Introduced by Vikram Singh, Dong-Eun Suh, Karl Walter, Dian Yu and Francis Yang
NBA Make-A-Team, a new project developed by UC Berkeley students, gives NBA coaches, general managers and everyday fans the opportunity to select hypothetical starting lineups and get accurate feedback on how that basketball team would perform.
This simple-to-use tool can project a team’s winning percentage and predict its statistical strengths and weaknesses. It can also suggest a player swap that could help a user build a stronger team.
The tool empowers the user to assemble their dream team – choosing from players that were part of the NBA between 2000 to 2017 – and compare that team against their friends’ squads-of-choice. Additionally, for those who hope to make their dream team a realistic team, there is a salary cap option that imposes a league’s spending limits on chosen players.
This project was developed by a team of students from Data-X, a project-oriented data and machine learning course at UC Berkeley. The largest hurdles the team faced were cleaning and processing the data and developing the models used to estimate a team’s winning percentage and to suggest player swaps.
This project is tremendously exciting because it provides instantaneous feedback on both real and hypothetical rosters and gamifies roster selection by highlighting team weaknesses and strengths.
See how it works here.
Introduced by Deep Dave, David Lin, Sharon Ng, Vanessa Salas and Alexandre Vincent
As you can imagine, there is a wealth of data on most topics in the Internet Age. Data on stock prices, the housing market and energy consumption are just a few of the areas with a large amount of data available to the public. But how can you use this data to make informed decisions in your own life?
Four students from Data-X, a data and machine learning course at UC Berkeley, set out to answer this question and make data-driven decision a little bit easier for travelers.
Originally, the students planned to work with available data sets to answer the questions “When is the best time to buy plane tickets?”. They found, however, that data in relation to this specific inquiry was limited. After weeks of searching for right data set, they found that data sets from Kaggle, a website dedicated to providing open data, only included flight times and Department of Transportation data sets only included yearly flight prices.
So the students decided to alter their project to make a data set themselves. They scraped flight data on a daily basis in order to create a workable database that can help others pinpoint when the right time is to book their flight.
Introduced by Brandon Chou, Djavan De Clercq, Andrew Gonzalez, Charles Li and Rinitha Reddy
Urbanization, one of the prevailing trends of the 21st century, places great stress on the water resources of cities across the globe. This stress is particularly pronounced in China, a country that has seen rapid urban growth in the past few decades.
This problem prompted a team of UC Berkeley students to begin a statistical learning project that could aid researchers by providing them with greater insight into urban water supply patterns.
They applied statistical learning methods to 12 years of urban water supply data for 627 cities across China in order to identify the factors most responsible for variance in patterns of urban water distribution and management. For instance, they found that Chinese cities have consistently suffered water loss and leakage rates above 20 percent since 2001, and water prices are closely associated with leakage.
Additionally, developed an urban water sustainability index in order to compare cities. From there they were able to identify the cities and regions in China that face sustainability issues.
Aside from their research, they also provided a general, systems-level perspective of major urban drinking water use trends in China for the benefit of public-sector stakeholders.
Introduced by Souhail Bentaleb, Guillaume Drugeot, Luna Izpisua Rodriguez, Farbod Nowzad and Ajay Shah
Using data science and machine learning, a team of UC Berkeley students have created a program that can tell an accurate and informative story about a person’s life — all from the photos that they post on social media.
Typically, advertisers target consumers by analyzing a person’s most frequented online sites or “clicks” to collect basic demographic information, such as age, gender and location. But they’re overlooking an entire other aspect of people’s online lives, which could be a rich source consumer information: personal photos.
The 1.8 billion images that are being posted to Facebook, Instagram, Flickr, Snapchat and WhatsApp every day hold valuable information about a person’s lifestyle, daily activities and consumer behaviors .
In order to address this problem, the team of students created a machine learning algorithm that tags photos from social media profiles to reveal the most prominent keywords that define each image. From there, the algorithm uses this information to classify each consumer profile into one of four categories: outdoorsy, sporty, family or foodie.
This classification can provide advertisers with personalized insights about a consumer’s preferences.
The general public can use the tool as well in order to gain insight into the image that they project online.
Introduced by Olabode Faleye, Kazuomori Lewis, Vicente Izquierdo and Pedro Pablo Correa
Bitcoin is one of the newest financial instruments to take the world by storm. This decentralized digital currency has been experiencing unprecedented levels of growth in recent months, proving itself as one of the most lucrative investment opportunities of our generation. As of Jan. 16, it was up more than 1000 percent.
Its extreme volatility, however, makes it difficult for even seasoned investors to predict market trends and maximize their return on investment. As a result, a team of UC Berkeley students from Data-X aimed to develop strategies to predict market trends to inform even a novice investor when it is best to buy, sell or hold Bitcoin with their new tool, CryptoTitans.
To develop informed trading strategies, the team used Bayesian Regression to predict prices on the minute scale, Recurrent Neural Networks for prediction on a day scale and the more traditional Bollinger Bands for trading strategies in accordance with long-term market trends. With these different approaches, the tool could help investors with all levels of experience to maximize the profits they make from Bitcoin.
Introduced by Claudia Iriondo, Devina Jain, Raouf Muhamedrahimov, Vasileious Papanikolau, Kosta Trotskovsky and Leo Sun
More than 25 percent of adults in the United States suffer from knee pain, creating an urgent need for an efficient way to match patients who have cartilage and bone lesions with effective treatments. Patients undergo Magnetic Resonance Imaging (MRI) scans across the country, but they require a large amount of time and access to a radiology expert.
