Source-linked AI summary
Geo-SAGE: A Geographical Sparse Additive Generative Model for Spatial Item Recommendation
Weiqing Wang, Hongzhi Yin, Ling Chen, Yizhou Sun, Shazia Sadiq, Xiaofang Zhou
TL;DR
Spatial item recommendation must contend with sparse check-in histories, travel locality, and interests that vary by region. Geo-SAGE combines user interests, item content, and regional crowd preferences, smoothing those preferences over a spatial pyramid. The authors report that it outperforms state-of-the-art methods on both out-of-town and home-town recommendations.
Problem
Spatial item recommendation is difficult because user-item histories are sparse, visits cluster near users’ homes, and interests may change across regions.
Source: §1 · p. 1Method
Geo-SAGE combines user interests learned from item co-occurrences and content with native or tourist crowd preferences smoothed over a spatial pyramid.
Source: §1 · p. 2Results
Geo-SAGE outperforms state-of-the-art methods for out-of-town and home-town recommendations on two real large-scale datasets.
Source: §1 · p. 2Takeaways & Limitations
Ancestor grids can guide native and tourist preference estimates for regions with few or no visits, and allow geographic scale changes without relearning parameters.
Source: §1 · p. 2The problem assumes each user’s home location is permanent; when it is unavailable, the model infers it from the spatial-pyramid cell with the most check-ins.
Source: §3.1 · p. 3
Abstract
from arXiv · showhide
With the rapid development of location-based social networks (LBSNs), spatial item recommendation has become an important means to help people discover attractive and interesting venues and events, especially when users travel out of town. However, this recommendation is very challenging compared to the traditional recommender systems. A user can visit only a limited number of spatial items, leading to a very sparse user-item matrix. Most of the items visited by a user are located within a short distance from where he/she lives, which makes it hard to recommend items when the user travels to a far away place. Moreover, user interests and behavior patterns may vary dramatically across different geographical regions. In light of this, we propose Geo-SAGE, a geographical sparse additive generative model for spatial item recommendation in this paper. Geo-SAGE considers both user personal interests and the preference of the crowd in the target region, by exploiting both the co-occurrence pattern of spatial items and the content of spatial items. To further alleviate the data sparsity issue, Geo-SAGE exploits the geographical correlation by smoothing the crowd's preferences over a well-designed spatial index structure called spatial pyramid. We conduct extensive experiments to evaluate the performance of our Geo-SAGE model on two real large-scale datasets. The experimental results clearly demonstrate our Geo-SAGE model outperforms the state-of-the-art in the two tasks of both out-of-town and home-town recommendations.
1. INTRODUCTION
Geo-SAGE targets sparse histories, travel locality, and changing interests across regions with a model that combines user interests, item content, and crowd preferences. It smooths region preferences over a spatial pyramid and outperforms state-of-the-art methods for both out-of-town and home-town recommendations.
1. INTRODUCTION
- Spatial item recommendation faces sparse user histories, home-centered visits that hinder out-of-town recommendations, and interests that shift across regions.The paper illustrates regional variation with the same users preferring casinos in Las Vegas and beaches in the Gold Coast.Source: §1 · p. 1
- Prior approaches discussed in the introduction address sparsity or travel locality but leave interest drift unresolved; social ties also provide limited same-place check-in clues beyond 100 km [7].The paper says the local experts in Bao et al. [2] cannot address interest drift because they are selected for interests similar to the target user.Source: Table 1 · p. 2
- Geo-SAGE adds generative facets in exponential space rather than inferring personalized latent switching variables, which the paper describes as difficult to learn from sparse data.The design is inspired by the Sparse Additive Generative model (SAGE).Source: §1 · p. 2
- Geo-SAGE combines item content and crowd preferences with user interests to address sparsity, travel locality, and regional interest drift.It models user interests as distributions over latent topics using item co-occurrences and content, which can transfer interests from a home town to unfamiliar regions.Source: §1 · p. 2
- For each region, Geo-SAGE distinguishes native from tourist preferences and smooths both over a spatial pyramid to address sparse regional activity.Preferences from ancestor grids guide inference when a region has few or no visits; the hierarchy also supports changing geographic scale without relearning parameters.Source: §1 · p. 2
- Geo-SAGE outperforms state-of-the-art methods on out-of-town and home-town recommendation across two real large-scale datasets.The introduction reports the overall comparison without providing metric values.Source: §1 · p. 2
2. PRELIMINARIES ABOUT SAGE
SAGE combines generative components additively in exponential space rather than mixing their distributions with an explicit switching variable. The section uses LCA-LDA as an example of the traditional probabilistic mixture approach.
