Well if you liked Gigli...

April 5, 2009 5:28 AM

Clerkdogsworks surprisingly well versus other web-based recommendations, partly because paid enthusiasts are involved, and partly for its intuitive interface.

An interview with CEO Stuart Skorman, previously of "dotshot" fame for founding and selling Reel.com.

posted by hypersloth (51 comments total)

If Clerkdogs doesn't give full access to my data I'm not interested. I've already used Amazon, IMDB and Criticker and they all deny access to your own data; I'm not about to add yet another recommendation service to the list.
posted by Foci for Analysis at 5:51 AM on April 5

LOL, the recommendation game has been going for 15+ years now, this is a capitulation/acknowledgment that even the best computer algorithms and data minding (Amazon, Google) can't match the wisdom of the crowd (or individual in this case). A brilliant simple idea, a knowledge base of video store clerks.
posted by stbalbach at 5:55 AM on April 5

Foci for Analysis, what data on the site is yours to access? It appears to be a closed system, only pre-approved (paid) ex-video store clerks can make recommendations (thus "Clerk" dogs).
posted by stbalbach at 5:57 AM on April 5

Really limited choices. Very few foreign films.
Way back when I was a video clerk at Kim's in nyc b/c I was a movie hound - this seems like clerks from Blockbusters in small-town Ohio. Who needs a search engine to find similar mainstream movies? Netflix suggestions does a better job and it sucks.
posted by hooptycritter at 6:07 AM on April 5

It's never heard of The Inspector General and wants to know if I mean Inspector Gadget.
posted by swift at 6:08 AM on April 5

Just to be clear and then I'll gtfo -
I wasn't trying to go pepsi blue in here, but I entered a few searches last week, thinking I'd predict the matches, was pleased to find alien results, watched some films I'd probably never have known of otherwise, and found the recommendations (and reasons for the choices) to be spot on.

On preview, yeah it's limited but it's still beta - I think one of the articles said they had 5000 movies catalogued, but hey, they're still hiring.
*scurries off to demand payment for the shilling if nothing else
posted by hypersloth at 6:27 AM on April 5

If you like Excalibur, you'll love Ladyhawke!

what
posted by Baby_Balrog at 6:28 AM on April 5

> Really limited choices. Very few foreign films.

Which isn't surprising with it being a relatively new site with recommendations made by just a few dozen people. For the movies the site does have, they seem to have gotten good recommendations (at least the movies I checked). It's a bit like Pandora but for movies, with it being based on item-content rather than user-ratings.
posted by bjrn at 6:30 AM on April 5

I actually did put Gigli in there to see what it would give, and it gave me Out of Sight.

I don't think I necessarily agree with some of these comparisons and I'm not sure I understand some of values for these data sets. Like Upbeat, Downbeat, or Hollywood Feel. If you go to Compare you can see them in action. In Supermatch you can tweak the values yourself.
posted by P.o.B. at 6:42 AM on April 5

tried solaris 1972 - star trek original movie came up. That's like comparing a Porsche to a Yugo.
posted by hooptycritter at 6:45 AM on April 5 [1 favorite has favorites]

If you liked: Old Boy
You'll love: Boys in the Band, Boys Shorts, Boys Town

Needs some fine tuning, apparently.
posted by PeterMcDermott at 6:46 AM on April 5

Really? "Old Boy" isn't even a recognized movie when I try it. And for Oldboy I get Mr. & Lady Vengeance, the Kill Bills, Izo as matches. Or are you just looking at the list of partial title matches for "Old Boy"?
posted by bjrn at 6:54 AM on April 5 [1 favorite has favorites]

stbalbach, I want to be able to export my movie data in several formats. Is the data is in a propietary format in use, I expect it to be fully documented. Better yet, stick to a existing format.
posted by Foci for Analysis at 7:00 AM on April 5

I put in Dune (1984) and got out Jodorowsky's The Holy Mountain ("Our Clerk Cory Says: More Psychedelic"), so that's a win.
posted by Artw at 7:18 AM on April 5 [1 favorite has favorites]

Foci for Analysis, from what I understand, they are re-selling the data to companies who need a recommendation engine. I doubt they would open up the data since it's a source of revenue.
posted by stbalbach at 7:22 AM on April 5

Commando... 'Clerk's pick' = Cobra. FAIL
posted by fearfulsymmetry at 7:42 AM on April 5

Bring Me the Head of Alfredo Garcia brought up The Three Burials of Melquiades Estrada as top match - with No Country for Old Men as "Equally Bleak" and To Live and Die in L.A. as "Equally Cynical".

