IJSRD - International Journal for Scientific Research & Development| Vol.

8, Issue 2, 2020 | ISSN (online): 2321-0613

Optimal Stopping Algorithm for Modern Interview Process using

Machine Learning
Husain A. Shaikh1 Ankit A. Pandey2 Asim M. Raj3 Premraj V. Zambre4
1,2,3,4
Student 5Assistant Professor
1,2,3,4,5
Department of Computer Engineering
1,2,3,4,5
Keystone School of Engineering, Pune, India
Abstract— This Research Paper Aims to Provide the Even if we take n-1 interviews we cannot ignore
Balance of Time and Quality in modern interview process the possibility of nth candidate being better suited than
using Optimal Stopping Algorithms coupled with machine everyone else.
learning algorithm called as decision tree classifier. In Any The balance of time and quality is an important
optimal stopping problem, the crucial dilemma is not which problem that we face also in our daily lives. In the modern
option to pick, but how many options to even consider. We interview process it becomes necessary to handle this
try to answer this question by combining optimal stopping situation carefully as the candidate pool is quite large.
algorithm and machine learning with human intuition.
Keywords: Optimal stopping algorithm, Decision Tree II. LITERATURE SURVEY
Classifier, 37% rule, Interview Process This section of the literature survey eventually will reveal
some known facts based on the thoughtful analysis of the
I. INTRODUCTION author’s works as follows. While writing this research
In Any optimal stopping problem, the crucial dilemma is not paper, there are more than 170 papers available on Google
which option to pick, but how many options to even Scholar referring to the famous Optimal Stopping Algorithm
consider. These problems turn out to have implications not or Secretary Problem & its extensions.
only for lovers and renters, but also for drivers, Brian Christian [1] in his book “Algorithms to Live
homeowners, burglars, and beyond. By” says Optimal stopping is the science of serial
The 37% Rule derives from optimal stopping’s monogamy. Optimal stopping algorithm offers solutions not
most famous puzzle, which has come to be known as the only for the search for an apartment, but also for all those
“secretary problem.” life situations in which we are faced with the question of an
The establishment of the secretary problem is as optimal stay. People face these problems every day,
follows. Suppose you are a Hiring manager who has a task although the poets certainly spilled more ink on the court's
to hire a secretary for your boss. You take up the challenge difficulties than on the parking lot, and in some cases with
but face a problem later on with number of candidates. As considerable anguish. But anguish is not necessary. At least
candidates who are interested in the job role are quite large mathematically, these are solvable problems. The Optimal
in number it might take quite a lot of time to interview Stopping algorithm tells the balance is 37%.
everyone. According to the Researcher Zishuo Zhao [2], the
We might employ following two approaches: “37% balance” for Optimal Stopping Algorithm has largely
1) Interview everyone and select the best among them or influenced people’s opinion on decision making strategies
2) Select the first candidate who meets the requirements about making a choice. In his research paper he mentioned
The problem with first approach is it takes quite a as an online decision problem, Optimal Stopping Algorithm
lot of time to interview everyone. While your company requires a Decision Maker to go through a series of n
might stop at nothing to get only the best employees it might candidates and then decide whether to hire or immediately
be challenging for small companies or startups to spend that reject after observing one candidate, and the end goal is to
amount of time on hiring. maximize the chances of selecting the best candidate
The second approach then seems promising and (Ferguson 1989 [3]). In the original problem, the ideal
intuitive. It makes sense to select the one candidate who solution is to skip first n/e candidates and hire one
meets the requirement at the earliest. But just as first immediately if it is best. The maximized probability is
approach it has a fundamental flaw. The flaw can be simply 1/e. Meanwhile author also suggests, in general, it is
emphasized by following questions- often more desirable to choose a “good” applicant, not
1) How do we rank them to appear for the interview? precisely the “best”.
2) How can we be sure that this candidate is best among According to Ferguson [3] optimal stopping
all the candidates? algorithm or secretary problem has a certain appeal. It is
3) Can we take some more interviews to acquire a better easy to state and has a conspicuous solution. It was instantly
suited candidate? taken up and advanced by certain well-known statisticians
Above questions are harder to answer than we like Lindley, Dynkin, Chow, Moriguti, Robbins & Samuels,
think. As it is apparent that our first approach which is to and Gilbert & Mosteller. From that time, the Optimal
take every interview and then select the best performer takes Stopping Algorithm or Secretary problem has been
a lot of time, while the second approach which is to select developed and became widespread in many different areas
the first candidate who meets the requirements compromises so that now one can say that it establishes a field of study
with quality. within statistics-probability-optimization. One can read from
the review paper by Freeman(1983) how extensive and vast

