Bayesian Models for Directed Graphs

From MaRDI portal





A digraph (or directed graph) consists of a set of g nodes connected to one another by some relationship and a \(g\times g\) adjacency (data) matrix \(X=(X_{ij})\) where \(X_{ij}=1\) if node i relates to node j and \(X_{ij}=0\) if node i does not relate to node j. In this paper, a Bayesian model is recommended for the determination of the matrix of cell probabilities P, where \(p_{ij}=\Pr (X_{ij}=1)\). This model extends the \(p_ 1\) model proposed by \textit{P. W. Holland} and \textit{S. Leinhardt} [ibid. 76, 33-64 (1981; Zbl 0457.62090)], namely, \[ p_ 1(x)=K \exp (\rho m+\theta x_{++}+\sum_{i}\alpha_ ix_{i+}+\sum_{j}\beta_ jx_{+j}). \] Here \(\rho\) is some measure of the reciprocity of relationships. For example, if \(\rho\) is positive and \(X_{ji}=1\), it is more likely to observe \(X_{ij}=1\) than \(X_{ij}=0\). The quantity \(m=\sum_{i<j}x_{ij}x_{ji}\) is the number of observed mutuals. The parameter \(\alpha_ i\) represents the expansiveness of the ith node, that is its propensity for relating to other nodes, while \(\beta_ j\) represents the attractiveness of the jth node to other nodes. The likelihood that node i relates to node j given that node j does not relate to node i, is defined to be, \(\theta_{ij}=\theta +\alpha_ i+\beta_ j\), \(i\neq j\). Further, \(\alpha_+=\beta_+=0\), where \(+\) denotes summation across the appropriate index, and the quantity K is a function of \(\theta_{ij}\) and \(\rho\). From the Bayesian viewpoint the fixed \(p_ 1\) model is unattractive as it ignores any prior information regarding the expansiveness and attractiveness of the nodes and furthermore does not recognise the association between these two characteristics of a node. (For example, a node that likes many other nodes may be more disposed to be well- liked.) In this paper, a normal Bayes prior for the \(p_ 1\) model is adopted. Specifically, it is assumed that \(\theta\), \(\rho\), \((\alpha_ i,\beta_ i)\) are mutually independent a priori and \((\alpha_ i,\beta_ i)\sim N(0,\Sigma).\) A combined EM-Newton algorithm is used to estimate the parameters \(\theta\), \(\rho\), \(\alpha_ i\) and \(\beta_ i\), denoted collectively by the \((2g+2)\) column vector \(\pi\), and to find the maximum likelihood estimates of the covariance components of \(\Sigma\), collectively denoted \(\Phi\). The algorithm, which seeks the maximum likelihood estimate of \(\Phi\), uses the Newton-Raphson method at the estimation step (E-step) of the EM algorithm. [See \textit{A. P. Dempster}, \textit{N. M. Laird} and \textit{D. B. Rubin}, J. R. Stat. Soc., Ser. B 39, 1-38 (1977; Zbl 0364.62022) for a description of the EM algorithm.] The Bayesian model is compared to the fixed effects \(p_ 1\) model by means of an example. Residuals and plots are considered and it is concluded that the two approaches can and do yield very different matrices P. The paper is concluded by some remarks on possible extensions to the two models.











This page was built for publication: Bayesian Models for Directed Graphs

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4721391)