For decades, the foundation of information retrieval was built on a simple, one-dimensional assumption: the cascade model. This model posits that users scan search results like a grocery list, starting at the top and proceeding linearly downward until they find what they need or give up.
However, modern interfaces have outgrown the list. From the image grids of Pinterest to the horizontal carousels of Netflix and the interactive maps of Airbnb, the "rank" of an item is no longer a single integer. It is a coordinate in a two-dimensional space. Researchers have found that traditional models fail in these environments because they ignore the spatial geometry of human attention.
The Failure of the Linear Assumption
In a standard list, position bias is straightforward—attention decays as you move down. But as Ryen W. White and colleagues demonstrated in Improving searcher models using mouse cursor activity, user interaction is far more dynamic. By tracking "invisible" signals like mouse movements and scrolling, they showed that vertical movement redistributes attention in ways that linear models cannot predict.
When results are arranged in a grid, the linear assumption collapses entirely. Jiaxin Mao and Xiaohui Xie have shown that users in 2D environments exhibit non-linear scanning patterns, such as a "Z-shaped" trajectory—moving horizontally across a row before dropping to the next. In these layouts, an item's relevance is filtered through a "middle-bias" or a "golden triangle" effect, where the center or top-left of the screen captures disproportionate focus regardless of the item's actual rank.
Modeling the Geometry of Attention
To capture these behaviors, researchers have moved toward models that treat the interface as a graph or a coordinate plane.
- Grid-based Browsing: In Constructing an Interaction Behavior Model for Web Image Search, Jiaxin Mao introduced the Grid-based User Browsing Model (GUBM). Unlike list models, GUBM accounts for the 2D distance between items and the "Z-shaped" path users take, allowing the system to infer which images were actually seen versus simply skipped.
- Mobile Multi-block Layouts: Mobile screens often mix vertical lists with horizontal "blocks." Lingyue Fu and Jianghao Lin proposed the F-shape Click Model (FSCM), which uses Directed Acyclic Graphs (DAGs) to represent how attention branches and merges across these complex blocks.
- Radial Decay on Maps: On spatial interfaces like Airbnb, the concept of "rank 1" is meaningless. Malay Haldar and the Airbnb team found that attention follows a radial decay centered on the map's viewport. In Learning to Rank for Maps at Airbnb, they argue that the density of high-quality results near the center of the map is more important than their specific ordering.
Evaluating Spatial Utility
If the way users look at a page changes, the way we measure success must change too. Traditional metrics like Normalized Discounted Cumulative Gain (NDCG) assume a 1D decay. If a relevant item is at the start of the second row, NDCG might penalize it as being "rank 5," even if the user’s eyes jump there immediately after "rank 1."
To fix this, Nicolò Felicioni and others proposed NDCG2D and N2DCG. These metrics, detailed in Measuring the Ranking Quality of Recommendations in a Two-Dimensional Carousel Setting, apply discount factors based on both row ($y$) and column ($x$) coordinates. This allows researchers to model the "interaction cost" of swiping a carousel versus scrolling down a page. As Mária Bieliková discovered through eye-tracking in Riding the Carousel, users often exhibit a "reverse F-pattern" after swiping, fixating on new items on the right before scanning back to the left.
Learning to Rank in Two Dimensions
The ultimate goal of this research is to optimize the layout itself. Traditional "Learning to Rank" (LTR) produces a list, but new frameworks like Panel-MDP, introduced by Xiao Zhang in Reinforcement Re-ranking with 2D Grid-based Recommendation Panels, treat the 2D grid as a sequential decision-making problem.
Similarly, the Double-Rank Model (DRM) by Harrie Oosterhuis and Maarten de Rijke attempts to solve for two things at once: which items are relevant, and which display positions (like a featured hero slot versus a side banner) the user prefers. By moving from a 1D list to a 2D coordinate system, these models ensure that the most relevant content is placed not just "first," but where the eyes are actually going to land.