Surrounding Item Bias

Definition

Surrounding Item Bias

Surrounding Item Bias is the violation of the Examination Hypothesis in which the probability of examining (and clicking) the item at rank $k$ depends not only on its own position and relevance, but on the features of the items displayed around it (ranks $k-1,k,k+1,\dots$). The relative appearance of an item — how it contrasts with its neighbours — modulates the user’s attention, so two identical items can receive different click rates purely because of who their neighbours are.

It is the context-dependent generalisation of Outlier Bias: instead of one distinctive item attracting clicks in absolute terms, the bias is framed as a function of the local list context, where any item that stands out from its immediate surroundings gains examination probability and any item that blends in loses it.

Intuition

The classic Position-Based Click Model assumes examination is a property of the rank slot alone: $P({\text{Exam}}_{d,k})=P({\text{Exam}}_k)$, identical for every document placed there. Surrounding item bias says this is false — the eye is drawn by contrast, not by absolute position.

Same item, two contexts (item X is identical in both):

Context A (X blends in):        Context B (X stands out):
[ Blue $20 ]                    [ Blue $20 ]
[ Blue $22 ]                    [ Blue $22 ]
[ Blue $21 ]  ← X, $21, blue    [ Red  $90 ]  ← X' neighbours differ
[ Blue $19 ]                    [ Blue $19 ]

X here looks ordinary           ...but if X is the Red $90 outlier,
→ baseline examination          its neighbours' uniformity makes
                                it pop → examination boosted

Two mechanisms are at play, mirroring the two factors of the Examination Hypothesis:

1. Attention contrast (examination): an item that is visually or feature-wise different from its neighbours is examined more — the Outlier Bias effect, but defined relative to the local window.
2. Comparative relevance (context bias / Trust Bias-like): an item’s perceived relevance shifts when it is shown next to much better or much worse alternatives (a contrast/anchoring effect), violating $P({\text{Rel}}_d\mid q,\text{context})=P({\text{Rel}}_d\mid q)$.

Because examination now depends on neighbours, clicks are no longer a clean signal of position $\times$ relevance, and naive Inverse Propensity Weighting (which assumes position-only propensities) becomes biased.

Mathematical Formulation

The PBM factorises a click as examination times relevance. Surrounding item bias replaces the position-only examination term with a context-conditioned one:

$P({\text{Click}}_{d,k}\mid {\mathcal{C}}_k)=\underset{\text{context-dependent}}{\underbrace{P({\text{Exam}}_k\mid {\mathcal{C}}_k)}}\cdot P({\text{Rel}}_d\mid q)$

where:

• ${\mathcal{C}}_k=(x_{k-w},\dots ,x_{k-1},x_k,x_{k+1},\dots ,x_{k+w})$ — the context window of feature vectors of the items surrounding rank $k$ (window half-width $w$)
• $x_j$ — feature vector (price, rating, colour, visual descriptors, $\dots$) of the item at rank $j$
• $P({\text{Exam}}_k\mid {\mathcal{C}}_k)$ — examination probability modulated by the surrounding items, no longer a constant for slot $k$
• $P({\text{Rel}}_d\mid q)$ — true (context-free) relevance of document $d$

A common additive parameterisation isolates a baseline position term plus a distinctiveness bonus:

$P({\text{Exam}}_k\mid {\mathcal{C}}_k)=\min (1,{\beta }_k+\alpha \cdot s_{\text{dist}}(x_k,{\mathcal{C}}_k))$

where:

• ${\beta }_k$ — the standard position-bias examination probability for rank $k$ (the PBM term)
• $\alpha$ — sensitivity of attention to local distinctiveness ($\alpha =0$ recovers the PBM)
• $s_{\text{dist}}(x_k,{\mathcal{C}}_k)$ — a distinctiveness score of item $k$ relative to its window, e.g. a normalised deviation from the local neighbours:

