At what scale are you building the recommendation engine, what kind of data are storing and do you need it to be real-time or offline recommendations?
Neo4j is certain a good option for doing real-time recommendations or Cassandra if you plan on doing your own map-reduce algorithms, but you can build a recommendation engine using MySQL as well, or not using any database at all or use a Recommendation Engine As A Service type of service like Google and Amazon are offering.
It all depends on what your requirements are.
Thanks for commenting. I would like it to be semi-realtime. As this is for an internal social network, users should be able to view new feed items every time they refresh. Feed shouldn't necessarily be realtime. I am aiming to serve around 10k users once we start. So, I was wondering what should be a good solution here. What do you think about MongoDB + Redis implementation?