The Edge Spectrum of Choice-Derived Item Graphs: Strong and Weak Edges Encode Different Relations in Collaborative Filtering

1 Introduction

• Choice-derived graphs have an edge spectrum: strong edges link in-slate competitors that smoothing pulls together against ranking, while co-click cannot share this sign mismatch.The paper reports that drop-in choice-derived operators do not improve over co-click and scalar fixes fail, whereas edge-magnitude-aware operators recover the predicted ordering. It does not claim choice modeling is unhelpful or that choice-aware operators can never outperform co-click; the diagnosis predicts modest absolute gains under attenuated propagation.Source: §1 · pp. 1–2

2.1 Graph CF and item-side smoothing

• Graph CF developed from NGCF [32] and LightGCN [8] generally uses positive smoothing over co-occurrence graphs, even when graph sources vary.Signed graph neural networks use explicit signs when they are directly observable, while this work derives the relevant sign from interaction with the ranking objective.Source: §2.1 · p. 2

2.2 Choice modeling in recommendation: the expectation we test

• Prior choice-modeling work treats choice-derived relations as richer than co-occurrence, but their use as graph-propagation sources had not been directly compared under controlled training.This paper’s distinction is driven by graph weights interacting with the ranking gradient, rather than by an a priori taxonomy such as similar versus complementary relations [15].Source: §2.2 · p. 2

2.3 Rigorous re-evaluation and impression-aware recommendation

• The paper applies controlled re-evaluation to impression-aware news recommendation and tests whether propagation edges should derive from slate-level choice context.MIND and EB-NeRD [14] are the principal benchmarks; prior slate-aware rankers model competition in the loss or policy, while this study tests it in the item graph.Source: §2.3 · p. 2

3 Setup: Item-Side Operators in Graph CF

• The setup isolates item-graph choice by replacing LightGCN’s implicit item-side operator while holding all other components fixed.The bipartite structure, layer combination, scoring rule, training objective, and hyperparameters remain unchanged.Source: §3 · p. 3

3.1 LightGCN’s Implicit Item-Item Operator and Four Candidates

• The four candidates are LightGCN’s S^T S, co-click, co-exposure, and MNL-diversion substitution; all receive identical degree normalization and top-K pruning.The parameterized item-side operator is symmetric and nonnegative, and substituting S^T S recovers LightGCN’s propagation polynomial at the operator level, not numerical equivalence with the locked baseline.Source: §3.1 · p. 3

3.2 Building the Substitution Operator

• The substitution operator aggregates symmetrized MNL diversion weights across co-appearing impressions, and the fitted MNL model is used only for graph construction.The MNL embeddings are not shared with the propagation framework; the model is fitted on impression data using slate-level cross-entropy.Source: §3.2 · p. 3

3.3 Controlled Comparison Protocol

• On the internal development split, co-click is strongest on both datasets, while substitution is at or near the bottom within a band of at most 0.01 NDCG@10.The comparison freezes the data splits, evaluator, and baseline pack; all four operator variants use the same within-impression BPR loss and three random seeds.Source: §1 · p. 1, §3.3 · p. 3, Table 2 · p. 4

4.1 Finding 1: Operator Choice Matters, But Within a Narrow Band

• The four operators span 0.4036–0.4088 NDCG@10 on MIND (spread 0.0052) and have an EB-NeRD spread of 0.0083; the simpler operator wins where scores move.The narrow spread is on the order of seed standard deviation, with ordering consistent across seeds within each dataset and across datasets at the top.Source: §4.1 · p. 4, §4 · p. 4

4.2 Finding 2: The Drop-In Choice-Derived Operator Does Not Outperform Co-Click

• Co-click beats substitution by 0.0029–0.0067 NDCG@10 on MIND and 0.0071–0.0089 on EB-NeRD; substitution never wins across four split-dataset comparisons.On the official validation split, co-click and substitution are statistically indistinguishable on EB-NeRD (0.5984 vs. 0.5971, within seed standard deviation), while co-click leads on MIND.Source: §4.2 · p. 4, §4 · p. 4

