I work at a public health department that takes in and stores lots of medical data every day. I've written a program that uses regular expressions to determine if particular fields in the incoming data are valid or invalid. Ex: DOBs come in as YYYYmmDD, so they should match regex ^[0-9]{8}$
I want to analyze the "invalid" data to help identify problems in our system (we get way too much data to go through each 'bad' record row-by-row). Can anyone suggest AI techniques/machine learning techniques that can 'monitor' the bad data and find patterns in what is wrong? I think that coming up with a bunch of regular expressions for possible ways the data could be invalid (ex. not enough or too many characters) and then keeping track of those results might work. But instead of me thinking up all of the ways the data could be invalid, I'm curious about ways to 'learn' the patterns from the bad data using AI.
Are there any known techniques that do this?
I think that coming up with a bunch of regular expressions for possible ways the data could be invalid (ex. not enough or too many characters) and then keeping track of those results might work. But instead of me thinking up all of the ways the data could be invalid, I'm curious about ways to 'learn' the patterns from the bad data using AI.
What's funny is I'm reminded of a quotation usually attributed to Jamie Zawinski:
Some people, when confronted with a problem, think "I know, I'll use
regular expressions." Now they have two problems.
Except, in this case, I think the hand-crafted regex route is actually your best bet!
Irony of ironies.
Anyway.
The point of this saying is that people tend to overcomplicate their solutions. Here, regexs are actually a fairly simple solution to your problem, whereas creating a learner is something that will take you a lot more time than I think you realize.
There are fewer ways for this very constrained data representation (a date) to be expressed correctly, than there are ways for it to be expressed incorrectly. Because there are infinite ways to define bad data. You want to train a learner to detect all of them? It's a rabbit hole. Think of this AI learner instead as a coworker or a friend: how would you describe to them all the ways that dates can't be represented properly?
While your intention was to make less work for yourself in the long run -- and that's a good quality to have -- figuring out how to develop a learner, not to mention train and validate it, not to mention watch it carefully, outweigh any benefits that learner can provide you in such a narrow use case.
Bayesian filtering might be what you are looking for.
It sounds like you want to apply supervised learning to regular expressions. These fellows seem to be up to something of that sort.
Perhaps look for techniques of "outlier detection"?
Related
Questions
I want to classify/categorize/cluster/group together a set of several thousand websites. There's data that we can train on, so we can do supervised learning, but it's not data that we've gathered and we're not adamant about using it -- so we're also considering unsupervised learning.
What features can I use in a machine learning algorithm to deal with multilingual data? Note that some of these languages might not have been dealt with in the Natural Language Processing field.
If I were to use an unsupervised learning algorithm, should I just partition the data by language and deal with each language differently? Different languages might have different relevant categories (or not, depending on your psycholinguistic theoretical tendencies), which might affect the decision to partition.
I was thinking of using decision trees, or maybe Support Vector Machines (SVMs) to allow for more features (from my understanding of them). This post suggests random forests instead of SVMs. Any thoughts?
Pragmatical approaches are welcome! (Theoretical ones, too, but those might be saved for later fun.)
Some context
We are trying to classify a corpus of many thousands of websites in 3 to 5 languages (maybe up to 10, but we're not sure).
We have training data in the form of hundreds of websites already classified. However, we may choose to use that data set or not -- if other categories make more sense, we're open to not using the training data that we have, since it is not something we gathered in the first place. We are on the final stages of scraping data/text from websites.
Now we must decide on the issues above. I have done some work with the Brown Corpus and the Brill tagger, but this will not work because of the multiple-languages issue.
We intend to use the Orange machine learning package.
According to the context you have provided, this is a supervised learning problem.
Therefore, you are doing classification, not clustering. If I misunderstood, please update your question to say so.
I would start with the simplest features, namely tokenize the unicode text of the pages, and use a dictionary to translate every new token to a number, and simply consider the existence of a token as a feature.
Next, I would use the simplest algorithm I can - I tend to go with Naive Bayes, but if you have an easy way to run SVM this is also nice.
Compare your results with some baseline - say assigning the most frequent class to all the pages.
