Two-Dimensional Position Bias in Information Retrieval

36 concepts18 papers1 proposal

For decades, search and recommendation engines operated under a convenient simplification: users scan results like a vertical grocery list. This is the linear scanning assumption. In this traditional view, position bias is one-dimensional—an item's probability of being examined decays monotonically as its rank drops from top to bottom.

However, modern digital interfaces have evolved far beyond the single-column list. E-commerce platforms display products in grids; streaming services organize content into rows of horizontal carousels; mobile search results combine vertical text blocks with horizontal swipeable elements; and travel platforms embed interactive maps.

When a two-dimensional interface is flattened into a one-dimensional list to calculate position bias, the underlying mathematics breaks down. To build accurate ranking models and evaluation metrics, we must replace linear rank discounts with models that understand the geometry of two-dimensional attention.

The Collapse of the Linear Scan

The mismatch between linear models and non-linear interfaces manifests clearly in user behavior. As Aleksandr Chuklin et al. demonstrate in Incorporating Clicks, Attention and Satisfaction into a Search Engine Result Page Evaluation Model, modern layouts featuring sidebars and heterogeneous content blocks cause user attention to move non-linearly, making traditional click models poor predictors of true item examination. Furthermore, Ryen W White et al. show in Improving searcher models using mouse cursor activity that vertical navigation and scrolling redistribute attention across a page in ways that standard top-to-bottom models fail to capture.

When recommendation systems ignore these spatial nuances, item placement becomes sub-optimal. Choon Hui Teo et al. note in Adaptive, Personalized Diversity for Visual Discovery that standard positional corrections fail in grid-based visual feeds. Similarly, Yingfei Wang et al. argue in Efficient Ordered Combinatorial Semi-Bandits for Whole-Page Recommendation that whole-page layouts create complex visual hierarchies and spatial biases that cannot be modeled by simple sequential rankings.

The Geometry of 2D Attention

Eye-tracking and interaction studies reveal that user attention in 2D viewports follows distinct geometric patterns rather than uniform gradients.

Z-Shaped and F-Shaped Scanning

In grid layouts and multi-block mobile pages, users often read horizontally across a row or block before dropping down vertically to the next segment. In Constructing an Interaction Behavior Model for Web Image Search, Jiaxin Mao models this multi-directional movement as a "Z-shaped" browsing path. On mobile devices, Lingyue Fu et al. capture this in An F-shape Click Model for Information Retrieval on Multi-block Mobile Pages, demonstrating that layouts combining vertical lists with horizontal scrolling elements violate standard top-to-bottom assumptions.

The Golden Triangle and Corner Bias

In interfaces featuring multiple horizontal carousels (such as Netflix or Spotify), attention concentrates heavily in the top-left corner and decays diagonally. In Measuring the Ranking Quality of Recommendations in a Two-Dimensional Carousel Setting, Nicolò Felicioni et al. reveal a striking consequence of this geometry: a highly relevant item placed at the beginning of a secondary row often yields higher utility to the user than a moderately relevant item placed at the end of the primary row.

Middle Bias and Row Skipping

Users navigating image or product grids do not always treat rows uniformly. Sirui Chen et al. observe in Reinforcement Re-ranking with 2D Grid-based Recommendation Panels that users exhibit "middle bias" (focusing on the center of a row) and "row skipping" (bypassing entire rows if the initial items fail to capture interest). This behavior is echoed by Jiaxin Mao et al. in Grid-based Evaluation Metrics for Web Image Search, confirming that users scan both horizontally and vertically to evaluate grid rows.

Radial Decay

Map interfaces present a completely different spatial paradigm. Instead of decaying from a corner, user attention follows a radial pattern centered on the viewport's middle coordinates. In Learning to Rank for Maps at Airbnb, Malay Haldar et al. show that engagement is highly concentrated at the center of the map display and diminishes symmetrically toward the periphery.

