Jinni: Wants to be Pandora for Movies
We're currently running a series of posts about recommendation technologies and in the comments of our last post about the Netflix Prize, a company called Jinni made itself known. Jinni is a kind of 'Pandora for movies', because it aims to recommend movies and tv shows to you based on its Movie Genome (aping Pandora's Music Genome Project). Jinni's genome project contains over two thousand "genes" that describe plot, mood, style, setting, soundtrack and more. Jinni says that its ontology was created by film professionals - much like Pandora employs people to create its unique music database.
How it Works
Jinni says that its video content is automatically indexed, using a mixture of metadata and reviews. It has a strong semantic technologies component, as it uses a proprietary Natural Language Processing solution to assign semantic tags to content and users. The company claims that this allows Jinni to "rapidly index more titles, becoming the universal catalog for professional video." When it launched in December, Jinni had 10,000 movie, TV and video titles. It also offers APIs for Internet and TV content providers.
In terms of its recommendations philosophy, Jinni believes that a mix of algorithms and human selection is the best solution. Although the initial data set comes from humans entering movie information into a computer, the actual recommendations come from its algorithm - which "can deeply analyze the type of content you like" and hence learn about your tastes in movies. Jinni gives you recommendations by "comparing your Taste Types and the genes of all the titles in our catalog".
Does it Work?
I took Jinni for a test drive, with a search on 'mood'. I quite like an "offbeat" movie on a Friday night, so it was an appropriate place to start. Also 'offbeat' movies like Napoleon Dynamite are the type of films that have caused the Netflix Prize contestants a lot of problems. So I clicked on that mood to see what came up.
The default video selections included ones like Twin Peaks, Donnie Darko, Hot Fuzz, Monty Python, and so on. It was a fairly predictable selection, but where Jinni promises to come into its own is when you filter down. There is a 'Story Tuner', which presented some interesting filter options.
I filtered based on the story tuners (little known, light, realistic, fast-ish) and got recommended a 1990 movie called 'Mr Destiny' starring James Belushi. I can't recall ever seeing it (little known? check!) and while I'm not really a James Belushi fan, it'd be worth a try. I tweaked the light/serious meter up to 'serious' and that gave me a neat selection of 9 films, few of which I'd seen - but they all looked interesting. This is what you want from a recommendation engine - to be told about products you didn't know about before. So Jinni appears to work quite well. There are many other filters other than mood; rating, date, length, plot, genres, and more.
In his review, Chris Gampat concluded that Jinni isn't quite Pandora for movies. That may be true, but we think Jinni is worth a try if you're a movie or tv show buff. It's also similar to ClerkDogs, which we reviewed recently. So if you've tried either or both, let us know in the comments what you thought.
ReadWriteWeb Resources for Recommendation Technologies
We will be profiling other recommendation companies in upcoming posts. We also invite you to explore using our custom ReadWriteWeb Resources:
- RWW Recommendation Industry Feed Favorites OPML file (save link)
- RWW Recommendation Industry Feeds - Best of Feed (copy and paste to your reader)
- Click to preview the above feeds before subscribing (pop-up window)
- RWW Recommendation Site Search (Visit and Bookmark)
Share
Comments
Subscribe to comments for this post OR Subscribe to comments for all ReadWriteWeb posts
Jinni is a pretty fun tool to use. A very comprehensive tool that even includes short films.
Thanks for the insightful perspective on Jinni. With meaning-based search, “more like” titles you’ve enjoyed, and personalized recommendations (jumpstart them by taking the Taste Test), we hope to offer useful tools for all tastes, moods and stages of decision-making about what to watch. As noted, filtering is key – such as “Squeeze your search,” a menu that appears once you’ve begun a search on Jinni to suggest terms for refinement. We welcome feedback!
Assuming this is the quickest route to the people behind this service, I have the following questions:
1. Have you considered DBpedia lookups ?
2. Have you considered Linked Movies Database lookups?
3. Have you considered Freebase lookups?
Even if you don't perform any of the above, if you are going to accurately be asociated with "Semantic Web" as opposed to "Semantic Technology" then at the very least there should be structured descriptions of the "Things" in your data space (your web presence) that have GUIDs/URIs/Entity IDs that allow Web User agents to negotiate representations of descriptions of movies, genres, actors etc..
Note to Richard: to keep things a little clearer for your readership, it is important to distinguish between "Semantic Web" and "Semantic Technologies". When "Web" is part of the description, the use of HTTP to negotiate and obtain representations of resource/entity/object/thing descriptions is assumed :-)
This post was written after taking a quick look at Jinni with the "Semant
Kingsley, duly note! Thanks for that feedback and I'll update the post accordingly.
Netflix needs to offer more educational videos. They could be a huge online database of educational videos, a gigantic resource for both self-education and individual education, with some for rental through the mail. Or maybe google should just do it for free.
Phoebe said (from the other thread)
To clarify, our recommendations are based on semantic information and not solely on ratings.
