A user with a long history suddenly clicks on an item that almost nobody has touched before. A recommender has to decide quickly what that click means. It may be a passing detour that should barely move the user's profile. It may also be the first sign of a lasting change in taste that should reshape what the user sees next. Today's systems handle this badly. Sequence models either overreact or let old history drown the new signal. Cold-start methods, which are built for items with very few interactions, assume the user's taste stays fixed while they adapt. Neither family asks the question that matters: does this new item sit inside a tight cluster with what the user has been doing recently, or is it an isolated outlier?
We propose to answer that question with local community search. These algorithms find a small, densely connected group of nodes around a few query nodes in a large graph without scanning the whole graph. Concretely, we run a fast, time-weighted community search around the cold item and the user's last few items. We then use how compact the resulting community is as the signal for "pivot" versus "detour." The test is purely behavioral. It works because the cold item's few early interactors, and what those people did recently, give second-order evidence even when the item's own embedding is unreliable and no text profile exists.
Core Claim
Our central claim is that a compactness score from local community search predicts whether a user's interaction with a cold item will be followed by more interactions in that item's neighborhood. The search runs on a graph whose edges fade with age and is bounded by distance from the query. We call such a lasting change "persistent drift," as opposed to a one-off exploration. We further claim that the score predicts this better than embedding similarity, content similarity, or a cold-start matching score. Finally, using the score to decide how far to update the user's short-term profile should improve next-item accuracy specifically in the window right after cold-item interactions. If the first claim fails, the second has no basis, so the proposal is built around testing the first claim directly.
Where This Sits
The search mechanism comes from query-dependent community search, which so far has been studied only on static graphs. Approximate Closest Community Search in Networks finds a minimum-diameter cohesive subgraph, called a closest truss community, around query nodes. Its local variant first builds a Steiner tree, the cheapest tree connecting the query nodes, and then expands outward. The authors show that diameter bounds keep out "free riders," meaning large dense cliques that satisfy density rules but are unrelated to the query. Community Focusing (FLCF) scores each node by its degree combined with its distance to the queries. This counters marginalization, where query nodes end up on the edge of a large unrelated blob. FLCF reports 0.018 seconds on Orkut against 38 seconds for the local closest-truss method. We reuse this method on a time-weighted user–item graph, which neither paper considers.
FCS-HGNN reports that accuracy saturates at a search depth of about 6. This result constrains us: it tells us to keep the search radius small and to report results as a function of depth.
The motivating evidence comes from drift and cold-start work, where the gap is visible. BISTRO separates real career pivots from exploration noise by filtering an interaction hypergraph. However, it filters over the whole hypergraph and depends on resume and job text. We drop the text dependence and replace whole-graph filtering with a query-local test. PISA separates long-term from short-term taste using ACT-R, a memory-decay model from cognitive psychology. That model looks only backward, so a cold item has no memory trace in it. Our test needs no trace for the cold item itself. S2PNM adds a dynamic preference vector, built from a learned dictionary, to a static one. It gives no rule for how much a single anomalous interaction should move that dynamic component; our score supplies one.
Mecos treats cold items as a few-shot problem, matching them against a support set of 1 to 5 interactions. It notes that more matching steps cause hubness, where embeddings crowd together and stop discriminating. It never asks whether those support interactions reflect drift or exploration. MAMO personalizes the starting point for meta-learning using memory, but it relies on demographic profiles and assumes the user's target taste is stationary.
Finally, Our Model Achieves Excellent Performance on MovieLens constrains the evaluation. It shows that much of the apparent "sequential drift" on MovieLens comes from the platform's onboarding stages. Shuffling the order of interactions costs SASRec 41.6% of its HR@10. Any drift claim we make must therefore survive a control for this effect.
Research Program
We maintain a bipartite user–item graph whose edge weights fade with age. An interaction between user $u$ and item $i$ at time $t_e$ has weight $w_e(t)$ at evaluation time $t$:
$$w_e(t) = \exp\big(-(t - t_e)/\tau\big),$$
where $\tau$ is a decay timescale tuned per dataset. Old co-interactions therefore contribute less to "distance" and cohesion.
When user $u$ interacts with an item $c$ that has at most $K$ prior interactions ($K \le 5$, following Mecos), we form a query set. It contains $c$ together with $u$'s last $m$ items. We then run an FLCF-style local search on the time-weighted graph, limited to $h$ hops. The search first builds a Steiner tree whose edge lengths are inversely related to decayed weight, then expands to an attention core and prunes negligible bridge vertices.
From the returned community we compute a compactness score $s \in [0,1]$. It combines three quantities: whether $c$ survives pruning, the community's diameter, and the share of $c$'s connection to $u$'s recent items that runs through high-degree hub nodes rather than through ordinary co-interactors. The hub-path term is the free-rider guard. A cold item that reaches the user's recent items only through a mega-popular item gets low compactness. In the graph setting, this is where hubness would otherwise show up.
The score gates a standard two-part user representation of the S2PNM or PISA kind. With $z_u$ the short-term vector and $e_c$ the cold item's representation, the update is
$$z_u \leftarrow (1 - \alpha s)\, z_u + \alpha s\, e_c.$$
Here $\alpha$ is a learned step size, and $e_c$ is the mean of the warm item embeddings in $c$'s extracted community. This makes the community serve as both the drift test and the cold item's stand-in representation. Detours ($s \approx 0$) leave $z_u$ almost unchanged. Pivots move it.
