Wikilens vs Netflix

WARNING:  This is going to sound more like an op-ed piece than a scientific analysis

  After a couple of weeks of playing with WikiLens I can't help but compare it to Netflix.  Yes, it is very different from Netflix.  It allows new categories to be made by any user.  It is not confided to movies (though it does have a movie category which appears to have the most data).  Users can add their own entries without permission.  The engine does NOT wait to your input for it to give you a prediction value.  And life is great at Wikilens.  Wait a sec...

It isn't as fun.  I enjoyed filling in those little stars on Netflix.  I enjoyed seeing the results that came back as a results of my previous entry.  Somehow it doesn't feel the same way over at Wikilens.  Maybe because the interface doesn't have nice little pictures of the movie poster... No.  Maybe because I know I'm not going to get a red envelope in the mail... Nope.

  The reason is the manner in which WikiLens presents the next "item" for rating/prediction.  It feels more like I'm taking a test and am ready to bubble in the next answer.  Netflix's interface feels more like the end of a magic trick.  The prediction is magically plucked from the inside of a hat like a rabbit  (no 'Harvey' jokes please).  

  Earlier I mentioned that the prediction were given before I rated anything.  Perhaps this is where the magic is lost.  To me, the user, it doesn't feel like my tedious work of adding smiley faces is affecting the results.  I know that it is deep down in my subconscious.  It just doesn't recommend the feeling to me.

  It still remains to be a wonderful tool.  Now if I can just convince more cyclist in Charlotte to enter data for me.

Trust issues

After reading the article for today's class I thought about a project I worked on last semester in an HCI class that focused on security and privacy.  The goal of the project was to determine whether or not the average UNCC student was more likely to reveal private information, that could compromise their UNCC email account, to a computer or to a random stranger person.  The survey was disguised as a Facebook survey due to the high number of surveys that occur on campus that focus on Facebook. The results found that among the students surveyed they were equally likely to reveal information to either a computer survey or a student surveyor.  

The link below details the project.  The data doesn't appear to be there for some reason.  I'll work on that.

The survey.

http://hci.sis.uncc.edu:8080/itis6010-fall07/26

Interesting Slashdot post about Netflix Prize

This is a couple of weeks old but I thought it should at least be posted here.  It is about one of the leading competitors that emerged recently on the Netflix Prize scene.  They claim that he does not have a computing or math background.  A little research shows that he does.

http://developers.slashdot.org/developers/08/03/04/2348257.shtml

New Version of UC Berkley's Joke Recommender

  This article describes UC Berkley's joke recommender and the updates that were just released.  Two pieces of software were released into new versions.  EigenTaste is now in version 5 and Jester is now version 4.  The system now has over 4 million ratings.  The article is an elementary look at this collaborative filtering algorithm run system.  It does describe the features and some of the jokes. ;)  Prof. Ken Goldberg notes that the engine can be applied to anything with large inventories.  Donation Dashboard is one other application that he mentions.  I seem to remember talking about this earlier in the semester.  Is that right?

  I have noticed a trend in systems I've read about recently.  Taking into account the user's most recent input seems to be popular.

The article: http://www.networkworld.com/community/node/26480

The Slasdot article: http://idle.slashdot.org/idle/08/03/31/1633223.shtml 

The updated software: http://eigentaste.berkeley.edu/info.php

Wikilens Charlotte Bicycle Shops

I have started a new category on WikiLens.  It is "Charlotte Bicycle Shops."  The goal is to enter all bicycle shops in the area and get users to rate them.  Finding a shop that suites a cyclist's needs is not always easy.  Different shops sell different brands, offer different services, open different hours and attract different styles of riders.  Shops have a overlap in this areas.  As a cyclist I usually am loyal to one shop.  However, I do find the need to shop at multiple shops for various reasons.  Often when I try a new shop I am comparing to pat ones and looking for the same "feel."

To start on a new system on WikiLens you have to create a new category and new items(shops) for that category.  Not the easiest data insertion process... but what is?

Currently I have a reasonable number of the shops in the area entered and am twisting the arms of my friends to rate the shops.  It isn't producing recommendations yet.  But I don't have much user data so far and don't want to create fake data for it.

Transition from Youtube Journaling to WikiLens

It appears that YouTube is not a personalized recommendation system.  Instead YouTube is only keeping track of Favorites so the user can go back and re-watch the videos again.  The ratings are not being used to recommend videos with similarities and that are enjoyed by users with similar interests in videos.  The ratings are only being used to rank how popular a video is.  Popularity determines rank in search results.  If two users with dissimilar interests that have rated videos query the same thing, they will receive the same videos.  Those videos will always link to the same videos regardless of declared preference.  The linking is determined by the user that has submitted the video and by videos with similar words in the title.

Since this is the case, I will be switching the focus of my journal to WikiLens.  It is a recommender engine that I have covered in one of my earlier posts.  I will begin establishing a more in depth user profile and will also begin a new recommender topic using the engine.  I think that it will most likely have something to do with cycling.  At the moment I am testing the feasibility of routes, bikes, races, teams, shops and vehicles for cyclists.

Recommender system and method for generating implicit ratings based on user interactions with handheld devices

This recommends items based on how a user interacts with a handheld device.  Users are given a repetitive task to do on a mobile device.  Their interactions are recorded in a history.  This history is used to create recommendations.  Unlike Amazon's content based system, this system takes into account how recent an event or interaction occurred.  More recent interactions are considered to be more relevant.

According to the claims, the calculations for the recommendations are made in the following three ways:

• "rating(item)=number of interactions(item) since datetime(item acquired)/number of total interactions (item) since datetime(item acquired)."
• "rating(item)=total interaction time(item)/size(item)"
  

• rating(item)=[total interaction time(item)/size(item)*exp(−damping coefficient]*(date−time acquired).

