Predicting human mobility is a complex challenge that extends beyond simple historical patterns. Where a person goes next is deeply contextual: a check-in at a coffee shop ten minutes ago suggests a nearby destination, whereas a check-in from three days ago may have little bearing on current movement. Traditional sequence models often struggle to account for these irregular gaps in time and space. This paper introduces ST-LSTM, a specialized Long Short-Term Memory (LSTM) variant designed to bridge this gap in Next Point-of-Interest (POI) recommendation.
The Challenge of Human Mobility
Most sequential recommendation systems, such as those used for language modeling or "next-basket" shopping, assume that items in a sequence are separated by uniform steps. However, human mobility is fundamentally different because the relationship between successive check-ins is governed by two continuous variables:
- Time Intervals ($\Delta t$): The predictive value of a check-in decays as time passes.
- Geographical Distance ($\Delta d$): Moving a short distance within a neighborhood represents a different behavioral pattern than traveling across a city.
Existing Recurrent Neural Network (RNN) solutions often ignore these intervals or simplify them into discrete "bins" (e.g., "0-1 hours"), which leads to data sparsity and a loss of precision. The authors propose a model that treats time and distance as continuous variables that directly regulate the flow of information through the network.
Methods: The ST-LSTM Architecture
The core innovation of the Spatio-Temporal LSTM (ST-LSTM) is the addition of two pairs of gates to the standard LSTM architecture. These gates explicitly model the influence of spatio-temporal intervals on both short-term and long-term user interests.
- Short-term Interest Gates ($T1_t, D1_t$): These gates control the influence of the most recent POI. If the time or distance interval is large, these gates "filter out" the influence of the last check-in, as it is less likely to dictate the immediate next move.
- Long-term Interest Gates ($T2_t, D2_t$): These gates manage how the cell state (the model's long-term memory) is updated. They help the model memorize spatio-temporal patterns over a user's entire history.
To improve efficiency, the authors also introduced ST-CLSTM, a variant that uses coupled input and forget gates. By performing forget and input actions in a single step, the model reduces parameter count without sacrificing accuracy.
Experimental Setup and Datasets
The model was evaluated on four real-world datasets from Location-Based Social Networks (LBSNs):
- CA & SIN: Foursquare check-ins in California and Singapore.
- Gowalla & Brightkite: Standard public LBSN datasets.
Data cleaning involved removing inactive users (fewer than 10 check-ins) and unpopular POIs (visited by fewer than 10 users) to ensure a high signal-to-noise ratio for training.
Key Insights
- Continuous Modeling: Unlike previous models (like ST-RNN) that used transition matrices for fixed bins, ST-LSTM uses continuous functions to learn the nuances of varying gaps (e.g., 15 vs. 20 minutes).
- Enforced Decay Logic: The authors enforced a constraint ($W \leq 0$) on the weights of the short-term gates. This ensures that as time or distance increases, the influence of the previous location naturally decreases, mathematically encoding the intuition that "stale" data should matter less.
- Dual-Layer Context: By separating gates for the hidden state (short-term) and the cell state (long-term), the model can simultaneously track immediate activities (like a shopping trip) while remembering general user preferences (like a love for Italian food).
Results and Performance
ST-LSTM and ST-CLSTM consistently outperformed seven baselines, including standard LSTMs, GRUs, and the previous state-of-the-art ST-RNN.
- Accuracy: On the Singapore (SIN) dataset, ST-CLSTM achieved an Acc@1 of 0.2037, nearly double the performance of a standard LSTM (0.1261).
- Cold Start Resilience: The model showed significant strength in "cold start" scenarios. By understanding universal spatio-temporal constraints, it can make better predictions for new users than models relying solely on long historical patterns.
- Efficiency: The ST-CLSTM variant reduced parameter count and marginally improved performance, suggesting the coupled structure acts as a helpful regularizer.
Conclusion and Future Work
While ST-LSTM excels at handling "where" and "when," it currently treats POIs as abstract IDs. Future work may involve incorporating richer contextual information, such as POI categories (e.g., "Restaurant" vs. "Park"), social network ties, or user reviews, to further refine recommendations within the gated architecture.
Summary updated