Hello everyone, and welcome to this session. I am Nidal Albeiruti. I am a solutions architect with Amazon Web Services. Our topic is about one of the important concepts in machine learning, which is feature engineering, but with a special focus on how to extract more from a specific type of sensors, which are binary sensors. We use them in the Internet of Things context. Therefore, in our agenda, I will take these concepts and discuss them, starting from the general topic down to the specific. So I will start first by discussing IoT solutions in general and their different challenges that must be thought of. To build such solutions. We will funnel down to the role of devices used there, mainly sensors. After that, binary sensors will be defined to clarify what these are and what make them different from other sensors. Then we will move on to introduce what features engineering is and by knowing what the feature is first. Later, we will discuss the additional features that can be created or extracted to help us to build more efficient machine learning models when dealing with binary sensors. Finally, we will introduce Amazon Sagemaker data wrangler as an efficient tool that can be used to make feature engineering tasks easier and repeatable. Let's make a start. So, first topic is Internet of Things IoT IoT solutions build on data received from the deployed sensors, and to receive this data, a connectivity solution in a way or another, must be established. The received data will then be analyzed to act on the different outcomes found in this data. The data or the actions can be forwarded to the integrated solutions to draw insights from them, and those insights can then support decision making or go back to the edge devices to trigger actions through the actuators. And that shows us two things. First, how IoT solutions can be complex and multidimensional. And second is how important is the role of sensors in IoT solutions? The explosive growth in IoT use cases and the sheer number and diversity of devices out there has been phenomenal in this slide. These are just a few examples of how AWS IoT is helping a IoT of customers solve their business problems, to mention a few. There is the use case of optimizing manufacturing. Then we have the use case of remotely monitoring patients and healthcare in the context of the medical field, tracking inventory levels and managing warehouse operations connecting homes in these context of ambient intelligence or connecting buildings or cities growing healthier crops with greater efficiencies, managing energy resources transforming transportation enhancing safety in working environments such as the worker safety and of course to monitor and manage electricity and water networks to achieve energy efficiency. AWS. You can see our customers have different use cases, but all use cases are taking data from sensors. The different types of IoT devices available are powered by embedded processors. We call them microprocessors in the case of sensors, which is at the middle of the slide, or microcontrollers in the case of actuators, which appear at the left hand side of this slide. Both of these devices have the necessary device software that enables them to integrate with AWS IoT. So where do we get these sensors and devices from? AWS has a device qualification program and qualified devices get listed in the AWS partner device catalog, and that helps you discover qualified hardware that worlds with AWS IoT services, so you never have to worry if your selected device will work with AWS IoT. The URL for this catalog is devices amazonaws.com. So binary sensors is part of the bigger group of sensors used in IoT solutions. They report the state of the monitored entities back to the IoT solutions. The peculiarity of their readings or values is that they can be only one of two mutually exclusive values, hence the name binary, which is demonstrated in the examples on the slide. So now here are the examples of binary sensors that are widely used in IoT environments. First, we have the passive infrared sensors, PIR sensors. These are motion detectors and they can report back movement or no movement. Similarly, we have pressure sensors and they can report pressure or no pressure connectivity sensors. They can report if there is a connection or no connection. Same for vibration. And finally, we have the example of smoke sensors, which are smoke detectors. They can report if there's a smoke or no smoke. Now we'll move on to the topic of feature engineering, which is under the context of machine learning topic. So feature engineering is part of the process to get data ready for machine learning modeling. After locating data, you first need to work on the various formats from the different identified sources, such Aws databases or data warehouses, which may require creating complex queries. Alternatively, data may exist AWS CSV or compressed format files on s three, for example, in data lakes. As part of this step exploratory data analysis, EDA is executed, and it's about exploring and analyzing the raw data even without domain knowledge. Once data is collected, it needs to be transformed into a usable format. Transforming your data requires you to write code to do these tedious tasks, for example, converting numbers into floating point dates into timestamps, or converting category text labels into integers. And all of those are some well known feature engineering tasks. After the data is transformed, you then