Augmenting Data Beats Better Algorithms

eldavojohn writes "A teacher is offering empirical evidence that when you're mining data, augmenting data is better than a better algorithm. He explains that he had teams in his class enter the Netflix challenge, and two teams went two different ways. One team used a better algorithm while the other harvested augmenting data on movies from the Internet Movie Database. And this team, which used a simpler algorithm, did much better — nearly as well as the best algorithm on the boards for the $1 million challenge. The teacher relates this back to Google's page ranking algorithm and presents a pretty convincing argument. What do you think? Will more data usually perform better than a better algorithm?"

Depends on the Problem

[roadkill_cr (1155149)]

I think it heavily depends on what you're kind of data your mining.

I worked for a while on the Netflix prize, and if there's one thing I learned it's that a recommender system almost always gets better the more data you put into it, so I'm not sure if this one case study is enough to apply the idea to all algorithms.

Though, in a way, this is sort of a "duh" result - data mining relies on lots of good data, and the more there is generally the better a fit you can make with your algorithm.

Re:Depends on the Problem

[blahplusplus (757119)]

"I worked for a while on the Netflix prize, and if there's one thing I learned it's that a recommender system almost always gets better the more data you put into it, ...."

Ironically enough, you'd think they'd adopt the wikipedia model where their customers can simply vote thumbs up vs thumbs down to a small list of recomendations everytime they visit their site.

All this convenience comes at a cost though, you're basically giving people insight into your personality and who you are and I'm sure many "Recommendation engines" easily double as demographic data for advertisers and other companies.

Re:Depends on the Problem

[roadkill_cr (1155149)]

It's true that you lose some anonymity, but there is so much to gain. To be perfectly honest, I'm completely fine with rating products on Amazon.com and Netflix - I only go to these sites to shop for products and movies, so why not take full advantage of their recommendation system? If I am in consumer mode, I want the salesman to be as competent as possible.

Anyways, if you're paranoid about data on you being used - there's a less well-known field of recommender systems which uses implicit data gathering which can be easily setup on any site. For example, it might say that because you clicked on product X many times today, you're probably in want of it and they can use that data. Of course, implicit data gathering is more faulty than explicit data gathering, but it just goes to show that if you spend time on the internet, websites can always use your data for their own means.

Re:Depends on the Problem

[teh moges (875080)]

Think less in sheer numbers and more in density. If there are 200 million possible 'combinations' (say, 50,000 customers and 4000 movies in a Netflix-like situation), then with 10 million data samples, we only have 5% of the possible data. This means that if we are predicting inside the data scope, we are predicting into an unknown field that is 19 times larger then the known.
Say we were looking at 100 million fields, suddenly we have 50% of the possible data, and our unknown field is the same size as the known field. Much more likely to get a result then.

I think better is subjective...

[3p1ph4ny (835701)]

In problems like minimizing lateness et. al. "better" can be simply defined as "closer to optimal" or "fewer time units late."

Here, better means different things to different people. The more data you have gives you a larger set of people, and probably a more accurate definition of better for a larger set of people. I'm not sure you can really compare the two.

Um, Yes?

[randyest (589159)]

Of course. Why wouldn't more (or bettter) relevant data that applies on a case-y-case basis provide more improved results than a "improved algorithm" (what does that mean, really?) that applied generally and globally?

I think we need much, much more rigorous definitions of "more data" and "better algorithm" in order to discuss this in any meaningful way.

For the Sake of Discussion

[eldavojohn (898314)]

Well, for the sake of discussion I will try to give you an example so that you might pick it apart.

"more data"

More data means that you understand directors and actors/actresses often do a lot of the same work. So for every movie that the user likes, you weight their stars they gave it with a name. Then you cross reference movies containing those people using a database (like IMDB). So if your user loved The Sting and Fight Club, they will also love Spy Games which had both Redford & Pitt starring in it.

"better algorithm"

If you naively look at the data sets, you can imagine that each user represents a taste set and that high correlations between two movies in several users indicates that a user who has not seen the second movie will most likely enjoy it. So if 1,056 users who saw 12 Monkeys loved Donnie Darko but your user has only seen Donnie Darko, highly recommend them 12 Monkeys.

You could also make an elaborate algorithm that uses user age, sex & location ... or even a novel 'distance' algorithm that determines how far away they are from liking 12 Monkeys based on their highly ranked other movies.

Honestly, I could provide endless ideas for 'better algorithms' although I don't think any of them would even come close to matching what I could do with a database like IMDB. Hell, think of the Bayesian token analysis you could do on the reviews and message boards alone!

