We recap some of the major highlights in data science and AI throughout 2018, before looking at the some of the potential newest trends and technological advances for the year ahead.
2018 has passed. The insane pre-holiday shopping is behind us, along with celebrations, and personal to-do lists for the next 12 months. But it’s a great time for a retrospective. So, let’s analyze the data science and artificial intelligence accomplishments and events of the past year. We talked with experts from Booking.com, Wolfram Research, BetConstruct, and other data science specialists who shared their thoughts about opportunities as well as their influence on business, research, and everyday lives for both industries.
Experts have different points of view on whether 2018 was rich in important achievements and events.
Machine learning and data science advisor Oleksandr Khryplyvenko notes that 2018 wasn’t as full of memorable breakthroughs for the industry, unlike previous years. No recent achievements can compete with inventions of a multilayer perceptron (MLP), neural net training techniques like backpropagation and backpropagation through time (BPTT), residual networks, the introduction of Generative Adversarial Networks (GANs), and deep Q-learning networks (DQN). ”So, looking back to memorable ones I listed before, there weren’t ‘brand new’ accomplishments in 2018,” summarizes Oleksandr.
head of data science Dmytro Fedyukov thinks that the release of the new version of the PyTorch machine learning framework for deep neural networks contributes to the unification of modeling, deployment, and debugging with a huge set of third-party libraries that speed up building complex data science pipelines. However, generally not much has changed: “There aren’t advances in algorithmics, at least in the areas in which we’re working — table data, fraud detection, scoring, and computer vision. For our production tasks, we mainly use models and methods that were known at least 3-4 years ago with adding tricks for benchmark optimization, stacking or dynamic model selection. The main peculiarity is about optimized libraries allowing to do the job using minimum resources. In other words, the shift towards minimization of resource use is the most interesting for us because we deal with real-time modeling.”
Not all data scientists are skeptical about 2018 advancements as they mention developments and news in several fields.
Jérôme Louradour, a computer research scientist at Wolfram Research, notes ongoing progress in deep learning for computer vision, natural language processing (NLP), and audio signal processing. The expert’s hot topic list includes GANs, reinforcement learning, the use of invertible neural networks (INNs) for generative modeling, as well as connectivity patterns in neural networks architectures, in particular, attention-based Transformers that overtook recurrent neural networks in natural language processing. Although these developments have been going on over the last four years, the expert notes that significant advances were made in 2018. “It’s tricky to make GANs work, and 2018 brought a bag of new tricks to extend their applicability to real problems.”
“There were quite a lot of significant improvements in ML in 2018,” says data science competence leader at AltexSoft Alexander Konduforov. “First of all, after getting great results mostly in computer vision and sound recognition in recent years, this year, deep learning researchers got another breakthrough in natural language processing.”
Principal data scientist at Booking.com Lucas Bernardi admitted it’s hard for him to say which of the events was the most important because “so much happened this year.” In particular the specialist mentioned the NLP field news — the BERT language representation model, which we’ll discuss in more detail below. “I also liked One pixel attack for fooling deep neural networksand Adversarial Reprogramming of Neural Networks, they show how much work is still to be done towards robust machine learning. I also got the feeling that more and more people are thinking about Machine Learning, AI and Causality, I like this article from Judea Pearl Theoretical Impediments to Machine Learning With Seven Sparks from the Causal Revolution, I truly believe lots of breakthroughs in AI will come from Causality related fields,” shares Lucas.
Now let’s take a closer look at accomplishments, problems, and trends for DS and AI.
Head of data science at BetConstruct Dmytro Fedyukov notes a more precise transition of data science projects and processes from research- to production-orientation: “We start to put more effort into building data science pipelines. We build and deliver them to working production offers automatically.”
A data science pipeline is a set of processes for automating data collection, preparation, visualization, modeling, and interpretation (review) to answer business questions or get informed conclusions. Data pipeline is beneficial for organizations that generate, store (also in cloud), and use high volumes of data or multiple data sources, as well as those requiring real-time or complex data analysis.
“What has affected us greatly over the past year is the step towards unification when it comes to working with models, their deployment, and debugging. The reason for this change is a release of PyTorch 1.0 and a great promotion of ONNX (Open Neural Network Exchange Format) regarding that model developed in one deep learning framework can be reused in other frameworks,” the expert shares.
What does that mean for businesses? As ONNX allows data scientists to move models between frameworks more easily, specialists can put their developments into production faster. “So, it’s not the state-of-the-art that motivates businesses to use data science more but the standardized approach to machine learning model building.”
Dmytro also notes that the meaning of the term research and development (R&D) has changed slightly. Development these days is about the effective delivery of models on production, including deployment and operation. “That means that data scientists should be more of programmers,” concludes the data scientist.
