RouteRec: Behavior-Guided Sparse Routing for Sequential Recommendation

Junyeong Song; Jaemin Yoo

cs.IRSep 30, 2026

RouteRec: Behavior-Guided Sparse Routing for Sequential Recommendation

Authors: Junyeong Song, Jaemin Yoo

Organizations: Korea Advanced Institute of Science and Technology Daejeon, Republic of Korea · Seoul National University Seoul, Republic of Korea

Abstract

Sessionized interaction histories contain behavioral patterns that can improve sequential recommendation. However, existing models process all sessions through the same parameterized blocks, regardless of their behavioral differences. Mixture of Experts (MoE) enables conditional computation, but it leaves open what should guide expert allocation. We propose RouteRec, a sequential recommender that uses observed session behavior as the routing criterion. RouteRec summarizes four types of behavioral evidence from sessionized histories: interaction tempo, item-group focus, repetition and carryover, and popularity tendency. It uses these cues to route computation at macro, mid, and micro scopes. Cue-derived scores first select expert groups; within each selected group, the current backbone state then refines expert selection. Across six public datasets and 18 dataset-metric combinations, RouteRec ranks first in 12 and second in three, yielding the best overall average rank of 1.61 compared with 4.11 for the next-best baseline. Additional analyses suggest that the behavioral cues guide expert allocation beyond added capacity and produce routing patterns aligned with observed behavior. Our code is available at https://github.com/jy1559/RouteRec

Figures & tables

Figure 1 . Sessions with different behavioral patterns share one computation path in standard sequential recommendation ( top ), motivating behavior-guided expert allocation via MoE routing ( bottom ). Two-part motivation figure. The first panel shows sessions with different past and active-session behavior entering the same shared sequential recommendation backbone. The second panel shows a router and sparse expert allocation, asking what information from sessionized history should be used as the router input.

Panel of Figure 2 — Figure 2 . Sessions differ along cue-derived behavioral patterns within and across four representative datasets. Each point is a session; diamonds mark dataset medians. This variation motivates using behavioral cues as the routing criterion. Two subfigures: (a) interaction tempo (median gap vs. gap variability), (b) cross-session recency and movement (prev-session gap vs. category switch rate). Colored point clouds represent KuaiRec, Foursquare, LastFM, and Retail Rocket.

Figure 3 . An overview of RouteRec. Left : interaction histories become behavioral cue vectors at macro, mid, and micro scopes. Center : the SASRec backbone applies routed expert blocks at these scopes. Right : family-specific cues select expert groups; the corresponding cues and current hidden state select experts within chosen groups. The hidden state is used only in this within-group stage. A three-panel architecture figure of RouteRec. The left panel converts observed evidence into cue families and scope cue vectors. The center panel shows an active-session SASRec backbone with macro, mid, and micro routing points. The right panel shows cue-based expert-group selection followed by cue-and-hidden-conditioned expert selection and sparse weighting.

Family	Macro examples	Mid examples	Micro examples
Tempo	session gaps, pace trend	interval mean/std	last gap, local pace shift
Focus	theme consistency	category concentration, switching	local switch, suffix focus
Memory	repetition, carryover	repeat rate, novelty	recent reuse, longest run
Popularity	popularity level, drift	mean popularity, spread, trend	last-item pop., suffix spread, pop. delta

Table 1 . Example cue scalars showing what each behavioral family measures at each routing scope. The observation window differs across scopes. A cue-bank summary is in Appendix B.2 .