Is the simplest approach good enough? If not, start iterating over algorithms and features.
If you go the supervised route, then the fact that the web pages are in multiple languages shouldn't make a difference. If you go with, say lexical features (bag-o'-words style) then each language will end up yielding disjoint sets of features, but that's okay. All of the standard algorithms will likely give comparable results, so just pick one and go with it. I agree with Yuval that Naive Bayes is a good place to start, and only if that doesn't meet your needs that try something like SVMs or random forests.
If you go the unsupervised route, though, the fact that the texts aren't all in the same language might be a big problem. Any reasonable clustering algorithm will first group the texts by language, and then within each language cluster by something like topic (if you're using content words as features). Whether that's a bug or a feature will depend entirely on why you want to classify these texts. If the point is to group documents by topic, irrespective of language, then it's no good. But if you're okay with having different categories for each language, then yeah, you've just got as many separate classification problems as you have languages.
If you do want a unified set of classes, then you'll need some way to link similar documents across languages. Are there any documents in more that one language? If so, you could use them as a kind of statistical Rosetta Stone, to link words in different languages. Then, using something like Latent Semantic Analysis, you could extend that to second-order relations: words in different languages that don't ever occur in the same document, but which tend to co-occur with words which do. Or maybe you could use something like anchor text or properties of the URLs to assign a rough classification to documents in a language-independent manner and use that as a way to get started.
But, honestly, it seems strange to go into a classification problem without a clear idea of what the classes are (or at least what would count as a good classification). Coming up with the classes is the hard part, and it's the part that'll determine whether the project is a success or failure. The actual algorithmic part is fairly rote.
Main answer is: try different approaches. Without actual testing it's very hard to predict what method will give best results. So, I'll just suggest some methods that I would try first and describe their pros and cons.
First of all, I would recommend supervised learning. Even if the data classification is not very accurate, it may still give better results than unsupervised clustering. One of the reasons for it is a number of random factors that are used during clustering. For example, k-means algorithm relies on randomly selected points when starting the process, which can lead to a very different results for different program runnings (though x-means modifications seems to normalize this behavior). Clustering will give good results only if underlying elements produce well separated areas in the feature space.
One of approaches to treating multilingual data is to use multilingual resources as support points. For example, you can index some Wikipedia's articles and create "bridges" between same topics in different languages. Alternatively, you can create multilingual association dictionary like this paper describes.
As for methods, the first thing that comes to mind is instance-based semantic methods like LSI. It uses vector space model to calculate distance between words and/or documents. In contrast to other methods it can efficiently treat synonymy and polysemy. Disadvantage of this method is a computational inefficiency and leak of implementations. One of the phases of LSI makes use of a very big cooccurrence matrix, which for large corpus of documents will require distributed computing and other special treatment. There's modification of LSA called Random Indexing which do not construct full coocurrence matrix, but you'll hardly find appropriate implementation for it. Some time ago I created library in Clojure for this method, but it is pre-alpha now, so I can't recommend using it. Nevertheless, if you decide to give it a try, you can find project 'Clinch' of a user 'faithlessfriend' on github (I'll not post direct link to avoid unnecessary advertisement).
Beyond special semantic methods the rule "simplicity first" must be used. From this point, Naive Bayes is a right point to start from. The only note here is that multinomial version of Naive Bayes is preferable: my experience tells that count of words really does matter.
SVM is a technique for classifying linearly separable data, and text data is almost always not linearly separable (at least several common words appear in any pair of documents). It doesn't mean, that SVM cannot be used for text classification - you still should try it, but results may be much lower than for other machine learning tasks.
I haven't enough experience with decision trees, but using it for efficient text classification seems strange to me. I have seen some examples where they gave excellent results, but when I tried to use C4.5 algorithm for this task, the results were terrible. I believe you should get some software where decision trees are implemented and test them by yourself. It is always better to know then to suggest.
There's much more to say on every topic, so feel free to ask more questions on specific topic.
I develop code in our proprietry system using a scripting language that is unique to that system.
Our director has allowed us to request enhancements to this language, which currently lacks user definable arrays.