write more code to create visualizations to inspect and analyze data such as quickly detecting outliers or extreme values within a data set, which is part of feature engineering as well. Once you have prepared your data in your development environment, you must make the data preparation work in production. This requires help from it operations team to schedule the data preparation to occur. AWS needed, such as on a regular calendar schedule or when the new data is available, or to translate the data preparation code into a more scalable language. Machine learning algorithms take a representation of the reality as vectors of features, which are aspects of the reality over which the algorithm operates. Pedro Domenigos says, some machine learning projects succeed and some fail. What makes the difference easily is the most important factor, which is these features used. So the definition of the feature you see here concurs with what Pedro has said. And on the right hand side of this slide you can see the different types of features. So what makes a good feature? A good feature must be informative, describes something that makes sense to a human. A good features must be available. You must have as many instances as possible or you need to deal with the missing data. A good feature must be discriminant, which divides instances into the different target classes or correlates with the target value. Good features allow a simple model to beat a complex model. You want features also to be as independent from each other and simple as possible. AWS better features usually means simpler models. We have to make the difference clear between features with hyperparameters. Hyperparameters are a set of parameters that are not determined by the learning algorithm, but rather specified as inputs to the learning algorithm. Hopefully the difference is clear. So, back to features. Now let's see some examples of features here. In this slide you can see the examples of features that can be created or extracted of your collected data, whether it's coming from sensors or other sources. In noniot solutions, the examples include in the case of images, we can extract the colors there, the texture, the contours. In the case of signals, the frequency, these phase, the samples, the spectrum, time series we have text and trends and self similarity between different time windows of the time series, biomedical context, the dna sequence and genes for text. One of these well known features to be extracted is POS tags, which refers to the parts of speech, which is the process to apply word classes to words within a sentence. For example, you take nouns, verbs, prepositions and tag them within statements. Great, now we know what the features is. So what is feature engineering? Feature engineering is the process of representing a problem domain to make it amenable for learning techniques. This process involves the initial discovery of features and their stepwise improvement, all based on domain knowledge and the observed performance of a given ML algorithm over specific training data. So features engineering helps improve results by modifying the data's features to better capture the nature of the problem. And feature engineering tends to bring out performance gains beyond tweaking the algorithms themselves in machine learning context. Sometimes feature engineering is referred to as data munging or data wrangling. Regardless, features engineering is a process to select and transform variables when creating a predictive model using machine learning or statistical modeling. And future engineering is an art like engineering is an art, like programming is can art, like medicine, is an art. There are well defined procedures in future engineering, and these defined procedures are methodical and provable, and they are widely known and understood, and future engineering is sensitive to these machine learning. Algorithm being used. AWS there are certain types of features, for example, categorical, that fare better with some algorithms, for example, decision trees, than others, such as, for example, svms. This slide lists the following subproblems, and they are as well known as subdomains or subproblems of feature engineering processes. I will go through some of them just giving examples. Starting by feature creation. Feature creation identifies these features in the data set that are relevant to the problem at hand. Moving on to feature importance, which is related to wrapper methods, it is possible to take advantage of some of the algorithms that do embedded feature selection to obtain a feature importance value as measured by the ML algorithm. For example, random forests produce an out of these feature importance score for each features as part of their training process. Feature transformation manages replacing missing features or features that are not valid. Some techniques include forming cartesian products of features, nonlinear transformations such as binning numeric variables into categories, and creating domain specific features. Moving on to feature extraction, it is the process of creating new features from existing features, typically with the goal of reducing the dimensionality of the features and in features engineering worlds the curse of dimensionality is a well known term referring to having numerous features and not restricting or reducing these dimensionality of these features that you have. While feature selection is the filtering of irrelevant or