Our previous post talked about two components to build a real-time data pipeline. To recap:
Most data pipelines contain these components:
- Data querying and event subscription
- Data joining or aggregation
- Transformation and normalization
- Storage and delivery
In this post, we will talk about the last two components and introduce some tools and techniques that can achieve them.
This post is extracted from the book Hands-on Machine Learning with JavaScript byBurak Kanber. The book is a definitive guide to creating an intelligent web application with the best of machine learning and JavaScript.
Transformation and normalization of a data pipeline
As your data makes its way through a pipeline, it may need to be converted into a structure compatible with your algorithm’s input layer. There are many possible transformations that can be performed on the data in the pipeline. For example, in order to protect sensitive user data before it reaches a token-based classifier, you might apply a cryptographic hashing function to the tokens so that they are no longer human readable.
Types of Data transformations
More typically, the types of transformations will be related to sanitization, normalization, or transposition. A sanitization operation might involve removing unnecessary whitespace or HTML tags, removing email addresses from a token stream, and removing unnecessary fields from the data structure. If your pipeline has subscribed to an event stream as the source of the data and the event stream attaches source server IP addresses to event data, it would be a good idea to remove these values from the data structure, both in order to save space and to minimize the surface area for potential data leaks.
Similarly, if email addresses are not necessary for your classification algorithm, the pipeline should remove that data so that it interacts with the fewest possible servers and systems. If you’ve designed a spam filter, you may want to look into using only the domain portion of the email address instead of the fully qualified address. Alternately, the email addresses or domains may be hashed by the pipeline so that the classifier can still recognize them but a human cannot.
Make sure to audit your data for other potential security and privacy issues as well. If your application collects the end user’s IP address as part of its event stream, but the classifier does not need that data, remove it from the pipeline as early as possible. These considerations are becoming ever more important with the implementation of new European privacy laws, and every developer should be aware of privacy and compliance concerns.
Data Normalization
A common category of data transformation is normalization. When working with a range of numerical values for a given field or feature, it’s often desirable to normalize the range such that it has a known minimum and maximum bound. One approach is to normalize all values of the same field to the range [0,1], using the maximum encountered value as the divisor (for example, the sequence 1, 2, 4 can be normalized to 0.25, 0.5, 1). Whether data needs to be normalized in this manner will depend entirely on the algorithm that consumes the data.
Another approach to normalization is to convert values into percentiles. In this scheme, very large outlying values will not skew the algorithm too drastically. If most values lie between 0 and 100 but a few points include values such as 50,000, an algorithm may give outsized precedence to the large values. If the data is normalized as a percentile, however, you are guaranteed to not have any values exceeding 100 and the outliers are brought into the same range as the rest of the data. Whether or not this is a good thing depends on the algorithm.
Instagram Example
The data pipeline is also a good place to calculate derived or second-order features. Imagine a random forest classifier that uses Instagram profile data to determine if the profile belongs to a human or a bot. The Instagram profile data will include fields such as the user’s followers count, friends count, posts count, website, bio, and username. A random forest classifier will have difficulty using those fields in their original representations, however, by applying some simple data transformations, you can achieve accuracies of 90%.
In the Instagram case, one type of helpful data transformation is calculating ratios. Followers count and friends count, as separate features or signals, may not be useful to the classifier since they are treated somewhat independently. But the friends-to-followers ratio can turn out to be a very strong signal that may expose bot users. An Instagram user with 1,000 friends doesn’t raise any flags, nor would an Instagram user with 50 followers; treated independently, these features are not strong signals. However, an Instagram user with a friends-to-followers ratio of 20 (or 1,000/50) is almost certainly a bot designed to follow other users. Similarly, a ratio such as posts-versus-followers or posts-versus-friends may end up being a stronger signal than any of those features independently.
Text content such as the Instagram user’s profile bio, website, or username is made useful by deriving second-order features from them as well. A classifier may not be able to do anything with a website’s URL, but perhaps a Boolean has_profile_website feature can be used as a signal instead. If, in your research, you notice that usernames of bots tend to have a lot of numbers in them, you can derive features from the username itself. One feature can calculate the ratio of letters to numbers in the username, another Boolean feature can represent whether the username has a number at the end or beginning, and a more advanced feature could determine if dictionary words were used in the username or not (therefore distinguishing between @themachinelearningwriter and something gibberish like @panatoe234).