Dataset	Metric	SASRec	GRU4Rec	TiSASRec	FEARec	DuoRec	BSARec	FAME	DIF-SR	FDSA	RouteRec
KuaiRec	HR@10	.4612 $\pm$ .0028	.4352 $\pm$ .0108	.3299 $\pm$ .0201	.3653 $\pm$ .0309	.4779 $\pm$ .0107	.4174 $\pm$ .0098	.4809 $\pm$ .0017	.3201 $\pm$ .0655	.4671 $\pm$ .0092	.4836 $\pm$ .0035
	NDCG@10	.4319 $\pm$ .0028	.4020 $\pm$ .0115	.2682 $\pm$ .0200	.3101 $\pm$ .0419	.4555 $\pm$ .0168	.3739 $\pm$ .0155	.4617 $\pm$ .0015	.2802 $\pm$ .0887	.4377 $\pm$ .0137	.4656 $\pm$ .0041
	MRR@10	.4229 $\pm$ .0029	.3918 $\pm$ .0117	.2491 $\pm$ .0199	.2932 $\pm$ .0452	.4486 $\pm$ .0187	.3604 $\pm$ .0171	.4558 $\pm$ .0016	.2679 $\pm$ .0959	.4286 $\pm$ .0151	.4602 $\pm$ .0043
LastFM	HR@10	.3439 $\pm$ .0018	.3281 $\pm$ .0017	.3353 $\pm$ .0011	.2978 $\pm$ .0022	.3447 $\pm$ .0010	.2278 $\pm$ .0013	.2465 $\pm$ .0011	.2884 $\pm$ .0014	.3321 $\pm$ .0014	.3569 $\pm$ .0006
	NDCG@10	.2703 $\pm$ .0018	.2645 $\pm$ .0019	.2644 $\pm$ .0005	.2260 $\pm$ .0017	.2577 $\pm$ .0010	.1572 $\pm$ .0018	.1846 $\pm$ .0020	.2114 $\pm$ .0014	.2639 $\pm$ .0009	.2866 $\pm$ .0012
	MRR@10	.2474 $\pm$ .0018	.2448 $\pm$ .0020	.2425 $\pm$ .0004	.2038 $\pm$ .0015	.2306 $\pm$ .0010	.1353 $\pm$ .0021	.1654 $\pm$ .0023	.1875 $\pm$ .0015	.2427 $\pm$ .0008	.2646 $\pm$ .0015

Table 2 . Next-item prediction performance on six datasets against nine baseline methods. Each entry reports mean

\pm

sample standard deviation over seeds 42, 43, and 44. Hyperparameters and checkpoints are selected using validation performance, and the test set is evaluated only after selection. Bold and underline denote best and second-best competition ranks. Avg. Rank is computed over all 18 dataset–metric cells (lower is better).

Figure 4 . Routing profiles on KuaiRec, measured as scope-averaged route share (%). Sessions are partitioned by dominant behavioral family; the highlighted diagonal indicates each family’s matched expert group. Routing profile alignment matrix on KuaiRec, with Overall row.

Panel of Figure 5 — Figure 5 . Evaluation-time cue perturbation results (MRR@10). A trained, frozen RouteRec is evaluated under four cue-corruption strategies to test whether the router relies on cue semantics at inference. MRR@10 under cue corruptions on KuaiRec and LastFM.

Panel of Figure 6 — Figure 6 . Routing attribution and backbone portability on KuaiRec. Bars show MRR@10 and lines show HR@10: (a) SASRec-based attribution controls; (b) matched DuoRec variants. Backbone denotes SASRec in (a) and DuoRec in (b), and SASRec-wide appears only in (a). The y-axes are truncated to make differences visible. Two KuaiRec panels with MRR at 10 as bars and HR at 10 as lines. The left panel compares SASRec, SASRec-wide, Hidden MoE, and Behavior Route. The right panel compares DuoRec, DuoRec with Hidden MoE, and DuoRec with Behavior Route.

Panel of Figure 7 — Figure 7 . Train-time routing-source ablation results (MRR@10). Clean cue-only and hidden-only controls are compared with variants that misalign or remove cues during training, testing which signals support routing and whether cue–session alignment matters. MRR at 10 for full, cue-only, hidden-only, train-shuffle, and train-zero routing variants on KuaiRec and LastFM.

Panel of Figure 8 — Figure 8 . Design ablations on KuaiRec (MRR@10 bars, HR@10 lines). Left : scope-count variants test whether more routing scopes consistently help. Right : routing-organization variants compare hierarchical sparse against flatter or denser alternatives. Ablations for temporal routing scopes and routing organization.

Panel of Figure 9 — Figure 9 . Scope contribution ablation (MRR@10). Each routing scope is removed from the full macro–mid–micro stack in turn, with the other two remaining active, to isolate each scope’s individual contribution; a dense feed-forward baseline is included for reference. Scope contribution ablation on KuaiRec and LastFM.

Panel of Figure 10 — Figure 10 . Auxiliary-loss ablation results (MRR@10). We compare the full objective with variants that remove route consistency, z-loss, or both, and with a variant that adds load-balancing regularization. Auxiliary-loss ablation on KuaiRec and LastFM.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 11 — Figure 11 . Cue-family ablation results (MRR@10 bars, HR@10 line). RouteRec is retrained under three reduced-cue conditions to test whether behavior-guided routing remains effective when metadata sources are partially unavailable. Dual-axis bar and line charts comparing RouteRec under full and reduced cue conditions.

