Problem

Recent events in the U.S have highlighted the need for an effective system that is able to identify police officers who are at risk of serious misconduct in order to provide them with the proper training to mitigate potential misconduct. According to the Chicago Reporter, the officer in Chicago who shot and killed 17-year-old Laquan McDonald in October 2014 had 19 citizen complaints and two misconduct lawsuits against him, yet the system in place had not flagged him as at risk of misconduct. Many current systems do not use computer programming to identify potentially problematic officers, and instead rely on a threshold system that uses only a limited number of metrics. This threshold system often places an unnecessarily high number of officers in an at-risk category, failing to identify officers that are truly most at-risk of committing adverse actions.

 

According to FiveThirtyEight, the number of prior complaints against police officers is strongly correlated with future misconduct. Police departments can combine this data with other police officer attributes to build a predictive model which identifies officers with a high-risk of misconduct prior to that misconduct actually taking place.

 

Solution

A team of researchers with the University of Chicago’s Center for Data Science and Public Policy developed an early intervention system (EIS) predictive model that can foresee adverse interactions between officers and civilians. In addition to flagging risk, the model also provides suggestions for preventative measures based on the individual’s risk profile.

In order to develop the predictive model, the researchers analyzed a comprehensive data set of interactions between cops and the public gathered by Charlotte police officials for more than ten years. The data included information on officer attributes, officer activities, and internal affairs investigations provided by the police department, as well as publically-available data such as weather and quality-of-life survey information. They tested the data using historical information to see which pieces of data would have predicted officers committing misconduct.

 

They found that cops with many instances of adverse interactions in one year were the most likely to have them in the next year, so they use this, combined with other indicators, to predict potential issues in the police force. Some of the identified additional indicators include excessive tardiness, misuse of medical leave, or a low grade on an annual performance review.

 

Results and Potential Improvements

Since police departments are not yet actively using this EIS, the researchers do not know if the predictions will successfully enable interventions that will reduce the likelihood of adverse interactions. However, three studies of police agencies that have implemented similar early warning systems have shown that targeted interventions can reduce citizen complaints against officers as much as 66 percent over two to three years. For example, if EIS identifies officers that are at high risk of committing misconduct, the police department can implement interventions that address the causes of the risk, such as training for working with youth.

 

The model developed by University of Chicago researchers correctly flagged 15 – 20 percent more officers who were involved later in adverse incidents than the current system in Charlotte, while also reducing the number officers incorrectly flagged by 50 percent or more. By improving the ability to correctly identify at-risk officers and the causes behind adverse incidents, this model will enable police departments to save time and resources while also improving their effectiveness in keeping their jurisdictions safe.

 

While other early intervention systems exist, the strength of this model is that it uses multiple indicators, rather than a single indicator threshold. By using multiple indicators, police departments will better be able to identify the root causes of police officer misconduct, which will allow them to implement more effective interventions to prevent this misconduct.

 

The success of the EIS algorithm is dependent on the quality and relevance of the inputted data. The Center for Data Science and Public Policy can improve the EIS by finding new sources of data that come from outside a police department. An organization called OpenOversight in Chicago has attempted to democratize police accountability by providing a digital gallery that allows the public to identify the name and badge number of a police officer they would like to file a complaint about. From March 2011 – March 2015, 28 percent of police complaints in Chicago were immediately dropped due to no officer identification. These are complaints that likely would not show up in the EIS so crowdsourced data from an organization like OpenOversight could help to fill this gap. Additionally, data crowdsourced from the public could serve as a way to check potential biases currently impacting a police department’s data.

By: Ex Machina Learners

Sources:

“About OpenOversight.” OpenOversight – a Lucy Parsons Labs Project. N.p., n.d. Web. 21 Apr. 2017.

 

Arthur, Rob. “We Now Have Algorithms To Predict Police Misconduct.” FiveThirtyEight. FiveThirtyEight, 10 Mar. 2016. Web. 21 Apr. 2017.

 

Gregory, Ted. “U. of C. Researchers Use Data to Predict Police Misconduct.”Chicagotribune.com. N.p., 18 Aug. 2016. Web. 21 Apr. 2017.

