Abstract:Multilayer networks describe the rich ways in which nodes are related by accounting for different relationships in separate layers. These multiple relationships are naturally represented by an adjacency tensor. In this work we study the use of the nonnegative Tucker decomposition (NNTuck) of such tensors under a KL loss as an expressive factor model that naturally generalizes existing stochastic block models of multilayer networks. Quantifying interdependencies between layers can identify redundancies in the structure of a network, indicate relationships between disparate layers, and potentially inform survey instruments for collecting social network data. We propose definitions of layer independence, dependence, and redundancy based on likelihood ratio tests between nested nonnegative Tucker decompositions. Using both synthetic and real-world data, we evaluate the use and interpretation of the NNTuck as a model of multilayer networks. Algorithmically, we show that using expectation maximization (EM) to maximize the log-likelihood under the NNTuck is step-by-step equivalent to tensorial multiplicative updates for the NNTuck under a KL loss, extending a previously known equivalence from nonnegative matrices to nonnegative tensors.
What problem does this paper attempt to address?
### Problems the Paper Attempts to Solve
This paper attempts to address the problem of quantifying and identifying interdependence between layers in multilayer networks. Specifically, the authors investigate the expressive power of Nonnegative Tucker Decomposition (NNTuck) under KL divergence loss to naturally extend existing multilayer network stochastic block models (SBMs). By quantifying interdependence between layers, it is possible to identify redundancy in network structures, reveal relationships between different layers, and potentially inform survey tools for collecting social network data.
### Main Contributions
1. **Definition of Interlayer Independence, Dependence, and Redundancy**: Based on likelihood ratio tests (LRTs) between nested Nonnegative Tucker Decompositions, definitions of interlayer independence, dependence, and redundancy are proposed.
2. **Algorithm Equivalence**: It is proven that maximizing the log-likelihood using the Expectation Maximization (EM) algorithm under NNTuck is stepwise equivalent to minimizing NNTuck under KL divergence loss using the multiplicative update algorithm.
3. **Empirical Analysis**: The use and interpretation of NNTuck as a multilayer network model are evaluated through synthetic and real-world data.
4. **Classification of Real Networks**: Using NNTuck and likelihood ratio tests, various empirical networks are classified as layer-independent, dependent, or redundant.
### Background and Motivation
- **Multilayer Networks**: Multilayer networks describe how a set of units are connected through different types of relationships, represented in different layers. For example, a social network can be defined by different types of social interactions such as "spending time together," "friends," and "family."
- **Existing Methods**: Existing multilayer network stochastic block models mainly focus on community detection and link prediction but lack systematic studies on interlayer interdependence.
- **Technical Challenges**: Traditional likelihood ratio tests have some issues in latent factor models, so the authors introduce split likelihood ratio tests (split-LRT), which do not require regularity conditions.
### Methods
- **Nonnegative Tucker Decomposition**: NNTuck is a method that decomposes the adjacency tensor of a multilayer network into multiple nonnegative factor matrices. The third factor matrix describes the interlayer dependencies.
- **Likelihood Ratio Tests**: By comparing the likelihood ratios of different models, interlayer independence, dependence, and redundancy are defined.
- **Algorithm Implementation**: The multiplicative update algorithm is used to minimize KL divergence, which is equivalent to using the EM algorithm to maximize the log-likelihood.
### Applications
- **Social Networks**: By identifying interlayer redundancy, the data collection and analysis process can be optimized to avoid redundant information collection.
- **Biological Networks**: In biological multilayer networks, identifying interlayer independence and dependence helps understand complex interactions between genes or proteins.
- **Cognitive Social Structures**: In cognitive social structures, identifying interlayer dependence can reveal associations between different types of social relationships.
### Conclusion
By proposing NNTuck and likelihood ratio tests, this paper provides a systematic method to quantify and identify interdependence between layers in multilayer networks. This method not only theoretically extends existing multilayer network stochastic block models but also offers new insights for various types of real-world networks in practice.