Next Point-of-Interest (POI) recommendation systems are essential for predicting user movement based on historical check-ins. However, most existing models are static, failing to adapt to the evolving nature of human behavior, such as seasonal shifts or changing personal routines. Researchers have introduced GIRAM (Generative Key-based Interest Retrieval and Adaptive Modeling), a model-agnostic framework designed for continual learning in POI recommendation.
The Challenge of Dynamic Interests
Traditional POI recommendation models are typically trained on a fixed historical dataset. Once deployed, their performance often degrades as user interests shift over time—a phenomenon known as concept drift. For instance, a user's preference for indoor gyms in winter may shift to outdoor parks in summer.
As shown in the figure above, static models (green line) experience a significant drop in accuracy over time. While retraining from scratch (red line) maintains performance, it is computationally expensive. Simple fine-tuning (blue line) often suffers from "catastrophic forgetting," where new data overwrites valuable historical knowledge. The goal of GIRAM is to provide an efficient, adaptive solution that balances new trends with long-term preferences.
The GIRAM Framework
GIRAM acts as a flexible wrapper around existing recommendation architectures (e.g., RNNs, Transformers). It employs a generative memory retrieval system and an adaptive fusion mechanism to handle continual updates without full retraining.
The framework is built on four key components:
- Interest Memory: Instead of raw data, GIRAM stores behavioral patterns as key-value pairs. Keys represent contextual embeddings, while values store the model's predicted POI probability distributions.
- Context-aware Key Encoding: This module integrates geographic coordinates, time (including periodic cycles), and POI categories into a unified vector. It even leverages LLMs like ChatGPT to derive richer category information.
- Generative Key-based Retrieval: Unlike standard deterministic searches, GIRAM uses a Conditional Variational Autoencoder (CVAE) to generate multiple diverse query keys. This allows the system to retrieve a broader spectrum of relevant historical interests matching the current context.
- Adaptive Update and Fusion: A "consistency score" is calculated by comparing a collective model (general trends) with a personalized model (individual habits). This score determines whether the system should prioritize recent trends or sustained historical interests during the final recommendation.
Experimental Evaluation
The framework was evaluated using three real-world datasets: NYC (Foursquare), TKY (Foursquare), and CA (Gowalla). To simulate a continual learning environment, the data was split chronologically into five blocks, with the model updating and predicting sequentially.
Key Results
- Superior Accuracy: GIRAM consistently outperformed state-of-the-art continual learning baselines (like ReLoop2 and ADER). On the NYC dataset with a Flashback backbone, GIRAM achieved an MRR of 0.3610, surpassing both the static baseline (0.2688) and the full retraining approach (0.3453).
- Efficiency: GIRAM's update time is significantly lower than full retraining and only slightly higher than simple fine-tuning, making it practical for real-time applications.
- Memory Optimization: By storing compressed "top-K" probability distributions instead of raw trajectories, GIRAM maintains a manageable memory footprint.
Conclusion and Future Directions
GIRAM provides a robust, model-agnostic solution for keeping POI recommendations fresh. By combining generative memory retrieval with adaptive fusion, it effectively mitigates catastrophic forgetting while remaining computationally efficient. Future work may explore hierarchical memory management to handle millions of users and strategies for "cold-start" scenarios where new users enter the system without historical data.
Summary updated