 

Joseph, George. “Crowdsourcing Police Accountability.” CityLab. N.p., 25 Oct. 2016. Web. 21 Apr. 2017.

 

Mitchum, Rob, Jacqueline Genova, and Lin Taylor. “Police Project Update: Expanding and Implementing the Early Intervention System.” Data Science for Social Good. N.p., 12 Jan. 2017. Web. 21 Apr. 2017

 

Newman, Jonah. “Program That Flags Chicago Cops at Risk of Misconduct Misses Most Officers.” Chicago Reporter. N.p., 04 Mar. 2016. Web. 21 Apr. 2017.

 

Smith, Megan. “Launching the Police Data Initiative.” National Archives and Records Administration. National Archives and Records Administration, 18 May 2015. Web. 21 Apr. 2017.

Money laundering destroys value. It facilitates economic crime and nefarious activities such as corruption, terrorism, tax evasion, and drug and human trafficking, by holding or transferring the funds necessary to commit these crimes. It can be detrimental to an organisation’s reputation – and its bottom line. – PWC Global Economic Crime Survey, 2016

As a percentage of global GDP, money laundering transactions are estimated to be 2 to 5%, of which less than 1% of violations are detected, according to the PWC study mentioned above.  This is a staggering figure, not just because of the numbers, but because of the activities propagated by these illicit transactions – terrorism and human trafficking among the most loathsome.

Governments around the world, aware of and concerned by this problem, have taken action in many ways, the most direct of which likely involve projects and missions hidden from the public view.  More visibly however, governments have also put the onus on the financial institutions that unknowingly (one would hope) facilitate this illicit activity.  Strict regulations are in place for both customer onboarding (know your customer or “KYC” practices) and transaction monitoring (anti-money laundering or “AML” practices).  One example of a recent penalty is the $16.5M fine levied on Credit Suisse in 2016 for deficiencies in its practices.  As described on amlabc.com:

FINRA found that Credit Suisse’s suspicious activity monitoring program was deficient in two respects. First, Credit Suisse primarily relied on its registered representatives to identify and escalate potentially suspicious trading, including in microcap stock transactions. In practice, however, high-risk activity was not always escalated and investigated, as required. Second, the firm’s automated surveillance system to monitor for potentially suspicious money movements was not properly implemented. A significant portion of the data feeds into the system were missing information or had other issues that compromised the system’s effectiveness. The firm also chose not to utilize certain available scenarios designed to identify common suspicious patterns and activities, and it failed to adequately investigate activity identified by the scenarios that the firm did utilize.

While governments may be focused on “catching the bad guys”, banks are mainly incentivized to comply with regulations, not necessarily to minimize illegal activity (taking a strict Lockean view of self- interest) – the financial penalty (as well as a reputational one) is designed for banks to comply with a set of dynamic (e.g. ever evolving) rules set forth by governments.  These kind of measures ensure that banks will remain focused on compliance – and to do this properly, banks have and will continue to spend massive amounts of capital on avoiding violations.

Each bank takes separate steps to comply with these complex regulations, but what is consistent across firms is that the processes are incredibly manual.  While many solutions are used to flag bad actors in both the KYC and AML spaces, the workflow is fragmented and antiquated.  Hundreds of analysts spending thousands of hours downloading and parsing through information manually in order to comply with these regulations is a byproduct of banks trying to keep up with regulations by building on legacy systems.  There are great products out there for specific tasks (e.g. voice recognition and ID fraud detection services), but the overall workflow is broken.

The solution to this is a “human-in-the-loop” workflow platform.  A service that can integrate with current systems (and/or replace other systems) already in place at banks (this optionality is sure to decrease sales cycles) that can use machine learning models to decrease the costs of compliance and to increase the performance of compliance practices.

Based on conversations with industry experts, it is believed that at least 70% (a conservative estimate) of cases are clear “flag” or “no flag” situations – yet analysts must still parse through each and every transaction in the current workflows.  By automating these easy-to-solve cases, banks can employ fewer higher performing analysts whose performance should be improved by focusing only on the “interesting” cases, which comprise a minority of the overall transaction base.

Models can be trained on historical data using machine learning models, and then continue to adapt as new information (the manual inputs from the 30% of cases that analysts assess), providing an ever-improving solution.  While data security will clearly be an issue, and banks may opt for on-premise solutions, the more data that can be anonymized and aggregated across banks, the better the models will perform and the more effective the solution will be.

