Abstract:We introduce a unified framework, formulated as general latent space models, to study complex higher-order network interactions among multiple entities. Our framework covers several popular models in recent network analysis literature, including mixture multi-layer latent space model and hypergraph latent space model. We formulate the relationship between the latent positions and the observed data via a generalized multilinear kernel as the link function. While our model enjoys decent generality, its maximum likelihood parameter estimation is also convenient via a generalized tensor decomposition <a class="link-external link-http" href="http://procedure.We" rel="external noopener nofollow">this http URL</a> propose a novel algorithm using projected gradient descent on Grassmannians. We also develop original theoretical guarantees for our algorithm. First, we show its linear convergence under mild conditions. Second, we establish finite-sample statistical error rates of latent position estimation, determined by the signal strength, degrees of freedom and the smoothness of link function, for both general and specific latent space models. We demonstrate the effectiveness of our method on synthetic data. We also showcase the merit of our method on two real-world datasets that are conventionally described by different specific models in producing meaningful and interpretable parameter estimations and accurate link prediction. We demonstrate the effectiveness of our method on synthetic data. We also showcase the merit of our method on two real-world datasets that are conventionally described by different specific models in producing meaningful and interpretable parameter estimations and accurate link prediction.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: how to study and analyze the interactions among multiple entities in complex higher - order networks within a unified framework. Specifically, the author introduced a general latent space model to deal with the high - order interaction problems in multi - layer networks, hypergraph networks and dynamic/sequential networks.
### Main problems
1. **Limitations of existing models**:
- **Strict model assumptions**: Traditional Stochastic Block Model (SBM) and its variants have strong assumptions about the relationships between nodes, such as transitivity and stochastic equivalence of nodes within the same cluster. These assumptions may not hold in practical applications, especially in higher - order networks.
- **Low - rank assumption**: High - order SBM usually assumes that the adjacency tensor has a low - rank structure, but the actually observed adjacency tensors often present many moderately - sized singular values, making the low - rank assumption doubtful.
2. **Need for more flexible models**:
- In order to better capture the complex interactions in higher - order networks, a model that can handle nonlinear relationships and heterogeneity is required.
- Existing models cannot handle the heterogeneity of latent positions between different layers well. For example, in multi - layer networks, the similarity between nodes in different layers may change significantly.
### Solutions
The author proposed a unified higher - order Latent Space Model (hLSM), which solved the above problems in the following ways:
- **Latent space model**: Each node has a latent position in the low - dimensional latent space, and these positions are used to describe the interaction probabilities between nodes. By introducing a kernel function, the nonlinear relationships between nodes can be flexibly modeled.
- **Generalized tensor decomposition**: To estimate the latent positions, the author proposed a method based on generalized tensor decomposition and developed a projected gradient descent algorithm on Grassmannians. This algorithm can converge linearly under mild conditions and provides statistical error bounds for the final estimate.
- **Special cases**: hLSM covers a variety of existing models, such as the mixture multi - layer latent space model, the hypergraph latent space model and the dynamic latent space model, and can handle specific problems in these models.
### Specific application scenarios
- **Multi - layer networks**: Allows different latent positions between different layers, thereby better capturing the heterogeneity between different layers.
- **Hypergraph networks**: Used to model high - order interactions between nodes, for example, in co - author networks or legislator networks.
- **Dynamic networks**: Used to model network structures that change over time, such as friendship relationships in social networks or email networks.
In conclusion, this paper aims to provide a unified and flexible framework through the introduction of hLSM to more accurately model and analyze the interactions in complex higher - order networks.