While the number of “We’re hiring” posts from tech companies, other businesses, and government agencies increases, fewer specialists are available to fill the new vacancies. And machine learning tasks still must be performed, at least the routine ones. Vendors introduced tools for automating machine learning workflows (autoML) to solve this problem. While MLaaS solutions help novice data scientists and developers solve more complex and typical problems, autoML tools make it possible for users without ML expertise like analysts or software engineers to complete custom and relatively simple tasks.
These may be data preparation (feature engineering, feature extraction, and feature selection) and modeling tasks, for instance, algorithm selection, hyperparameter tuning, or model assessment.
“Automated machine learning was a popular topic in 2018: Auto-Keras open source software released, and many companies are working on their own products and libraries. This work will definitely proceed further,” thinks Aleksander Konduforov.
It’s important for heavily regulated organizations, such as healthcare providers or financial institutions, to understand the logic behind AI-based systems’ predictions. These entities must know all the facts that influenced a particular decision and provide them if needed to prove they didn’t discriminate against a customer.
Understanding how a model makes decisions — model interpretation — has been on the front burner since the end of 2017. Decision support systems and models they are based on that don’t explain which features influenced their decisions were known as black boxes.
Data scientists working with neural networks are obviously the ones with the most severe headaches. “Deep learning models have been black boxes for us — we couldn’t exactly tell why they made forecasts they made. In our case, the problem is about the interpretation of numerous scoring models in fraud detection. Sometimes we have to explain why we deny financial transactions or don’t agree with some game situations using regulatory principles,” shares Dmytro Fedyukov.
Model interpretability is not only important for companies that need to fulfill legal obligations to customers. It serves a technical purpose as well. Every ML model considers input features (problem properties) to predict results (outputs). The more relevant features we create and use to train an ML model during feature engineering, the more accurate results we can get and the simpler our model is. That’s why the ability to understand how the model makes predictions is crucial for its debugging.
“The good news is that there are tools providing data scientists with more information for model debugging and validation [evaluation to achieve its best performance.] Some such tools are included in TensorFlow (i.e., TensorBoard) and PyTorch,” adds Dmytro.
Another nifty and relatively new tool is Featuretools — an open source Python library for automated feature engineering created by Feature Labs in 2017. Featuretools has a built-in LIME (Local Interpretable Model-Agnostic Explanations) module, which is a technique that helps to explain predictions of any model.
The purpose of LIME and use case examples
For the sake of simplicity, we won’t describe how the mechanism works. However, you can learn more about LIME here.
“So, this technique becomes applicable in production systems, and that’s what businesses and regulators need,” summarizes Dmytro Fedyukov from BetConstruct.
The popularity of Featuretools has grown significantly in 2018 among data scientists and developers, according the vendor’s report.
Natural Language Processing (NLP) is a branch of AI that focuses on machine understanding, interpretation, processing, and usage of human language. NLP tasks include sentiment analysis, semantic search, summarization, dialogue state tracking, language translation, and others.
The NLP community has been discussing how natural language understanding has to be done, shares Oleksandr Khryplyvenko. “What does it mean to ‘understand’ a sentence?”
Google’s response to this question is the BERT technique for pre-training NLP applications. Pre-training entails using a huge volume of unlabeled generic text from the web to train a language representation model before fine-tuning it on a small dataset prepared specifically for a task. BERT (Bidirectional Encoder Representations from Transformers) is intended to help developers and researchers working with deep neural networks to deal with the shortage of training data (labeled/annotated public datasets).
BERT models can predict each input word in a context (based on previous words in a sentence), as well as understand the relationships between sentences. Google also optimized Cloud TPU (Cloud Tensor Processing Units used to streamline ML workloads). The company researchers note these improvements will allow for training an NLP model in 30 minutes using a single TPU or in a few hours with a single GPU (graphics processing unit.)
Principal data scientist at Booking.com Lucas Bernardi thinks that BERT will have a huge impact on the industry: “The BERT language representation model not only gives state of the art results on many NLP tasks but also makes rich word representations accessible to more researchers and engineering without the need of end to end training of large and complex neural networks.”
The introduction of BERT, as well as Universal Language Model Fine-tuning (ULMFiT), ELMo, and other proposed solutions brought the power of transfer learning to texts and were able to improve many state-of-the-art results, according to Aleksander Konduforov of AltexSoft. Transfer learning is the use of knowledge gained while solving a former problem to deal with the latter, related problem. In this case, transfer learning entails pre-training an NLP model on a generic text. “This can lead to improvements in many applications. For instance, virtual assistants (Siri, Alexa, and Google Assistant) will be able to better understand the speech and give more relative answers to a user; the capability of chatbots to define the context of requests might increase as well,” the expert adds.
