[IEEE Access] SoCaST*: Beyond Simple Aggregation—Personalizing Event Recommendations via Multi-Criteria Decision Making

TL;DR

The paper presents SoCaST*, a recommendation framework that treats event suggestion as a complex decision-making problem. Unlike traditional systems that treat all factors (location, category, social, time) as equally important, SoCaST* learns personalized weights for each user and employs Dominance Intensity Measures to rank events. Tested on massive Meetup.com datasets, it sets a new SOTA for both regular and cold-start users.

Problem & Motivation: The Fallacy of Equal Weights

In Event-Based Social Networks (EBSNs) like Meetup, why do you attend an event? Is it because it's nearby? Or because the topic is "Machine Learning"?

Existing SOTA methods usually follow a "one-size-fits-all" fusion rule. They calculate scores for different criteria and sum them up or multiply them. This assumes that if User A cares deeply about location and User B cares only about the topic, the system should treat those preferences identically. The authors argue that this linear aggregation fails to capture the "trade-off" logic humans use when deciding whether to leave their house for an event.

Methodology: The MCDM Architecture

SoCaST* breaks the recommendation pipeline into three distinct, mathematically rigorous phases:

1. Multi-Dimensional Influence Modeling

The framework models four distinct "influences":

• Geographical (GI): Uses Adaptive KDE on 2D coordinates to handle sparse data (noise) in some areas and dense clusters in others.
• Categorical (CI): Combines individual interest (TF-IDF) with the popularity of that category within the hosting group.
• Social (SI): Measures the relevance of the group to the user and their friends.
• Temporal (TI): Employs KDE to model continuous time probabilities, avoiding the information loss common in discrete time-slotting.

2. Personalized Weight Estimation

This is a critical innovation. Instead of manual tuning, the system uses a distance-based method. It looks at a user's history and measures how far their attended events deviate from "optimistic" and "pessimistic" utility values for each criterion. If a user's attendance shows a very tight distribution around specific categories, that criterion receives a higher weight ($w_k$).

3. Dominance Intensity Ranking

Rather than a simple score, SoCaST* uses Pairwise Dominance.

Figure 1: Conceptual overview of EBSN influences.

An event $e_i$ is said to dominate $e_j$ if it performs better across the weighted criteria. The Dominance Intensity ($D_i$) is calculated by subtracting a "dominated ratio" from a "dominating ratio." This ensures that the recommended events aren't just "good on average" but are truly competitive alternatives that align with individual user biases.

Experiments & Results: Proving the Value of MCDM

The authors validated SoCaST* on two large datasets (New York and San Francisco) and compared it against several baselines (SRE, CFM, CAER, PAAT, and the original SoCaST).

Figure 2: Precision and Recall comparisons on NY and SF datasets.

Key Findings:

• SOTA Achievement: SoCaST* consistently outperformed all baselines in Precision and Recall.
• Weight Matters: In ablation studies, "SoCaST* w/o PW" (without personalized weights) showed a significant drop in performance, proving that understanding what a user values is as important as understanding what they like.
• Fusion Superiority: Comparison with simple Product and Sum rules showed that the MCDM approach is significantly better at "breaking ties" and identifying high-value events for cold-start users.

Critical Analysis & Conclusion

SoCaST* is a sophisticated evolution in EBSN research. By moving from prediction (will they like it?) to decision-making (will they choose it?), it aligns more closely with human psychology.

Limitations: The computational complexity of pairwise comparisons for the dominance matrix ($|E^{\ast }{|}^2$) might be a bottleneck for real-time systems with thousands of candidate events. Future work could explore approximate dominance measures or pruning techniques to scale this further.

Takeaway: If you are building high-stakes recommendation systems where users must commit real-world resources (time/travel), move beyond simple score averages. Weights and dominance intensity provide a much clearer signal of user intent.