4.3 The Two Graphs Are Structurally Distinct

• Substitution and co-click index largely disjoint neighborhoods (Jaccard@50 0.014/0.044; weighted cosine 0.026/0.139), while co-exposure overlaps more (Jaccard@50 0.20–0.22).The paired co-click values are for MIND and EB-NeRD, respectively; substitution weights co-appearances by MNL choice probabilities, whereas co-exposure weights them equally.Source: §4.3 · p. 4, Table 3 · p. 4

4.4 Finding 3: The Gap to Locked Baselines Is Outside the Item-Side Operator Axis

• The best operator variant trails LightGCN++ by 0.083 NDCG@10 on MIND and 0.054 on EB-NeRD, far more than the 0.005–0.008 operator spread.LightGCN-aligned axes jointly explain less than 0.002 NDCG@10 of the gap; the residual is approximately 0.04 and is attributed to training-time hyperparameter and architecture choices outside the operator axis.Source: §4.4 · p. 5

4.5 Finding 4: The Operator Ordering Is Robust to Framework Knobs

• Under an alternative LightGCN-compatible framework, co-click and S⊤S remain statistically indistinguishable at the top, while substitution and co-exposure rank below.The narrow band remains ≤0.012 NDCG@10, and substitution still does not surpass co-click.Source: §4.5 · p. 5

4.6 Why the Band Is So Narrow: A Mixer Collapse Observation

• Under within-impression BPR, the learnable mixer collapses onto the un-propagated user term, with even-parity mass above 0.9996 on both datasets.A mixer floor modestly improves dev NDCG@10 (+0.009 on MIND, +0.006 on EB-NeRD) but does not exceed uniform coefficients and preserves the operator ordering. Mixer collapse explains the magnitude, not direction, of operator effects; the diagnosis predicts the observed ordering of positive smoothing, uniform sign flip, then edge-magnitude-aware operators.Source: §4.6 · p. 5

5.1 Setup and Notation

• The analysis considers a first-order smoothing step with symmetric non-negative item-graph weights, bilinear scores, and within-slate BPR on clicked–unclicked pairs.Source: §5.1 · p. 5

5.2 Sign Mismatch for Choice-Derived Edges

• For top-K-dominant in-slate pairs, strong MNL substitution edges make smoothing shrink the clicked–unclicked score gap while BPR grows it; outside that regime, the sign is not predicted.MNL diversion concentrates strong weights on jointly plausible choices in the same slate; the lift profile decays toward zero or reverses beyond the strong-edge regime.Source: §5.2 · pp. 5–6

5.3 Why Co-Click Does Not Satisfy the Same Misalignment

• Co-click weights measure copositivity across user history rather than in-slate clicked–unclicked pairs, so they cannot satisfy the same misalignment by construction.The same propagation operator can therefore have different geometric effects depending on which pairs carry large graph weights.Source: §5.3 · p. 6

5.4 Empirical Confirmation: Within-Slate Lift Across Cutoffs

• At k=10, substitution lift is +0.22 on MIND and +0.36 on EB-NeRD; it decays or reverses with larger k, unlike nearly flat co-click lift.On MIND, substitution has the largest lift and lowest NDCG, while co-click pairs modest lift with highest NDCG; on EB-NeRD, co-click leads in both, and substitution’s misalignment is not a headline penalty. Co-exposure lift is negative across cutoffs.Source: §5.4 · p. 7, Table 6 · p. 7

6 When Does the Edge Spectrum Surface in Headline Metrics?

• On MIND, substitution has the largest top-K lift and lowest NDCG; on EB-NeRD, co-click leads in both lift and NDCG.The within-slate lift persists on both datasets, but its headline consequence differs.Source: §6 · p. 7

6.1 Three Quantities That Modulate Whether Misalignment Surfaces

• Three candidate moderators are in-slate neighborhood density, choice-context predictability, and candidate-pool construction, but the two-dataset study cannot disentangle them.Mean unclicked in-slate substitution-neighbors per clicked item are 5.49 on MIND and 2.84 on EB-NeRD; MNL held-out NLL is 3.22 and 1.92, respectively. The factors should be read as a bundle of conditions, not independently identified causes.Source: §6.1 · p. 7