redundant features from your data sets, this is usually done by observing variance or correlation thresholds to determine which features to remove. So let's discuss now what are these suggested feature engineering techniques? And to do that, we'll introduce these first set of the well known feature engineering techniques, such AWS imputation handling outliers binning log transform one hot encoding grouping operations feature split scaling extracting dates there are other ways for feature engineering, but those are a good and well known examples. Feature feature feature engineering techniques binary IoT sensors now into the related features engineering techniques that we use with sensors. So what can we do with the data that we get from sensors in the IoT world? These slide shows an example of applying Fourier transform which is a mathematical transformation. It is widely used to take the sensors reading, which are distributed along the time domain, and transport them to a frequency domain. This is a very helpful transformation in the case of any sensors, including binary sensors. But what can we do for binary sensors? Is it more than on off indication only? Let's see specifically for binary sensors, what I'm suggesting here are these techniques in this slide and upcoming slides. First is a way to cut up the observations into a series of time windows. These inside each window you ignore the temporal parts of your sensor data. These technique can include same or different types of sensors in one bag and hence its name as in bag of features. Another way is for a series of sensor events can be time windowed and the challenge then is to select the appropriate length of this time window. So we are segmenting the series of sensor events. Please note that you should do some exploratory data analysis to find the typical sensor activity durations, and then you can select the time window length. It can be 5 seconds, 10 seconds or more after segmenting the series into time windows, which is out of the values, the available values that you have in the time window, you take each time window and then you can abstract the whole of it by representing it into a single value. For example, you can take the median of the values available within the time window, or in other cases you can see the presence of one value can be taken to represent the whole of the time window, for example by applying minimum or maximum operations. Let's see more techniques that we can apply to binary sensors. In this case as what we see in the slide. Binary sensors can be used to identify the location of the monitored entity by applying localizations in regions. Technique binary sensors can detect the presence or absence of a particular target in their sensing regions. They can be used to partition a monitored area and provide localization functionality. Even more, if many of these sensors are deployed to monitor an area, the area itself can be partitioned into subregions, and each subregion is characterized by the sensors detecting targets within that region. Let's see more techniques as well. Some of these new features that can be extracted to help your machine learning models. We have the location area, which may be associated with a unique sensor number within the ensemble of deployed sensors. The total elapsed time of each continuously happening event when these sensor has switched to the on state. Because we're speaking about binary sensors here, this will indicate the length of detecting continuous sensor state we spoke about time window, but time window aggregation where we group events together. We can then compare similar periods of the day, for example, together. Finally, merging the binary sensors events with other sensor events to add context to the binary sensor event to extract new features. For example, having a binary sensor associated with a light sensor, for example, to differentiate between an event happening in daytime and another event happening at night, for example. Now, taking the PIR sensor case specifically, and when detecting activity using passive infrared sensors, PIR sensors here are some examples of the features that can be extracted for PIR sensors. For example, these activity level within a region, an area which is the number of events firing or happening or coming from a PIR sensors can be one of the extracted features. For example, as well, we have the elapsed time between sensor events. So if we have two consecutive same value events from the same sensor, this means that the entity in the same region or subregion is active within the region. So those are two consecutive same value events from the same sensors. And the time between those two events can be extracted aws, well as inactivity time within the same region or subregion. In another way, if we have two consecutive same value events from different sensors, this means that the entity being monitored here is moving from one region to another, and this is active time or changing these region can be another feature and a very helpful feature to identify activity between different regions. The frequency of movement events within the same region, as we said, can be extracted as an activity time or no activity time. This is very useful in the case of ambient intelligent intelligence solutions, which is in the home monitoring contexts or environments. Finally, we have a very important event which I found it to be very helpful in ambient intelligence and behavioral modeling