2. PRELIMINARIES ABOUT SAGE
- Traditional mixture models combine component distributions using an explicit indicator, while SAGE combines their natural parameters instead.The section characterizes the latter design as robust with limited training data because it avoids inferring a complex indicator variable.Source: §2 · p. 2
- LCA-LDA serves as an illustrative traditional probabilistic mixture model for spatial-item choice.It is introduced as a location-content-aware model intended to mimic human decision-making on spatial items.Source: §2 · p. 2
- SAGE combines generative facets through addition in exponential space, avoiding latent switching variables that are difficult to infer from sparse data.The model draws on the Sparse Additive Generative Model, whose design avoids inferring an indicator that selects among causes.Source: §2 · pp. 2–3
3. GEOGRAPHICAL SAGE MODEL
Geo-SAGE models spatial-item visits using user interests and role-specific crowd preferences, combining these factors additively in exponential space. A spatial pyramid smooths regional preferences across locations and supports changes in geographic granularity without retraining.
3.1 Problem Definitions
- Spatial items have identifiers, locations, and descriptive words, while user activities record visits in a local or tourist role.The problem formulation uses user activity profiles and a distance threshold to cover both home-town and out-of-town recommendations.Source: §3.1 · p. 3
- Recommendation queries pair a target user with a location, and distance from the user’s home determines whether the task is home-town or out-of-town recommendation.The model assumes a permanent home location and infers unavailable home locations from the spatial-pyramid cell containing the most check-ins.Source: Table 2 · p. 3, §3.1 · p. 3
3.2 Model Description
- Geo-SAGE combines user topic interests, role-specific native and tourist preferences, and item content and co-occurrence in exponential space rather than using latent switching variables.Item words and item co-occurrence patterns jointly inform topic discovery, while background distributions make learned topics more discriminative.Source: §3.2 · pp. 3–4
3.3 Model Inference
- Model inference alternates Gibbs sampling of latent topic assignments with parameter optimization to maximize the observed-data likelihood.The M-step uses projected scaled sub-gradient learning (PSSG) to update parameters.Source: §3.3 · p. 4
3.4 Spatial Smoothing
- Hierarchical spatial preferences depend on a location’s ancestors, smoothing estimates across neighboring areas and supporting sparse locations and granularity changes without retraining.The spatial pyramid partitions geographic space into regions of varying sizes; the authors relate shared ancestors to location proximity.Source: §3.4 · pp. 5–6, §3.5 · p. 6
3.5 Spatial Item Recommendation
- For each unvisited item at the target location, Geo-SAGE computes the probability that the user will visit it, using a geometric mean for the item’s word-set probability.The recommendation task is to return the top-k items; the geometric mean accounts for differing numbers of content words across items.Source: §3.5 · p. 6
4. EXPERIMENTS
Experiments on Foursquare and Twitter evaluate Geo-SAGE for home-town and out-of-town recommendation, with comparisons against existing methods and model variants. Geo-SAGE performs best in both recommendation settings, and further tests examine component contributions and cold-start behavior.
