Given the May 1 deadline for students to enroll, colleges and universities have been carefully watching the numbers of students who put down deposits and commit to a school. Like nearly every other sector in the U.S., colleges and universities have been hit hard by the pandemic and economic downturn, making enrollment numbers more important than ever to maintain financial health and meet student body goals. Technology and AI-based models are providing a way for institutions to manage their admissions data efficiently and cost-effectively.

For example, an admissions officer may want to offset drops in international or out-of-state enrollment by attracting local talent so they can achieve their target class profile. Additionally, institutions have sought to increase the availability and access of higher ed and are intentionally recruiting more low-income and underrepresented populations. New technologies allow admissions officers to have a stronger grasp of the incoming class’ details. With the press of a button, users can switch between overall, in-state, out-of-state, and international enrollment figures. The “How Do Regions Compare to Last Year” tile lets users view the percentage of change and the total number of candidates by region, comparing 2020 and 2021 side-by-side. With access to data, regional recruiters can focus on individual candidates and develop a plan to change the numbers.

An innovative enrollment strategy at MSU

Michigan State University was looking to improve how they managed the recruiting process. They implemented an enrollment strategy that saw their out-of-state enrollment increase by over 20% and generate $5M in additional net tuition revenue in a single year. “We knew we needed better real-time and predictive insights about our enrollment pipeline, and we understood the value of bringing in external data on prospects, but we couldn’t do that by ourselves. 

The solution really helped us be innovative with our analytics and improve our enrollment outcomes,” said someone close to the project at Michigan State University.

Data-driven insights support informed enrollment decisions

Student admissions and marketing groups are turning to Google Cloud and Deloitte to help support their enrollment decisions with predictive, actionable insights across recruiting and admissions processes. The partnership resulted in Candidate360, a solution that helps institutions process and analyze large amounts of data and develop meaningful insights to support their enrollment goals and mission. Part of Google’s Student Success Services offerings, Candidate360 uses artificial intelligence and predictive analytics to help higher education institutions improve enrollment and matriculation, as well as optimize financial aid decisions.

With nine AI/ML models, Candidate360’s capabilities help university leaders, enrollment teams and recruiters do things like:

• Identify regions with clusters of candidates or see dips in application numbers
• Deploy marketing resources and admissions staff to the right geographic areas 
• Understand applicants’ needs in terms of academics, safety, and social activities
• Make intelligent financial aid decisions to recruit and retain talented students from underrepresented communities
• Respond faster to the students who are most likely to be accepted and then go on to enroll, stay enrolled, and graduate
• Work towards the ideal class composition and increase competitiveness with peer institutions

Enhancing the student experience

Enrollment is the first key touchpoint with incoming students, and tools like Candidate360 can help institutions build a student preferences profile throughout the application process. They can understand each student’s prospective major, areas of interest, dream job, and more, allowing colleges and universities to personalize recommendations and advising for enrolled students.

With integrated toolsets, Candidate360 and Google Cloud’s Student Success Services can help colleges and universities emerge from the pandemic stronger than ever. 

To see a demonstration of Candidate360 in action, contact our sales team. 

Editor’s note: PedidosYa is the market leader for online food ordering in Latin America, serving 15 markets and over 400 cities. It’s also one of the largest brands within the German multinational company Delivery Hero SE. With over 20 million app downloads, PedidosYa provides the best online delivery experience through 71,000+ online partners, including restaurants, shops, drugstores, and specialized markets. 

Having constant access to fresh customer data is a key requirement for PedidosYa to improve and innovate our customer’s experience. Our internal stakeholders also require faster insights to drive agile business decisions. Back in early 2020, PedidosYa’s leadership tasked the data team to make the impossible possible. Our team’s mission was to democratize data by providing universal and secure access while creating a comprehensive information ecosystem across PedidosYa. We also had to achieve this goal while keeping costs under control— even during the migration stage and removing operational bottlenecks. 

Challenges with legacy cloud infrastructure

PedidosYa first built its data platform on top of AWS. Our data warehouse ran on Redshift, and our data lake was in S3. We used Presto and Hue as the user interfaces for our data analysts. However, maintaining this infrastructure was a daunting task. Our legacy platform couldn’t keep up with the increasing analytics demands. For example, the data stored on S3 complemented by Presto/Hue required high operational overhead. This was because Presto and our IAM (identity access management) didn’t integrate well in our legacy ecosystem. Managing individual users and mapping IAM roles with groups and Kerberos was operationally time-consuming and costly. Further, sharding access on the S3 files was far too complicated to enable seamless ACLs (access control lists).  

