Abstract
Identifying "anchor users"—individuals who maintain accounts across multiple social media platforms—is essential for cross-network recommendations and understanding information diffusion. This task is hindered by high manual labeling costs and the "interlock property," where one identification affects the probability of others. We present DALAUP (Deep Active Learning for Anchor User Prediction), a framework that integrates deep representation learning with an ensemble of active learning strategies. DALAUP utilizes a shared-parameter neural network to learn a unified embedding space and an adaptive multi-armed bandit algorithm to select the most informative user pairs for labeling. Experimental results on real-world datasets demonstrate that DALAUP significantly outperforms state-of-the-art methods while requiring far fewer labeled examples.
Introduction
In social media analytics, linking accounts belonging to the same individual across different platforms is a foundational challenge. Traditional supervised methods face two primary obstacles: the prohibitive cost of manual verification and the non-i.i.d. nature of social network data. Specifically, the interlock property dictates that confirming a match between User A and User B implies that User A cannot match any other user on the second platform. Most standard machine learning models fail to account for this structural constraint.
Methodology
DALAUP addresses these challenges through a three-part architecture: embedding, active learning, and strategy selection.
1. The Embedding Component (AUP)
To capture structural context, DALAUP employs Random Walks with Restart. These contexts are processed by a "coupled" neural network architecture featuring convolution and deconvolution layers. As shown in Figure 1, these networks utilize parameter sharing, which forces the model to map anchor users from disparate networks into a similar vector representation space.
2. Active Learning Query Strategies
DALAUP avoids random sampling by employing three targeted query strategies to identify the most informative pairs for human labeling:
- Structure-Aware Information Entropy: Prioritizes pairs that maximize information gain while accounting for the interlock property.
- Cosine Similarity: Targets pairs already close in the embedding space, identifying high-probability candidates.
- Expected Error Reduction: Selects pairs expected to minimize overall label uncertainty across the remaining unlabeled data.
3. Adaptive Selection Mechanism
The effectiveness of query strategies evolves during training. DALAUP uses a Multi-Armed Bandit ($\epsilon$-greedy) algorithm to adaptively choose the most rewarding strategy at each iteration, ensuring the model remains robust and efficient.
Experimental Results
The framework was evaluated using a real-world dataset comprising 5,313 Foursquare users and 5,120 Twitter users, with 3,141 known anchor pairs.
Performance Against Baselines
DALAUP's supervised component (AUP) consistently outperformed state-of-the-art methods like IONE, PALE, and DeepWalk. For instance, with only 50% of anchor pairs used for training, AUP achieved approximately 97% Precision@30, doubling the performance of its nearest competitor, IONE.
Active Learning Efficiency
The active learning component allows DALAUP to reach high performance levels with significantly fewer labels than random sampling. Figure 2 illustrates how different query strategies, particularly the structure-aware variants, lead to steeper learning curves compared to random selection.
Key Insights and Conclusion
DALAUP demonstrates that network structure is a highly reliable signal for identity linkage, often surpassing the need for potentially unreliable content-based features. By explicitly modeling the interlock property and using meta-learning via bandits, the framework provides a robust solution for cross-network user alignment.
Future work will focus on reducing the computational complexity of the Expected Error Reduction strategy and exploring how these deep deconvolutional layers can scale to massive graphs with millions of nodes.
Summary updated