In the previous section, we saw how a system could benefit from predictions and user preferences to implement suggestions, maximize opportunities, minimize risks and take action to fully leverage its potential.

Some of the fundamental concepts when implementing intelligent systems are: prediction, predictive models and predictive analytics. These are related to the creation of mathematical and statistical models which link observed values to a variable of interest (sometimes called the dependent or response variable) with contextual information, in other words, the data available (also called predictors or independent variables which explain what was observed).

That is why identifying what information is important is crucial. Data mining and feature selection/extraction help in this sense, as they simplify the complexity of the problem. This weeding out of the unimportant information is known as Dimensionality Reduction.

To get familiar with these techniques, the beginner’s guide by S. Ray, K. Jain and D. Gupta is a good read, as well as Seven Techniques for Data Dimensionality Reduction. In the Recommendation System domain, techniques such as Matrix Factorization and Single Value Decomposition (SVD)  have been stressed during recent years as they boast great results, as seen in the famous Netflix competition.

One of the main advantages of dimensionality reduction is that it allows us to simplify its complexity, thus eliminate insignificant variables for a marked improvement in performance and scalability when working with large amounts of data. Furthermore, if the amount of dimensions could be reduced to 2 or 3, the data visualization and the understanding would improve. In this sense, a technique of Dimensionality Reduction which stands out regarding projection and visualization of data is t-SNE, winner of the Merck Visualization Challenge by Kaggle in 2012.

Recently, I have used this technique in the research Vector representation of Internet Domain Namesusing a Word Embedding technique presented in SLIOIA 2017 (Latin American Symposium on Operations Research and Artificial Intelligence) with the purpose of finding semantic similarities between internet web pages. It allowed the minimizing and visualizing of the vectorial representation, learned from Internet domain, of more than 100 dimensions to just 2 or 3. Figure 1 shows an example of 2 dimensional projections after implementing t-SNE for some vectors from Internet domains. Thus, similar sites are crafted by close sites and different sites are modeled by far sites.
‍

Now we can start to link these concepts to understand that a smart system is, usually, heavily based on predictions. These predictions are generated from:

• The recognition of patterns in data, which often times come from multiple heterogeneous sources and need preprocessing or transformation towards a common model.
• The selection and extraction of the most significant characteristics which allow for reducing the complexity of the problem and relating it in some way to generate a predictive model. This will then be displayed and used from an application or a system to maximize opportunities and minimize risks.
  

All this is what is known as the process of Machine Learning, which is summarized in Figure 2.

Machine Learning (ML) is a term that, though it has existed for many decades, has recently become a hot topic, mainly due to the progress made in Artificial Neural Networks (ANN) and Deep Learning (DL). These are specific techniques from Machine Learning which have enabled us to move significantly forward in the cognitive areas such as: image processing, text analysis and natural language processing (NLP). Furthermore, as it’s shown in Figure 3, Machine Learning is a field inside of Artificial Intelligence (AI).

In the information age, the Internet of Things (IoT) and Big Data have empowered the development of Machine Learning and smart applications, due to their great capacity to generate, process and store large amounts of data. It is the fuel that Machine Learning needs in order to work.
.

Fig 3 - Deep Learning and Machine Learning assubset of artificial intelligence

Just like people learn and improve their skills by gathering more experience and information, programs based on Machine Learning improve the accuracy of the results through the constant use of the system, data about the interaction of the users and the contextual information obtained from multiple external sources.T

he algorithms from Machine Learning can be supervised or unsupervised, depending on their training mode. The algorithm is supervised if, when training our model, for each input we indicate the expected output – also called label, target or response. In this case, if the output has a numeric value among an infinite amount of possible values, we are talking about a regression problem. On the contrary, if the value is discrete or finite, each output is called class and the problem to solve is that of classification. If there were only 2 outcomes (yes/no, true/false, etc.) the problem is called binary classification.

In contrast, the algorithm is unsupervised if, when training our model, the expected outputs for the different inputs are not taught. In this case, the unsupervised algorithms are useful to understand the structure and the organization of the data. They are usually used to group the inputs into groups or clusters (clustering) or to apply dimensionality reduction. The unsupervised algorithms are difficult to evaluate because there are no expected results with which to compare a prediction.

To evaluate the models generated from the supervised algorithms, we divide the dataset in two parts, one called the training set, used to train the model (80% or 70% of the total of the entire dataset), and another called the evaluation/validation set, used to evaluate the accuracy of the generated model (20% or 30% of the dataset). In this way, we can observe for each input in the evaluation set the difference between the model’s prediction and the actual expected result. Figure 4 shows some of the main algorithms of Machine Learning and in which category they could be employed. We can see that some may be used in more than one category.

In the news site example, predicting the number of users is a regression problem, knowing to which category they belong is a classification problem and studying them to know if there are groups/segments or not is a clustering problem.

Fig.4 - Algorithms of Machine Learning

We already know that data is the fuel needed to make Machine Learning work. In other words, the more the system is used and the more data available for pattern identification there is, the better. Hence, if you are about to launch or already have systems in production to which you would like to add features of Predictive Analysis and Machine Learning in the future, make sure you are registering all the relevant information. When in doubt about the relevance of something, it is preferable to store that information. It’s better to overestimate than underestimate and then end up needing it. Many times this information is represented in access logs, web pages visited, usage time, etc. Due to the amount that this information can generate, in most cases it does not make sense to keep it in the transactional database of the system. Thus, in an early stage of any Machine Learning project it is crucial to create a good logging strategy, decide what kind of storage will be used, if they would be centralized or distributed, define a backup plan, etc.