Humans may never be completely eliminated from the process, not only because regulations are unlikely to allow it, but also because banks would certainly be hesitant to completely leave their compliance fate to a series of machine learning algorithms, but the massive cost savings and performance improvement provided by such a solution would be a huge boon banks’ compliance practices.  Beyond this, one could imagine similar systems being equally useful to governments, whose goals are more focused on catching the bad guys – though that is a use case beyond the scope of this post!

Sources:

http://www.pwc.com/gx/en/services/advisory/forensics/economic-crime-survey/anti-money-laundering.html

https://www.protiviti.com/sites/default/files/guide-to-us-aml-requirements-6thedition-protiviti_sec_0.pdf

http://www.geracilawfirm.com/Events-Insights-News/2015/October/Protect-Your-Business-Potential-Penalties-for-Fa.aspx

https://amlabc.com/aml-timelines-on-historical-scale/

http://amlabc.com/aml-category/aml-sanctions-fines/finra-fines-credit-suisse-securities-usa-llc-16-5-million-for-significant-deficiencies-in-its-anti-money-laundering-program/

The Problem: Science Research and Scientific Journals Are Inaccessible

Science Ease is a crowdsourced platform that aims to translate and aggregate scientific research, making it more accessible and more likely to be implemented by practitioner and policymaker alike.

If innovation is the engine that drives economic growth, basic science is the fuel.  In the U.S. alone, government agencies annually spend ~$130B on science funding. However, despite its centrality to the economy and the huge resources invested in it, research too often sits on the shelves for reasons of accessibility. The typical layman with limited knowledge of research design and jargon gains little from reading them. Unfortunately, only weak and informal mechanisms exist to turn new knowledge into practical gain.

The problem is especially acute where the public value of the research far exceeds the private, appropriable value.  To illustrate: suppose a Booth professor came up with a new auction design that, if implemented by Amazon, would be worth $400M.  There is little doubt that somehow, whether via the professor’s private consulting firm or media reports, Amazon would manage to find the idea and implement it.  In contrast, take an idea like Booth professor Eric Budish’s discrete-time fix for high frequency trading. It has high social returns—ones that are highly distributed—and only a few firms (those specializing in HFT) suffer from it.  Scientific findings like these, where the benefits are mostly to those who would not read the research, can be lost in translation.

The Solution: A Crowdsourced Platform for Translating and Sharing Research

Science Ease is designed to alleviate these issues, through the power of the crowd. Science Ease will serve as a centralized resource through which scientists, innovators, and everyday people will interact to translate state-of-the-art scientific findings into language that everyone can understand. It will operate on a non-profit model, funding itself primarily through donations and running at a low cost. Instead of an economic profit, the site will aim to raise awareness about cutting-edge scientific findings for average, everyday people.

The platform operates on a Wiki 2.0 model, using editors at the top of the stack, responsible for pushing publishable content to the public site, and contributors below, constantly working on the backend to refine the content in an iterative fashion.  This Editor/Contributor structure is critical in ensuring that the quality is implementation-ready and to avoid things like political fights over the minimum wage (as would happen in a Wiki 1.0 model).  Moreover, unlike traditional handbooks or meta-research articles, using the crowd allows the tone to step away from the jargon that dominates academia while also ensuring that the content can be updated real-time, rather than on a publisher’s schedule.

To attract contributors, the site will begin by partnering with scientists who are looking to increase the profile of their work (particularly research whose findings would benefit the general public). We would then work with their teams (and potentially hire some initial contributors) to translate that research into easy-to-understand, interesting web pages. Once enough publically beneficial research is made accessible, we would focus on attracting users through press outreach, partnerships with policy-makers, and potentially advertising. Once enough readers and contributors joined the site, a critical threshold would be reached where the most excited readers will become contributors, and the quality would improve, attracting additional readers.

Demonstration: Turn a Publicly Beneficial Discovery into Action

There are two ways we can demonstrate the product. To convince investors or critical decision makers, we would do a simple experiment: show them sections of text from research articles as well as some from Science Ease (without identifying which is which). We would then ask them to evaluate how well they understand the texts and how important they are. These results would show just how powerful Science Ease can be, both from the perspective of making things accessible and showing how research can be important for everyone.

