Hi All, I’ve been working through this MOOC and I have a question about recommender systems. Almost all of the literature I have read on recommender systems is on generating recommendations when you already have a set of ratings. But what if you are just starting out and have NO ratings yet?
Specifically, the problem I am trying to find a solution for has a huge set of products – where we have images and text descriptions and a ontology for the products. In addition, we have a set of 5 multiple choice questions that we ask of the user about their “personality”. The goal is to match the set of products to these personality answers. Of course in the medium and long term I want to use the implicit feedback to recommend products. But starting out I don’t have that. How can I build a recommender system that matches the answers to these questions to the right set of products in a meaningful way? Seems like I have to hard code some rules?
Well, there is no free lunch You can’t make something out of nothing If there is no mapping to infer here than there is not a whole lot you can learn either in a supervised nor unsupervised fashion.
One thing that comes to my mind - and take it with a grain of salt please as I am a newb when it comes to these sort of things - you could ask a subset of users what products they like while at the same time asking them those ‘personality’ questions. Than based on that you could learn something and then you could try to predict what a person might like based on their answers to the multiple choice questions. Still, quite a tall order.
Another consideration is to what extent those 5 multiple choice questions can really tell something about a person’s personality. That will be the crux of the matter in the long time to my mind. There was an interesting research done by this one guy not so long ago (he received a lot of press so it should be easy to google) when based on a very small amount of postings to social media sites he was able to learn something about a person’s personality and he could - on some measure - be able to tell more about a person then their friends / close relatives.
Either way, will let more experienced people chime in, just my two cents Best of luck - seems like a fun endeavor.
Just view the video of lesson 4 and got some questions.
1 : How should we use the model?The recommendation model we trained could help us find out the latent factors value of movies and users, so I guess we could use the trained weights to find out how many scores the users would rate on the movie they haven’t viewed, if the score is high(like 4), we can recommend the movies to the users?
2 : How could we deal with new users and movies with zero records? Do we need to collect the data and train the model again? In the real world, the data could be quite big, lots of new users would join and leaves, and every year got tons of movies release, how could we get the model up to date with efficient way?
3 : How many data you need to get reasonable results if you train the recommendation from scratch?
4 : Not every users would like to rate the movies but they view lots of them, we know the movies they view but do not know their rating. In this case, what kind of training approach we could use?
Last week seen a great article on your topic. Guy who wrote it has been developing and supporting movie and series recommender systems for a while. I’m sure it’ll bring some answers to your questions.
You can’t make something out of nothing 
If there is no mapping to infer here than there is not a whole lot you can learn either in a supervised nor unsupervised fashion.