Setting	Value
Backbone layout	self-attn $\to$ macro block $\to$ mid block $\to$ self-attn $\to$ micro block
Routing granularity	Macro/mid group gate: session-wise; conditional expert gate: position-wise; micro: both position-wise
Expert groups	$G=4$ (Tempo, Focus, Memory, Popularity)
Experts per group	$K$ selected per dataset from $\{3,4,5,6\}$
Expert activation	Top- $k_{G}{=}3$ groups; top- $k_{E}{=}2$ experts per selected group (6 active)
Cue dimensionality	16 scalars per scope; 4 families $\times$ 4 summaries each

Appendix

Table 3 . RouteRec hyperparameter configuration used in the main experiments. The router layout and activation pattern are fixed across datasets; capacity and optimization values are tuned per dataset.

Family	Macro (cross-session, H=5)	Mid (active session)	Micro (recent $w$ items)
Tempo	$\bullet$ valid-history ratio $\bullet$ inter-session gap $\bullet$ mean pace $\bullet$ pace trend	$\bullet$ valid active prefix $\bullet$ mean interval $\bullet$ interval std. $\bullet$ session age	$\bullet$ valid local suffix $\bullet$ last gap $\bullet$ mean local gap $\bullet$ gap shift vs. mid
Focus	$\bullet$ theme entropy $\bullet$ top-theme mass $\bullet$ theme repeat $\bullet$ theme shift	$\bullet$ category entropy $\bullet$ top-category mass $\bullet$ switch rate $\bullet$ category uniq.	$\bullet$ current switch $\bullet$ last-category mismatch $\bullet$ suffix cat. entropy $\bullet$ suffix cat. uniq.
Memory	$\bullet$ repeat intensity $\bullet$ adj.-category overlap $\bullet$ adj.-item overlap $\bullet$ repeat trend	$\bullet$ item uniqueness $\bullet$ repeat rate $\bullet$ novel-item rate $\bullet$ longest item run	$\bullet$ last reconsumed $\bullet$ suffix reconsume rate $\bullet$ suffix item uniq. $\bullet$ suffix longest run
Popularity	$\bullet$ mean popularity $\bullet$ pop. variability $\bullet$ pop. entropy $\bullet$ pop. trend	$\bullet$ mean popularity $\bullet$ pop. variability $\bullet$ pop. entropy $\bullet$ pop. trend	$\bullet$ last-item popularity $\bullet$ suffix pop. variability $\bullet$ suffix pop. entropy $\bullet$ pop. shift vs. mid

Appendix

Table 4 . Full cue bank used by RouteRec, grouped by family and routing scope. Macro uses the five-session history window; mid uses the observed active session; micro uses the recent local window.

Panel of Figure 12 — Figure 12 . Behavioral cue diversity across six datasets, projected onto the first two principal components ( $2{\times}2$ grid, one panel per family). Diamonds mark per-dataset medians. This extends Figure 2 to all four cue families. Principal-component scatter plots for four cue families.

Dataset	Domain	Inter. (M)	Sess. (K)	Items (K)	Avg. len.
KuaiRec	Video	3.44	183.3	6.5	18.8
LastFM	Music	14.23	658.6	426.2	21.6
Retail Rocket	Browse	0.37	41.8	26.6	8.7
ML-1M	Movie	0.57	14.2	3.1	40.4
Foursquare	POI	0.05	4.6	3.6	9.9
Beauty	E-com.	0.02	2.5	1.9	7.5

Appendix

Table 5 . Processed dataset statistics used in the main experiments.

Dataset	Model	Exec./ logical	Train examples/s $\uparrow$	Full-sort ms/batch $\downarrow$	Peak memory train / infer $\downarrow$
Foursquare	SASRec-wide	1.00	4,913.4	2.887	363.1 / 163.3
Foursquare	Hidden MoE	1.62	3,220.7	16.259	593.7 / 198.0
Foursquare	RouteRec	1.62	2,978.5	17.491	593.7 / 197.4
KuaiRec	SASRec-wide	1.00	5,121.5	4.722	652.7 / 288.3
KuaiRec	Hidden MoE	1.98	2,975.6	19.150	1,318.5 / 356.9
KuaiRec	RouteRec	1.98	2,841.4	20.289	1,319.4 / 356.0

Appendix

Table 6 . Same-GPU efficiency results. Exec./logical is the ratio of actually executed to logically active non-item parameters; peak memory reports training/inference in MiB.