Derived features can be of any level of sophistication or simplicity. Another simple feature could be whether the Instagram profile contains a URL in the profile bio field (as opposed to the dedicated website field); this can be detected with a regex and the Boolean value used as the feature. A more advanced feature could automatically detect whether the language used in the user’s content is the same as the language specified by the user’s locale setting. If the user claims they’re in France but always writes captions in Russian it may indeed be a Russian living in France, but when combined with other signals like a friends-to-followers ratio far from 1, this information may be indicative of a bot user.
There are lower level transformations that may need to be applied to the data in the pipeline as well. If the source data is in an XML format but the classifier requires JSON formatting, the pipeline should take responsibility for the parsing and conversion of formats.
Other mathematical transformations may also be applied. If the native format of the data is row-oriented but the classifier needs column-oriented data, the pipeline can perform a vector transposition operation as part of the processing.
Similarly, the pipeline can use mathematical interpolation to fill in missing values. If your pipeline subscribes to events emitted by a suite of sensors in a laboratory setting and a single sensor goes offline for a couple of measurements, it may be reasonable to interpolate between the two known values in order to fill in the missing data. In other cases, missing values can be replaced with the population’s mean or median value. Replacing missing values with a mean or median will often result in the classifier deprioritizing that feature for that data point, as opposed to breaking the classifier by giving it a null value.
What to consider when transforming and normalizing
In general, there are two things to consider in terms of transformation and normalization within a data pipeline. The first is the mechanical details of the source data and the target format: XML data must be transformed to JSON, rows must be converted to columns, images must be converted from JPEG to BMP formats, and so on. The mechanical details are not too tricky to work out, as you will already be aware of the source and target formats required by the system.
The other consideration is the semantic or mathematical transformation of your data. This is an exercise in feature selection and feature engineering, and is not as straightforward as the mechanical transformation. Determining which second-order features to derive is both art and science. The art is coming up with new ideas for derived features, and the science is to rigorously test and experiment with your work. In my experience with Instagram bot detection, for instance, I found that the letters-to-numbers ratio in Instagram usernames was a very weak signal. I abandoned that idea after some experimentation in order to avoid adding unnecessary dimensionality to the problem.
At this point, we have a hypothetical data pipeline that collects data, joins and aggregates it, processes it, and normalizes it. We’re almost done, but the data still needs to be delivered to the algorithm itself. Once the algorithm is trained, we might also want to serialize the model and store it for later use. In the next section, we’ll discuss a few considerations to make when transporting and storing training data or serialized models.
Storing and delivering data in a data pipeline
Once your data pipeline has applied all the necessary processing and transformations, it has one task left to do: deliver the data to your algorithm. Ideally, the algorithm will not need to know about the implementation details of the data pipeline. The algorithm should have a single location that it can interact with in order to get the fully processed data. This location could be a file on disk, a message queue, a service such as Amazon S3, a database, or an API endpoint. The approach you choose will depend on the resources available to you, the topology or architecture of your server system, and the format and size of the data.
Models that are trained only periodically are typically the simplest case to handle. If you’re developing an image recognition RNN that learns labels for a number of images and only needs to be retrained every few months, a good approach would be to store all the images as well as a manifest file (relating image names to labels) in a service such as Amazon S3 or a dedicated path on disk. The algorithm would first load and parse the manifest file and then load the images from the storage service as needed.
Similarly, an Instagram bot detection algorithm may only need to be retrained every week or every month. The algorithm can read training data directly from a database table or a JSON or CSV file stored on S3 or a local disk.
It is rare to have to do this, but in some exotic data pipeline implementations you could also provide the algorithm with a dedicated API endpoint built as a microservice; the algorithm would simply query the API endpoint first for a list of training point references, and then request each in turn from the API.
Models which require online updates or near-real-time updates, on the other hand, are best served by a message queue. If a Bayesian classifier requires live updates, the algorithm can subscribe to a message queue and apply updates as they come in. Even when using a sophisticated multistage pipeline, it is possible to process new data and update a model in fractions of a second if you’ve designed all the components well.
