logs.
Addressing Complex Biases: Trust and Visibility
The field has evolved to handle more than just position. Ali Vardasbi, Harrie Oosterhuis, and Maarten de Rijke identified that standard Inverse Propensity Scoring (IPS) fails when users have "trust bias"—the tendency to click a top result because they assume the search engine is always right. They proposed affine estimators that use EM to simultaneously estimate examination probabilities and rank-based trust offsets.
Similarly, in eCommerce, Yibo Wang and Yanbing Xue treat item visibility as a latent variable. By neutralizing this visibility during the "inference" phase (assuming all items are equally visible), the model can rank products purely on their intrinsic utility to the shopper.
The Reality Gap
Despite the mathematical elegance of these "unbiased" models, real-world application remains challenging. A large-scale study on Baidu's dataset by Romain Deffayet and Philipp Hager revealed a sobering truth: being better at predicting what a user will click (which includes predicting their biases) does not always translate to finding what is actually relevant. This suggests that while EM-based decoupling is powerful, the ultimate goal of ULTR requires a careful balance between modeling user behavior and identifying objective utility.