In the world of organizational management and collaborative research, building the "dream team" is often treated as an art. However, network science has increasingly turned it into a data problem, using complex algorithms to suggest who to hire, who to let go, or how to expand. The problem? These algorithms are often "black boxes"—they might tell you that Candidate A is the best fit, but they rarely explain why.
At RecSys ’18, researchers from Arizona State University, Tongji University, and ARL presented EXTRA, a system designed to move team recommendation from "what" to "why." By leveraging the mathematics of graph kernels, EXTRA provides intuitive, visual explanations for team-based decisions.
TL;DR
EXTRA is an interactive prototype that explains networked team recommendations (replacement, expansion, and shrinkage) by identifying which specific connections and skills drive an algorithm's decision. It uses a novel approach to calculate "influence scores" for network elements, turning complex graph math into intuitive visual maps for decision-makers.
What Problem Does This Paper Address?
Modern team recommendation systems rely heavily on Random Walk Graph Kernels (RWGK). These are powerful mathematical tools that measure the similarity between two networks by looking at how "walks" (paths) through the graphs overlap. While effective at finding a candidate who "looks like" a departing team member in terms of skills and social structure, RWGKs are mathematically dense and opaque to the end-user.
If a team leader is told to replace a veteran engineer with a specific newcomer, they need to know if that recommendation is based on the newcomer's Python skills, their previous collaboration with the lead designer, or their central position in a specific professional sub-community. Without this "sense-making" step, human leaders are unlikely to trust or adopt algorithmic suggestions.
Methods
The authors introduce a framework that decomposes the recommendation process into three explainable perspectives: Edges (relationships), Nodes (team members), and Attributes (skills).
The technical "secret sauce" is the use of Influence Scores. Instead of just calculating the final similarity score between a candidate and a team, EXTRA calculates the derivative of the graph kernel function with respect to specific network elements.
- Edge Influence: Measures how much the overall team similarity would change if a specific collaboration (edge) were added or removed.
- Node Influence: Aggregates the influence of all edges connected to a specific person to see how vital they are to the team's structural integrity.
- Attribute Influence: Measures the impact of specific skills (like "Data Mining" or "Java") on the recommendation.
By identifying the elements with the highest influence scores, the system can highlight the "top-k" reasons why a recommendation was made.
Datasets
The paper demonstrates the system using a co-authorship network. This dataset represents researchers as nodes, their collaborations as edges, and their research areas (e.g., "Databases," "Machine Learning") as attributes. Specifically, the authors showcase a case study involving the replacement of a prominent researcher (Dr. Yu) within a specific academic team.
Key Insights and Ideas
- Influence as Sensitivity: The core insight is that "explanation" can be mathematically defined as "sensitivity analysis." If removing a specific skill from a candidate's profile causes the recommendation score to plummet, that skill is a primary reason for the recommendation.
- Structural vs. Skill Match: EXTRA distinguishes between who a candidate knows and what they know. In the team replacement scenario, a "good" replacement isn't just someone with the same skills, but someone whose "influence map" (their key collaborators) mirrors that of the person leaving.
- The "Absent Skill" Flag: For team shrinkage (deciding who to remove), the system identifies "insignificant" members—those whose removal has the least negative impact on the team’s collective skill set and structural connectivity.
Key Results
The authors successfully implemented these influence algorithms into an interactive UI. In their case study:
- Qualitative Accuracy: When looking for a replacement for Dr. Yu, the system identified Dr. Jiawei Han as the best candidate.
- Visual Proof: The system visually demonstrated that Han’s top-5 most influential edges overlapped almost perfectly with Yu’s, proving he could step into the same collaborative role.
- Attribute Alignment: The system correctly identified "databases" and "data mining" as the critical shared attributes that justified the replacement, presented via intuitive pie charts on the nodes.
Limitations and Future Work
While EXTRA provides a significant leap in transparency, there are a few areas for growth:
- Scalability: Random walk graph kernels are computationally expensive. While the authors propose an "effective and fast" algorithm, applying this to massive social networks with millions of nodes remains a challenge.
- User Evaluation: The paper focuses on the algorithmic and systemic implementation. A formal user study with actual team managers would be necessary to prove that these visual explanations actually lead to better or faster human decision-making.
- Broader Application: The authors suggest that this "influence-based explanation" could be applied to other fields, such as explaining why a specific product is recommended to a user in a standard e-commerce setting.
Summary updated