Linear 1D Decay      Diagonal 2D Decay      Radial Map Decay
   [ 1 ]                [ 1 ] [ 2 ] [ 3 ]         [ . ] [ . ] [ . ]
   [ 2 ]                [ 2 ] [ 3 ] [ . ]         [ . ] [ X ] [ . ]
   [ 3 ]                [ 3 ] [ . ] [ . ]         [ . ] [ . ] [ . ]
(Top-to-Bottom)        (Golden Triangle)         (Center Outward)

Mathematically Modeling 2D Position Discounts

To formalize these patterns, researchers replace the traditional 1D rank discount factor with a coordinate-based discount function $D(x, y)$, where $x$ and $y$ represent horizontal and vertical positions (or row and column indices).

In multi-carousel interfaces, Nicolò Felicioni et al. in A Methodology for the Offline Evaluation of Recommender Systems in a User Interface with Multiple Carousels and Maurizio Ferrari Dacrema et al. in Evaluating Recommendations in a User Interface With Multiple Carousels decouple the user's physical effort by applying separate discount factors for vertical scrolling and horizontal swiping:

$$D(x, y) = f(x) \cdot g(y)$$

By tuning the weights of $f(x)$ and $g(y)$, the model can adapt to the specific interaction costs of different devices, such as desktop clicks versus mobile swipes.

For map viewports, the discount is modeled as a function of Euclidean distance from the center coordinate. Malay Haldar et al. optimize "spatial utility" by multiplying an item's relevance probability by this radial attention decay function. This allows the system to algorithmically determine the optimal viewport boundaries and pin density.

However, 2D attention is not purely geometric. In image search, Jiaxin Mao et al. note in Preference-based Evaluation Metrics for Web Image Search that visual saliency plays a major role; users frequently skip across a layout based on the visual prominence of an image rather than strict physical proximity.

Evaluation and Optimization in 2D Layouts

Accounting for two-dimensional position bias requires updating both how we evaluate systems offline and how we optimize rankings online.

2D Evaluation Metrics

Traditional metrics like Normalized Discounted Cumulative Gain (NDCG) assume a linear sequence. To evaluate grid and carousel interfaces accurately, the authors of Measuring the User Satisfaction in a Recommendation Interface with Multiple Carousels introduced NDCG2D (and its variant, N2DCG), which applies joint horizontal and vertical position discounts. Similarly, Jiaxin Mao et al. establish grid-based attention models in Modeling User Behavior for Vertical Search: Images, Apps and Products to calculate position bias across rows and columns simultaneously, aligning offline metrics with actual human satisfaction.

Combinatorial Layout Optimization

When position bias operates in two dimensions, ranking algorithms must evolve from sorting a list to solving a combinatorial slot-allocation problem.

Sirui Chen et al. address this with Panel-MDP in Reinforcement Re-ranking with 2D Grid-based Recommendation Panels, a reinforcement learning framework that treats item placement in a 2D grid as a sequential decision-making process, explicitly accounting for row skipping and middle bias. Taking this a step further, Shuai Xiao et al. introduce Tile Networks in Tile Networks: Learning Optimal Geometric Layout for Whole-page Recommendation. This architecture uses reinforcement learning to jointly select items and determine their optimal geometric layout on a page, moving entirely past static, fixed-order placement strategies.

Key Foundational Work

Incorporating Clicks, Attention and Satisfaction into a Search Engine Result Page Evaluation Model (2016) — contributed 1 concept elements to this direction.
Improving searcher models using mouse cursor activity (2012) — 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_two_dimensional_position_bias_5a2e3314,
  title = {two-dimensional position bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/two-dimensional-position-bias/rdr_5a2e3314e3ca454d984a12d01e6102f4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_two_dimensional_position_bias_5a2e3314_v1,
  title = {two-dimensional position bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/two-dimensional-position-bias/rdr_5a2e3314e3ca454d984a12d01e6102f4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake