DualClusterNet combines shared mobility patterns with personal context to predict the next place

A shared routine does not imply a shared destination. Two people might both leave work, get a meal and visit a shop, yet never enter the same establishments. In their preprint, Ziyang Su, Guanjie Zheng, Jilin Hu and Linghe Kong introduce DualClusterNet to combine reusable behavioral structure with local context and personal preferences.

The task is narrowly defined: given a person’s observed check-in history, rank the place they will visit immediately next. The model does not predict a route, require a future visit time or restrict recommendations to places the person has never visited.

Two shared dictionaries, personalized combinations

DualClusterNet first processes the observed sequence with a recurrent neural network. Its inputs include place, category, hour, weekday, geographic region, elapsed time, distance from the preceding check-in and user identity. The encoder is causal in the sequential sense: it cannot inspect later check-ins when making a prediction. This is not a claim about identifying cause and effect.

The resulting representation selects weighted mixtures from two learned dictionaries. One is intended to provide reusable trajectory-pattern components; the other provides spatially conditioned components. Assignments change with the available history rather than permanently placing each person in one cluster.

The model multiplies transformed versions of these mixtures, allowing spatial context to change how a pattern contributes. This factorized operation combines pattern–space pairs without storing a separate unrestricted representation for every pair.

Personalization then takes two forms. User-specific scaling and shifting adjust the shared representation, while a separate branch carries information that need not pass through the dictionaries. A learned gate blends an additive combination of these branches with a transformed multiplicative interaction. The final score also includes a learned transition term connecting the most recent place to each candidate, plus a candidate bias.

The dictionaries are not verified labels for activities such as commuting or shopping. Their components receive no semantic supervision, and the authors do not establish that they are independent factors or geographically coherent clusters.

Preserving what “next” means

The authors evaluate chronological check-in partitions for New York City and Tokyo. The training data define 4,980 candidate places in New York and 7,793 in Tokyo. Every evaluated method ranks the complete corresponding catalog, including previously visited places.

Their filtering policy avoids manufacturing transitions. If an unknown place interrupts a held-out sequence, the evaluation does not delete it and join the visits on either side. Instead, it keeps only the continuous known suffix ending at the original final check-in. An unknown final destination is rejected, not replaced with an earlier known place. Unknown users are also excluded.

This leaves 1,300 test queries for New York and 4,711 for Tokyo. It is a closed-set experiment involving known users and places, rather than a test of recommendations for newcomers or newly added destinations.

Competitive results, small margins

The main comparison uses eight baseline methods implemented within a common evaluation framework. With random seed 42, DualClusterNet places the correct destination first on 24.00% of New York queries, compared with 23.77% for DCHL. Its mean reciprocal rank, which rewards placing the correct answer nearer the top, is 0.3516 versus 0.3509.

For Tokyo, DualClusterNet reaches 26.64% first-choice accuracy and a mean reciprocal rank of 0.3679, also the highest among the displayed methods. But it does not dominate broader recommendation lists. Other models lead New York’s top-five and top-ten accuracy, while DCHL leads both measures in Tokyo.

Controlled component-removal experiments use a different seed, fixed across variants. In those runs, removing the complete compositional and personalized stack lowers first-choice accuracy by 2.94 percentage points in New York and 2.56 points in Tokyo. Removing the separate personal branch also reduces performance in both cities. Other effects are smaller: removing user-specific scaling and shifting slightly improves Tokyo results. These comparisons support contributions from particular components within the reported runs, not uniformly positive effects from every design choice.

What remains unresolved

Each reported configuration uses only one seed, so the paper does not establish statistical significance or robustness across repeated training runs. The controlled-study scores should not be treated as direct improvements over the differently seeded main comparison.

All controlled variants retain the transition term, leaving its independent contribution unresolved. The work also does not test unseen users, unseen places, catalog growth or cross-city transfer. Its computational analysis is not a measured speed or memory advantage over competitors.

The evidence therefore supports a competitive approach to combining shared and personal information under this specific protocol—not universal superiority or demonstrated deployment benefits.