Wolfram Research‘s Jérôme Louradour says that a big breakthrough in machine translation was made in 2018: “State-of-the-art deep learning models have reached human performance in such tasks as Chinese to English translation. Besides, tremendous progress was made in unsupervised machine translation for which specialists don’t need to build a labeled training dataset with translation examples.”
Specialists also introduced new methods for training translation models using two independent sets of text in the source and the target language, without any alignment. “These advances unlock the possibility for training machine translation models on rare languages, such as Urdu,” the expert notes.
Machines now can compete with humans in question answering, which is another application of NLP. OneConnect, a subsidiary of Ping An Insurance has developed a modelwith an accuracy of 83.435 percent, which is close to the human average performance (86.831 percent). The model was trained on the latest version of Stanford Question Answering Dataset (SQuAD).
SQuAD2.0 is a reading comprehension dataset that consists of more than 100,000 question-answer pairs from its previous version and over 50,000 unanswerable questions. The dataset publishers’ goal is to force models to understand when a question can’t be answered given a context.
Ping An representatives note that the model can be applied to enhance customer service.
Oleksandr Khryplyvenko notes that NLP will continue transforming customer support through automated virtual assistants. Numerous companies deliver messaging automation solutions, for instance, Agent.ai. Intra-corporate information and contact search is another common use case for the technology that will remain in 2019. One of such services, the People.ai platform, can be used for CRM contact management to ease workflows for marketing, customer success, and sales teams.
Brands are looking for ways to engage with clients in relationships that would be mutually beneficial. Businesses strive to personalize interaction with customers showing content that may be meaningful for each of them. Data on customer preferences, their individual details, online behavior, or device must be collected and processed to deliver a personalized experience.
Entertainment is one of the industries leveraging DS and AI to recommend media content (think of suggestions on Spotify, Netflix, or YouTube.)
Dmytro Fedyukov notes that personalization remains the industry trend: “Even in our case case [BetConstruct provides online gaming and sports betting solutions], we see more granular personalization aimed at targeting a particular user with their needs, as well as the attempt to create a customized, comfortable environment. We start relying less on classic recommender systems and try to communicate with a user solely via a personal channel providing tailored information and content that are relevant for a user at a given time. For example, if a user likes the 2018 FIFA World Cup, the environment must fully inform him or her about this event. But in a not very intrusive way.”
Travelers appreciate offers with a personal touch. According to Google/Phocuswright Travel Study 2017, 57 percent of US travelers feel that brands should consider personal preferences or past behaviors when providing information.
What’s more, 36 percent of customers are likely to pay more for services if companies tailor information and overall travel experience. We discussed approaches to customer experience personalization in travel and hospitality in a dedicated article, so feel free to check it if you want to know more.
Global online travel brands respond to traveler preferences putting high volumes of data to work to improve customer experience. Principal data scientist at Booking.com Lucas Bernardi notes that the personalization is the main opportunity for machine learning in the travel industry: “More and more people want personalized services, these services are becoming a must, so we need to make sure we can offer that.
The travel company specialists deal with plenty of data of all types. “In particular, we’re working a lot with what we call transactional data (user searches, clicks, reservations, etc.) We’ve been also working with computer vision systems for a couple of years now. Text data is also present across the whole funnel, from searches, to destinations and accommodations descriptions and guest reviews. We try to select content that’s relevant and useful for our customers.” Bernardi also talks about major success with machine translation used for such tasks as translating hotel descriptions.
Personalization at glance: Booking.com tailors products/services based on traveler type, past searches from an endpoint device, and offers a more personalized search to registered users
However, building and developing a recommender system for the travel sector poses several challenges. One of the problems is caused by the exactitude of services/products provided via an online travel agency. Unlike media content (songs or movies), accommodation options are limited and dynamic: “You cannot really recommend one room to everyone because of a limited number of rooms that we can sell.” Moreover, recommending a wrong hotel or an apartment may cause more customer frustration than suggesting a video, for instance. So, many attributes for booking options (location, facilities, refund policies, etc.) must be taken into account to make the suggestion that meets the preferences and requirements of a particular person. Data sparsity is another problem: “The interaction of users with our platform is sparse. People don’t get to us every day, they use the website only a couple of days a year. So, it’s very tricky for us to construct a proper user profile.”
All these interconnected issues are related to the general match and supply problem: Travelers are looking for the best accommodations and suppliers need to showcase their properties in ways to attract the right guests. You can read more on this topic in the Learning to Match research paper. Although many of the techniques and methods developed for the Movies Recommendations case are also relevant for the travel sector, one can’t look at this challenge as recommendations systems alone, stresses Lucas. Booking.com addresses this problem by looking at it from various angles (i.e. information retrieval, econometrics, game theory). “The real challenge is how to combine the relevant methods from each field into a system that makes sense as a whole.”