[LNCS] INAMA: Bridging Social Identities through Intra and Inter Attention Mechanisms

TL;DR

Connecting user identities across fragmented social platforms (e.g., matching a DBLP profile to a LinkedIn account) is a critical challenge for personalized services. This paper presents INAMA, a model that utilizes Matched Neighbors to filter topological noise and a dual Intra/Inter Attention mechanism to weight the influence of social connections. It yields a massive performance boost, especially in attribute-rich environments like academic co-author networks.

Problem & Motivation: The Noise in the Crowd

Most existing "User Alignment" methods fall into two traps:

1. Topology Noise: They try to approximate the entire network structure. However, a user’s 1,000 random followers on Twitter don't necessarily provide a "signal" for matching them to their 50 professional connections on LinkedIn.
2. Neighbor Homogeneity: They treat all neighbors as equal. In reality, a neighbor who is already matched across both platforms is a far stronger indicator of identity than one who is only active on a single network.

The authors' insight is simple yet powerful: Focus only on the "Matched Neighbors" (nodes known to exist in both networks) and dynamically weight their importance using attention.

Methodology: The Core Mechanics

1. Matched Neighbors Construction

Instead of using the raw adjacency matrix $G_s$ and $G_t$, INAMA defines Matched Neighbors. For any candidate pair $(v,u)$, the model only considers neighbors that are part of the set of labeled aligned pairs. This effectively prunes the "noisy" edges that don't contribute to the alignment evidence.

2. The Dual-Attention Architecture

The model processes node attributes through an embedding layer and then applies two specific attention types:

• Intra-Attention: Measures how much a neighbor within the same network contributes to the user's representation.
• Inter-Attention: Measures how much a neighbor from the target network influences the user's representation in the source network.

Figure 1: The INAMA Framework, showing the transformation from raw features to attentional embeddings.

Experiments & Results

The authors conducted benchmarks across six datasets, including SNS pairs (Flickr-Lastfm) and Academia pairs (DBLP temporal splits).

Key Performance Wins

• DBLP Superiority: On the DBLP 18-16 dataset, INAMA achieved an MRR of 55.35, significantly outperforming MEgo2Vec (34.33).
• Attribute Sensitivity: The model performs best when rich node attributes (like conference names in DBLP) are available, allowing the attention mechanism to find meaningful semantic overlaps.

Table 2: Comparison of MRR across various datasets and baselines.

Ablation Insights

The ablation study revealed that removing either Intra or Inter attention leads to a consistent drop in performance (approx. 2% MRR), proving that the mutual influence across networks is just as important as the local social context.

Deep Insight & Conclusion

INAMA demonstrates that in the world of social data, less is more. By ignoring the vast majority of "unmatched" neighbors and focusing exclusively on the bridge nodes (Matched Neighbors), the model avoids the error propagation that plagues unsupervised or global topology methods.

Limitations: The model relies heavily on a seed set of labeled pairs to construct the "Matched Neighbors." In a cold-start scenario where no prior alignments exist, the model's performance would likely degrade. Future work might explore using high-confidence pseudo-labels to bootstrap this process.

Takeaway for Practitioners: When aligning entities, don't just look at the neighbors—look at the neighbors you already know are "correct." Selective attention is the key to filtering social noise.