$s_{\text{dist}}(x_k,{\mathcal{C}}_k)={\max }_{f\in \text{features}}\frac{|x_k^{(f)}-{\mathrm{median}}_{j\in {\mathcal{C}}_k}{x}_j^{(f)}|}{{\mathrm{std}}_{j\in {\mathcal{C}}_k}{x}_j^{(f)}+ϵ}$

In neural click models the whole context map is learned directly, the window of features being fed to a network whose output is the examination logit:

$P({\text{Exam}}_k\mid {\mathcal{C}}_k)=\sigma (f_{\theta }(x_{k-w},\dots ,x_k,\dots ,x_{k+w}))$

where $\sigma$ is the sigmoid and $f_{\theta }$ a learned function (e.g. an MLP or attention layer over the window) that captures arbitrary neighbour interactions.

Debiased Listwise Objective

To learn true relevance from context-biased clicks, weight each click by the inverse context-dependent propensity: ${\hat{\mathcal{L}}}_{\text{ULTR}}={\sum }_d\frac{c_d}{P({\text{Exam}}_{k_d}\mid {\mathcal{C}}_{k_d})}\ell (f(x_d),d)$ where:

• $c_d\in \{0,1\}$ — observed click on document $d$
• $P({\text{Exam}}_{k_d}\mid {\mathcal{C}}_{k_d})$ — the context-aware propensity (not the position-only ${\beta }_{k_d}$ of standard IPS)
• $\ell$ — the per-document ranking loss; $f(x_d)$ — the scoring model being trained

Key Properties / Variants

• Generalises Outlier Bias: a single outlier is the special case where one item’s $s_{\text{dist}}$ is large; surrounding item bias treats distinctiveness as a continuous, window-relative quantity for every item.
• Two violated assumptions: it breaks both PBM assumptions at once — examination is no longer position-only (attention contrast) and relevance is no longer context-free (comparative/anchoring contrast).
• Symmetric effect: an item gains examination by standing out and loses examination by blending into a uniform neighbourhood, so the bias can both over- and under-expose items.
• Identifiability cost: adding context terms enlarges the parameter space, worsening the identifiability problem already noted for the Examination Hypothesis — diverse logging policies (rank/context randomisation) are needed to separate $P(\text{Exam}\mid \mathcal{C})$ from $P(\text{Rel})$.
• Display-side mitigation: because the bias is relative, presenting a diverse result page (so every item is roughly equally distinctive) flattens $s_{\text{dist}}$ and reduces the bias without changing the ranking model.

Estimation procedure (EM over a context-aware click model):

Algorithm: EM for Context-Aware (Surrounding Item) Click Model
──────────────────────────────────────────────────────────────
Input: click logs {(query q, ranked list L, clicks c)},
       context window half-width w
Initialize examination network θ (params of f_θ over windows)
Initialize relevance estimates rel_d for all (q, d)
 
Repeat until convergence:
  E-step:  for each impression and each rank k:
             build context C_k = features(L[k-w : k+w])
             exam_k ← σ( f_θ(C_k) )                 # context-dependent
             infer P(Exam_k = 1 | c_k, exam_k, rel_d) via Bayes
  M-step:  update θ to maximise click log-likelihood
             given inferred examination events
           update rel_d from clicks de-weighted by exam_k
 
Output: relevance estimates rel_d (debiased for surrounding context)
        examination model f_θ (context propensities for IPS)

Connections

• Special case / sibling of: Outlier Bias (single distinctive item; window-free)
• Violation of: Examination Hypothesis, Position-Based Click Model (position-only examination)
• Related biases: Position Bias, Trust Bias, Cascading Position Bias
• Affects estimation: Inverse Propensity Weighting needs context-aware propensities; Doubly Robust Estimation for added robustness
• Lives within: Click Models, Unbiased Learning to Rank

Appears In

• Examination Hypothesis
• Outlier Bias
• Unbiased Learning to Rank
• Click Models
• IR-L11 - Unbiased Learning to Rank