Which Graph Properties are Useful for Predictive Analytics?

Question

Let's assume I'm building a content recommendation engine for online content. I have web log data which I can import into a graph, containing a user ID, the page they viewed, and a timestamp for when they viewed it. Essentially, it's a history of which pages each user has viewed.

I've used Neo4J and Cypher to write a simple traversal algorithm. For each page (node) I want to build recommendations for, I find which pages are most popular amongst other users who have also visited this page. That seems to give decent results. But, I'd like to explore alternatives to see which method gives the most relevant recommendations.

In addition to my simple traversal, I'm curious if there are graph-level properties I can utilize to build another set of recommendations with this data set. I've looked at SNAP, it has a good library for algorithms like Page Rank, Clauset-Newman-Moore community detection, Girvan-Newman community detection, betweenness centrality, K core, and so on.

There are many algorithms to choose from. Which algorithms have you had success with? Which would you consider trying?

Answer 1

Using Community detection you can build a recommendation system. The most commonly used algorithm in this field is Blondel Algorithm which u have probably seen in SNAP. Blondel is almost the fastest Community Detection algorithm among widely accepted ones and its result is pretty acceptable (at least according to modularity score). As a side comment, you may also want to have a look at NetworkX and igraph libraries for more graph algorithms and visualizations.

Hope it helps, good luck.