Personalized PageRank

PageRank is a metric developed by Google founders Larry Page and Sergey Brin. PageRank is defined by the on a directed graph $G=(V,E)$ as the probability that a random walk will be on any particular node $v_i$ when walking over the edges randomly. As there may be vertices that only point to each other (spider traps) they introduced the idea of a random restart, so with probability $\beta$ the random walk with jump randomly to a new vertices which is also known as the dampening factor.

Specifically using $|V|$ as the number of vertices in the graph $G$ and $|E_j|$ as the number of outgoing edges from vertices $v_j$ also known as the out degree of $v_j$.

$$ P(v_i) = \frac{1-\beta}{|V|} + \beta\sum_{e_{ij}\in E}\frac{P(v_j)}{|E_{j}|} $$

Let’s create a matrix formulation of this by defining a matrix $\mathbf{M}$, let $v_i$ have out degree $d_i$. If $e_{ij} \in E$ then $M_{ji} = \frac{1}{d_i}$. Thus the columns of $\mathbf{M}$ sum to 1 and $\mathbf{M}$ is a stochastic matrix, which gets to Markov properties and we’ll leave for another day (when I’m more familiar). In another notation if $\mathbf{A}$ is the adjacency matrix of graph $G$ and $\mathbf{K}$ is the a diagonal matrix with out degrees of each vertice on the diagonal them $\mathbf{M} = (\mathbf{K}^{-1}\mathbf{A})^{T}$. For notation we’ll also define $\mathbf{R} := P(v)$

$$ \begin{align} P(v_i) &= \frac{1-\beta}{|V|} + \beta\sum_{e_{ij}\in E}\frac{P(i_j)}{|E_{j}|} \\ \mathbf{R} &= \frac{1-\beta}{|V|}\mathbf{1} + \beta\mathbf{M}\mathbf{R} \\ \mathbf{R} &= \frac{1-\beta}{|V|}\mathbf{1} + \beta\mathbf{M}\mathbf{R} \\ \mathbf{R} &= (\frac{1-\beta}{|V|}\mathbf{1} + \beta\mathbf{M})\mathbf{R} \end{align} $$

With the equation (4) being true as $\mathbf{R}$ is a probability vector and sums to 1 as do the columns of $\mathbf{M}$. To make sure that $\mathbf{M}$ is stochastic it is necessary for vertices with no edges out (sinks) to split probability to $1/|V|$ for all vertices.

If we define $\widehat{\mathbf{M}} = \frac{1-\beta}{|V|}\mathbf{1} + \beta\mathbf{M}$ then we can reformulate as

$$ \widehat{\mathbf{M}}\mathbf{R} = \lambda\mathbf{R} $$

Indicating that $\mathbf{R}$ is the eigenvector of $\widehat{\mathbf{M}}$ with an eigenvalue $\lambda$ of 1 and since $\widehat{\mathbf{M}}$ is stochastic that is the biggest eigenvalue.

In general calculating eigenvectors and eigenvalues is hard but one example to get only the first $k$ eigenvectors then we can use the power method which through the miracle of eigenvectors any random vector multiplied enough times by $\widehat{\mathbf{M}}$ will converge to the eigenvector of that matrix. If the eigenvalue of that eigenvector is significantly larger than the next eigenvalue then it should converge quickly.

$$ \begin{split} \mathbf{r}^{(0)} := [1/|V|,...,1/|V|]^T \\ while: |\mathbf{r}^{(t+1)} - \mathbf{r}^{(t)}|_1 < \epsilon \\ do: \mathbf{r}^{(t+1)} = \widehat{\mathbf{M}}\mathbf{r}^{(t)} \end{split} $$

In Personalized PageRank instead of restarting from any random vertice we only restart with probability $\beta$ from particular vertices that we want to personalize from. In the random walk framework we are seeing what are the likely vertices a walker will be at if the keep restarting from the personalized set. So instead of adding $\frac{1-\beta}{|V|}$ to every entry in $\mathbf{M}$ we only add it to particular vertices that we want to personalize from.

In particular

$$ \widehat{\mathbf{M_{ij}}} = \left \{ \begin{split} (1-\beta)/|\mathbf{S}| + \beta\mathbf{M_{ij}} & \quad i \in \mathbf{S} \\ \beta\mathbf{M_{ij}} & \quad i \notin \mathbf{S} \end{split} \right \} $$