This reminds me

[FredFredrickson (1177871)]

This reminds me of those articles who say that the amount of data humanity has archived is so much data that nobody could possibly use it in a lifetime. I think what people fail to remember is this: the point is to have available data just-in-case you need to reference it in the future. Nobody watches security tapes in full. The review the day or hour that the robbery occured. Does that mean we should stop recording everything? No. Let's keep archiving.

Combine that with the speed at which computers are getting more efficient - and I see no reason to just keep piling up this crap. More is always better. (More efficient might be better- but add the two together, and you're unstoppable)

Too a large extent ...

[haluness (219661)]

I can see that more data (especially more varied data) could be better than a tweaked algorithm. Especially in machine learning, I see many people publish papers on a new method that does 1% better than preexisting methods.

Now, I won't deny that algorithmic advances are important, but it seems to me that unless you have a better understanding of the underlying system (which might be a physical system or a social system) tweaking algorithms would only lead to marginal improvements.

Obviously, there will be a big jump when going from a simplistic method (say linear regression) to a more sophisticated method (say SVM's). But going from one type of SVM to another slightly tweaked version of the fundamental SVM algorithm is probably not as worthwhile as sitting down and trying to understand what is generating the observed data in the first place.

More vs Better

[Mikkeles (698461)]

Better data is probably most important and having more data makes having better data more likely. It would probably make sense to analyse the impact of each datum on the accuracy of the ruslt, then choose a better algorithm using the most influential data. That is, a simple algorithm on good data is better than a great algorithm on mediocre data.

All things being equal...

[Just Some Guy (3352)]

One team used a better algorithm while the other harvested augmenting data on movies from the Internet Movie Database. And this team, which used a simpler algorithm, did much better nearly as well as the best algorithm on the boards for the $1 million challenge.

And the teams were identically talented? In my CS classes, I could have hand-picked teams that could make O(2^n) algorithms run quickly and others that could make O(1) take hours.

Hold on a sec...

[peacefinder (469349)]

"What do you think? Will more data usually perform better than a better algorithm?"

I need more data.

Five stars

[CopaceticOpus (965603)]

If more data is helpful, then Netflix is really hurting themselves with their 5-star rating system. I'd only give 5 stars to a really amazing movie, but to only give 3/5 stars to a movie I enjoyed feels too low. Many movies that range from a 7/10 to a 9/10 get lumped into that 4 star category, and the nuances of the data are lost.

How to translate the entire experience of watching a movie into a lone number is a separate issue.

This is assuming...

[jd (1658)]

...that algorithms and data are, in fact, different animals. Algorithms are simply mapping functions, which can in turn be entirely represented as data. A true algorithm represents a set of statements which, when taken as a collective whole, will always be true. In other words, it's something that is generic, across-the-board. Think object-oriented design - you do not write one class for every variable. Pure data will contain a mix of the generic and the specific, with no trivial way to always identify which is which, or to what degree.

Thus, an algorithm-driven design should always out-perform data-driven designs when knowledge of the specific is substantially less important than knowledge of the generic. Data-driven designs should always out-perform algorithm-driven design when the reverse is true. A blend of the two designs (in order to isolate and identify the nature of the data) should outperform pure implementations following either design when you want to know a lot about both.

The key to programming is not to have one "perfect" methodology but to have a wide range at your disposal.

For those who prefer mantras, have the serenity to accept the invariants aren't going to change, the courage to recognize the methodology will, and the wisdom to apply the difference.

Recommendations Systems and subjectivity

[mlwmohawk (801821)]

I have written two recommendations systems and have taken a crack at the Netflix prize (but have been hard pressed to make time for the serious work.)

The article is informative and generally correct, however, having done this sort of stuff on a few projects, I have some problems with the netflix data.

First, the data is bogus. The preferences are "aggregates" of rental behaviors, whole families are represented by single accounts. Little 16 year old Tod, likes different movies than his 40 year old dad. Not to mention his toddler sibling and mother. A single account may have Winnie the Pooh and Kill Bill. Obviously, you can't say that people who like Kill Bill tend to like Winnie the Pooh. (Unless of course there is a strange human behavioral factor being exposed by this, it could be that parents of young children want the thrill of vicarious killing, but I digress)

The IMDB information about genre is interesting as it is possibly a good way to separate some of the aggregation.

Recommendation systems tend to like a lot of data, but not what you think. People will say, if you need more data, why just have 1-5 and not 1-10? Well, that really isn't much more added data it is just greater granularity of the same data. Think of it like "color depth" vs "resolution" on a video monitor.

My last point about recommendations is that people have moods are are not as predictable as we may wish. On an aggregate basis, a group of people is very predictable. A single person setting his/her preferences one night may have had a good day and a glass of wine and numbers are higher. The next day could have had a crappy day and had to deal with it sober, the numbers are different.

