The Edge Spectrum of Choice-Derived Item Graphs: Strong and Weak Edges Encode Different Relations in Collaborative Filtering
Keigo Sakurai, Takahiro Ogawa, Miki Haseyama
cs.IR
2026-08-30
Strong choice-model edges link in-slate competitors, so positive smoothing fights BPR. Drop-in substitution scores 0.4036 NDCG@10 on MIND, below co-click at 0.4088.
Graph collaborative filtering treats a stronger item-item edge as more of the same relation. The neighbor cutoff k is then a sparsification knob. That fails when weights come from a choice model. Strong edges concentrate on in-slate competitors of the clicked item; weak edges are occasional co-appearance. Positive smoothing pulls those competitors together; the within-slate ranking gradient pushes them apart.
Hokkaido University calls this an edge spectrum and tests it on the news datasets MIND and EB-NeRD. Dropping a choice-derived graph into LightGCN's item-side slot does not beat co-click. A uniform sign flip or a competition margin loss does not fix it. Operators that treat strong and weak edges differently recover the predicted order.
Data splits, the evaluator, and the baseline pack are frozen. Only the item-side even-hop operator inside LightGCN is swapped. Training is within-slate BPR throughout. Four candidates:
The diagnostic plots within-slate lift against k, locates the strong/weak boundary (around k=10), and pre-registers an edge-partitioned operator: negative on the strong head, positive or dropped on the weak tail. The claim is which interventions fail and why, not a large headline gain. The propagation channel is attenuated, so absolute lifts are predicted to be small.
Internal dev split, three seeds, uniform layer mix. On MIND the four operators span 0.4036–0.4088 NDCG@10 (spread 0.0052); on EB-NeRD the spread is 0.0083. Co-click wins (0.4088 / 0.5985); substitution sits at or near the bottom (0.4036 / 0.5905). Jaccard@50 versus co-click is 0.014 and 0.044, so this is not a dressed-up co-click graph.
On the official validation split, in-house co-click is 0.3926 and substitution 0.3857. LightGCN++ sits at 0.4754, an order of magnitude above the operator band, and that gap is in the training framework. A learned softmax mixer collapses onto the un-propagated user term, with even-hop mass above 0.9996.
The pre-registered partition Anchor-B (negate the k=10 head, drop the tail) reaches 0.3941 on MIND and 0.6045 on EB-NeRD, +0.0042 / +0.0076 over positive substitution, and slightly over co-click 0.6018 on EB-NeRD. Uniform sign flip hurts on MIND and roughly ties on EB-NeRD. Margin-loss cells trail co-click by at least 0.006.
For anyone wiring choice models, diversion, or impression slates into graph CF, this is a "do not drop in" checklist. On these graphs k is a semantic switch: the strong head is in-slate opponents; the weak tail is a different relation. The repair is not a global sign on the loss. It is a split on edge magnitude.
Headline numbers are small, and the paper treats that as a consequence of the diagnosis: mixer collapse almost shuts the propagation channel, so a correct operator cannot move NDCG much. The evidence is that the intervention order matches the mechanism, not that LightGCN++ was beaten by 0.02.
News impression slates are a clean testbed for choice models; whether the same spectrum appears in e-commerce logs without slates is unknown. The main runs attenuate propagation, so the mechanism can hold while absolute gains stay tiny, and readers may misread "small lift" as "no value." Anchor-B versus co-click on MIND sits inside seed standard deviation; the authors do not claim significance. After mixer collapse, how much the item graph still does during training is open. A complementary hard-negative sampler is sketched, not run.