I need to write a concept brief on why we need arrays and how they can benefit us, however I need to explain it in a fashion that someone who has no understanding of code will understand.
I'm a programmer, therefore I suck at documentation and explaining things in a non-technical manner. I tried banging my head on the desk to see if anything useful would come out but it hasn't. Can anyone help?
I love analogies.
Much easier to have a 100 DVD holder that sits neatly on your floor and holds 100 dvds in order than 100 individual DVDs scattered around your house where you last used them
Especially relevant when you need to move the collection from one place to another or share it with a friend.
What's your application area? To speak the users' language you need to know that. Suppose it's stocks trading: then what to you is an array, to the users may be a portfolio -- get the quotes for several stock at once rather than having to do it repeatedly for one at a time. If your application area is CRM, then the array will let the users check on a group of customers at once, rather than do it one at a time. And so on, and so forth.
In every application area there will be cases in which users may want to deal with a bunch of things at once, it being easier than dealing with one thing at a time. Phrase it in the appropriate vocabulary, and you have the case for arrays!
You might want to see if you can move the business away from your custom scripting environment and into a standard scripting environment like LUA or Python. You might be surprised at how much easier it is to get LUA up and running than it is to :
Support an in house system
Create tools for it (do you have an IDE?)
Train new programmers in it
Live without modern features that you lack the time/skills to impliment.
Key to getting that to happen would be to make LUA interoperable with your standard scripting system or writing a translation from your old scripts to LUA scripts.
The benefit is that it makes the code shorter, and thus less money is spent coding and debugging. You can then present some example code that you could make it shorter had the language supported arrays.
Sounds like you've been asked to create code in the past (or anticipate having to create code in the future), where your job would have been faster/easier/cheaper if the system that you used had arrays.
That's the issue: you want to do more for your director and you need arrays to help you.
Your director will understand the business benefits of you having a better toolkit--you'll be able to do more for him or her. And that's how you increase business efficiency.
Tell your director: I want to improved my productivity for you and our team. To do so, arrays would be very helpful.
I like Alex's answer - it has to be put in terms of the user's problems. What problem (that they care about) can they do with it that they cannot do without it?
I used to teach introductory programming in college, and arrays are simply not something that comes easily to non-programmers. They need to understand some other basics first, like the sequential nature of programs, the lego-block way programs are constructed, the idea of run-time (as opposed to write-time) and really importantly the concept of a variable as a container of a value, and how that is different from its name, and how its contents changes with time while its name does not.
I found a useful way to get into this area is to let them program a very simple, decimal, simulated computer, in "machine language". They get the notion of memory address vs. memory contents, and that address is just a number. That makes it a lot easier to introduce arrays in a more "real" language.
Another approach is to have them work on a kind of problem where they really start wishing they could invent variables on-the-fly. Like they don't want to just have a variable A, but they feel a need for A1, A2, etc. and then they would really like to say Ai where i is a another variable. Once they feel the need for that, then they will grasp arrays. (For example, they could take a simple program that asks for their name and has a simple conversation with them, and then extend it to talk to two people at once, then three, and so on.)
Then, a useful next step is "parallel arrays" which can serve as rudimentary arrays of structures. i.e. N$(i) can be name of student i, while A(i) can be age of student i. This makes the idea useful.
Only then would I dare to start to introduce algorithms like sorting, merging, table lookup, and so on.
I think to fully realize the potential of arrays, you must somehow mention two things:
1) Array Algorithms
Sort, Find, etc. All the basics. Equate this in your business brief as structured data that can organize itself. No extra query language. No variable naming conventions. All you need is good standards.
2) Multi-Dimensional Arrays
The power of arrays seems fully realized to me with matrices. With these you can practically hold limitless data.
Plus, depending on the power of the propriety language you are using, arrays can store objects.
The power of an array is that it allows you to put a group of things together so that you can perform the same operation on all of them with less code.
Sorting is one example of an array operation, and is like having a box of index cards that you are putting in order.
Or if you had a collection of letters that need to go out, being able to write a loop that stamps each letter and then sends it is better than writing out
Take first letter, stamp it.
Mail it.