6.2 A Diagnostic Protocol

• The four-step protocol uses a trained model and held-out impressions to assess whether the spectrum is present and whether it will surface in headline metrics.It is intended for practitioners to run before deploying a choice-derived operator.Source: §6.2 · p. 7, §6 · p. 7

Box 1: Diagnostic Protocol for Choice-Derived Item-Side Operators

• Compute Δo(k) for k ∈ {10, 20, 50, 100, all}, partitioning non-clicked slate items into operator neighbors and their complement, then average partition scores across contributing pairs.Skip clicked-item and cutoff pairs whose neighbor partition is empty.Source: Box 1: Diagnostic Protocol for… · p. 7
• A spectrum is present if candidate lift changes sign between k=10 and all, or its absolute value at k=10 exceeds three times its value at all, while co-click lift is flat.Source: Box 1: Diagnostic Protocol for… · p. 7
• On EB-NeRD, substitution lift flips from +0.36 at k=10 to −0.17 at k≥50 while co-click remains flat; MIND’s spectrum surfaces in the headline, but EB-NeRD’s impact is attenuated.The protocol verdict attributes the attenuation to lower Quantity ii.Source: Figure 2 · p. 8, Box 1: Diagnostic Protocol for… · p. 8
• The pattern additionally surfaces in NDCG when in-slate operator-neighborhood density is high and held-out MNL NLL, measuring choice predictability, is low.Source: Box 1: Diagnostic Protocol for… · p. 8
• The protocol costs one inference pass per (k, operator) plus one MNL fit per dataset; it diagnoses misalignment, not magnitude, assumes positive smoothing, and awaits validation on more choice-context datasets.The passage also names hard-negative sampling as an assumption context (§8).Source: Box 1: Diagnostic Protocol for… · p. 8

7.2 Edge-Partitioned Substitution: A Diagnostic-Guided Operator

• The operator partitions each row’s top 50 nonzero substitution edges by diversion magnitude into strong and weak sets, then symmetrizes and degree-normalizes each part.The pre-registered anchors use K_strong=10, the diagnosis-located strong-to-weak boundary, rather than a cutoff chosen by post-hoc tuning.Source: §7.2 · pp. 8–9
• Anchor-B exceeds co-click by +0.0027±0.0019 NDCG@10 on EB-NeRD and is statistically indistinguishable on MIND (+0.0013 ± 0.0029).The EB-NeRD margin is outside seed standard deviation, but the limited seed count does not support a claim of statistical significance.Source: Table 8 · p. 9, §7.2 · p. 9
• Both pre-registered anchors exceed positive substitution on both datasets, while Anchor-A gains +0.0026 on MIND and +0.0046 on EB-NeRD.Anchor-B gains +0.0042 on MIND and +0.0076 on EB-NeRD; the multi-seed results confirm the predicted benefit of negatively signing strong edges relative to positive smoothing on the same graph.Source: §7.2 · p. 9, Table 8 · p. 9
• The sweep favors K_strong=10 over 5 or 20 and β=0.0 over 0.5 or 1.0, indicating the weak tail is best discarded.The co-click improvement is only +0.001 to +0.003 NDCG@10, within the operator-equivalence band; the four pre-registered predictions match, including scalar-remedy failure and the regime boundary at k=10.Source: Table 7 · p. 8, §7.2 · p. 9

8 Discussion

• The concentration mechanism extends beyond MNL: choice-model-weighted item graphs [13] [38] inherit the property, and the §6.2 diagnostic applies unchanged.Source: §8 · p. 9
• Hard-negative sampling from strong-edge competitors can use the misalignment as a signal: BPR pushes these negatives away, while positive smoothing pulls them toward clicked items.The §6.2 diagnostic identifies substitution-graph neighborhoods suitable for this complementary loss-side use.Source: §8 · p. 9
• Four operator variants span 0.005–0.011 NDCG@10, versus LightGCN → LightGCN++ gains of 0.041 on MIND and 0.017 on EB-NeRD.At matched compute, item-graph construction is not where the largest gains are found, consistent with prior re-evaluations [6] [10].Source: §8 · p. 9

