- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes the Latent Asymmetric Bidirectional Diffusion Bridge (LAB-Bridge) for translation across heterogeneous data domains. Modality-specific encoders map both domains into a common latent space, where an entropy-regularized optimal transport objective aligns their distributions. A shared score network, conditioned on either endpoint via Bernoulli masking, supports bidirectional generation, with modality-specific decoders mapping generated latent states back to the respective domains.
The paper presents theoretical claims regarding topology preservation, bidirectional transition kernels, stochastic differential equations (SDEs), and the relationship between the training objective and path-space KL divergence. The framework is applied to human mobility modeling, translating discrete historical preference sequences to continuous geographic coordinates and vice versa. On five Foursquare city datasets, the authors report improvements over recommendation and diffusion baselines in spatial prediction and discrete ranking metrics. Additional evaluations include component ablations, latent geometric distortion measurements, and bidirectional cycle-consistency comparisons.
Critical strengths
- The paper addresses a meaningful problem: bidirectional translation between discrete mobility histories and continuous spatial information.
- Experiments across five city datasets report consistent improvements and include component ablations, providing encouraging empirical evidence, subject to the evaluation concerns below.
Critical weaknesses
The transition kernels and SDEs are not established by the stated probabilistic arguments.
In Appendix A.2, the step from Eq. (25) to Eq. (26) assumes that the noise terms vanish after conditioning on . However, depends on , and is generally nonzero. For example, fix , let be independent standard Gaussian variables, and use Eqs. (22)–(23). Gaussian conditioning gives
rather than as in Eq. (26). The analogous conditional-expectation step in Eq. (29) has the same issue.
There is also a mismatch between the theorem statement, which conditions on both endpoints, and the proof, which switches to conditioning on only one endpoint. Algebraically eliminating an endpoint is not equivalent to marginalizing over its conditional distribution. Under the independent-noise construction used in the proof, conditioning on both endpoints makes and independent, yielding a forward conditional mean of , inconsistent with Eq. (20). More generally, the one-time marginals alone do not uniquely determine the temporal coupling or transition kernels.
Appendix A.3 introduces a further issue: Eq. (36) treats as Gaussian with covariance , although the specified Gaussian distribution conditions on both endpoints. Marginalizing over generally yields a Gaussian mixture, not necessarily a single Gaussian, with conditional covariance
Thus, the score expression in Eq. (36) does not follow from the stated assumptions. Separately, the coefficient compatibility required to equate Eqs. (34) and (37), and the score identity needed to obtain Eq. (40) from Eq. (39), are not established.
These are central rather than peripheral issues: Theorem 1 is presented as the theoretical cornerstone of the method, Theorem 2 derives the sampling dynamics from its transition kernels, and the inference algorithms use those dynamics. Consequently, the current manuscript does not establish its main claimed theoretical contribution. A satisfactory resolution would require a corrected probabilistic construction and a demonstration that the resulting dynamics agree with the implemented sampler. The reported empirical improvements do not, by themselves, resolve these theoretical concerns.
Questions for authors
-
What temporal coupling or reference process defines the transition kernels in Theorem 1? Please clarify whether Eqs. (20)–(21) condition on both endpoints or only one, and explain how Eqs. (26) and (29) account for the nonzero conditional expectations of the noise terms identified above.
-
Can you provide a corrected derivation connecting the bridge marginals to the SDEs in Theorem 2? In particular, please justify the single-endpoint score in Eq. (36), establish the compatibility conditions required for Eqs. (34)–(37), and derive Eq. (40) from the time-reversal identity in Eq. (39). Do the resulting dynamics and schedules agree with those used in the implemented sampler?
-
What exactly are the inputs, outputs, prediction horizons, and data splits for each evaluation direction? Please reconcile next-destination prediction with the trajectory-based ADE/FDE definitions, and historical-sequence recovery with the single-POI ranking metrics. How were the baseline methods adapted to these tasks?