Storing encoded Features in the Feature Store? - machine-learning

I am currently in the design phase of a feature store pipeline to train and serve ML models. One thing which confuses me is the "state" in which features are supposed to be stored in the feature store.
Let's assume I have a data source which has an entries of the following schema:
ID
Value1
Value2
123
0.1
20
During my data analysis I found that a feature Value2 / Value1 is a meaningful feature for my ML model. This would yield:
ID
Value1
Value2
V2_by_V1
123
0.1
20
200
Now to the heart of my question: do I store this table in my feature store OR do I additionally apply transformations like scaling / binning / one hot encoding / ... and store the encoded features in the store?
A consultant from one of the big cloud providers opted for storing encoded features in the store which I find strange because different models will use the same type of feature but will need different encoding procedures and storing encoded value complicates model monitoring and debugging in production.

Related

Azure Machine Learning Studio Conditional Training Data

I have built an Microsoft Azure ML Studio workspace predictive web service, and have a scernario where I need to be able to run the service with different training datasets.
I know I can setup multiple web services via Azure ML, each with a different training set attached, but I am trying to find a way to do it all within the same workspace and passing a Web Input Parameter as the input value to choose which training set to use.
I have found this article, which describes almost my scenario. However, this article relies on the training dataset that is being pulled from the Load Trained Data module, as having a static endpoint (or blob storage location). I don't see any way to dynamically (or conditionally) change this location based on a Web Input Parameter.
Basically, does Azure ML support a "conditional training data" loading?
Or, might there be a way to combine training datasets, then filter based on the passed Web Input Parameter?
This probably isn't exactly what you need, but hopefully, it helps you out.
To combine data sets, you can use the Join Data module.
To filter, that may be accomplished by executing a Python script. Here's an example.
Using the Adult Census Income Binary Classification dataset, on the age column, there's a minimum age of 17.
If I wanted to filter the data set by age, connect it to an Execute Python Script module and here's the filtering code with the pandas query method.
# The script MUST contain a function named azureml_main
# which is the entry point for this module.
import pandas as pd
def azureml_main(dataframe1 = None, dataframe2 = None):
# Return value must be of a sequence of pandas.DataFrame
return dataframe1.query("age >= 25")
And looking at that output it filters out the data set where the minimum age is now 25.
Sure, you can do that. What you would want is to use an Execute R Script or SQL Transformation module to determine, based on your input data, what model to use. Something like this:
Notice, your input data is cleaned/updated/feature engineered, then it's passed to two different SQL transforms which will tell it to go to one of two paths.
Each path has it's own training data.
Note: I am not exactly sure what your use case is, but if it were me, I would instead train two different models using the two different training data, then try to just use the models in my web service, not actually train on the web service as that would likely be quite slow.