• Figure 2 shows substitution’s lift changes sign across cutoffs on EB-NeRD, while co-click’s lift stays flat.Source: §9 · p. 9
• Choice-derived graphs have an edge spectrum: strong edges concentrate on in-slate competitors of clicked items, while weak edges do not; co-click graphs cannot exhibit this misalignment by construction.Source: §9 · p. 9
• The neighbor cutoff k is a semantic switch, not a sparsification knob; uniform loss-level fixes cannot repair the graph’s encoding, but edge-magnitude-aware constructions can.Source: §9 · p. 9
• The protocol in §6.2 returns a verdict from one MNL fit and a handful of inference passes, and applies to non-parametric models [13], richer utility functions [38], KG-augmented variants [31], and hard-negative sampling.Source: §9 · p. 9

A Parity Decomposition: Formal Details

• The item-side block of A^(2k) is exactly (S^T S)^k, and the ℓ-hop item representation factors into item-side and cross-type propagation.The representation is (S^T S)^floor(ℓ/2)(S^T)^(ℓ mod 2) X_user.Source: A Parity Decomposition: Formal Details · p. 9
• The intervention replaces (S^T S)^floor(ℓ/2) with I, leaves the cross-type factor unchanged, and uses a uniform layer-wise average for the final item embedding.Source: A Parity Decomposition: Formal Details · pp. 9–10
• Setting C_I = S^T S recovers LightGCN’s uniform-mixer propagation polynomial term by term, but matched item-side operators do not imply end-to-end training equivalence.The variant differs numerically from locked LightGCN because of training-time framework choices around the operator.Source: A Parity Decomposition: Formal Details · p. 10

B Framework-Axis Ablation: Decomposing the Gap to Locked LightGCN

• The residual ≈0.04 NDCG@10 gap on both datasets is attributed to training-time hyperparameters and architecture, not the item-side operator axis.The identified differences include embedding dimension (32 vs 64), training epochs (5 vs 20), learning rate, and layer-combination architecture.Source: B Framework-Axis Ablation: Decomposing… · p. 10, Table 9 · p. 10
• Binary R and no pruning each change NDCG@10 by at most 0.001 and jointly explain less than 0.002 after removing the sampling confound.The ablation re-ran the S⊤S operator with three framework axes moved toward LightGCN-compatible settings one at a time.Source: B Framework-Axis Ablation: Decomposing… · p. 10, Table 9 · p. 10
• Random-negative BPR changes NDCG@10 by −0.030 on MIND and −0.072 on EB-NeRD, widening the gap rather than explaining it.This is not a LightGCN-versus-ours difference because locked LightGCN also uses within-impression sampling.Source: B Framework-Axis Ablation: Decomposing… · p. 10

C Mixer Collapse Diagnostics

• The mixer learns coefficients through a global softmax jointly with the embeddings.Table 10 reports learned mixer coefficients under the global softmax.Source: C Mixer Collapse Diagnostics · p. 10, Table 10 · p. 10
• +0.009 on MIND and +0.006 on EB-NeRD are the dev NDCG@10 gains from imposing coefficient floors of 0.25 or 0.5 over three seeds.These constrained configurations do not exceed the uniform-coefficient configuration.Source: C Mixer Collapse Diagnostics · p. 10
• The coefficient-floor constraint leaves the operator ordering reported in Table 2 unchanged.This pattern holds across all four operator variants.Source: C Mixer Collapse Diagnostics · p. 10

D Sign-Flip and Competition-Loss Full Sweep

• Every margin-loss configuration trails co-click by at least 0.006 NDCG@10 on both datasets.The full sweep is reported in Table 11 for the internal dev split, using NDCG@10 mean ± std over three seeds.Source: D Sign-Flip and Competition-Loss Full… · p. 10, Table 11 · p. 10
• Only sign-flip γ=1.00 on EB-NeRD marginally exceeds co-click; that setting hurts on MIND.No sign-flip configuration improves over co-click on both datasets.Source: D Sign-Flip and Competition-Loss Full… · p. 10, Table 11 · p. 10
• No configuration in the sign-flip and competition-loss sweeps improves over co-click on both datasets.The sweep confirms that these uniform scalar interventions miss the misalignment described in §7.1.Source: Table 11 · p. 10, D Sign-Flip and Competition-Loss Full… · p. 10