techniques is the no sensors events, which means that we don't have any PIR sensor firing coming from any deployed or ensemble of sensors. So any no sensor firings at all can be considered a features, which means that the monitored entity is not available as well, which helps behavioral modeling. So we spoke about feature engineering. We defined what a feature is, and we defined what feature engineering is. Now we're going to introduce Amazon Sagemaker data Wrangler as part of the Amazon Sagemaker studio. Amazon Sagemaker is a service with a lot of different features and capabilities in it. We typically talk about those capabilities as falling into four categories. We have on the left hand side of the slide the data preparation and then these model build phase. Then we move on to training and tuning and deployment and management or hosting of the model. These four categories really address the need that machine learning builders have when dealing with each stage of a model's lifecycle. Since we are discussing feature engineering, the highlighted column on the left shows the prepare these and available services that are used there. Related to that, we can see the data angler which is Sagemaker data angler and Sagemaker feature store. Amazon Sagemaker Data Wrangler is the fastest and easiest way to prepare data for machine learning. Sagemaker Data Wrangler gives you the ability to use a visual interface to access data, perform EDA exploratory data analysis and feature engineering, and seamlessly operationalize your data flow by exporting it into an Amazon sagemaker pipeline. Amazon Sagemaker Pipeline is one way of exporting your data, but you can export as well to Amazon Sagemaker data Wrangler job or a Python file, or to SageMaker feature group. Amazon Sagemaker Data Wrangler also provides you with over 300 built in transforms, custom transforms using Python Pyspark or Spark SQL runtime. Built in data analysis such as common charts, custom charts, and the useful model analysis capabilities such as feature importance, target leakage, and model explainability all are available for you as part of the SageMaker studio. Finally, SageMaker data Angular creates a dataflow file that can be versioned and shared across the teams for reproducibility. With SageMaker data angular data selection tool, you can quickly select data from multiple data sources such as Amazon Athena, Amazon Redshift, AWS, Lake Formation, Amazon S three, and Amazon Sagemaker feature store. Recently, Snowflake has been added as a data source for Amazon Sagemaker data Wrangler. You can now quickly and easily connect to Snowflake without writing a single line of code for other sources. You can write queries for data sources and import data directly into SageMaker from various file formats such as CSV files, Parquet files, and database tables. This data is imported into a secure central data preparation environment where users will have access to a variety of pre built tools to prepare their data. To transform your data, Sagemaker data Wrangler offers a rich selection of pre configured data transformed. For example, you can convert a text file column into a numerical column with a single click, or author custom transforms in Pyspark, SQL or pandas to provide flexibility across your organization. This streamlines the process of cleaning, verifying, and visualizing data without writing a single line of code. Once the data is transformed, Sagemaker data angle makes it easy to clean and explore the data with data visualization in SageMaker Studio. These visualizations allow you to quickly identify inconsistencies in data preparation workflow and diagnose issues before models are deployed into production. Finally, data Wrangler makes it easy to create a pipeline for data preparation. You can export your data preparation workflow to a notebook or code script with a single click, efficiently brings your data preparation workflows into production without manually sifting through and translating hundreds of lines of data preparation code. So this slide summarizes all the steps taken by Sagemaker data wrangler and provide details on what functions data Wrangler performs at each step. The key takeaway here is that these steps can be iterated on in whole or in part, to quickly build a strong set of data transformation code and features to improve the performance and quality of your machine learning models. As we mentioned before, you import from multiple data sources, but here we mentioned Amazon Sagemaker feature store, which is just mentioned as an example only at the right hand of the slide you can see that to operationalize your data you can integrate with these other Amazon services available, or you can export to pipeline your tasks that you have created in data Wrangler to pipeline them and convert them to a script or python file as what we have described before. So we have reached the end of this session. Thank you for taking the time to listen to this session. We have highlighted the different additional features that can be extracted and created based on binary sensors data, and of course all has happened utilizing different feature engineering techniques. So binary sensors are not about zeros or ones, only much more can be extrapolated and extracted to help your machine learning models. Please use these additional features to enhance your models and get more accurate results depending on the selected ML algorithm. Thank you very much. Have a nice time. Thank you.