Beam/Dataflow design pattern to enrich documents based on database queries

Evaluating Dataflow, and am trying to figure out if/how to do the following.
My apologies if anything in the above is trivial--trying to wrap our heads around Dataflow before we make a decision on using Beam, or something else like Spark, etc.
General use case is for machine learning:
Ingesting documents which are individually processed.
In addition to easy-to-write transforms, we'd like to enrich each document based on queries against databases (that are largely key-value stores).
A simple example would be a gazetteer: decompose the text into ngrams, and then check if those ngrams reside in some database, and record (within a transformed version of the original doc) the entity identifier given phrases map to.
How to do this efficiently?
NAIVE (although possibly tricky with the serialization requirement?):
Each document could simply query the database individually (similar to Querying a relational database through Google DataFlow Transformer), but, given that most of these are simple key-value stores, it seems like there should be a more efficient way to do this (given the real problems with database query latency).
SCENARIO #1: Improved?:
Current strawman is to store the tables in Bigquery, pull them down (https://github.com/apache/beam/blob/master/sdks/python/apache_beam/io/gcp/bigquery.py), and then use them as side inputs, that are used as key-value lookups within the per-doc function(s).
Key-value tables range from generally very small to not-huge (100s of MBs, maybe low GBs). Multiple CoGroupByKey with same key apache beam ("Side inputs can be arbitrarily large - there is no limit; we have seen pipelines successfully run using side inputs of 1+TB in size") suggests this is reasonable, at least from a size POV.
1) Does this make sense? Is this the "correct" design pattern for this scenario?
2) If this is a good design pattern...how do I actually implement this?
https://github.com/apache/beam/blob/master/sdks/python/apache_beam/io/gcp/bigquery.py#L53 shows feeding the result to the document function as an AsList.
i) Presumably, AsDict is more appropriate here, for the above use case? So I'd probably need to run some transformations first on the Bigquery output to separate it into key, value tuple; and make sure that the keys are unique; and then use it as a side input.
ii) Then I need to use the side input in the function.
What I'm not clear on:
for both of these, how to manipulate the output coming off of the Bigquery pull is murky to me. How would I accomplish (i) (assuming it is necessary)? Meaning, what does the data format look like (raw bytes? strings? is there a good example I can look into?)
Similarly, if AsDict is the correct way to pass it into the func, can I just reference things like a dict normally is used in python? e.g., side_input.get('blah') ?
SCENARIO #2: Even more improved? (for specific cases):
The above scenario--if achievable--definitely does seem like it is superior continuous remote calls (given the simple key-value lookup), and would be very helpful for some of our scenarios. But if I take a scenario like a gazetteer lookup (like above)...is there an even more optimized solution?
Something like, for every doc, writing our all the ngrams as keys, with values as the underlying indices (docid+indices within the doc), and then doing some sort of join between these ngrams and the phrases in our gazeteer...and then doing another set of transforms to recover the original docs (now w/ their new annotations).
I.e., let Beam handle all of the joins/lookups directly?
Theoretical advantage is that Beam may be a lot quicker in doing this than, for each doc, looping over all of the ngrams and doing a check if the ngram is in the side_input.
Other key issues:
3) If this is a good way to do things, is there any trick to making this work well in the streaming scenario? Text elsewhere suggests that the side input caching works more poorly outside the batch scenario. Right now, we're focused on batch, but streaming will become relevant in serving live predictions.
4) Any Beam-related reason to prefer Java>Python for any of the above? We've got a good amount of existing Python code to move to Dataflow, so would heavily prefer Python...but not sure if there are any hidden issues with Python in the above (e.g., I've noticed Python doesn't support certain features or I/O).
EDIT: Strawman? for the example ngram lookup scenario (should generalize strongly to general K:V lookup)
Phrases = get from bigquery
Docs (indexed by docid) (direct input from text or protobufs, e.g.)
Transform: phrases -> (phrase, entity) tuples
Transform: docs -> ngrams (phrase, docid, coordinates [in document])
CoGroupByKey key=phrase: (phrase, entity, docid, coords)
CoGroupByKey key=docid, group((phrase, entity, docid, coords), Docs)
Then we can iteratively finalize each doc, using the set of (phrase, entity, docid, coords) and each Doc
Regarding the scenarios for your pipeline:
Naive scenario
You are right that per-element querying of a database is undesirable.
If your key-value store is able to support low-latency lookups by reusing an open connection, you can define a global connection that is initialized once per worker instead of once per bundle. This should be acceptable your k-v store supports efficient lookups over existing connections.
Improved scenario
If that's not feasible, then BQ is a great way to keep and pull in your data.
You can definitely use AsDict side inputs, and simply go side_input[my_key] or side_input.get(my_key).
Your pipeline could look something like so:
kv_query = "SELECT key, value FROM my:table.name"
p = beam.Pipeline()
documents_pcoll = p | ReadDocuments()
additional_data_pcoll = (p
| beam.io.BigQuerySource(query=kv_query)
# Make row a key-value tuple.
| 'format bq' >> beam.Map(lambda row: (row['key'], row['value'])))
enriched_docs = (documents_pcoll
| 'join' >> beam.Map(lambda doc, query: enrich_doc(doc, query[doc['key']]),
query=AsDict(additional_data_pcoll)))
Unfortunately, this has one shortcoming, and that's the fact that Python does not currently support arbitrarily large side inputs (it currently loads all of the K-V into a single Python dictionary). If your side-input data is large, then you'll want to avoid this option.
Note This will change in the future, but we can't be sure ATM.
Further Improved
Another way of joining two datasets is to use CoGroupByKey. The loading of documents, and of K-V additional data should not change, but when joining, you'd do something like so:
# Turn the documents into key-value tuples as well[
documents_kv_pcoll = (documents_pcoll
| 'format docs' >> beam.Map(lambda doc: (doc['key'], doc)))
enriched_docs = ({'docs': documents_kv_pcoll, 'additional_data': additional_data_pcoll}
| beam.CoGroupByKey()
| 'enrich' >> beam.Map(lambda x: enrich_doc(x['docs'][0], x['additional_data'][0]))
CoGroupByKey will allow you to use arbitrarily large collections on either side.
Answering your questions
You can see an example of using BigQuery as a side input in the cookbook. As you can see there, the data comes parsed (I believe that it comes in their original data types, but it may come in string/unicode). Check the docs (or feel free to ask) if you need to know more.
Currently, Python streaming is in alpha, and it does not support side inputs; but it does support shuffle features such as CoGroupByKey. Your pipeline using CoGroupByKey should work well in streaming.
A reason to prefer Java over Python is that all these features work in Java (unlimited-size side inputs, streaming side inputs). But it seems that for your use case, Python may have all you need.
Note: The code snippets are approximate, but you should be able to debug them using the DirectRunner.
Feel free to ask for clarification, or to ask about other aspects if you feel like it'd help.

