Data scientists choose models based on various tradeoffs when solving machine learning (ML) problems that involve tabular (i.e., structured) data, the most common data type within enterprises. Among such models, decision trees are popular because they are easy to interpret, fast to train, and can obtain high accuracy quickly from small-scale datasets. On the other hand, deep neural networks offer superior accuracy on larger datasets, as well as the benefits of end-to-end learning, but are black-box and difficult to interpret. TabNet, an interpretable deep learning architecture developed by Google AI, combines the best of both worlds: it is explainable, like simpler tree-based models, and can achieve the high accuracy of complex black-box models and ensembles.

We’re excited to announce that TabNet is now available in Vertex AI Tabular Workflows! Tabular Workflows provides fully managed, optimized, and scalable pipelines, making it easier to use TabNet without worrying about implementation details, and to deploy TabNet with the MLOps capabilities of Vertex. TabNet on Vertex AI Tabular Workflows is optimized for efficient scaling to massive tabular datasets. Moreover, TabNet on Vertex AI Tabular Workflows come with machine learning improvements on top of the original TabNet, yielding better accuracy for real-world data challenges.

TabNet on Vertex AI is well-suited for a wide range of tabular data tasks where model explainability is just as important as accuracy, such as financial asset price prediction, fraud/cyberattack/crime detection, retail demand forecasting, user modeling, credit/risk scoring, diagnosis from healthcare records, and product recommendations.

Overview of Tabnet

TabNet has a specially-designed architecture (overviewed in Fig. 1), based on sequential attention, that selects which model features to reason from at each step. This mechanism makes it possible to explain how the model arrives at its predictions and the thoughtful design helps with superior accuracy. TabNet not only outperforms alternative models (including neural networks and decision trees) but also provides interpretable feature attributions. More details, including results on academic benchmarks, can be found in our AAAI 2021 paper.

Since its publication, TabNet has received significant traction from various enterprises across different industries and a variety of high-value tabular data applications (most of which include the ones for which deep learning was not even used a priori). It has been used by numerous enterprises like Microsoft, Ludwig, Ravelin, and Determined. Given high customer interest on TabNet, we’ve worked on making it available on Vertex given the real-world deep learning development and productionization needs, as well as improving its performance and efficiency.

Highlights of TabNet on Vertex AI Tabular Workflows

Scaling to Very Large Datasets
Fueled by the advances in cloud technologies like BigQuery, enterprises are increasingly collecting more tabular data, and datasets with billions of samples and hundreds/thousands of features are becoming the norm. In general, deep learning models get better learning from more data samples, and more features, with the optimal methods as they can better learn the complex patterns that drive the predictions. The computational challenges become significant though when model development on massive datasets is considered. This results in high cost or very long model development times, constituting a bottleneck for most customers to fully take advantage of their large datasets. With TabNet on Tabular Workflows, we’re making it more efficient to scale to very large tabular datasets.

Key Implementation Aspects: The TabNet architecture has unique advantages for scaling: it is composed mainly of tensor algebra operations, it utilizes very large batch sizes, and it has high compute intensity (i.e., the architecture employs a high number of operations for each data byte transmitted). These open a path to efficient distributed training on many GPUs, utilized to scale TabNet training in our improved implementation.

In TabNet on Vertex AI Tabular Workflows, we have carefully engineered the data and training pipelines to maximize hardware utilization so that users can get the best return for their Vertex AI spending. The following features enable scale with TabNet on Tabular workflows:

• Parallel data reading with multiple CPUs in a pipeline optimized to maximize GPU utilization for distributed training, reflecting best practices from Tensorflow.
• Training on multiple GPUs that can provide significant speedups on large datasets with high compute requirements. Users can specify any available machine on GCP with multiple GPUs, and the model will automatically run on them with distributed training.
• For efficient data parallelism with distributed learning, we use Tensorflow mirrored distribution strategy to support data parallelism across many GPUs. Our results demonstrate >80% utilization with several GPUs on billion-scale datasets with 100s-1000s of features.

Standard implementations of deep learning models could yield a low GPU utilization, and thus inefficient use of resources. With our implementation, TabNet on Vertex, users can get the maximal return on their compute spend on large-scale datasets.

Examples on real-world customer data: We have benchmarked the training time specifically for enterprise use cases where large datasets are being used and fast training is crucial. In one representative example, we used 1 NVIDIA_TESLA_V100 GPU to achieve state-of-the-art performance in ~1 hour on a dataset with ~5 million samples. In another example, we used 4 NVIDIA_TESLA_V100 GPUs to achieve state-of-the-art performance in ~14 hours on a dataset with ~1.4 billion samples.