Takes second letter, stamp it,
Mail it.
Basically anything you'ld use to refer to the first, second, third, fifth, etc, is basically like an array.
And then indexed/hashed arrays are like having an index in the book - you know the author describes the Defenistration of Prague somewhere in the volume, but looking in the index shows that it's on page 255.
Here's the easiest-to-understand benefit: It lets you refer to things by a number. Try to emphasize the importance of this.
do you know any good algorithms that match two strings and then return a percentage in how many percent those two strings match?
And are there some, that work with databases too?
The Levenstein distance is such a measure. It basically tells you how many characters need to be edited, deleted or added, to get from the first to the second string. I'm not sure whether some database systems support that.
But I know for sure that a much more simplified algorithm named Soundex is supported in some database systems.
It depends upon your criteria for similarity. Other people have already referred you to Levenstein distance (edit distance is the same thing). That's usually pretty good, and definitely more language-independent than something like soundex. However, be aware that Levenstein difference does not handle transposition very well. Thus:
Levenstein("copy", "cpoy") == 2
If you're trying to deal with human input, transpositions are fairly common. Whether that's a problem or not depends on your metrics for similarity.
It's been a while, but I believe Postgresql has levenstein() either built-in or available as a contrib C module.
I think the problem you're looking for is called Edit Distance. It is expensive to compute in general, but if you are looking for strings within small edit distance of other strings, it is not so bad. There is more information in the Wikipedia article.
How to best match two strings? Have them go out for coffee, and if they hit it off, dinner and a movie. Or maybe they could do some peer programming? It depends on the strings, really. Even coffee can often be tricky.
Would this be of help? I just ran into it. Comparing Two Strings producing a numeric delta
I need to implement some kind of metric space search in Postgres(*) (PL or PL/Python). So, I'm looking for good sources (or papers) with a very clear and crisp explanation of the machinery behind these ideas, in such way that I can implement it myself.
I would prefer clarity over efficiency.
(*) The need for that is described better here.
Especially for geographical data, look at PostGIS first to see if you need to implement anything. If you do, start with the papers listed in the Wikipedia entry on GiST.
Looking at your link, it seems your metric space is strings with some sort of edit distance as the metric. A nice but oldish overview of some solutions is given by Navarro, Baeza-Yates, Sutinen, and Tarhio, IEEE Data Engineering Bulletin, 2001; the related papers on Citeseer could also be useful. Locality Sensitive Hashing is a newer technique that might be useful, but a lot of the papers are heavy on math.
BK-Trees are useful for indexing and searching anything that obeys the triangle inequality, metric spaces included. The canonical example is searching for strings within a given edit distance of a target. I wrote an article about that here.
Unfortunately, there's no built in support for this in Postgres. You could implement it yourself using GIST, but obviously that'll be a lot of work. I can't think of any way to implement it without writing your own indexes short of storing the tree in a table, which obviously isn't going to be very efficient.
You can try http://sisap.org where many modern metric indexes are listed, including BK-trees. You can find code in C to try different alternatives.
Some techniques that involve space search that might help you are Hill-Climbing, Neural Network Training, Genetic Algorithm, and Particle Swarm.
You will also need to define a distance metric over your metric space. Have you done so?(& out of curiosity, what is it, if you have done so)
I have a variety of time-series data stored on a more-or-less georeferenced grid, e.g. one value per 0.2 degrees of latitude and longitude. Currently the data are stored in text files, so at day-of-year 251 you might see:
251
12.76 12.55 12.55 12.34 [etc., 200 more values...]
13.02 12.95 12.70 12.40 [etc., 200 more values...]
[etc., 250 more lines]
252
[etc., etc.]
I'd like to raise the level of abstraction, improve performance, and reduce fragility (for example, the current code can't insert a day between two existing ones!). We'd messed around with BLOB-y RDBMS hacks and even replicating each line of the text file format as a row in a table (one row per timestamp/latitude pair, one column per longitude increment -- yecch!).
We could go to a "real" geodatabase, but the overhead of tagging each individual value with a lat and long seems prohibitive. The size and resolution of the data haven't changed in ten years and are unlikely to do so.