There were also challenges with workload management. Our data warehouse had batch data loaded overnight. If one analyst scheduled a query to run during the overnight ETL (extract, transform, load) workload, it would disrupt the current ETL task. This could stop the entire data pipeline. We’d have to wait until data engineers intervened with a manual fix.

It was also difficult to understand whether a query error was due to performance issues or platform resource exhaustion. This lack of clarity affected our data analysts’ ability to autonomously improve querying efficiency. Data team members needed to manually inspect personal queries looking for performance issues. Also,  the current architecture was prone to a ‘tragedy of the commons’ situation; it was seen as an unlimited and free resource. As a result, it was impossible to disentangle the infrastructure from different stakeholder teams, as all had very different needs. 

The decision to modernize our data warehouse

Given the growing challenges from our legacy platform, our tech team decided to transform our analytics environment with a modern data warehouse. They required the following key criteria from their next data platform: 

• Scalability – The ability to grow with elastic infrastructure.
• Cost control – Cost management and transparency. These factors promote efficiency and ownership—both key aspects of data democratization.
• Metadata management – Intuitive data platform focusing on users’ previous SQL knowledge. Plus, being able to enrich the informational ecosystem with metadata,  to diminish data gatekeepers.
• Ease of management – The team needed to reduce operational costs with a serverless solution. Data engineers wanted to focus on their key roles rather than acting as database administrators and infrastructure engineers. The team also wanted much higher availability, and to reduce the impact of maintenance windows and vacuum/analysis.
• Data governance and access rights – With a growing employee base with varying data access requirements, the team needed a simple yet comprehensive solution to understand and track user access to data.

Migrating to Google Cloud

After exploring other alternatives, we concluded Google Cloud had an answer to each of our decision drivers. Google Cloud’s serverless, managed, and integrated data platform, coupled with its seamless integration across open-source solutions, was the perfect answer for our organization. In particular, the natural integration with Airflow as a job orchestrator and Kubernetes for flexible on-demand infrastructure was key.  

We  used Dataflow together with Pub/Sub and Cloud Functions for our data ingestion requirements, which has made our deployment process with Terraform seamless. Because we set up everything in our environment programmatically, operation time has diminished. Google Cloud reduced the deployment process from about 16 hours in our legacy platform to 4 hours.  This is partly due to the friendliness of automating the deployment (such as schema check, load test, table creation, build.) process with Terraform, Cloud Functions, Pub/Sub, Dataflow, and BigQuery on GCP. Input messages processed with Dataflow allow us to abstract and plan the schema changes according to the needs of the functional team. For example, schema changes raise an alarm, and then we can modify the raw layer table schema. By doing this, we ensure that backend modifications that we don’t control do not affect upper layers.

A key reason why we picked Google Cloud was because of its advanced cost and workload management coupled with its transparent log analytics. This information gives us a complete view into any query performance issues to make improvements on the fly. Further, we achieved a significant amount of cost savings by consolidating multiple tools to BigQuery.With BigQuery, we’ve been able to reduce our total cost per query by 5x.

This was due to a number of reasons:

• Automating pipeline deployment made it much simpler to maintain the data processing processes. 
• Analysts are conscious about what queries they’re running, resulting in running better, more optimized queries. 
• Analysts use a Data Studio dashboard to see their queries and all the associated costs. As a result, there’s a lot more transparency for each persona.

 With these changes, we can easily manage and assign costs associated with each workload with their own cost centers using specific Google Cloud projects.

Change management is always challenging. However, BigQuery is intuitive and doesn’t have a steep learning curve from Hue/Hive on SQL basics. BigQuery also allowed the team to expand its capabilities and enabled them to properly work with nested structures, avoiding unnecessary joins and improving query efficiency. Additionally, we now use Data Catalog as our unique point of truth for metadata management. This allows our team to break the data access barriers and enable federation of data across the organization. By using Airflow to orchestrate everything, we keep track of every data stream. With this information, each end user can see their regularly used data entities’ status via the dashboard. This also adds transparency to our everyday data processes.