Storing and Delivering Data in a Spam Filter
Returning to the spam filter example, we can design a highly performant data pipeline like so: first, an API endpoint receives feedback from a user. In order to keep the user interface responsive, this API endpoint is responsible only for placing the user’s feedback into a message queue and can finish its task in under a millisecond. The data pipeline in turn subscribes to the message queue, and in another few milliseconds is made aware of a new message. The pipeline then applies a few simple transformations to the message, like tokenizing, stemming, and potentially even hashing the tokens.
The next stage of the pipeline transforms the token stream into a hashmap of tokens and their counts (for example, from hey hey there to {hey: 2, there: 1}); this avoids the need for the classifier to update the same token’s count more than once. This stage of processing will only require another couple of milliseconds at worst. Finally, the fully processed data is placed in a separate message queue which the classifier subscribes to. Once the classifier is made aware of the data it can immediately apply the updates to the model. If the classifier is backed by Redis, for instance, this final stage will also require only a few milliseconds.
The entire process we have described, from the time the user’s feedback reaches the API server to the time the model is updated, may only require 20 ms. Considering that communication over the internet (or any other means) is limited by the speed of light, the best-case scenario for a TCP packet making a round-trip between New York and San Francisco is 40 ms; in practice, the average cross-country latency for a good internet connection is about 80 ms. Our data pipeline and model is therefore capable of updating itself based on user feedback a full 20 ms before the user will even receive their HTTP response.
Not every application requires real-time processing. Managing separate servers for an API, a data pipeline, message queues, a Redis store, and hosting the classifier might be overkill both in terms of effort and budget. You’ll have to determine what’s best for your use case.
Storage and Delivery of the model
The last thing to consider is not related to the data pipeline but rather the storage and delivery of the model itself, in the case of a hybrid approach where a model is trained on the server but evaluated on the client. The first question to ask yourself is whether the model is considered public or private. Private models should not be stored on a public Amazon S3 bucket, for instance; instead, the S3 bucket should have access control rules in place and your application will need to procure a signed download link with an expiration time (the S3 API assists with this).
The next consideration is how large the model is and how often it will be downloaded by clients. If a public model is downloaded frequently but updated infrequently, it might be best to use a CDN in order to take advantage of edge caching. If your model is stored on Amazon S3, for example, then the Amazon CloudFront CDN would be a good choice.
Of course, you can always build your own storage and delivery solution. In this post, I have assumed a cloud architecture, however if you have a single dedicated or collocated server then you may simply want to store the serialized model on disk and serve it either through your web server software or through your application’s API. When dealing with large models, make sure to consider what will happen if many users attempt to download the model simultaneously. You may inadvertently saturate your server’s network connection if too many people request the file at once, you might overrun any bandwidth limits set by your server’s ISP, or you might end up with your server’s CPU stuck in I/O wait while it moves data around.
No one-size-fits-all solution for data pipelining
As mentioned previously, there’s no one-size-fits-all solution for data pipelining. If you’re a hobbyist developing applications for fun or just a few users, you have lots of options for data storage and delivery. If you’re working in a professional capacity on a large enterprise project, however, you will have to consider all aspects of the data pipeline and how they will impact your application’s performance.
I will offer one final piece of advice to the hobbyists reading this section. While it’s true that you don’t need a sophisticated, real-time data pipeline for hobby projects, you should build one anyway. Being able to design and build real-time data pipelines is a highly marketable and valuable skill that not many people possess, and if you’re willing to put in the practice to learn ML algorithms then you should also practice building performant data pipelines. I’m not saying that you should build a big, fancy data pipeline for every single hobby project—just that you should do it a few times, using several different approaches, until you’re comfortable not just with the concepts but also the implementation. Practice makes perfect, and practice means getting your hands dirty.
In this two part series, we learned about data pipelines and various mechanisms that manage the collection, combination, transformation, and delivery of data from one system to the next. Next, to exactly choose the right ML algorithm for a given problem, read our book Hands-on Machine Learning with JavaScript.
Read Next
Create machine learning pipelines using unsupervised AutoML [Tutorial]