Two analyses accompany the build. The first defines a persistence label: whether at least one of $u$'s next $N$ interactions, within a fixed time window, falls inside $c$'s community. We then test whether $s$ predicts that label before any recommendation metric is computed. The second reports how accuracy and latency vary with hop depth $h$ and decay timescale $\tau$, which tests whether the FCS-HGNN saturation finding carries over.
Baselines
The naive reference point is a sequence model that treats a cold item as just another token with a content-initialized or averaged embedding and applies no gating. We use SASRec, retrained per dataset with the same tuning budget. For the persistence-prediction task, its counterpart is "always predict persistent" together with a recency-only predictor.
The strongest cold-start competitor is Mecos, because it is the only cited method that addresses cold items inside sequential recommendation. Mecos reports average relative HR@10 gains over SR-GNN of 99% on Steam, 91% on Amazon Electronic, and 70% on Tmall. No absolute HR@10 values for Mecos are reported on Lacuna, so we will re-run it on its own cold-item protocol (bottom 20% of items by interactions, $K \in \{1,3,5\}$). For persistence prediction, its matching score acts as a direct competitor to $s$.
The strongest drift competitor on listening data is PISA, which reports NDCG of 12.16% on Last.fm and 11.20% on Deezer. We evaluate PISA restricted to the windows that follow cold-item interactions; its overall numbers are not the comparison. BISTRO cannot run without text profiles, so we include its wavelet filter applied to the behavioral hypergraph alone, as a whole-graph denoising alternative.
All methods share the same chronological splits, the same item embeddings where relevant, and the same hyperparameter search budget, counted in trials.
The key ablations target three ingredients. The first replaces $s$ with cosine similarity between $e_c$ and $z_u$, which isolates the graph test from simple embedding proximity. The second removes edge decay ($\tau \to \infty$), which isolates the role of time. The third removes the hub-path term, which isolates the free-rider guard.
We count a win in two parts. First, $s$ must beat every alternative score on persistence AUC by a margin larger than the seed-to-seed spread. Second, the gated model must beat both Mecos and PISA on HR@10 and NDCG@10 in the post-cold-interaction windows, without losing overall accuracy.
Why not just…?
Why not just let a sequence model with attention absorb the cold click? Attention can down-weight an outlier, but it needs a usable embedding for the cold item to judge it, and with $K \le 5$ that embedding is mostly noise. Mecos's large gains over SR-GNN on cold items show that untreated sequence models fail in this regime. Fan et al. also show that what such models learn about "order" can be platform artifact. The SASRec baseline, evaluated only on post-cold windows, measures how much ground the gate actually adds.
Why not just use embedding or content similarity between the cold item and recent items? This is the real threat, because it is far cheaper. The arguments against it are partial. PISA found that its semantic-matching component collapsed to zero weight during training, and CLCRec found that content and behavior spaces match poorly under direct distance constraints. Neither result directly shows that similarity fails at predicting persistence. The cosine ablation and a content-similarity predictor (on Amazon and Steam, where text exists) settle the question head on. If either matches $s$ on persistence AUC, we will report that the graph search does not earn its cost and narrow the claim to the hub-guard term, if that term still helps.
Why not just run a global graph neural network such as LightGCN over the whole interaction graph? Global propagation gives cold items representations through their few edges. It also passes signal through hubs, which is exactly the free-rider and marginalization behavior shown by the closest-truss and FLCF papers, and it gives no per-interaction pivot-or-detour decision. We include LightGCN-initialized cold embeddings as an input to the cosine predictor so that the comparison isolates the decision rule rather than embedding quality. We also report latency per query against the FLCF-style search.
Proposed Experiments
We use Steam, Amazon Electronic, and Tmall, following the Mecos cold-item protocol, together with Last.fm and Deezer from PISA, where timestamps and repeat behavior support real drift labels. All splits are chronological. Metrics are persistence AUC, HR@10 and NDCG@10 on the $N$ interactions after each cold interaction, overall HR@10 and NDCG@10, and query latency.
The first decisive check is the persistence-prediction table. Unless $s$ beats the cosine, content, Mecos-score, and recency predictors, the downstream recommendation results are not interpretable.
The second is a time control. Shuffling each user's past order before computing $s$ should erase most of its predictive power on Steam and Last.fm, which would show that the score reads temporal structure rather than static co-occurrence.
The third is a confounder control on MovieLens1518. Following Fan et al., we check whether persistence labels and $s$ line up with onboarding-stage boundaries rather than with later behavior. A method whose MovieLens gains vanish after removing early-stage interactions, while its Steam and Last.fm gains hold, is measuring drift. If the reverse holds, it is measuring the platform's interface.
References
- Approximate Closest Community Search in Networks
- Community Focusing: Yet Another Query-Dependent Community Detection
- FCS-HGNN: Flexible Multi-type Community Search in Heterogeneous Information Networks
- Adapting Job Recommendations to User Preference Drift with Behavioral-Semantic Fusion Learning
- PISA: Combining Transformers and ACT-R for Repeat-Aware Sequential Listening Session Recommendation
- Modeling Dynamic User Preference via Dictionary Learning for Sequential Recommendation
- Cold-start Sequential Recommendation via Meta Learner
- MAMO: Memory-Augmented Meta-Optimization for Cold-start Recommendation
- Our Model Achieves Excellent Performance on MovieLens: What Does It Mean?