That's pretty impressive.
posted by Joe Beese at 7:42 AM on April 5 [1 favorite has favorites]

Plugged-in Them and got Kiss Them for Me. Ummmm...no.
posted by Thorzdad at 7:45 AM on April 5

Read Heat gives me dead Heat ("More Horror Oriented ").
posted by Artw at 7:45 AM on April 5

One of the recommendations for Bladerunner was a Paprika, an anime I've never heard off but it's now on my UKEquivalentToNetflix queue...
posted by fearfulsymmetry at 8:07 AM on April 5

Tremors gives Night of the Comet. WIN!!!
posted by P.o.B. at 8:11 AM on April 5

Paprika is AWESOME!
posted by P.o.B. at 8:11 AM on April 5

Somehow Robocop gave me this, way down the list. WTF? Is it like a Belguian version of The Wrestler?
posted by Artw at 8:17 AM on April 5

A brilliant simple idea, a knowledge base of video store clerks.

Trying it out a few times confirms this. However, the video store clerks in question are apparently all of the "hey, if you liked that, you should totally rent Armageddon!" mode.
posted by mightygodking at 8:18 AM on April 5

Plugged-in Them and got Kiss Them for Me. Ummmm...no.

Eh?
posted by Artw at 8:20 AM on April 5

Seriously JCVD is a good watch, check it out ArtW.
posted by P.o.B. at 8:22 AM on April 5 [1 favorite has favorites]

partly for its intuitive interface

At least two people in this thread have confused the search results with the movie matching, so I'm going to say maybe not quite as intuitive as it could be.
posted by ook at 8:40 AM on April 5

I don't take movies very seriously (I mean, to me they're just entertainment right, not life-changing or anything), and have often been told I have bad taste in movies. I'm the kind of person who usually doesn't even know about a movie's existence until like 5 years after it was released. This thing is pretty much perfect for me. Thanks!
posted by greasepig at 8:43 AM on April 5

This is really good. Anything that leads a fan of Ghost in the Shell to Serial Experiments Lain is doing it right.
posted by 5imian at 9:00 AM on April 5

I doubt they would open up the data since it's a source of revenue.

Also, really, why would they bother spending development time on this?
posted by smackfu at 9:08 AM on April 5

I like the "More Pathetic" category.
posted by odinsdream at 9:15 AM on April 5

Artw: Weird. You got completely different results than I did. My results weren't nearly as large, not did they include any other big bug movie.
posted by Thorzdad at 9:36 AM on April 5

"Also, really, why would they bother spending development time on this?"

Maybe for the million dollar Netflix prize and associated $50,000 progress prizes? (www.netflixprize.com/rules). From a quick scan of the rules, it appears that Netflix is looking for a purely algorithmic solution, but might accept a human mediated method. Clerkdogs also might help find the extra dimensions needed to form a better algorithm.
posted by fydfyd at 9:41 AM on April 5

I really like that it gives reasons for the recommendations, and those reasons are short and useful.

The ajax on the search field is a bit glitchy though, leading to an iPhone text prediction like level of overeager helpful unhelpfulness.
posted by Artw at 9:45 AM on April 5

The Aristocrats returned nothing, confirming my belief that it is peerlessly funny and one of the best movies ever.

I've even looked on IMDB to see if the editing was Academy nominated. It should have been.
posted by geekyguy at 9:59 AM on April 5

Maybe for the million dollar Netflix prize and associated $50,000 progress prizes?

I was referring to Foci's need to export his data, which is just one of those niche demands that you might have in a mature product, but not worth the development time in a new app.