In order to test the broad efficacy of this platform, we will experiment first with just a single discrete idea – Budish’s high-frequency trading fix, mentioned above.  The solution, to be undertaken by a trading exchange like NASDAQ or BATS, is already outlined in Budish’s work.  Therefore, this experiment will test our central theory that the main barriers to the idea taking hold in reality are the barriers to understanding it, and by extension the current inability of anyone but Budish himself to advocate for the idea.  Over a six-month period, with Eric and two of his colleagues acting as Editors, we will invite Contributors from around the world to contribute to a centralized article summarizing Eric’s research, and outlining the specific policy and practice steps necessary to implement it.  Should our theories hold, the outcome to be measured is whether indeed any trading exchanges implement the idea.

 

 

In today’s corporate America, lawyers are a necessity. Corporations spend billions of dollars a year on legal expenses. In 2013, Bank of America alone spent $6B on legal fees.

Yet advances in Artificial Intelligence (“AI”) and machine learning in recent years are beginning to disrupt this huge market. One example is Ross, a robotic attorney powered by IBM’s Watson artificial intelligence to perform legal research. Ross was “hired” by 10 companies in the past year to help prepare legal cases. It understands questions in plain English and provides specific analytic answers. Another example of these early robot-lawyers is the DoNotPay chatbot. Created by a 19-year-old and currently operating in the UK and US, DoNotPay helped overturn $4m in ticket fines since 2014 by appealing parking tickets with an AI-powered chat.

These examples demonstrate what is becoming possible in the legal field, and how technology is contributing to make a dramatic change in the law profession. With computer’s abilities to replace humans by performing their tasks better, faster and more efficiently, robot lawyers are soon to displace human lawyers. We are predicting a lawyer-less future for humanity.

We are happy to introduce Grisham™, the Robo-Lawyer of tomorrow. Grisham™ is a cloud-based robot that learns from millions of historical rulings, precedents, natural language processing of court decisions, and the legal code itself. This AI will be enhanced through an organization-specific learning that will train Grisham™ on the organization interests, financial limitation, and more.

Grisham™’s main functions will range from (1) “simple” legal tasks, such as research, and getting relevant paperwork based on inputs to support legal support, through (2) answering legal questions to clients and new associates through a robo-chat that will predict the questions and generate answers. This capability will be similar to Amazon Connect that was launched in March 2017 (see reference below). Moreover, (3) Grisham™ will function as a fully autonomous corporate attorney by learning the company’s interests and negotiating on its behalf with other human/robot lawyers. Finally, (4) Grisham will prepare lawyers to trail by playing the opposing lawyer and finding the pitfalls in the case. This function will allow the human lawyer to prepare better to trail and will in parallel train Grisham™ to become a fully functional lawyer.

The effectiveness of Grisham™ is huge. In 2016 the demand for corporate and real estate lawyers grow by 2.5% and 4% respectively, while lawyers declined in their productivity (also affected by overall demand growth and lawyer growth). Looking at the largest law firms in the US by revenues we can also evaluate the potential market size Grisham™ is facing. Latham & Watkins revenues in 2016 were $2.65B with profit margins of 50%. Grisham™ be able to provide at least 50% of tical law firm services for a fraction of a price. This signals on a high commercial promise. The Report on the state of the legal market 2016 also indicates that the hourly rate for lawyer has grown from ~$350 in 2005 to ~$500 in 2015, a 33% increase. This means that Grisham™ will also help to reduce costs in the legal system. Nevertheless, the anticipated competition is high. As of today Ross by IBM already serves a significant number of customers.

Yet the functions described above are only the beginning. In the future, Grisham™ will be able to entirely replace human lawyers. With Grisham’s advanced legal capabilities, human lawyers will no longer be needed. Every citizen will have access to Grisham™ and it will represent every person in court. Courts will no longer be run by human judges. Powerful government-held computers will act as intermediaries, assessing the evidence and arguments from both sides’ lawyers to make a ruling in the case.

One interesting aspect of this system is its positive social implications. Grisham™ would not only make good legal representation affordable and obtainable by all citizens, it would also make the entire system fair and unbiased.