Parameter	Candidate sets
Shared optimization
Learning rate	Log-uniform: KuaiRec $[3{\times}10^{-4},5{\times}10^{-3}]$ ; Foursquare/Retail Rocket/ML-1M $[1.5{\times}10^{-4},2.2{\times}10^{-3}]$ ; LastFM $[8{\times}10^{-5},1.2{\times}10^{-3}]$ ; Beauty $[1.5{\times}10^{-4},2{\times}10^{-3}]$
Weight decay	$\{5{\times}10^{-7},10^{-6},10^{-5},5{\times}10^{-5},10^{-4},1.5{\times}10^{-4}\}$
History length	$\{10,20,30\}$ or $\{10,20,30,50\}$
Hidden / inner width	$\{64,96,112,128,160\}$ ; inner $\{128,192,224,256,320\}$ (or $+32$ each)
Layers / heads	$\{1,2,3,4\}$ ; $\{1,2,4,8\}$

Appendix

Table 7 . Model-selection search spaces.

Explore similar work

Modality-Guided Mixture of Structured Experts with Entropy-Triggered Routing for Multimodal Recommendation
Feb 24, 2026Ji Dai, Quan Fang, DeSheng CaiMultimodal Recommendation Model Recommendation
Beijing University of Posts and Telecommunications, China · Beijing University of Posts and Telecommunications, Beijing, China · Tianjin University of Technology, China
RecRec: Recursive Refinement for Sequential Recommendation
Jul 12, 2026Pervez Shaik, Prosenjit Biswas, Abhinav Thorat +2Multimodal Recommendation Model Sequential Recommendation
Sony Research India India
Beyond the Previous Layer: Residual Predictive Structure in Sparse MoE Routing
Sep 15, 2026Hao Li, Yasuyuki Tahara, Yuichi SeiLarge Language Model Routing
Department of Informatics, Graduate School of Informatics and Engineering, The University of Electro-Communications, Tokyo, Japan

Feb 24, 2026, cs.AIJ/K move · Enter open · S save

Modality-Guided Mixture of Structured Experts with Entropy-Triggered Routing for Multimodal Recommendation

Ji Dai, Quan Fang, DeSheng Cai

Beijing University of Posts and Telecommunications, China · Beijing University of Posts and Telecommunications, Beijing, China · Tianjin University of Technology, China

Multimodal Recommendation Model Recommendation Experts Strongest Prior Asynchronous Baseline Magnetic Tunnel Junction

Open paper AskPDF arXiv

Multimodal recommenders combine collaborative behavior with visual and textual item evidence, whose usefulness varies across user-item interactions. Independently trained source-specific diagnostic probes partition held-out interactions into behavior-, appearance-, semantics-, and mixed-evidence regimes across five benchmarks, within which capacity-matched fixed fusion rules exhibit systematic regime-dependent performance crossovers. This diagnostic observation motivates MAGNET, a multimodal graph recommender with two core mechanisms. First, a calibrated expert bank organizes trainable experts by anchor source (behavior, appearance, or semantics) and fusion family (dominant, balanced, or complementary), while an interaction-conditioned router selects a sparse composition using all three evidence sources. Second, an entropy-triggered, coverage-aware progressive schedule decouples population-level routing-mass coverage from per-instance decisiveness, transitioning from broad routing exploration to confident specialization once sufficient coverage is sustained, while preserving coverage thereafter. MAGNET separately encodes the observed interaction graph and a filtered content-induced structural view, applying cross-view alignment after independent propagation. We evaluate MAGNET on four Amazon domains and the non-Amazon MicroLens-100K benchmark, including tail-item and low-history warm-start user evaluation alongside matched expert-design controls. Under the shared-feature, fixed-split protocol, MAGNET-DV exceeds the strongest protocol-compatible non-MAGNET baseline in every reported main-table cell, supported by seed-wise difference tests. Both the fixed and KL-anchored structured variants achieve higher five-seed mean NDCG@20 than matched homogeneous and free-mixture alternatives across all five datasets. MAGNET supports route-level diagnostics through explicit expert semantics.

Abstract

Explore similar work

Modality-Guided Mixture of Structured Experts with Entropy-Triggered Routing for Multimodal Recommendation

RecRec: Recursive Refinement for Sequential Recommendation

Beyond the Previous Layer: Residual Predictive Structure in Sparse MoE Routing