Slashdot Mirror
When Profiling Goes Wrong
huskymo writes "This morning's Wall Street Journal is carrying a funny story on TiVo and Amazon's automatic customer profiling. As most Slashdot readers probably know, TiVo keeps track of which programs you record and--if you haven't told it not to--records other programs it thinks you'd like. The article describes users that TiVo's mistaken for Korean, for gay, even for "a pregnant gay man.""
Funny as hell.
i dunno...i'm guessing...
"Sex and the City"
"Will and Grace"
"Dr. Phil"
"Facts are meaningless. You could use facts to prove anything that's even remotely true." - Homer Simpson
It looks to me as if they simply look at the genre of the program you rate high and then take that to be your preference.
I found out that the hard way, one day I went home and I found the tivo filled with idiotic shows like: "Price is right" and "Spend $1000 in 1 minute", "Blind date" etc... upon investigating I realized that I've have rated "Junkyard Wars" (a competition of building things from junk) and "BattleBots" (remote controlled robot fight show) high the previous day, this triggered the game-show category to be recorded.
As Larry David would say: pretty-pretty-pretty dumb.
"I'm just counterprogramming because TiVo thinks I'm gay."
That really is quite funny. I think we've hit on a new tech-term: counterprogramming - noun - to use the front-end of a software program to perform operations with which the backend program should have been able to do in the first place.
"This isn't a study in computer science, its a study in human behavior"
...give a checkbox in the user preferences, "I {do,do not} have an interest in stories from subscription-only sites."
You cannot apply a technological solution to a sociological problem. (Edwards' Law)
It was funny until I realized this is the sort of system the US gubmint wants to use label people as enemies of the State.... All Hail the Ministry of Homeland Security!
During the Final Four last year, my beloved Kansas Jayhawks were playing. I came home from a happy hour the Friday night before the final four to find that TiVo had recorded the Final Four practices for all four teams. I didn't even know anybody would be nuts enough to cover that non-event. Needless to say, I was thrilled to see KU and their opponent's practices appear, unbidden, on my TiVo.
For that occasional miracle, I'll take all the Univision soap operas, shopping channel dreck, and Korean news I can delete, and I'll thank TiVo every time for trying.
ceci n'est pas un sig.
Here's an interesting question for fair use. By posting it in the forum, he enables other people to comment - and they've commented. So while on an individual level, there's no commentary, on a group or site level, there definitely is.(Me not lawyer.)
I find it remarkable (and alarming) that nearly every person who felt they'd been mis-profiled responded first by altering their viewing habits, and second by *buying more*. It's as if the profiler encourages the viewer to increase their consumption by playing off their insecurities-- I wonder if this is by design, or just a happy accident for the people in marketing.
Then again, come to think of it, I suppose the entire advertising industry operates this way- alter people's behaviour (and boost their inclination to consume) by exploiting their insecurities. The moral of the story- turn off your TV!
So long, and thanks for all the Phish
I find it interesting that people are talking about their Tivo's like they're sentient beings crouched on top of the TV, casting judgement on the crap you watch and recommending new crap to watch. It's just a computer program, people, and likely a fairly simple one at that.
I don't have a Tivo (or a TV, for that matter), but my Amazon profile still hasn't recovered from when my wife was in graduate school studying developmental psychology, specializing in childhood trauma. More books about child sexual abuse, just what I wanted. =:-O The programming books are staging a comeback, though.
What I find particularly interesting are the "people who ordered this also ordered these" selections. On infrequently-ordered titles, it only takes one or two wackos with bizarre profiles to generate some really peculiar results.
Regardless of linking to pay sites. I find your idea for a website that anonomously tracks the habits of a Korean gay pregnant man intriuging. I'd like to option it for a screen play. With Arnold Schwarzenegger apparently taking a break from politics, I think the time is right for just this sort of story.
I'm thinking of something along the lines of The Manchurian Canidate meets The Net meets Raw Deal. Karen Mistal could be the vapid love intrest who puts him in a family way. John Ashton the unscrupulous villain. George Clooney the dashing rival. And a cameo by Gary Condit.
Well it certainly couldn't be worse than extreme ops, or half past dead.
--Jimmy has fancy plans; and pants to match.
Please ensure your brain is fully in place before engaging your loud, rude mouth, next time.
Like woodworking? Build your own picture frames.
Which is why we should all fight against the lame brains in Washington that believe in it. You can't predict an act until the act has been carried out. This of course isn't prediction.
Innocent until proven guilty, that used to be the law.
Do unto others as you would have done to yourself, don't let America become like Israel. It is un-American to support human rights violations, support justice in Palestine.