The legal system of the future is just around the corner. Grisham is coming.

 

References:

• The Atlantic – Rise of the Robolawyers
• The Atlantic – Here’s How Much America’s Biggest Banks Spent on Legal Bills This Quarter
• List of largest law firms by revenue

• ROBOT LAWYER OVERTURNS $4 MILLION IN PARKING TICKETS
• New York Times – A.I. Is Doing Legal Work. But It Won’t Replace Lawyers, Yet.
• Report on the state of the legal market 2016

Problem

With population growing exponentially around the world, the FAO estimates that world food production would need to rise by 70% by 2050. In particular, developing countries would have to increase output by 2X. In contrast, agricultural production in developing countries is lagging far behind the rest of the world. An estimate of corn production shows that yield in the United States are 5X of those in Africa.

A key factor driving this disparity is the low levels of mechanization in emerging economies across the agricultural value chain. In particular, farming practices tend to be reliant on unpredictable traditions (what does this mean? if it is what I think it is, maybe unreliable is a better word?). While high tech and precision farming are growing in the United States, where effective soil maps and understanding of environmental factors is enabling higher yield through lower resources, many resource-poor countries are limited by inconsistent rainfall, poor input quality, and historically ineffective farming practices.

A lack of basic understanding of the soil being cultivated on is a key deterrent. With most agricultural work in developing economies being done by smallholder farmers with less than 1-2 hectares, soil quality can vary significantly from one farm to another. However, practices being adopted tend to be standard across communities.

To be able to ensure that food production rises to meet the increasing demand from growing population, as well as ensure yield parity across the globe – technology must be harnessed.

Solution

SoilMapr will collect multidimensional data about soil through a low-cost device with sensors that can measure soil quality and nutrient mix. Small sensor devices will be placed in farmland, and through a combination of optical, electrochemical, and mechanical sensors will be able to create a soil profile.

The use of SoilMapr would include the following steps:

Before cultivation:

o   Place the SoilMapr sensor in the ground for 30 minutes

o   SoilMapr will give a complete read of the nutritional qualities of the soil

o   Input the crop to be cultivated

o   SoilMapr will give a recommendation of nutrients that require additions, and specific fertilizers that could work

o   SoilMapr would also predict expected yield based on current conditions, versus improved conditions – allowing users to make a judgment on need for investment.

 

 

o   Place the SoilMapr sensor in the ground for 30 minutes to effectively monitor soil quality throughout

o   SoilMapr would give a read of the nutrient quality, and also expected yield at current levels

SoilMapr has three sensors:

•        Optical sensors: Measure infrared levels, organic matter, and moisture content
•        Electrochemical sensors: Measure ions, pH levels and nutrient mix
•        Mechanical sensors: Measure soil resistance and compaction

Data collected from these sensors is modelled against datasets collected over time of ideal crop production, optimum soil mix, expected yields, harvest time, etc. Analysis will focus on leveraging robust data from evolved agrarian systems in developed markets to developing countries where this data is inaccessible. For example, historical corn production in the USA with irrigation of X liters per day grown on soil with specific characteristics, can serve as a reference and benchmark for what Ethiopian production could be under similar or slightly different conditions. These ballpark figures would create predictability, and over time feed back into the dataset as reference for other production centers – establishing average performance metrics, as well as best-in-class.

 

Demonstration

With large cycle times in agriculture, SoilMapr may be able to effectively display the ability to measure soil quality in real time, but would need a long window to validate its predictive qualities. The ideal demonstration would involve working on two identical plots of land, one where conventional farming methods are used, and another where SoilMapr is used. This would allow results to be compared easily to identify the value SoilMapr brings.

The farmer could then leverage SoilMapr’s interface to understand (a) what nutrients need to be added to the soil to be effective for their intended crop, (b) predict expected yield based on those specific environmental factors.