Obfuscation of sensitive data for machine learning

I am preparing a dataset for my academic interests. The original dataset contains sensitive information from transactions, like Credit card no, Customer email, client ip, origin country, etc. I have to obfuscate this sensitive information, before they leave my origin data-source and store them for my analysis algorithms. Some of the fields in data can be categorical and would not be difficult to obfuscate. Problem lies with the non-categorical data fields, how best should I obfuscate them to leave underlying statistical characteristics of my data intact but make it impossible (at least mathematically hard) to revert back to original data.
EDIT: I am using Java as front-end to prepare the data. The prepared data would then be handled by Python for machine learning.
EDIT 2: To explain my scenario, as a followup from the comments. I have data fields like:
'CustomerEmail', 'OriginCountry', 'PaymentCurrency', 'CustomerContactEmail',
'CustomerIp', 'AccountHolderName', 'PaymentAmount', 'Network',
'AccountHolderName', 'CustomerAccountNumber', 'AccountExpiryMonth',
'AccountExpiryYear'
I have to obfuscate the data present in each of these fields (data samples). I plan to treat these fields as features (with the obfuscated data) and train my models against a binary class label (which I have for my training and test samples).
There is no general way to obfuscate non categorical data as any processing leads to the loss of information. The only thing you can do is try to list what type of information is the most important one and design transformation which leaves it. For example if your data is Lat/Lng geo position tags you could perform any kind of distance-preserving transformations, such as translation, rotations etc. if it is not good enough you can embeed your data in lower dimensional space while preserving the pairwise distances (there are many such methods). In general - each type of non-categorical data requires different processing, and each destroys information - it is up to you to come up with the list of important properties and finding transformations preserving it.
I agree with #lejlot that there is no silver bullet method to solve your problem. However, I believe this answer can get you started thinking about to handle at least the numerical fields in your data set.
For the numerical fields, you can make use of the Java Random class and map a given number to another obfuscated value. The trick here is to make sure that you map the same numbers to the same new obfuscated value. As an example, consider your credit card data, and let's assume that each card number is 16 digits. You can load your credit card data into a Map and iterate over it, creating a new proxy for each number:
Map<Integer, Integer> ccData = new HashMap<Integer, Integer>();
// load your credit data into the Map
// iterate over Map and generate random numbers for each CC number
for (Map.Entry<Integer, Integer> entry : ccData.entrySet()) {
Integer key = entry.getKey();
Random rand = new Random();
rand.setSeed(key);
int newNumber = rand.nextInt(10000000000000000); // generate up to max 16 digit number
ccData.put(key, newNumber);
}
After this, any time you need to use a credit card num you would access it via ccData.get(num) to use the obfuscated value.
You can follow a similar plan for the IP addresses.