What is worth development time? You probably get 1-2 minutes of attention from the casual browser following a link from someplace like here, so you want to wow them and make it stupidly easy to use the tool. I think they actually do a good job of that -- you can get a recommendation with one click on the front page. Although it would be smarter of them to bypass the search results page, and just show the best guess and offer the option to see others. Because getting a search results page back is a killer. You end up showing your users how good your search algorithm works, not how well your recommendation algorithm is.
posted by smackfu at 10:00 AM on April 5

LOL, the recommendation game has been going for 15+ years now, this is a capitulation/acknowledgment that even the best computer algorithms and data minding (Amazon, Google) can't match the wisdom of the crowd (or individual in this case).

Well, no, it's not. Netflix hasn't said "Whoops, we give up on algorithmic solutions, people are better." This is simply someone trying an alternative approach; time - and maturity of the solutions - will tell us which works better.
posted by Tomorrowful at 10:05 AM on April 5 [1 favorite has favorites]

> You got completely different results than I did.

I think you are looking at the search results page, which it displays because the site isn't sure which movie you meant. If you click on "Them! (1954)", you get to the actualy movie matches page Artw linked to.
posted by bjrn at 10:09 AM on April 5

Works reasonably well for me, and I have crazy taste.

In fact it's so clever I may suspend my 20-year war on the term "intuitive interface" (they mean "intuitable", right?)... for today only.
posted by rokusan at 11:18 AM on April 5

argh, intuitable. I even spent a few minutes thinking about what term to use and I still anthropomorphized a website.
posted by hypersloth at 1:35 PM on April 5

oh I just noticed they said it first, so I feel slightly vindicated.
posted by hypersloth at 1:42 PM on April 5

It's interesting that movie recommendations and music recommendations went on 2 diverging paths.

Movie recommendations, as exemplified by Netflix (but see also Movielens, which is set up by the University of Minnesota as a research project - Foci for Analysis, I'm not sure if they have exportable lists but if anybody would be willing to put it in they would) and Imdb have gone the route of automatic links between videos, based on user clustering. So User X voted similarly to User Y on a bunch of movies, so let's recommend a video to X that Y has seen and liked. Now we have this entry that uses expert recommendations.

Music recommendations, as far as I can remember, got their first big break with Pandora. Pandora, through their Music Genome Project, hired qualified musicians to analyze songs for dozens if not hundreds of characteristics, and built a map out of this information. It doesn't care what other people do, but what the songs are like. However, the market moved towards the clustering & nearest neighbour approach with Last.fm, or iTunes' Genius.

It's kind of a criss-cross, and I wonder if there's a reason to be found in peoples' different view on what movies are to them compared to music. Movies are more generally a given event, for entertainment, whereas music is on while you walk or work out, while you're reading or romancing. I think it took longer for people to give up the notion of personal recommendations for music, which Pandora feels like.

Or I'm just talking out of my ass at the end there.
posted by Lemurrhea at 2:22 PM on April 5

LOL, the recommendation game has been going for 15+ years now, this is a capitulation/acknowledgment that even the best computer algorithms and data minding (Amazon, Google) can't match the wisdom of the crowd (or individual in this case).

In addition to Tommorowful's point about this not being a retreat - what do you think the best computer algorithms mine for data? The wisdom of the crowds. I'm not saying there is no place for expert advice - but when you use a phrase like 'wisdom of the crowd', you're referring mainly to Nearest-Neighbour Clustering measuring against a root mean square error. You're referring to singular vector decomposition, because without these tools to take millions upon thousands (as in, 10 millions users rating 10 thousand videos) of entries, we wouldn't have the slightest clue about what the crowd is thinking.
posted by Lemurrhea at 2:27 PM on April 5

> Music recommendations, as far as I can remember, got their first big break with Pandora.

Last.fm/Audioscrobbler was around quite a while before Pandora was launched. I also remember iRate, which was a sort of open-source streaming radio thing using the listeners' ratings to pick what you'd hear next. And then of course there's Gnoosic/Music-map which also takes the clustering approach.


