User 5242 | 6/4/2016, 9:45:07 AM
I am developing a travel recommender and would love to have some ideas how and what kind of models should be combined and used.
So for user, each user choose however many personalities they associate with from a list of 20 categories. For city, each city has a number between 0 and 1 to denote its likelihood to attract certain personality from the list of 20 aforementioned.
Now I don't have any data from each user except for their associated personality and want to recommend cities to them. So I would call it a normal cold start problem. I attempted to solve it by dot product the user personality vector with the city personality. However I would like to use a more rigorous method that just dot product.
Also I want to collect their implicit interactions with cities and improve the recommendations over time. So it would be great to make a hybrid model using both user personality and their implicit interaction.
Furthermore, I would like to incorporate side information such as user demographics. I would also like to incorporate contextual information with time stamp such as user origin city information, temperature and weather at the time of implicit interaction, user browsed city meta information at the time of implicit information.
Finally it would be great to incorporate user social information from Facebook if they grant access, including their friends' favorite places, have been places, etc.
Also it would be good to incorporate the fact that users preference change over time.
Does graph lab support the aforementioned request I listed? If so, how to implement the hybrid model?
I know this is a really long post, but thanks a lot for any idea;)