Abstract
This paper introduces UCLAF (User-Centered Collaborative Location and Activity Filtering), a framework designed to provide personalized "where to go" and "what to do" recommendations using sparse mobile sensor data. By modeling user-location-activity interactions as a 3D tensor and incorporating auxiliary information through regularized tensor decomposition, the approach effectively mitigates the extreme data sparsity inherent in mobile contexts and significantly improves recommendation accuracy.
Introduction
As GPS-enabled devices have become ubiquitous, the demand for intelligent mobile recommendations has grown. Users frequently seek answers to two fundamental questions: "I want to go sightseeing; where should I go in this city?" (Location Recommendation) and "I am at the Forbidden City; what else can I do nearby?" (Activity Recommendation).
However, individual user data is often "grossly insufficient and sparse." Most users visit only a tiny fraction of a city's locations and rarely provide explicit annotations for their activities. Traditional collaborative filtering (CF), which typically links two entities like Users and Items, struggles with the three-party relationship between Users, Locations, and Activities, especially when over 98% of potential data points are missing.
Methodology
The authors propose a User-Centered approach that treats the user as a first-class entity within a multi-dimensional space.
1. The Tensor Representation
Data is modeled as a 3D tensor $\mathcal{A}$ of size (Users $\times$ Locations $\times$ Activities). An entry $a_{ijk}$ represents the frequency of user $i$ performing activity $k$ at location $j$. This structure captures the complex dependencies between the three entities.
2. Regularized Tensor Decomposition
To address the sparsity of the tensor, the framework employs a PARAFAC-style decomposition to break the tensor into low-dimensional latent factors. To enhance the accuracy of these factors, the process is "regularized" using four auxiliary matrices that incorporate side information:
- User-User Similarity (B): Encourages similar users to have similar latent representations.
- Location-Feature Matrix (C): Uses Point of Interest (POI) data to inform location profiles.
- Activity-Activity Correlation (D): Incorporates common-sense knowledge about which activities typically co-occur.
- User-Location Visiting (E): Integrates raw GPS visits even when the specific activity is unknown.
By sharing low-dimensional representations across the tensor and these auxiliary matrices (Collective Matrix Factorization), the model effectively "fills in the blanks" of the sparse tensor.
Dataset and Experimental Setup
The study utilized a substantial real-world GPS dataset collected in Beijing, China, involving 164 users over a 2.5-year period.
- Scale: 12,765 GPS trajectories totaling over 139,310 kilometers.
- Processing: Raw GPS points were clustered into 168 meaningful locations, and activities were categorized into five groups (e.g., Food & Drink, Tourism & Amusement).
- Sparsity: The user-location-activity tensor was extremely sparse, with only 1.04% non-zero entries.
Key Results
The UCLAF model was evaluated against several baselines, including standard User-based CF and High-Order Singular Value Decomposition (HOSVD).
- Accuracy: UCLAF achieved a superior RMSE of 0.006, indicating highly accurate reconstruction of missing data.
- Ranking Quality (nDCG):
- For Location Recommendation, UCLAF outperformed the best memory-based baseline by approximately 19%.
- For Activity Recommendation, it achieved an nDCG of 0.959, a 22% improvement over simple collaborative filtering.
- Robustness: The model demonstrated stability across various settings of the "low dimension" ($k$) parameter.
Conclusion and Future Work
UCLAF demonstrates that location and activity recommendations are mutually reinforcing. While the primary tensor is sparse, the inclusion of "side information" acts as a vital guide for decomposition, preventing overfitting.
Future research will focus on transitioning the model to an online setting for real-time updates and incorporating additional contextual dimensions such as time and weather to further refine recommendation relevance.
Summary updated