Finally, with Google Cloud’s IAM rules applied across the different products, data sharing and access is close to a noOps experience. We have programmatically implemented access according to roles and level access within the company. This allows certain pre-validated roles to view more sensitive information. These solutions help drive a more automated data governance experience. 

Up next: Google Cloud AI/ML

The new stack based on BigQuery has created significant productivity gains. Freed from the burden of operational management, PedidosYa’s data team can now focus on adding value through data tools and products.  

• Our data engineers are better equipped to integrate constantly changing transactional and operational data.
• The dataOps team can automate the infrastructure and provide autonomy to the end user.
• Our data quality team can focus on bringing added value to data stakeholders. 
• Data scientists and data analytics can spend more time analyzing data and less time asking data gatekeepers for data access.

PedidosYa can now democratize data access with a well-governed architecture. We are still at the beginning of our journey, but we are closer to achieving our vision of building a data-driven organization. Up next: expanding our artificial intelligence and machine learning capabilities.

Tune in to Google Cloud’s Applied ML Summit on June 10th, 2021, or listen on-demand later, to learn how to apply groundbreaking machine learning technology in your projects.

Volkswagen strives to design beautiful, performant, and energy efficient vehicles. This entails an iterative process where designers go through many design drafts, evaluating each, integrating the feedback, and refining.

For example, a vehicle’s drag coefficient—its resistance to air—is one of the most important factors of energy efficiency. Thus, getting estimates of the drag coefficient for several designs helps the designers experiment and converge toward more energy-efficient solutions. The cheaper and faster this feedback loop is, the more it enables the designers.

Unfortunately, estimating drag coefficient is an expensive and time-consuming operation that involves either a physical wind tunnel or a computationally intensive simulation. This can be a bottleneck in the feedback cycle.

For this reason, Volkswagen and Google Cloud decided to collaborate on a joint research project to investigate using machine learning (ML) to get fast and inexpensive estimates of the drag coefficient. In this post, we’ll explore the challenges and approaches undertaken in this project.

The core principles of the project were simple. First, we needed to collect a dataset of existing car designs and their respective drag coefficients. Then, we needed to create a representation of the various cars that would be suitable for ML. The next step was to train a deep learning model to predict the drag coefficient, and then, finally, we would use that model to efficiently estimate drag for any new design.

Representing three-dimensional car designs

Design software recreates a physical object as a three-dimensional triangle mesh made up of three types of objects—faces, edges, and vertices. Figure 1, below, shows such a mesh for an Audi S6. Faces are flat surfaces, such as the window in a car door. An edge is where two faces meet (e.g., the side of the door), and a vertex is where two or more edges meet, such as the corner of the door.

Car bodies, however, come in all shapes and sizes. A Volkswagen Golf economy model is very different from a Tiguan SUV, and a single vehicle can have both large smooth surfaces as well as areas with delicately designed features. Consequently, there can be a huge variety from one polygonal mesh to the next.

ML models need consistent representation in order to form robust generalized rules. With such a dramatic variance between each polygonal mesh, the models would be compromised and the results could have huge margins of error.

We needed to find a way to create simple meshes that capture the shape of the car but are still suited for ML models.

Representing a car with digital shrink wrapping

Rather than building a representation of each car from the ground up, we applied a “shrink wrapping” method for the 3D meshes. The principle is very similar to vacuum-sealing a cucumber. The cucumber is placed in a plastic bag and the air is then gradually removed until the bag fits tightly around it, capturing its shape.

Our approach works similarly: we start with a base mesh, a simple shape that corresponds to the plastic bag, and we deform it until it captures the shape of the target mesh. For our purposes, the base mesh is a simplified representation of a car and the target mesh is the particular car we are designing for at that moment. Such meshes can be defined, managed, and presented to ML models for training using the Tensorflow Graphics and trimesh libraries.

Our “shrink wrapping” method mainly works by iteratively minimizing a measure of distance (e.g., chamfer distance) between the two meshes. Additionally we can regularize our mesh to preserve certain qualities, like smoothness, in the resulting mesh. This iterative optimization is analogous to the vacuum pump, gradually shrinking and fitting the vertices of the mesh as closely as possible to the complex shape of the car. With shrink-wrapping, we are able to produce cleaner meshes that are more suitable to our estimation task. An example of such a procedure is shown in Figure 2.