You can't make a system that will accurately predict responses of a single specific individual at an arbitrary time. Let alone based on an aggregated data set. That's why I haven't put much stock in the Netflix prize. Maybe someone will win it, but I have my doubts. A million dollars is a lot of money, but there are enough vagaries in what qualifies as a success to make it a lottery or a sham.

That being said, the data is fun to work with!!

One Trivial Result, One Big Assumption

[fygment (444210)]

Two things. The first is that it is tritely obvious that adding more data improves your results. But there are two possible mechanisms at work. On the one hand add more of the same data ie. just make your original database larger with more entries. That form of augmentation will hopefully give you more insight into the underlying distribution of the data. On the other hand you can augment the existing data. In the latter you are really adding extra dimensions/features/attributes to the data set. That's what seems to be alluded to in the article i.e. the students are adding extra features to the original data set. The success of the technique is a trivial result which depends very much on whether the features you add are discriminating or not. In this case, the IMDB presumably added discriminating features. However, if it had not, then "improved algorithms" would have had the upper hand.

The second thing about the claim seems to be that there is always additional information actually available. The comment is made that academia and business don't seem to appreciate the value of augmenting the data. That is false. In business additional data is often just not available (physically or for cost reasons). Consequently, improving your algorithms is all you can do. Similarly in academia (say a computer science department) the assumption is often that you are trying to improve your algorithms while assuming that you have all the data available.

This does not mean what I think you think it means

[aibob (1035288)]

I am a graduate student in computer science, emphasizing the use of machine learning.

The sound bite conclusion of this blog post is that algorithms are a waste of time and that you are better off adding more training data.

The reality is that a lot of really smart people have been trying to come up with better algorithms for classification, clustering, and (yes) ranking for a very long time. Unless you are already familiar with the field, you really are unlikely to invent something new that will work better than what is already out there.

But that does not mean that the algorithm does not matter - for the problems I work on, using logistic regression or support vector machines outperforms naive bayes by 10% - 30%, which is huge. So if you want good performance, you try a few different algorithms to see what works.

Adding more training data does not always help either, if the distributions of the data are significantly different. You are much better off using the data to design better features which represent/summarize the data.

In other words, the algorithm is not unimportant, it just isn't the place your creative work is going to have the highest ROI.

Re:attn computer scientists: stop renaming stuff

[by Anonymous Coward]

you guys are nothing more than glorified engineers.

Computer scientists are not glorified engineers. They're the butt of engineers' jokes too.

Re:attn computer scientists: stop renaming stuff

[Freeside1 (1140901)]

Say what you want about computer scientists, but without them you'd probably be complaining on a chalkboard.

Re:attn computer scientists: stop renaming stuff

[jank1887 (815982)]

Mathematics is physics without purpose, Chemistry is physics without thought, Engineering is physics - CliffsNotes edition.

Re:attn computer scientists: stop renaming stuff

[JasonKChapman (842766)]

Mathematics is physics without purpose, Chemistry is physics without thought, Engineering is physics

Mathematics is physics without purpose, Chemistry is physics without thought, Engineering is physics without tenure.

Re:attn computer scientists: stop renaming stuff

[Metasquares (555685)]

And nonlinear dimensionality reduction is just nonconvex trace optimization coupled with kernel principal component analysis (fine, call it "singular value decomposition") using Mercer's theorem to map the resulting dot product through a kernel function (usually represented as a Hermitian positive semidefinite Gram matrix), yielding an inner product space of higher (possibly infinite) dimensionality in which the original problem is linearly separable.

Now take this description and write an algorithm that performs it efficiently. And you use PageRank as an example, so let's call "efficient" "performs as well as Google on the entire web's worth of data".

If you can't do this, perhaps you should reconsider your view of computer scientists. There's no reason whatsoever to play up the boundaries between two very related fields. Arbitrary boundaries in knowledge are already bad enough; they need to be knocked down, not reinforced.

Re:attn computer scientists: stop renaming stuff

[Arthur B. (806360)]

"machine learning" is just statistical inference

Riiiht. And mathematical research is just finding a Hamiltonian cycle in a graph defined by the set of axioms used.

Re:Is it just me that is surprised here?

[gnick (1211984)]

The netflix challenge is to arrive at a better algorithm with the supplied data.

Actually, the rules explicitly allow supplementing the data set and Netflix points out that they explore external data sets as well.

Re:Heuristics??

[EvanED (569694)]

One would hope that the thing that calculates the heuristic is an algorithm. See wikipedia [wikipedia.org].