4.1.1 Data Sets
- Foursquare contains 4,163 users and 483,813 check-ins; Twitter contains 114,058 users and 1,434,668 check-ins, with venue categories and tags added from Foursquare.Twitter lacks social-network information and explicit home locations; the latter are inferred from the spatial-pyramid cell with the most check-ins.Source: §4.1.1 · p. 6
4.1.2 Comparative Approaches
- Comparisons include LCA-LDA, UPS-CF, and CKNN, alongside three Geo-SAGE variants designed to test preference roles and spatial smoothing.S1 omits native and tourist preferences, S2 merges those roles, and S3 removes spatial-pyramid smoothing.Source: §4.1.2 · pp. 6–7
4.1.3 Evaluation Methods
- The evaluation splits each user’s activities into training and test sets and ranks unvisited items near each test item, measuring whether the held-out item appears in the top k.Activities more than 100 km from a user’s home are treated as out-of-town; Recall@k averages hits across test cases.Source: §4.1.3 · p. 7
4.2 Recommendation Effectiveness
- The experiments assess recommendation effectiveness across increasing k values, reporting Recall at k = 2, 6, 10, 14, and 18.The reported recall values rise as k increases because larger recommendation lists are more likely to include items users would visit.Source: §4.2 · p. 7
- Geo-SAGE performs best on both datasets; on Foursquare out-of-town recommendation, its Recall is about 0.327 at k = 2 and 0.535 at k = 10.On Twitter, Geo-SAGE also ranks first, followed by LCA-LDA; UPS-CF was not evaluated because social-network data was unavailable.Source: §4.2 · pp. 7–8, Figure 4 · p. 8, Figure 5 · p. 8
4.3 Impact of Different Factors
- Geo-SAGE consistently outperforms S1, S2, and S3 in home-town and out-of-town recommendation; the variants’ gaps are smaller for home-town tasks.S2 and S3 also outperform S1, while the S3 comparison supports spatial-pyramid smoothing as a way to alleviate sparsity.Source: Figure 6 · p. 8, Figure 7 · p. 8, §4.3 · p. 8
4.4 Test for Cold Start Problem
- On Foursquare, Geo-SAGE remains the best-performing method for cold-start users, although recommendation effectiveness falls for all compared algorithms.Geo-SAGE can identify a user’s local or tourist role from known home-town information even when the user has no activity records.Source: Figure 8 · p. 9, §4.4 · p. 9
- For cold-start users, Geo-SAGE performs better out-of-town than at home, while increasing training histories from 0 to 5 improves home-town results more.Figure 9 relates this pattern to tourist preference having a more important role in out-of-town recommendation than native preference has in home-town recommendation.Source: Figure 9 · p. 9, §4.4 · p. 9
5. RELATED WORK
Prior spatial recommendation work explored geographical and social influence, temporal patterns, and POI content. The paper identifies travel locality and regional interest drift as challenges that prior work did not jointly address.
5. RELATED WORK
- Prior methods use geographical and social influence, temporal cyclic or sequential patterns, and POI content to improve spatial recommendation and address sparsity.Examples include combining preference with social and geographic influence, modeling time-aware check-ins, and using POI-associated content or tags.Source: §5 · pp. 9–10
- The authors state that prior work did not jointly address travel locality and interest drift, and that methods using local crowd preferences did not distinguish native from tourist preferences.They describe their contribution as a unified approach to the three challenges and preference inference over a spatial index structure.Source: §5 · p. 10
6. CONCLUSION AND FUTURE WORK
Geo-SAGE combines item co-occurrence and content to infer and transfer user interests, incorporates target-location native or tourist preference, and smooths those preferences over a spatial pyramid. Experiments report advantages over other spatial recommendation methods for both home-town and out-of-town recommendations; future work would integrate social influence and temporal effects.
6. CONCLUSION AND FUTURE WORK
- Geo-SAGE combines item co-occurrence and content, target-location native or tourist preference, and spatial-pyramid smoothing to address data sparsity, travel locality, and interest drift.Experiments on two large-scale datasets report advantages over other spatial item recommendation methods for both home-town and out-of-town recommendations.Source: §6 · p. 10
- The authors identify integrating social influence and temporal effects as future work.Source: §6 · p. 10