Improving Accuracy given Real-World Data Challenges

Compared to its original version, TabNet on Vertex AI Tabular Workflows has improved machine learning capabilities. We have specifically focused on the common real-world tabular data challenges. One common challenge for real-world tabular data is numerical columns having skewed distributions, for which we productionized learnable preprocessing layers (e.g. including parametrized power transform families and quantile transformations) that improve the TabNet learning. Another common challenge is the high number of categories for categorical data, for which we adopted tunable high-dimensional embeddings. Another one is imbalance of label distribution, for which we added various loss function families (e.g. focal loss and differentiable AUC variants). We have observed that such additions can provide a noticeable performance boost in some cases.

Case studies with real-world customer data: We have worked with large customers to replace legacy algorithms with TabNet for a wide range of use cases, including recommendation, rankings, fraud detection, and estimated arrival time predictions. In one representative example, TabNet was stacked against a sophisticated model ensemble for a large customer. It outperformed the ensemble in most cases, leading to a nearly 10% error reduction on some of the key tasks. This is an impressive result, given that each percentage improvement on this model resulted in multi-million savings for the customer!

Out-of-the-box Explainability

In addition to high accuracy, another core benefit of TabNet is that, unlike conventional deep neural network (DNN) models such as multi-layer perceptrons, its architecture includes explainability out of the box. This new launch on Vertex Tabular Workflows makes it very convenient to visualize explanations of the trained TabNet models, so that the users can quickly gain insights on how the TabNet models arrive at its decisions. TabNet provides feature importance output via its learned masks, which indicate whether a feature is selected at a given decision step in the model. Below is the visualization of the local and global feature importance based on the mask values. The higher the value of the mask for a particular sample, the more important the corresponding feature is for that sample. Explainability of TabNet has fundamental benefits over post-hoc methods like Shapley values that are computationally-expensive to estimate, while TabNet’s explanations are readily available from the model’s intermediate layers. Furthermore, post-hoc explanations are based on approximations to nonlinear black-box functions while TabNet’s explanations are based on what the actual decision making is based on.

Explainability example: To illustrate what is achievable with this kind of explainability, Figure 2 below shows the feature importance for the Census dataset. The figure indicates that education, occupation, and number of hours per week are the most important features to predict whether a person can earn more than $50K/year (the color of corresponding columns are lighter). The explainability capability is sample-wise, which means that we can get the feature importance for each sample separately.

Benefits as a Fully-Managed Vertex Pipeline

TabNet on Vertex Tabular Workflows makes the model development and deployments tasks much simpler – without writing any code, one can obtain the trained TabNet model, deploy it in their application, and use the MLOps capabilities enabled by Vertex Managed Pipelines! Some of these benefits are highlighted as:

• Compatibility with Vertex AI ML Ops for implementing automated ML at scale including products like Vertex AI Pipelines and Vertex AI Experiments.
• Deployment convenience: Vertex AI prediction services, both in batch and online mode, are supported out-of-the-box.
• Customizable feature engineering to enable the best utilization of the domain knowledge of users.
• Using Google’s state-of-the-art search algorithms, automatic tuning to identify the best-performing hyperparameters, with automatic selection of the appropriate hyper-parameter search space based on dataset size, prediction type, and training budget.
• Tracking the deployed model and convenient evaluation tools.
• Easiness in comparative benchmarking with other models (such as AutoML and Wide & Deep Networks) as the user journey would be unified.
• Multi-region availability to better address international workloads.

More details

If you’re interested in trying TabNet on Vertex AI on your tabular datasets, please check out Tabular Workflow on Vertex AI and fill out this form.

Acknowledgements: We’d like to thank Nate Yoder, Yihe Dong, Dawei Jia, Alex Martin, Helin Wang, Henry Tappen and Tomas Pfister for their contributions to this blog.

GOALS

• Develop smarter, more effective media buying models through data
• Drive relevant advertising
• Scale predictive modelling across all touchpoints in the customer journey

APPROACH

Combined contextual data to create a predictive model with granular layers

Activated data in real-time

RESULTS

• 40% lower cost per booking
• More than twice as many bookings at same spend
• 1.4 times higher click-through rate for test group than control

KLM partnered with Relay42, whose data management platform (DMP) empowers marketers to achieve data-driven personalisation at scale.