How to train a model

Shrink-wrapping the 3D car designs was an important first step, but the work was far from over. Our next challenge was to build and test the machine learning algorithms.

We wanted our algorithms to estimate the drag coefficient as accurately and quickly as possible each time it looked at a new design. To do so, we had to train the ML models on existing data.

From publicly available datasets, we calculated the drag coefficients for 800 different car meshes, which we trained the models on. Then, we evaluated the trained models on a further 100 meshes, seeing how accurate their estimates were on new data.

As we worked through this training, we refined our approach. Initially, we tested models based on convolutional neural networks – similar to PointNet – that observed only the vertices, i.e., the fixed points in each mesh. But when we tested mesh-convolutional models – similar to FeastNet – we found a slightly different focus improved the accuracy of the estimates. Rather than focusing on vertices alone, these models looked at a mesh of vertices and how they relate to each other. These models placed each vertex in a richer context, leading to more accurate estimates when air-flow hit particularly subtle design features.

Working in parallel and at scale
To collaborate across time zones and two organizations, we’ve used the Google Cloud Vertex AI platform.

Vertex AI Workbench serves as a central hub to interact with other services and infrastructure on the Vertex AI platform. It enables quick experiments and preparation of training packages for resource-intensive ML model training jobs, all in a Python notebook environment for immediate execution of code. The notebook environments allow code-based interaction with other services on Google Cloud and ML tools such as Vertex AI Training and Vertex AI Pipelines.

The process of training a new model is a seamless one. First, a dataset is prepared and stored in Google Cloud Storage, usually with the help of Tensorflow Datasets. Then, for every ML model we want to test, we package and store the training code as a container image with Google Cloud Build and Container Registry. This ensures that every job is fully documented, including the provided parameters, training code package, logs from the training task, and resulting artifacts such as metrics and model files.

From there, we submit the model to the Vertex AI Training service, which provides easy access to large scale infrastructure and hardware accelerators, such as GPUs and TPUs, by simply defining resource needs when submitting a job. By using Vertex AI Training’s hyperparameter tuning feature, we can run experiments in parallel with multiple neural networks to find the right one for our purposes.

With Vertex AI Tensorboard, we can capture metrics and visualize the results of our experiments. These are readily available to anyone in the team, wherever they are in the world, for a wider discussion.

The first milestone

This joint research effort between Volkswagen and Google has produced promising results with the help of the Vertex AI platform. In this first milestone, the team was able to successfully bring recent AI research results a step closer to practical application for car design. This first iteration of the algorithm can produce a drag coefficient estimate with an average error of just 4%, within a second.

An average error of 4%, while not quite as accurate as a physical wind tunnel test, can be used to narrow a large selection of design candidates to a small shortlist. And given how quickly the estimates appear, we have made a substantial improvement on the existing methods that take days or weeks. With the algorithm that we have developed, designers can run more efficiency tests, submit more candidates, and iterate towards richer, more effective designs in just a small fraction of the time previously required.

Going forward, faster and more accurate estimates could even enable more automated searching for efficient designs, which would help both engineers and designers to hone in on the areas of the vehicle body where they could have the most impact. An important next step will be integrating the results into 3D design software to let designers benefit from the output and provide feedback.

As we continue, our focus is on improving the accuracy of the models. Firstly, we will build a larger, better quality dataset. Secondly, we will improve our shrink-wrapping algorithm to capture more details. Finally, we will enhance our existing models by experimenting with Vertex AI Neural Architecture Search to explore and experiment with different neural architecture options.

Moreover, we believe that our results for drag coefficient estimation is only a starting point for further exploration. There could potentially be numerous use cases in the space of physical simulations and assessments where cost and time savings could be achieved through ML-based estimators.

Acknowledgements

This work wouldn’t have been possible without the contributions from Volkswagen Data:Lab, Google Research, and Google Cloud. Thanks to Ahmed Ayyad, Dr. Andrii Kleshchonok, Dr. Daniel Weimer, Gülce Cesur, Henrik Bohlke, Andreas Müller from Volkswagen, Ameesh Makadia, Ph.D., and Carlos Esteves, Ph.D., from Google Research, and Daniel Holgate, Holger Speh, and Dr. Michael Menzel from Google Cloud.