Am I a terrorist or a Patriot? What about George Washington?
If your country were being invaded by religious fanatics with a Nazi-like chosen race mentality would you give your life to defend your country?
Consider me an anti-Semitic now. What if I am Arab and therefore a Semite?
Profiling is just another simple solution for simple minds. It assumes some minimum set of categories or some nearest neighbor over some minimal set of dimensions and does not take into account the broad reach of the human mind.
How many people in DC might have been saved if the police had of kept an open mind instead of looking for a 20-40 year old white male sniper?
Of course, the reality is that profiling says more about the profiler than the profiled. What are the preconceptions of the creator of the profiles and their predilection to pigeon hole people into categories. And given this predilection to error their other knowledge is suspect and generally junk knowledge. For instance, I would venture to guess that many people would believe that if a person said that there were more than one "begotten" sons of God that they would be a polytheist. Ever read Psalms 2:7? So can I be a Christian and believe that "begotten" actually means something different from what other Christians mean? Or was it just a very bad translation from a person that already had a fixed paradigm. Can I be a Christian and a Muslim at the same time? What exactly do the Muslims say about Jesus? Do you think you know?
How would you profile me now?
They probably don't have categories such as "gay" "Korean" etc. at all. More likely it correlates your preferences with the preferences of other TiVo users. e.g., you recorded X, a lot of other people who recorded X also recorded Y, so it will recommend Y for you. (The actual statistical algorithm is probably more complex, but that's the basic idea.) No explicit categories necessary at all.
Never take moderation advice from sigs, including this one.
- Suggestions are optional. There is a menu setting under "Messages and Setup ==> My Preferences. Don't care for suggestions? Turn them off. (like they are on my TiVo)
- TiVo will never delete something you have as a Season Pass in order to record a suggestion. If stuff is expiring too quickly, that's just a sign you need to put in some bigger hard disks.
(or watch less TV)
- If TiVo records something you don't like, give it one thumb down before you delete it. That's all it takes. Seriously! The only thing you should give three thumbs down to is Paid Programming.
- If your suggestions get too out of whack then you can clear all of them. Go into the "System Resets" under "Messages and Setup" and there is an option to "Reset Thumb Ratings and Suggestions".
- A good resource for all things TiVo is the Tivo Community forum at
http://www.tivocommunity.com/tivo-vb/
You'll learn more than you ever wanted to know about TiVo there.
(they'll probably kill me now for linking to them from Slashdot.)-bwillcox-
(owner of a Philips S1 TiVo with 249 hours + turbonet and tivoweb)
The really stupid part is when people have to go out of their way to search for things they have no interest in just so it will stop making bad recommendations. Of course, if I had the source I'd probably waste far more time tweaking it :)
They that quote Benjamin Franklin on liberty and safety deserve neither.
Remember oliver, the electronic personality extender predicted by Alvin Toffler in "Future Shock" ...?
... I predict that The Sims will fuse with Counter-Strike into a new game where heavily-armed psychopaths massacre hapless suburban clones ....
There's an interesting passage about olivers in John Brunner's excellent novel, "The Shockwave Rider":
"... so-called olivers, electronic alter-egos designed to save the owner the strain of worrying about all his person-to-person contacts. A sort of twenty-first-century counterpart to the ancient Roman nomenclator, who discreetly whispered data into the ear of the emperor and endowed him with the reputation of a phenomenal memory." (pp. 41-42)
More than a few of those emperors went crazy from all that power, which makes me wonder:
What happens when tens of millions of 21-century citizens have their personalities extended -- and some of them already crazy?
Well, for a start
-kgj
I hate to agree, but I do. I was initially of the "Oh. man, that stupid profiling nonsense" mindset. However, I, too, have discovered new authors as well as new areas of scientific and technological interest thanks to Amazon's suggestions. Haven't had any major howlers, though.
I learned if I add a book for someone else to one of my orders, I should always go in and manually remove it. I ordered a political book for my dad once, and started getting recommendations for all sorts of ideological fluff from Michael Moore to whatsherface... the blond conservative woman I want to bang... Ann Coulter, I think. Ooo, boy! The world according to various clogged political filtration systems! I'll just pass by the latest book on superstring theory or AI for one of those!
--- Ban humanity.
While I can't really know what kind of effort was put into these systems, it seems unlikely that Amazon or TiVo hired a team of veteran AI developpers to build these features. That being said, this problem still underlies a trend in all AI systems, no matter how good. That is that they are all really quite dumb when you compare them to anything we would call "reasonable" intelligence. They are incredibly fallible, incredibly silly machines in terms of their output to a large extent. Sure they can be made to do wonderful things, but it always has to be done with a group of human "moderators" to judge and assess the machine's performance and output, and that with a large grain of salt.
