Mag-Mamba: Modeling Coupled spatiotemporal Asymmetry for POI Recommendation

The Big Problem: City Traffic is One-Way (and Time-Dependent)

Imagine you are trying to predict where a person will go next in a city.

The Morning Rush: At 8:00 AM, everyone flows from Residential Areas (Home) to Business Districts (Work). The streets are crowded in that direction.
The Evening Rush: At 6:00 PM, the flow reverses. Everyone goes from Work back to Home.
The One-Way Street: Even if two places are close to each other, going from A to B might be easy, but going from B to A might be impossible (due to a one-way street) or very slow.

The Challenge: Most computer models for predicting these trips treat the city like a static, symmetrical map. They think "A is close to B, so B is close to A," and they forget that the direction and the time of day completely change the rules. They struggle to understand that the "force" pushing you from Home to Work is totally different from the force pushing you from Work to Home.

The Solution: Mag-Mamba (The Magnetic Compass)

The authors created a new AI model called Mag-Mamba. Instead of just looking at a flat map, they treat the city like a magnetic field where time acts like a compass needle that spins.

Here is how it works, broken down into three simple parts:

1. The Magnetic Phase Encoder (The "Time-Sensitive Compass")

Imagine the city map is a giant, invisible magnetic field.

Static Map: Usually, maps just show distance.
Mag-Mamba's Map: This model adds a "magnetic charge" to the streets.
- In the morning, the magnetic field points strongly from Home to Work.
- In the evening, the field flips and points from Work to Home.
- It uses a mathematical trick called a Magnetic Laplacian. Think of this as a special lens that looks at the city and sees not just where places are, but which way the wind is blowing at that specific hour.

2. The Complex-Valued Mamba (The "Spinning Top")

Most AI models update their memory by adding or subtracting numbers (like a calculator). Mag-Mamba is different; it uses Complex Numbers (numbers that have a real part and an imaginary part).

The Analogy: Imagine a spinning top.
- Real AI: Just moves the top forward or backward.
- Mag-Mamba: Spins the top.
- Why spin? In the world of complex numbers, "spinning" (rotation) is the perfect way to represent direction. If you spin 180 degrees, you are facing the opposite way.
- The model uses the "magnetic compass" from Step 1 to tell the spinning top exactly how much to turn. If the traffic is heavy toward the CBD, the top spins toward the CBD. If it's evening, it spins back home.

3. The Time-Interval Modulator (The "Speed Dial")

Time isn't just a label; it changes how fast things happen.

If you leave work at 5:00 PM and check in again at 5:05 PM, you probably just walked to the subway.
If you check in again at 8:00 PM, you might have gone home, cooked dinner, and are now going out.
Mag-Mamba uses the time gap to adjust the speed of the spin. A short gap means a small, quick adjustment. A long gap means a big, sweeping turn in the model's memory.

Why is this better than the old way?

Old Models (Symmetric): They think the road from Home to Work is the same as Work to Home. They get confused when the traffic reverses.
Mag-Mamba (Asymmetric): It understands that the "force" of the city changes with time. It doesn't just memorize "Home is near Work"; it learns "At 8 AM, the magnetic pull is strong toward Work, but at 8 PM, it pulls toward Home."

The Results: A Super-Predictor

The authors tested this on real data from New York, Tokyo, and California.

The Test: They asked the model to predict the next stop for thousands of people.
The Winner: Mag-Mamba beat every other existing model.
The Secret Sauce: It was especially good at predicting asymmetric trips (like the morning commute vs. the evening return), which are the hardest for other models to get right.

Summary in One Sentence

Mag-Mamba is like giving a GPS a magnetic compass that spins based on the time of day, allowing it to perfectly predict not just where you are going, but which way the city's traffic flow is pushing you at that exact moment.

1. Problem Definition

The paper addresses the challenge of Next Point-of-Interest (POI) Recommendation, which involves predicting a user's next location based on historical check-in trajectories. The core difficulty lies in modeling coupled spatiotemporal asymmetry inherent in urban mobility:

Spatial Asymmetry: Transition probabilities between two locations are often non-reciprocal (e.g., $P(B|A) \neq P(A|B)$ ) due to one-way streets, functional zoning (residential to commercial), and traffic patterns.
Temporal Conditioning: These directional flows are dynamic; a route dominant in the morning (commuting to work) may reverse in the evening (returning home).
Limitations of Existing Methods:
- Graph-based methods often rely on symmetric operators (undirected graphs) or decouple directions, failing to unify reciprocal relationships under a single parameterization.
- Sequence models (LSTM/Transformer) typically use real-valued aggregations, treating directionality as an external feature rather than an intrinsic dynamic.
- State-Space Models (Mamba) offer efficient long-sequence modeling but generally operate in the real domain, lacking explicit mechanisms to encode geographic directionality and time-varying flow fields.

2. Methodology: Mag-Mamba

The authors propose Mag-Mamba, a framework that models spatiotemporal asymmetry as phase-driven rotational dynamics in the complex domain. The architecture consists of three main components:

A. Time-conditioned Magnetic Phase Encoder

This module constructs a time-varying directional field on the geographic graph.