It sounds like you're already using tensor calculus type recommendation algorithms? I am just assuming here, since I know tensor is a very complex topic which is still common only in the field of Physics/Engineering, it is only recent that it has captured the attentions of data-analysts because of its multidimensional capability. User Item-ratings based recommendation is 2D dataset, ie, a matrix (rows & columns). The 2 dimensions are (user,items) where every user rates an item. It looks to me that your app is based on extra dimensions, such as semantic taggings, which makes your recommendation system a 3D one, ie, the algorithm computes box dataset such as (user,items,semantic-tags).
I pointed to you on the other thread about the following paper, in which I am cutting & pasting its abstracts below. This 3D tensor recommendation engine ie, (user,items,tags) improves its accuracy over the traditional 2D (user,items), as stated by the authors in their abstract.
Abstract:
--------
Social tagging is the process by which many users add metadata in the form of keywords, to annotate and categorize information items (songs, pictures, web links, products etc.). Collaborative tagging systems recommend tags to users based on what tags other users have used for the same items, aiming to develop a common consensus about which tags best describe an item. However, they fail to provide appropriate tag recommendations, because: (i) users may have different interests for an information item and (ii) information items may have multiple facets. In contrast to the current tag recommendation algorithms, our approach develops a unified framework to model the three types of entities that exist in a social tagging system: users, items and tags. These data is represented by a 3-order tensor, on which latent semantic analysis and dimensionality reduction is performed using the Higher Order Singular Value Decomposition (HOSVD) technique. We perform experimental comparison of the proposed method against two state-of-the-art tag recommendations algorithms with two real data sets (Last.fm and BibSonomy). Our results show significant improvements in terms of effectiveness measured through recall/precision.
Tag recommendations based on tensor dimensionality reduction
Pheobe said...
Semantic approaches have several advantages compared to collaborative filtering, including eliminating cold start problems
Yes, Phoebe, I have a depth of knowledge in recommendation engines algorithms since I have been consulting on the subject over the last 4 years and even implemented various algorithms in recommendations in certain projects that I got involved in.
As I have stated previously, the tensor type recommendation will make a huge impact because one can go higher than 2D and 3D, such as a 4D (user,items,tags,time-stamp) , etc. Such 4D recommendation will have the capability to recommend only recent items rather than old ones (it the old item and new item are similar). The tensor algorithm HOSVD described in the abstract above is one one type. There are a number of tensors that have appeared in the literatures lately. In tensors, there is potential for still higher dimensions (5D, 6D, 7D or more). Heh, heh, when you see science fiction movies talking about multidimensional universe or hyperspace universe, you know that they mean tensors (Einstein brought the world's attention to tensors when he developed his General Relativity Theory using tensors where science fiction writers have jumped on with phrases such as hyperspace time-warping, etc...).
Phoebe said...
The Jinni team includes world-class scientists who are engaged with new developments in their fields.
Good on Jinni team. I think that Jinni deserves millions of VCs funds (I am serious). The reason is that Jinni is doing real research. My definition of real R&D is researches that involve brains, in contrast to startups that don't need brains and there are heaps of them covered here at RWW and TechCrunch, but they receive millions for something that you could recruit dropout high school students to develop/write. There was the $13 millions funding for a geneology web start up company covered recently here at RWW and it is beyond me that something so so simple like creating a geneology web app can cost that much. These are the types of applications that don't need a brains and anyone who argue otherwise is dishonest to him/herself.
Can you tell us the accuracy of your recommendation engine? What's its classification error rate? Is your engine lower (ie, more accurate) than NetFlix? Is your classification error rate lower than the HOSVD tensor described in the abstract of that paper?
Don't really understand the comparison to Clerkdogs. At least right now, they have a sophisticated "similar movies" feature that's interesting but not the same as personalized recommendations of the kind that Netflix and Pandora - and Jinni, for that matter - give.
Phoebe/Marissa,
Here is another paper that I have come across on using tensor for recommendation, that your R&D team might be interested to take a look at it.
Probabilistic polyadic factorization and its application to personalized recommendation
Cheers.
I found that one of the authors of the paper has a freely downloable PDF copy (see below) from her site:
Probabilistic polyadic factorization and its application to personalized recommendation
Kingsley, thanks for your comments. This is (obviously) a good way to reach us and you're also welcome to email me at phoebe.spanier@ jinni.com.
You're right to note that while we use semantic technologies, our efforts are not - so far - part of the semantic web.
We don't use the sources you mention, we rely on sources focused specifically on film. As we have our own techniques to extract concepts from unstructured data, we don't rely on what people extracted e.g. from Wikipedia in the DBpedia effort. However, as our catalog of titles grows over time, we expect to consider additional sources, possibly including those you listed.
Falafulu, thanks very much for your thoughts and generous knowledge-sharing. I'd be happy to be in touch with you by email, at (as above) phoebe.spanier@ jinni.com.
Does pandora have something like Jinni's tuner? If so, I couldnt find it when I use Pandora.