The idea that a machine is objective and not biased like people is absurd. They are more biased. They can only follow mindlessly the rules set down for them by their designers, and do not have the breadth or depth of experience that people have to know when the rules don't apply. Even the best dynamic systems and neural nets have these flaws somewhere or other. While it is funny to see them goof like this, it is scary to think that Governments are going to use similar techniques for vital things like law enforcement. This is a serious concern to all of our civil rights.
There are a thousand forms of subversion, but few can equal the convenience and immediacy of a cream pie -Noel Godin
I watch "Buffy the Vampire Slayer" every Tuesday, but have no interest in "Clueless", "Charmed" or "Sabrina the Teenage Witch", as I am neither a teenaged girl, nor am I obsessed with them.
A properly configured TiVo would see that I like "Buffy" & "The Sopranos" and realize that thoughtful writing is what captures my attention, making reccomendations like "The Industry" when its on PBS (That's the US title for "Made In Canada", the funniest and smartest damn sitcom I've ever seen.)
Information wants to be anthropomorphized.
It always seemed to me they are using a linear scaling system rather than a logarithmic system with a minimum threshold. I also wonder why they don't allow the user to easily reset the profile. The best question would be, with all the complaints, why are they attempting to fix the problem?
Define "does not work." I think profiling often does work, not with 100% accuracy, but it does work. The more data you give it, the better the output you get. Profiling with a very low n is bound to be silly, but as you increase it the results get better.
Profiling is just a statistical correlation (or probably a more complicated type of factor analysis) that says people who like X,Y,Z also tend to like A,B,C. Yes, aberrant data (buying a needlepoint kit for my aunt) can skew the results, but again as the amount of info goes up, the ability of such outlying points to skew decreases. I am more and more impressed with both tivo and amazon's picks as the amount of data it uses increases (even more by tivo, as it is getting a rather split personality input between my picks & my gf's).
point out what people really want is contextual information at the moment
Sometimes. Somtimes people want a list related only to the current action, such as when buying presents. On other occasions, they want a recommendation based on previous likes as well.
Profiling could be used to predict what you want when you ask for it. For example if you are making a query for children's books then the top 10 could be presented. Or a couple of questions could be asked and then a top 10 is presented. But in either case the background of the user is not taken into account.
But why not use the background? Why should I have to repeatedly answer questions to give it information that could be obtained passively?
For different situations, different methods will work better. For things that tend to be more personal (tivo), profiling is great. For things that aren't (gifts from amazon), context would be better. But to say 'profiling doesnt work.' is just wrong. [evidence, look at the ultimate profiling: car insurance...]
-Ted
-=-=- Quantum physics - the dreams stuff are made of.
As someone who's done work in the field (both academic research and in the private sector), these systems don't "profile" the way, say, police officers do. I can't speak to the specific algorithms used by Tivo or Amazon, but the techniques are generally the same. (Though I can make really good guesses about Amazon, since they have a patent with some specific algorithm descriptions.)
These systems (generically, recommendation systems or collaborative filtering) don't use pre-defined genres or categories. They use correlations between actions to predict future actions. So your recommendations are essentially based on the sum of your past actions. In other words, you can't make it ignore "gay" stuff by selecting "macho" stuff-- it will just sum those together and you'll get both "gay" and "macho" stuff.
Worse, if enough people try to outsmart the system, it pollutes the correlations. "Gay" and "macho" items become linked together, and requesting one makes it recommend the other. This can work both ways. If most of the people who record "The Rise and Fall of the Third Reich" are doing so to counter their watching "Queer as Folk", then someone who watches "Third Reich" will get "Queer" as their first recommendation!
In case it's not clear how this works, let me describe one generic type of recommendation system. The system forms rules like "People who like A also like B, C, and D" based on analyzing its database, with some definition of "like." This being e-commerce, it's usually "like=buy". It might be more complicated, e.g. "35% of people who buy A also buy B."
These rules can be shown raw, as Amazon does, or they can be personalized. I've bought A, C, and D, so it combines the rules for A, C and D (using set intersections, sums, averages, etc. depending on how the rules are stored.) It decides, in essence "People who like A like B; people who like C like B; you like A and C, so you probably would like B."
So if you wanted to counter your watching "Queer as Folk", what you would want to do is not train it with "anti-gay" input, rather you would want to flood it with "everything but gay" input-- BUT you would probably have to do it IN EQUAL PROPORTION to the viewing habits of the average person in their database.