Graph Construction: An undirected geographic graph is built based on distance thresholds.
Antisymmetric Edge Signals: For each time bin (e.g., hour of the week), the model computes a smoothed log-ratio of transition counts ( $N_{i \to j}$ vs. $N_{j \to i}$ ) to quantify directional asymmetry.
Low-Rank Factorization: To avoid computational overhead, the time-conditioned edge direction matrix is compressed via Truncated SVD into a low-rank form: $S \approx \Pi \Psi$ . Here, $\Pi$ represents time-bin mixing coefficients, and $\Psi$ represents shared spatial direction bases.
Magnetic Laplacian: The model lifts these bases into a Hermitian complex adjacency matrix ( $H$ ) using a Magnetic Laplacian formulation. The phase $\Phi_{ij}$ is derived from the antisymmetric direction field, ensuring that reciprocal edges have opposite phases ( $\Phi_{ji} = -\Phi_{ij}$ ).
Phase Extraction: The eigenvectors of the Magnetic Laplacian are used to generate unit-modulus phase tokens ( $U$ ). These tokens are gauge-invariant (independent of arbitrary scaling) and capture global directional structures.
Feature Generation: For a user step moving from $p_{src}$ to $p_{curr}$ , the model computes a phase difference vector $\Delta U = U_{p_{curr}} \odot \overline{U_{p_{src}}}$ . This vector is mixed across time bases to produce the final magnetic phase feature $m_t$ .

B. Context Feature Embedding

Standard embeddings for POI, category, user ID, and time features (hour, day) are concatenated and projected into a $D$ -dimensional backbone sequence $x_t$ .

C. Complex-valued Mag-Mamba Layer

This is the core recurrence mechanism, extending the Mamba architecture to the complex domain:

Dual Dynamics: The state evolution is governed by two factors:
- Time Intervals ( $\Delta t$ ): Mapped to step-size vectors ( $\delta$ ) controlling the decay and update magnitude (standard Mamba behavior).
- Magnetic Phases ( $m_t$ ): Mapped to angular velocity ( $\theta_{mag}$ ), acting as a topological forcing term.
Decay-Rotation Recurrence: The hidden state $h_t$ $h_{t}$ is updated via a decay-plus-rotation mechanism. The state is rotated by an angle $\phi_t = \delta \odot (\theta_{tok} + \theta_{mag})$ $ϕ_{t} = δ ⊙ (θ_{t o k} + θ_{ma g})$ .
- Mathematically, the recurrence treats the $D$ -dimensional real space as $D/2$ interleaved 2D rotation pairs.
- The rotation operator $R_t$ applies a complex multiplication $e^{i\phi_t}$ , where the angle is explicitly driven by the geographic flow direction.
Unified Modeling: This allows the model to simultaneously learn how time intervals scale updates and which direction the state should rotate based on the underlying traffic flow.

3. Key Contributions

Time-conditioned Magnetic Phase Encoder: A novel module that explicitly models time-varying asymmetric geographic flows in the complex plane using low-rank spatiotemporal factorization and Magnetic Laplacian spectral theory.
Complex-valued Mamba Module: A generalization of State-Space Models (SSMs) that replaces scalar decay with decay-rotation dynamics. This explicitly modulates sequence updates using both time intervals and magnetic geographic priors.
Unified Asymmetry Modeling: The framework successfully unifies the modeling of time-conditioned direction fields and update scaling into a single recurrence mechanism, overcoming the limitations of symmetric operators in real-valued domains.

4. Experimental Results

The model was evaluated on three real-world datasets: NYC and TKY (Foursquare) and CA (Gowalla).

Performance: Mag-Mamba outperformed 11 state-of-the-art baselines (including LSTM, Transformer, GNNs, and other Mamba-based models) across all metrics (NDCG@k, MRR).
- On the TKY dataset, it achieved a 14.71% improvement in NDCG@1 over the best baseline.
- On the CA dataset, it showed a 13.85% improvement in NDCG@1.
Ablation Studies:
- Removing the Mamba structure caused the largest performance drop, confirming the necessity of sequential modeling.
- Removing Spatial priors (Magnetic Laplacian) significantly degraded performance, proving the importance of geographic context.
- Removing Phase information (collapsing to an undirected graph) caused a substantial drop, validating that simple proximity is insufficient for asymmetric flows.
- Replacing Complex Mamba with Real-Mamba resulted in lower performance, highlighting that "complex features" require "complex dynamics" (rotations) to be effective.
Asymmetry Analysis: In a stratified evaluation, Mag-Mamba showed superior robustness in Strongly Asymmetric subgroups (e.g., tidal commuting) compared to baselines, which struggled significantly with non-reciprocal flows.
Efficiency: The model maintains linear complexity $O(L \cdot D)$ relative to sequence length. The spectral decomposition is performed as a one-time offline precomputation, ensuring low inference latency comparable to GeoMamba.

5. Significance

Theoretical Advancement: The paper bridges the gap between spectral graph theory (Magnetic Laplacians) and modern sequence modeling (Mamba/SSMs). It demonstrates that modeling urban mobility as rotational dynamics in the complex domain is a more natural and effective approach for handling coupled spatiotemporal asymmetry than real-valued aggregation.
Practical Impact: By accurately capturing tidal patterns (e.g., morning commute vs. evening return), Mag-Mamba offers significant improvements for location-based services, intelligent navigation, and urban planning.
Generalizability: The proposed paradigm of using complex-phase rotation to encode directional constraints could be extended to other spatiotemporal tasks, such as traffic flow forecasting or wind prediction, where directionality and time-varying asymmetry are critical.