But Cartilage-X offers a new way of assessing knee injuries. The new tool uses AI-guided lesion detection to predict the location and severity of cartilage and bone lesions, reducing the amount of time and expertise needed to assess these injuries.
The team that created this tool consists of graduate and undergraduate students from UC Berkeley. With the help of Python, Matlab and Tensorflow, the team used a data set of more than 1,800 patients’ knee MRIs and integrated with existing UCSF research to create an automated pipeline for MRI analysis. The pipeline ingests MRI images stored in its database, segments cartilage compartments, automatically preprocesses to include only signal of interest and predicts the presence of a cartilage or bone lesion using neural networks.
The team’s future work will focus on improving model performance, introducing an image quality control step before training and testing model generalizability on a fresh data set from a different cohort.
By providing insight into lesion location and severity, and doing so in under a minute, Cartilage-X brings quantitative metrics for knee MRIs into the clinic.
Introduced by Nicholas Hirons, Julian Kudszus, Soham Kudtarkar and Spencer Lee
Mechanics spend a lot of time diagnosing car issues and performing tests, but oftentimes customers aren’t sure if their cars are getting the service they need.
But a new web app created by UC Berkeley students could change this. The algorithm takes standard information about the type of car a user owns and takes error codes from car telematics data, or wireless information transmitted from your car. From this information, the algorithm is able to predict what issues need to addressed and what services need to be performed.
The algorithm takes into account a variety of vehicle data, such as make, mileage and engine control units to give the most precise prediction possible.
Because it takes information from vehicle telematics, the tool could mean that dealerships and auto repair shops could preemptively reach out to customers to let them know they need service, to stop a potential breakdown on the road.
A feature of the app that suggests likely causes and solutions for car issues also means that customers could avoid paying for services they don’t need once they arrive at an auto repair shop.
While initial results are mixed, the team was able to predict when certain service operations were required fairly well. For example, the model accurately predicted the need for tire inflation about 70 percent of the time. With more data, the team believes that they could significantly improve prediction accuracy by building models specific to each car brand.
The students hope their project can help members of the vehicle service industry understand the benefit of using connected car telematics to improve efficiency and customer service.
Introduced by Jordan Cox, Rob Holbrook, Ramon Lim and Stephanie Zhu
Oftentimes, even just preparing data for analysis can take significant time, effort and knowledge. But now, a group of UC Berkeley students are introducing a solution to make data analysis a little bit easier: The Match ‘n’ Merge Application.
This unique application uses predictive models to match columns across your separate data sets and applies string pattern recognition algorithms to match the rows, ultimately resulting in a consistent, merged data set.
Version 1.0 allows merging of two CSV data sets and focuses on a specific set of alpha-numeric features.
Future releases will allow you to merge multiple data sets of varying forms. They also will automatically generate relevant features for stronger predictive modeling and will utilize probabilistic joins to fill in missing data.
Introduced by Jillian Chan, Yuyang Pan, Simarjeev Singh and Yin Zhang
The Anxiety and Depression Association of America estimates that 16.1 million people in the United States are impacted by major depression. Yet, depression is severely underdiagnosed and only a fraction of the people affected by it will ever receive treatment.
But now, a team of UC Berkeley students are using machine learning to to help predict a person’s likelihood of experiencing depression, making diagnosis easier. Using 12 years of data from national health organization Geisinger, the students created a model that uses personal information about a user to predict the likelihood that they could experience depression.
In particular, the students’ project analyzes the correlation between anxiety, insomnia and depression while also taking into account demographic factors, such as age, gender and ethnicity.
Through their investigation they verified that the relationships among depression, insomnia and anxiety are multi-directional, meaning each of these condition could trigger or worsen one of the other two conditions. Furthermore, the students confirmed previous research that concluded that females, anxiety patients and insomnia patients are at a greater risk of developing depression.
The project gives a general sense of which populations are more likely to be diagnosed as depressed and will eventually help doctors identify a patient’s likelihood of experiencing depression.
The team is now working to increase the accuracy of the classical models used for the project, which is currently around 70 percent. In order to improve accuracy, the students hope to integrate patient histories of drug use, alcohol consumption and smoking into the factors taken into account by their model.
Introduced by Bilal Halabi, Seung Woo Son, Ki Hyun Won and Avery Yip
At times, living in Berkeley can feel unsafe. Shootings, violent demonstrations and robberies are just a few among many incidents that have threatened Berkeley students living on campus this year.
To address this problem, a team of four UC Berkeley students who have experienced these threats first-hand set out create an app to help make residents in the area safer. The result is the web-app platform, Stayfe, which allows users to visualize and track in real-time where crime is occurring in their neighborhood and beyond.
Currently, a medium of this type does not exist in Berkeley. Alerts from Berkeley Police Department or the University of California Police Department can sometimes come hours or days late. But Stayfe uses a customized crawler, parser and crime type classifier to search through news articles in the Bay Area to get the most up to date information on local crime.
Additionally, the students created an algorithm for the app that can use this crime data to suggest the safest path for you take to your destination if you’re walking. This feature gives users an alternative to Google Maps, which may lead pedestrians on a quicker, but more dangerous path.
Though the creation of the app was challenging, the students say that their passion and belief in their mission helped them to create a successful product and led them to be selected as finalists at the 2017 SCET Collider Cup.
Moving forward, the team plans to deploy their web app as a mobile app, so people can use Stayfe’s safety path suggestion feature on the go.