The Relay42 DMP works by stitching together all touchpoints and data sources (including all Google solutions), orchestrating customer journeys in real-time and then activating the unified data to drive business and user experience results.

Relay42 and KLM created a data flow setup with the Relay42 DMP at the core. Thanks to Relay42’s tag management system, all customer interactions on KLM’s website and app, as well as relevant indicators from other channels and data sources such as email, social, CRM, call center and affiliates can be tracked and synced in Google Analytics 360.

Data gathered via Relay42’s own tags and DoubleClick Floodlight tags can also be sent from the DMP to the DoubleClick platform so that relevant ads can be targeted to appropriate audiences.

In order to perform deep-dive analyses, KLM simply exports raw-level data (from DoubleClick using Data Transfer and from Google Analytics 360 through seamless integration) to BigQuery. BigQuery can then correlate site behaviour with ad impressions, which enables Relay42 to activate the data and inform the build of predictive models.

“We believe that data in combination with technological innovation is bound to make advertising smarter and more relevant on every touchpoint. By combining data sources, leveraging first-party data and activating this in real-time, advertising can turn into a personal dialogue, a rewarding one-to-one interaction instead of one-to-many push messaging.”

– Kevin Duijndam, Cross Channel Marketing Manager, KLM

A Predictive Model to Improve Display Remarketing

KLM decided to test a new approach to remarketing.

“Our goal was to get rid of irrelevant ads, as they are simply annoying”, explains Kevin Duijndam, the airline’s Cross Channel Marketing Manager.

“Our assumption was that people who fly with us often already know us, so it would be irrelevant to tell them about flying with us again. However, we were wondering when exactly someone is a ‘frequent flyer’. The more we thought about it, the more complex the set of business rules became, so in the end we realized we couldn’t just focus on frequent flyers, but instead should use machine learning to understand when ads are irrelevant.”

KLM developed a real-time buying setup to include predictive modelling. In this setup, website and app interactions are tracked in the DMP thanks to the Relay42 tag management system.

Relevant consumer behaviour can be streamed in real-time to a prediction engine developed by KLM in the Google Cloud Platform, with the outputs then streamed straight back to the DMP. From here, the DMP can activate rule-based segments based on the prediction outcome. And by syncing this with DoubleClick, ads can be served and targeted to maximise relevance.

The team tested their new predictive model to assess any gains in performance. KLM deliberately chose to measure the results in an A/B setup within a defined period rather than measuring the differences year over year or month over month. Such comparisons are less reliable due to rapid changes relating to seasonality, internal capacities and external factors caused by competitors.

Business Gains and Customer Experience Wins

In the test, KLM linked the Relay42 DMP customer interaction data to their predictive model to predict how relevant their ads would be. The setup enabled a decision to be made in real-time whether or not to serve a specific ad to a user.

Through the A/B tests it became clear that the new model in fact did generate a significant uplift in bookings. With the cost per booking 40% lower during the test period, KLM was able to achieve more than twice as many bookings at the same spend.

 The test produced wins in terms of customer experience, too. The click-through rate for the test group was more than 1.4 times higher than for the control group, indicating that the new model was successfully reaching users with messages they found to be relevant rather than annoying.

 The success goes beyond improving KLM’s display remarketing efforts, though.

“Even more importantly, we’ve laid the IT data flow foundation in such a way that KLM is now able to execute on our data through all of our digital marketing channels and apply our prediction models at scale”, Kevin says. “So we can be flexible to plug in other models but can also now scale through other online media channels like search or video.”

It’s no secret that many organizations and job seekers find the hiring process exhausting. It can be time consuming, costly, and somewhat risky for both parties. Those are just some of the experiences we wanted to change when we started Gyfted, a pre-vetted talent marketplace for people who complete tech training or degree programs and are looking for the right career move. At the same time, we’re helping businesses save time and improve recruiting outcomes with our automated candidate screening and sourcing tools.

Our vision is clear: To take a candidate through one structured hiring process, and then put them in front of thousands of companies. It’s similar to the common app system in higher education. Sounds simple, but it is a herculean technical and UX task. To succeed we had to combine advanced psychometric testing, machine learning, the latest in behavioral design, and develop the highest quality structured, relational dataset to represent candidate and manager profiles and preferences on our network. Fortunately, we were cofounded by world-leading experts in these areas including Dr. Michal Kosinski, one of the world’s top computational psychologists, and Adam Szefer, a gifted young technologist. We’ve been joined by a group of equally talented employees, most of whom work remotely in Poland, US, Switzerland, UK, Israel, and Ukraine.