I think one problem with the numerical approach is that you're working with averaging numbers. As your dataset and userbase grows, everything will move closer and closer together towards the middle. Although I don't check often, I can't really remember when I last saw a predicted rating (for a movie that I hadn't seen before) that was above a 4 out of 5 stars. There was a short discussion about it here if you have a ML account, where one user notes that

Actually, the problem seems to be more significant than I at first thought. I've done some poking around and have yet to find a single prediction for me (with almost 1000 movies rated) that is more than 0.4 stars away from the average user rating. In other words, rather than an individualized prediction, it appears that I am merely getting averages rounded up or down to whole and half star increments.

As a side-note, Movielens doesn't use user-similarity anymore for predictions, but item-similarity instead. So instead of seeing which users are alike, they see which movies have similar ratings. They started out with the former but stepped over to the latter. In general the two approaches seem to perform about the same, but item-similarity is less resource intensive.
posted by bjrn at 4:10 PM on April 5

On the basis of Amazon or Netflix I like item-similarity better - user-similarity basically means that you glance at Akira once and you;re getting recommends for Anime titles of no interest to you forever.
posted by Artw at 4:14 PM on April 5

It's interesting that movie recommendations and music recommendations went on 2 diverging paths

My guess is that similarity in music is a simpler thing. It's only a 3-5 minute song, fairly regular in sound and speed from start to finish. If you pick the right attributes, which like you say may require a human to evaluate, you have a good chance of properly classifying the song. Then if you spit out another song that matches the attributes closely enough, that's a good recommendation.

In contrast, movies are just two different, and have too many attributes, for the same approach to work.

An analogy in music would be if I really like songs where the music drops out and the singer does a bit a capalla. A music recommendation engine has no chance of figuring that out, unless that is one of their attributes.
posted by smackfu at 4:14 PM on April 5

I think one problem with the numerical approach is that you're working with averaging numbers. As your dataset and userbase grows, everything will move closer and closer together towards the middle.

This is why you aren't just averaging numbers - you take different people's tastes into account. Basically, these algorithms try to identify not "what people like" but "what people like you like."

For example, due to my own taste patterns, Netflix's algorithm suggests a five-star rating for several MST3K volumes - despite the overall average being only 3.9. Somewhat more astoundingly, there's a sci-fi flick called Bad Channels that Netflix suggests as a 4.2 for me - when the average is only a 2.1. Nights in Rodanthe, on the other hand, averages 3.5 but is only a 1.9 for me. And my numbers are probably a little inaccurate because I just haven't watched too many movies I've savaged with ratings.
posted by Tomorrowful at 4:28 PM on April 5

Amazon Recommends: Would You Like Some Wolf Urine With That?
posted by rokusan at 6:22 PM on April 5

Is Amazon still held up as a providing good recommendations? They seem like the most obvious and poor choices whenever I look, and fall into a few categories:

1) You bought this popular item, you may like other popular items, because you are a tool.
2) You bought a book by X, you might like other books by X. Bet you couldn't figure that out.
3) You bought a book on subject Y, you might like this other book on subject Y that covers the exact same topics.
posted by smackfu at 9:02 PM on April 5 [1 favorite has favorites]

> This is why you aren't just averaging numbers - you take different people's tastes into account. Basically, these algorithms try to identify not "what people like" but "what people like you like."

And I'm saying that although you're not averaging over all users, you are still averaging over a subset of users and/or items, and this will result in predictions moving toward the middle. You might not be averaging all numbers, but you are still in essence "just averaging numbers" (with some bits of trickery added in).

With my six hundred odd ratings on ML I don't think it's quite as extreme for me as for that user I quoted from, but at the same time the lack of 4.5 and 5 star predictions is something that hasn't passed me by unnoticed.
posted by bjrn at 1:23 AM on April 6

It's interesting that movie recommendations and music recommendations went on 2 diverging paths.

Maybe there's no one-size-fits-all answer... The closest I've seen to "Pandora for Movies" is Jinni. Their Movie Genome Project looks to be very promising, with search by mood and recommendations based on each person's taste.
posted by astorygirl at 8:50 AM on April 12