Escaping the Euclidean Void: Manifold-Informed Flow Matching for Sequential Recommendation
Dengzhao Fang, Jingtong Gao, Yu Li, Xiangyu Zhao, Yi Chang
cs.IR, cs.AI
2026-07-27
Straight-line flow matching paths traverse empty regions of sparse item embedding space; MIRAGE fixes path geometry at training time using co-occurrence graph topology regularization while remaining graph-free at inference, achieving Beauty H@10 of 8.53 and 9%+ long-tail improvement over the best flow-matching baseline.
Flow matching is an emerging generative paradigm for sequential recommendation: train an ODE velocity field to transport Gaussian noise along a straight path to the target item embedding, then at inference generate a candidate embedding in one step and retrieve by inner product over the full catalog.
The problem is that item embedding spaces are discretely sparse — there are wide void regions between items containing no valid item semantics. A straight-line path crossing a void leaves its intermediate states without nearby item anchors; the velocity field predictor is forced to extrapolate without local geometric evidence. The paper calls this the Euclidean void. Long-tail items reside in the sparsest regions and suffer most. The failure is structural, not a discretization error — adding more ODE steps does not fix it.
MIRAGE's approach is to keep the straight probability path unchanged but reshape the surrounding embedding geometry so intermediate states find semantic support from neighboring items.
A co-occurrence graph G is built from training sequences as a discrete proxy for the semantic manifold. At each training timestep t, the K nearest co-occurrence neighbors NK(i+) of the target item i+ are identified, and the closest neighbor to xt becomes the active anchor jt. The topology regularizer Ltopo = E[w(t) × ‖xt - e{jt}‖²] pulls intermediate states toward valid item clusters.
The time-modulation weight w(t) = 4t(1−t) is a parabola: zero at both endpoints (t=0 and t=1), peak at t=0.5. This concentrates regularization on path interiors without touching the Gaussian source or the target embedding.
Backpropagation through Ltopo simultaneously pulls target embeddings toward anchor positions and anchor embeddings toward path positions, dynamically clustering semantically related items around generative trajectories, forming a semantic support corridor.
Inference is fully graph-free: single-step generation xhat1 = ftheta(x0, ε, Su), ranked by inner product over the full catalog.
Amazon benchmark datasets (Beauty, Sports, Toys, CDs), leave-one-out, full-catalog evaluation:
| Dataset | H@10 | H@20 | N@10 | N@20 |
| Beauty | 8.53 | 12.37 | 4.94 | 5.71 |
| Sports | 4.77 | 7.14 | 2.51 | 3.11 |
| Toys | 8.89 | 12.52 | 5.31 | 6.08 |
| CDs | 9.11 | 13.51 | 4.65 | 5.76 |
vs. best flow-matching baseline (FMRec): Beauty H@10 +4.02%, Sports +9.40%, Toys +6.98%, CDs +3.88%. Long-tail gains: Beauty H@20 +9.0%, Toys H@20 +9.9%. Inference latency is the lowest among all baselines.
MIRAGE resolves a conceptual confusion: flow matching failure in recommendation is a geometry problem, not a discretization problem. Adding ODE steps does not help; reshaping embedding geometry does. Because the graph is only used at training time, inference is as fast as the simplest one-step baseline. Long-tail gains above 9% are directly relevant for recommendation diversity.
The method operates in flat Euclidean space; hierarchical item taxonomies (category trees, parent-child relationships) are not captured. The authors propose hyperbolic embeddings as future work. Co-occurrence graph quality degrades with sparse or noisy interaction data. All experiments use Amazon review benchmarks; streaming or session-based settings are untested.