Screenshot of SoilMapr’s potential interface >>>>

 

Source

http://www.thisisafricaonline.com/News/Closing-Africa-s-agricultural-yield-gap?ct=true

https://www.populationinstitute.org/resources/populationonline/issue/1/8/

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2873448

http://cropwatch.unl.edu/ssm/sensing

Online shoppers are usually influenced by customer reviews posted when researching products and services. In a 2011 Harvard Business School study, a researcher found that restaurants that increased their ranking on Yelp by one star raised their revenues by 5 to 9 percent. Reviews can be useful especially when it comes to tourist destinations and “experience” products that you really need to try out. But with studies suggesting that 30% of online product reviews and 10-20% of hotel and restaurant reviews are fake, how do you know which reviews to believe?

Customers are in danger of being misled by millions of “fake” reviews orchestrated by companies to trick potential customers. But even experts are having a hard time identifying deceptive reviews. Websites such as Yelp and TripAdvisor are search engines for a specific category that rely heavily on ratings and reviews. For example, TripAdvisor hosts hundreds of millions of reviews written by and for vacationers. This site is free to use with revenues coming advertising, paid-for links, payments or commissions from the companies. Some businesses try to “buy their way in” to the search results, not by buying advertising slots but by faking online reviews. However, with traditional advertising, you can tell it’s a paid advertisement. But with TripAdvisor, you assume you’re reading authentic consumer opinions, making this practice even more deceiving.  

The research work conducted have used three different approaches including part of speech tags (POS), linguistic inquiry and word count (LIWC), and text categorization. The researchers, including a team at Cornell University, have developed sophisticated automated methods to detect the fake reviews. On the left, is an example of how a fake review is identified using strong deceptive indicators that obtained from the above mentioned theoretical approaches. Features from these three different approaches are used to train Naive Bayes and Support Vector Machine classifiers. Integrating work from psychology and computational linguistics, the solution develop and compare three approaches to detecting deceptive opinion spam, and ultimately develop a classifier that is 90%+ accurate. While previous conducted research work has focused primarily on manually identifiable instances of opinion spam, the latest solutions have the ability to identify fictitious opinions that have been deliberately written to sound authentic. While, the solution is quite robust, we believe that there is are possible areas of future work.

The best performing algorithm that the Cornell research team developed was 89.8% accurate (calculated based on the aggregate true positive, false positive and false negative rates).  This is in contrast to 61.9% accuracy from the best performing human judge.  These results suggests that TripAdvisor could realize substantial improvements if they implement this algorithm in weeding out false reviews.  However, it is important to note that the effectiveness shown in this study may not be replicated in an actual commercial setting.  As soon as the algorithm is put into production, it is likely that spammers will start to reverse-engineer the rules of the algorithm, creating ever more realistic fake reviews.  Additional study would be needed to determine if the algorithms are able to keep up.

TripAdvisor advertises “Zero Tolerance for Fake Reviews.”  They currently use a team of moderators who examine reviews; this team is added by automatic algorithms.  Though they do not publicly discuss the algorithms used, they state that their team dedicates “thousands of hours a year” to moderation.  To the extent that improved algorithmic tools are both more accurate and less costly than human moderation, the commercial upside for Trip Advisor is substantial.

While using text is a good starting point, other metadata could help flag fake reviews. For example, how many reviews the user has, and the authenticity of those other reviews. Filtering like this faces the problem of fakers adjusting their writing based on what gets screened in and out. If certain words appear more genuine, fakers may adopt those words. Companies screening need to keep their identification methods secret. Making the model adaptive to identify the changing behavior of fakers is more challenging than this static model, as the model needs to identify what fraudsters will do vs. they do today.

Sites like TripAdvisor may not have clear incentives to remove fake reviews, as even fake reviews provide more content to show to visitors. Misclassifying genuine reviews as fake would lead to complaints from users and businesses. Users and businesses may have an expectation reviews should be truthful, and may stop using a site if they rely on too many reviews they find to be fake, but how strict a threshold companies should apply when screening.