So, considering that data is of paramount importance, you are probably wondering: how could Machine Learning be applied to a new system if there is no data history? For example, during the first days using the news site, we could wonder: how can we identify patterns in the data to predict the amount of users that will access the site next weekend? Or which news should we suggest to a new user without a record of visits? These problems are known as “lack of data” and “cold start”. A common alternative to solve the initial lack of data is to access external datasets in the same application domain and use it as system input for the initial model. Then, periodically, we can update the generated model taking into consideration the new data in the system since the last update. UCI Machine Learning Repository and Kaggle Datasets are two points of reference where we can find a great variety of public datasets to use in Machine Learning projects. For the cold start problem, other techniques are often combined to minimize the lack of history. For example, we could infer a user profile from social networks and (if we have the news metadata or properties) apply algorithms to find which piece of news has the properties that better suit said profile. This is common in Recommender System, which is another interesting field inside Machine Learning.

A Recommender System tries to predict users’ preferences about the system to make suggestions. The personalization of the advertisement in a news web page or the prediction of the kind of news to suggest to a specific user are typical problems that can be solved with recommender systems. The most used techniques by recommender systems are collaborative filtering, content-based filtering and Matrix Factorization. However, the Netflix Price contest has shown that a Hybrid approach, which combines different techniques, is the one that achieves better results.

To get to know more about Recommender Systems we suggest reading the Recommender Systems Handbook or the research work I published in CLEI 2014 with my University of the Republic (UdelaR) colleagues, where we combined Slope One (one of the most simple and effective algorithms inside collaborative filtering) with a strategy known as Multiplicative Utilitarian to generate recommendations for transitory groups of people who go to the movies.

Cloud solutions work quite well in general; we don’t need to know a lot to start using them. Many times they have visual aids which help create our models, some even allow starting for free and paying after increased usage.

One of the Cloud’s main advantages is that it provides a highly scalable and safe infrastructure. Not only does it offer the possibility to create predictive models to solve common problems, but it also provides specific services to work in cognitive areas such as image processing or natural language. This simplifies the task of adding features such as face or object detection or voice recognition. In this way, we aim to maximize the user experience of our application.

Sometimes we cannot use cloud services because we need to manipulate our algorithms at a lower level or because we have the restriction of data confidentiality and we can’t train models in the cloud. In these cases it would be better to directly  implement our solution, whether it is to program a technique if it is simple, such as a simple collaborative filter or a naïve Bayes classifier; or using a framework or specialized library because they usually are optimized in their operations. Also, they are well tested, so we can considerably minimize the testing needed if we compare it to programming the techniques directly.

In the case of programming languages, if you have experience developing with NET and love LINQ as much as we do, it is possible that the frameworks, Accord.NET or NUml, are interesting options to consider, and also the library, math.net. Furthermore, we recommend using the functional language, F#, and reading the books, Machine Learning Projects for .NET developers and Mastering .NET Machine Learning. But, even though any general purpose programming language could be used, the two most popular languages are Python and R.

Furthermore, it is worth mentioning some of the principal products and tools that should be taken into account when implementing a deep learning solution.

My first approach to the deep learning libraries was with Tensorflow due to my Masters’ thesis which had an implementation of the word2vec family of algorithms. These algorithms (Skip-Gram and CBOW) are widely used in Natural Language Processing (NLP). Tensorflow was developed by Google and shared with the community as an open source project at the end of 2015. In less than two years, it has gained significant popularity, especially because of the fact that it’s used in the production environment by some Google products, which gives it good backing and reputation.

Among the main features of TensorFlow, we highlight its capacity to process in gpu cards, which optimizes both the vectorial & matrix operations and the deploy at great scale for distributed solutions. It comes with APIs for Python, C++ and java (among others) however, the Python API is the most complete and recommended. From my experience with this library, I’d like to point out the support for Windows systems which worked perfectly for me and was easy to use. The tools included in TensorBoard have also been of great help for the visualization of data generated in TensorFlow. Regarding its negative aspects, I believe the learning curve is not easy because it’s a low level library, and both the library and the community around it are starting to mature: the 1.0 version was launched in mid-February 2017.

Since it is so new, it can still be difficult to find documentation or examples to illustrate specific problems we might face. For this reason, it’s helpful to evaluate the usage of some wrapper over Tensorflow which provides a higher level of abstraction with the purpose of mitigating some of these problems. Among them, we highlight:

• Keras (it also includes support for Theano)
• TFlearn
  
• TF-Slim
  
• Skflow (recommended if you have experience with scikit-learn, originally it was a separated project, but it was later included in the Learn module inside TensorFlow

Due to all this, if you already know Python and have a background in Machine Learning and Deep Learning, Tensorflow (or one of its wrappers) is probably one of the best options for smart solutions development. Also, there are other active deep learning frameworks that you can consider nowadays, besides Tensorflow:

• Theano
• Caffe
  
• Caffe 2 (Facebook)
  
• Torch
  
• Microsoft Cognitive Toolkit (before CNTK)
  
• MxNet

On Wikipedia you can find a rather complete list (although not very updated) of deep learning frameworks and you could be interested in reading a very good analysis of R.G. Gomez (BEEVA) about several of them.