The influence of dating platforms and matching

When seeking inspiration, we were influenced by the success of dating platforms, especially Bumble with its focus on commitment. These platforms have done a great job using design to match people together.

We like to think we’re doing the same for recruiters and candidates in terms of not only role and culture fit matching, but also through a fundamental feature of Gyfted, which is that our job-seekers are anonymous. This helps recruiters meet one of their goals today, which is to minimize bias in the hiring process and enable objective, diversity-oriented recruiting.

Another of our unique selling points is that we conduct candidate screening that is gamified and automated, with a structured interview, where the interview remains with the candidate’s profile. We roughly estimate that just for the 15 million open jobs on LinkedIn, if companies fill in 50% of those via external recruiting, and conduct a screening interview with 10 candidates per job filled at $50/hour paid to an employee to do the screening, that’s $3.75 billion and 75 million hours in direct costs. This is on top of applying for jobs, selecting CVs, and coordinating the process, which takes an even bigger financial and time toll on both applicants and recruiters. Instead, it would be better to take one interview for 1000 companies. The impact of what we want to achieve with our vision is enormous.

We also offer career discovery and career search tools for job candidates. This includes free, personalized feedback for every job-seeker. Right now, we’re aiming the service at students, bootcamp graduates and juniors, helping them to land jobs in tech and the creative industry at large. Next, we’ll expand into mid and senior roles. In the long run we want to reshape how recruiting happens through a common app that saves everyone in the market significant time and resources, helping people find jobs not only faster, but jobs that truly fit them.

Developing advanced AI applications with Google Cloud

We obtained our original funding from angel and institutional investors, and we were selected into the Fall 2021 batch of StartX, the non-profit start-up accelerator and founder community associated with Stanford University. But like most startups our budgets are tight, and we need to find ways to operate as efficiently as possible, especially when building out our technology stack and developer environment.

That’s where Google Cloud comes in. It’s a lot more affordable and flexible than competing solutions, and our developers love it. We use Google App Engine for the hosting and development of our applications giving us enormous flexibility. Vertex AI enables us to build, deploy, and scale machine learning models faster, within a unified artificial intelligence platform. On top of that we use Google Vertex AI Workbench as the development environment for the data science workflow, which allows us to have everything that we need to host and develop innovative AI-based applications.

BigQuery, Google Cloud’s serverless data warehouse, is another stand-out solution for us. We use it to crunch big data from all our systems and the UX is very intuitive and easy to use, allowing us to use it across the business and get insights from a wide range of employees, not just technical experts.

Above all, Google Cloud helps us solve the main platform challenges facing Gyfted including scalability and identity management, so we are perfectly positioned for growth. Right now, we handle about 2 million candidate interactions, a volume we expect to grow exponentially. As that number grows, we rely on Google Cloud to help us scale securely and with reliability.

Eliminating bias from the hiring process

Our technology partners have also been integral to helping us get to an advanced stage of our beta program. MongoDB on Google Cloud takes the data burden off our teams and reduces time to value of our applications. We can stay nimble and can scale database capacity at the push of a button.

Our collaboration with the Google team has been fantastic. Our Startup Success Manager is an expert when it comes to Google Cloud solutions, and he also understands our business from his own experience as an entrepreneur and an investor. It’s great to have an internal point of contact who can help us navigate all of Google’s resources.

I’d also stress the extent to which Google Cloud values align with ours. For example, a key benefit for our customers is the ability to strip unconscious bias out of the hiring process. Google Cloud tools support this commitment to diversity, especially when we are building out our AI models.

On a team level, we also appreciate the support that Google Cloud has shown through its Google Support Fund for Start-ups in Ukraine. This has helped many Ukrainian businesses to continue to operate at a very challenging time, including startups with remote, distributed teams in Poland where most of us stem from.

If I had to sum up Google Cloud and our collaboration with Google for Startups in a phrase, I’d say that it adds enormous value to our business while removing much of the risk when scaling up a start-up. We’ve seen the addition of many new tools and features in the past two years and our Google mentors are always looking at the best way these can be integrated with Gyfted’s own roadmap. That means that we can continue to transform recruiting and hiring processes with the support of one of the world’s most advanced tech companies as a strategic growth partner.

If you want to learn more about how Google Cloud can help your startup, visit our page here to get more information about our program, and sign up for our communications to get a look at our community activities, digital events, special offers, and more.