how to get the hash value when using StaticWordValueEncoder in Mahout

I'm look at an example in the Mahout in Action book. It uses the StaticWordValueEncoder to encoder a text in the feature hashing manner.
When encode "text to magically vectorize" with a standard analyser and probe = 1, the vector is {12:1.0, 54:1.0, 78:1.0}. However, I can't figure out which word the hash index refers to.
Is there any method to get the [hash, original word] as a pair? e.g. hash 12 refers to the word "text"?
if you have read Mahout in Action paragraph:
"The value of a continuous
variable gets added directly to one or more locations that are allocated for the storage
of the value. The location or locations are determined by the name of the feature.
This hashed feature approach has the distinct advantage of requiring less memory
and one less pass through the training data, but it can make it much harder to reverse engineer
vectors to determine which original feature mapped to a vector location."
-----I am not sure how the reverse engineering can be done(which certainly a difficult task as Author has put) Perhaps some one might put some light on this.

"Smart" / Economical Data Storage Techniques?

I would like to store millions of data lines that looks like this:
key, value
key is an integer in the range of (0 to 5,000,000); all values are unique;
value is an unsigned int16 value (0 to 65535)
the key is to store the data while taking the LEAST AMOUNT OF DISK SPACE, and yet, be able to query the data. can you think of any algorithms / smart schemes for data storage that would be helpful?
just in case it matters, I use Linux.
One option would be, if the key values are not important data but rather just index data to utilize a flat file of bits ( with a descriptive header ). Every 16 bits is a value and the nth value would then be (n - 1) * 16 bits from the end of the header.
Additionally, if the key value does matter, a set flat file of about 10MB would allow for the entire key space to be stored without storing actual keys. The 16 bits that are at the (n - 1) * 16 offset would be that key's value.
That would probably be the least space intensive method for storage, as it would be only the data that is literally required. ( Though, if you are only interested in say 100k values and one has a key of 5 million you do end up with a lot of wasted space, which wouldn't be there with an actual key,value addressing system. So this methodology only achieves a minimum disk storage for sets of tightly grouped values or many many numbers (over about the 2 million mark ).
how do you plan to use stored data? with random or sequential access? for sequential access you can use any archiving algorithm, e.g. LZMA. Random access doesn't leave you a lot of space for improvements.
can you see any patterns of this data? e.g. if the difference between adjacent keys/values are often small you can store only packed differences. and million of other possible approaches.
[EDIT] also you can check techniques used for data compression in network communication
[EDIT1] and you can check this Google Code Integer Array Compression project
This depend upon the operation and data. I would first recommend "just using a database" (a simple key-value store such as BDB/EhCache [read: Key Value store], for instance :-)
Mimisbrunnr also has a good answer if all the keys are used.
If the keys are near constant/read-only and only a relatively small percent of the keys are used, consider the use of a (disk-based) Heap data-structure (very similar to an Array-based Heap; Heaps need not be Array-based). Robert Sedgewick had a good book from the late 80's that had a very lean implementation, but I forget the name. A Heap will be more beneficial when compared to a flat index with a smaller proportion of used keys and at full-load will have worse storage requirements.
(If abstracted, the used method could be switched and/or a hybrid heap with indexed/sequenced leaf-node values could be used [along with Huffman encoding or whatnot], but that is just adding far more complications. Keep it simple ... hence first suggestion of an existing key/value store ;-)
Happy coding.
Have you considered using a database designed for mobile devices such as SQL Server Compact, or another similar database? These will have a small footprint on the disk, while still providing the full search power you need.
Another example of a compact database engine is KeyDB for linux:
http://3d2f.com/programs/11-989-keydb-download.shtml

Resources