I've been noodling around with putting everything in NetCDF files, but think we need to get past the file mindset entirely -- I hate that all my software has to figure out filenames from dates, deal with multiple files for multiple years, etc.. The alternative, putting all ten years' (and counting) data into a single file, doesn't seem workable either.
Any bright ideas or products?
I've assembled your comments here:
I'd like to do all this "w/o writing my own file I/O code"
I need access from "Java Ruby MATLAB" and "FORTRAN routines"
When you add these up, you definitely don't want a new file format. Stick with the one you've got.
If we can get you to relax your first requirement - ie, if you'd be willing to write your own file I/O code, then there are some interesting options for you. I'd write C++ classes, and I'd use something like SWIG to make your new classes available to the multiple languages you need. (But I'm not sure you'd be able to use SWIG to give you access from Java, Ruby, MATLAB and FORTRAN. You might need something else. Not really sure how to do it, myself.)
You also said, "Actually, if I have to have files, I prefer text because then I can just go in and hand-edit when necessary."
My belief is that this is a misguided statement. If you'd be willing to make your own file I/O routines then there are very clever things you could do... And as an ultimate fallback, you could give yourself a tool that converts from the new file format to the same old text format you're used to... And another tool that converts back. I'll come back to this at the end of my post...
You said something that I want to address:
"leverage 40 yrs of DB optimization"
Databases are meant for relational data, not raster data. You will not leverage anyone's DB optimizations with this kind of data. You might be able to cram your data into a DB, but that's hardly the same thing.
Here's the most useful thing I can tell you, based on everything you've told us. You said this:
"I am more interested in optimizing my time than the CPU's, though exec speed is good!"
This is frankly going to require TOOLS. Stop thinking of it as a text file. Start thinking of the common tasks you do, and write small tools - in WHATEVER LANGAUGE(S) - to make those things TRIVIAL to do.
And if your tools turn out to have lousy performance? Guess what - it's because your flat text file is a cruddy format. But that's just my opinion. :)
I'd definitely change from text to binary but keep each day in a separate file still. You could name them in such a way that insertions in between don't cause any strangeness with indices, such as by including the date and possible time in the filename. You could also consider the file structure if you have several fields per location for example. Is it common to look for a small tile from a large number of timesteps? In that case you might want to store them as tiles containing data from several days. You didn't mention how the data is accessed which plays a big role in how to organise it efficiently.
Clarifications:
I'm surprised you added "database" as one of the tags, and considered it as an option. Why did you do this?
Essentially, you have a 2D, single component floating point image at every time step. Would you agree with this way of viewing your data?
You also mentioned the desire to insert a day between two existing ones - which seems to be a very odd thing to do. Why would you need to do that? Is there a new day between May 4 and May 5 that I don't know about?
Is "compression" one of the things you care about, or are you just sick of flat files?
Would a float or a double be sufficient to store your data, or do you feel you need more arbitrary precision?
Also, what programming language(s) do you want to access this data with?
your answer on how to store the data depends entirely on what you're going to do with the data. for example, if you only ever need to retrieve by specifying the date or a date range, then storing in a database as a BLOB makes some sense. but if you need to find records that have certain values, you'll need to do something different.
please describe how you need to be able to access the data/
Matt, thanks very much, and likewise longneck and jirv.
This post was partly an experiment, testing the quality of stackoverflow discourse. If you guys/gals/alien lifeforms are representative, I'm sold.
And on point, you've clarified my thinking considerably. Mind, I still might not necessarily implement your advice, but know that I will be thinking about it very seriously. >;-)
I may very well leave the file format the same, add to the extant C and/or Ruby routines to tack on the few low-level features I lack (e.g. inserting missing timesteps), and hang an HTTP front end on the whole thing so that the data can be consumed by whatever box needs it, in whatever language is currently hoopy. While it's mostly unchanging legacy software that construct these data, we're always coming up with new consumers for it, so the multi-language/multi-computer requirement (gee, did I forget that one?) applies to the reading side, not the writing side. That also obviates a whole slew of security issues.
Thanks again, folks.