By: Team Codebusters

Sources:

http://www.hbs.edu/faculty/Publication%20Files/12-016_a7e4a5a2-03f9-490d-b093-8f951238dba2.pdf

https://sha.cornell.edu/centers-institutes/chr/research-publications/documents/anderson-social-media.pdf

http://www.nytimes.com/2011/08/20/technology/finding-fake-reviews-online.html

http://aclweb.org/anthology/P/P11/P11-1032.pdf

https://www.tripadvisor.com/vpages/review_mod_fraud_detect.html

http://www.eater.com/2013/9/26/6364287/16-of-yelp-restaurant-reviews-are-fake-study-says

The Problem and the Opportunity

The current college admissions process is flawed. Despite the best efforts of admissions officers, there is little proof to suggest that current application review processes are robust and result in ideal class compositions. A 1996 Northwestern University study found that the likeliest determinants of admission were the standard test scores, grades, and information found in one’s application – not the unique characteristics admissions officers claim to look for in essays and interviews as more of a complete evaluation of the candidate. Perhaps even more worrisome is that admissions offices, outside of those at some of the top universities, do not evaluate the outcomes of application decisions – such as the success of an admitted student or the future financial benefits to the institution of that student’s acceptance – and incorporate that feedback into their admission evaluation criteria.

In addition, the application review process at most universities is still largely manual, requiring admissions officers to read and evaluate tens of thousands of applications multiple times each year. Significant time is spent on even the most clear-cut of cases . In a University of Texas study of how a machine learning solution could aid PhD application review processes, it was observed that reviewers spent over 700 hours reviewing a pool of 200 applicants. Now, consider that many of the top universities receive tens of thousands of applicants.

We see a clear opportunity for a machine learning-based solution to address the existing flaws in the college admissions process. Not only is there a large capturable market with over 3,000 institutions of higher education in the U.S. that all face this admissions evaluation issue, the number of college applications continues to increase, which will only exacerbate the issues described above.

 Our platform, HiPo!, will help provide significant time and human resource savings to university admissions offices. In addition, it will be trained using historical application and student performance data to help admission officers optimize their admissions evaluation process across a number of outcome-based factors (e.g., future earnings potential, yield likelihood, philanthropic giving potential).   

The Solution

The proprietary algorithm will utilize a semi-supervised machine learning model. The supervised elements that the model will optimize for are the quantifiable outcomes such as future earnings potential, yield likelihood, and philanthropic giving potential. However, given the vast expanse of data that the model will be trained on through years of qualitative data from student essays and interview transcripts, there are other elements that in an unsupervised way, the algorithm can make associations and clusters from to derive additional predictive value. These elements are not as measurable such as creativity or diversity of thought – both things that an admissions committee would value in a class. However, over time if the algorithm can add additional measurable information to admissions officers on these dimensions, they would provide additional evaluative data.

The inputs into the core product are traditional quantitative metrics, such as GPA, standardized testing scores, etc., in addition to qualitative inputs such as essays, interview transcripts, recommendations, and resumes. By creating a robust feedback loop by measuring the success of various students over time based on the HiPo! evaluation criteria, the algorithm will be able to estimate outcomes and provide admissions officers with quantifiable score reports:

This is merely a prediction of the likelihood of future outcomes based on historical results of similar profiled candidates, not a pure measure of an individual’s current attributes.

Empirical Demonstration

To validate the effectiveness of the machine learning algorithm and to validate the hypothesis that certain characteristics and patterns present in an candidate’s application, essays, and interviews are reflective of future outcomes, the HiPo! team would perform the following demonstration pilot. Partnering with an institution of higher education, such as the University of Chicago Booth School of Business, HiPo! would collect historical applicant records from 1950-1992. Half of this data would be randomly selected to train the algorithm, under the supervised and unsupervised learning methods described above. The algorithm would then be applied to the remaining half of the data. The predictive output of the solution would be measured against actual outcomes of the students evaluated in the sample. For instance, if Michael Polsky MBA ‘87 was evaluated as part of the sample, a successful algorithm would predict that he would have both high career earnings potential and strong likelihood of philanthropic behavior. It is critical that the data used for the demonstrated be at least 20 years old, so that outcomes such as future career success and philanthropic giving can be accurately assessed.

 
Sources

Cole, Jonathan R. “Why Elite-College Admissions Need an Overhaul.” The Atlantic. Atlantic Media Company, 14 Feb. 2016. Web. 13 Apr. 2017.

Fast Facts. N.p., n.d. Web. 13 Apr. 2017.

Miikkulainen, Risto, Waters, Austin. “GRADE: Machine Learning Support for Graduate Admissions.” Proceedings of the 25th Conference on Innovative Applications of Artificial Intelligence, 2013.