The first method takes into account the recency of the interaction.  The second method takes into account the amount of time that the user spent on a particular interaction.  The third method takes into account something that I don't understand.  I can't figure out what the difference is between claim 2 and claim 3.  They appear to be the same with the exception of the calculations.

  Mobile devices that the patent covers in its data gathering include cell phones, mp3 players, PDAs, and electronic book readers.

**WARNING this was probably written by a lawyer***

http://www.freepatentsonline.com/6947922.html

http://www.google.com/patents?id=QkEWAAAAEBAJ&dq=09/596070

Photoree

http://www.photoree.com/photos/tour

 The site describes itself as being "to photos what Last.fm is to music and StumbleUpon is to websites."  It is an image recommendation system that uses a thumbs up or thumbs down method of rating.  Preferences are recorded and used on the fly to recommend other images that the user might prefer.  The recommendations are user-to-user based and might cross some privacy boundaries if not careful.  The user is allowed to see the stats on how popular an image is.  It goes on to tell the user WHO liked or disliked the image.  There are security settings that the user can opt into to decide not to share their preferences to others.

***WARNING***

There are no guarantees that the images made available for rating or suggestion will not be nudity.

***WARNING***

Youtube Doppleganger

For three weeks now I have been rating youtube videos under to different accounts.  One account prefers to watch music videos and movie trailers while the other account prefers to watch stand up comedy and magic tricks. 

For both accounts ratings were done by adding the videos to the "Favorites" list and by using a 1 to 5 star system.  

Searching for 'comedy' and 'card tricks' in each account yields mostly the same results with roughly 10% difference.  'Comedy' yields the same results for both.  

The main page, channels, and 'video' section for each always shows the same items.  

The interesting thing is that the 'Promoted Videos' are different for each always.  

Youtube --bubbles

When watching a video on YouTube in full screen mode there is a button that appears next to the play button.  This button, which I'll call the bubbles button, reveals an interactive network of "related" videos.  I say "related" because they appear to related mostly by submitting user and popular vote.  The video that you are watching appears as a slightly larger bubble or node in the network in the center.  As the user mouses over the other bubbles, more bubbles appear.  Mousing over one bubble reveals all bubbles that are directly linked to that bubble.  The network plain is expandable in the X and Y directions.  

I still have found no evidence, other than popular vote, to show that the recommendations are at all personalized to the user.

For next week I'll set up another account and make very different 'favorites.'  Hopefully I'll be able to do the same search from each account and compare the results to see how personalized they are.

CarWale.com recommended that I get new interests in cars

Carwale.com

This recommendation system uses car attributes and car shopper needs and preferences to recommend cars.  I attempted to find something that matches what I want in my new car.  I had to try several times before it was able to give me something.  I ended up giving it very little information.  

The content based system "works" by collecting information on preferences through a series of check boxes, fill-in the blanks and fancy sliding values.  It then attempts to find a vehicle or vehicles that matches those values.  It does not ask questions that deal with your interests or hobbies.  The attributes are all specific to the cars themselves.

No special account is required to use the system... only special interests.

Link

YouTube -- YouRecommend???

I've begun looking into YouTube's "recommendation" system.  At first I'm quite confused about what is going on under the hood.  YouTube definitely has a content based system.  At the same time however it allows users to rate videos Netflix style.   It also allows users to declare something as a "Favorite."   It is unclear to me whether or not YouTube uses my ratings and favorites to "recommend" videos to me.  The Netflix style star system is being used for sorting.  Videos can be sorted into a 'Top Rated" category.  I have not found a system the appears to be using these ratings other than "Top Rated."  To test I began rating videos.  I rated mostly music video, trailers, and videos of Santa Claus falling of the roof.  I hoped to sway any possible recommendation system towards videos similar to them.  Once I returned to the main page I found that I was being should videos of "Super Tuesday."  This either means that not enough input has been given, no system is in place using the stars, or the JUNO soundtrack is closely related to Barrack O'Bama.

GoogleTechTalks on WikiLens - User generated recommendation data

A talk by Dan Frankowski on BEER!  More spefically Frankowski talks about WikiLens.org, a site which allows users to start their own recommendation systems on any topic including beer.  This system works similar to a wiki but does require some level of user rating before changes are allowed to be made.  Like a wiki, users can start a new topic or add content to an existing topic.  The goals of project are to allow users to FIND, ADD, create DEEP CHANGE, MICRO-CONTRIBUTE, and SEE OTHERS. 

FIND: Members should be able to find items that interest them

ADD: Members should be able to add items immediately

DEEP CHANGE: Members should be able to uniquely identify items, and define and redefine their attributes and organization

MICRO-CONTRIBUTE: Members should be able to make small contributions

SEE OTHERS: Members should be able to see each other and their contributions

Frankowski is  software engineer for Google Groups and his paper is called "Recommenders Everywhere."

The video: (~35 minutes)

http://www.youtube.com/watch?v=bsgLSb9Dbz8

The slides:

http://www.cs.umn.edu/~dfrankow/files/wikilens12.ppt

Microsoft's Behavioral Recommender System's Patent

http://www.google.com/patents?id=XVh4AAAAEBAJ&dq=recommendation+systems

View patent at USPTO (will require a special plug-in)

Filing date: Jul 15, 2003
Issue date: Mar 14, 2006
Inventor: Christopher B. Weare

Claim Summary:

This system is unique from attribute based or collaborative filtering based systems in that it recalculates to make predictions based on current behavior of the user.  Users rate items and rate the recommendations made by the system.  This is still similar to Netflix to me.

I have not been able to determine *exactly* where MS is using this.  It is possible that has been shelved.