All rights reserved by www.ijsrd.com 1139

 Optimal Stopping Algorithm for Modern Interview Process using Machine Learning
(IJSRD/Vol. 8/Issue 2/2020/263)

the field has become, furthermore, the field has sustained its efficient in terms of time and number of candidates.
exponential growth in the years since that paper appeared. Although we are assuming candidate pool is large, it might
As stated by Theodore Hill [4] in his research not be always the case. Therefore, decision tree classifier
paper, every decision humans make involves risk factors. caters our need when it comes to small training set. It’s
Selecting the best time to stop and hire a candidate is accuracy only improves as we add more candidates to the
critical. Bad timing can be critical. We face negligible pool. It takes attributes and their values while the interview
consequences while stopping decisions all time, when is in progress. As we’ve made interviewer to be the central
responding to a job offer, hunting for better housing prices. entity of decision, the interviewer will rate each candidate in
The basic framework of all these Optimal Stopping terms of their qualities based on organization’s
problems is same. A decision maker observes a process that requirements, some qualities/skills might be Speaking
evolves over time and is somewhat random. Based on what fluency, clarity of thoughts, effective communication,
is known, a person must take decision on the way of leadership traits, etc. Based on this 37% of the candidates
maximizing the reward or minimizing cost. will be interviewed and interviewer will provide his/her
decision to the algorithm as label for training. This training
III. PROBLEM set will be used to then predict whether the candidate should
How do we balance time and quality in the modern be selected or even interviewed based on available data prior
interview process? to and after the interview.
As stated earlier, we will be employing optimal C. Human Intuition
stopping algorithm more specifically the 37% rule along As stated earlier human decision maker is the central
with Machine learning and common intuition to solve this decision maker, we will not give full command to the
problem. machine learning algorithm.
Although the classifier will provide suggestion in
IV. SOLUTION binary (0 or 1) whether we should select or reject the
Many of the thinkers earlier have tried to automate the candidate, the final decision will be of the interviewer.
process of hiring. Although it seems like we are progressing As we are dealing with someone’s entire career, we
in terms of technology at a pace at which everything seems should deal this situation with utter care as to take unbiased
possible to automate the simple answer to automated hiring decision and record the parameters based on which the
is “We are not there yet.” decision was taken.
While taking an interview the interview panel looks
for their requirements, candidate’s experience, interests, V. IMPLEMENTATION
energy, confidence, ability to express ideas clearly, etc. Implementation will be using python programming language
apart from their aptitude scores. on Django framework for GUI because after years of
These analog entities cannot be translated to a experience with different web technologies, we believe
single value so that a bot can take an interview and select Django framework offers more features as compared to the
candidates who meet the requirements. other frameworks available out there. Machine Learning
Therefore, in our research we arrived at the Algorithm will be provided by the scikit-learn Library
conclusion that current interview technique is better - at which is commonly known as sklearn. A simple algorithm is
least for now. used to implement 37% rule.
In this paper we will optimize this current Storage of data will be done by using PostgreSQL
interview technique which involves taking interview face to for its simplicity and flexibility.
face by a panel and based on intuition select a candidate.
First, we will employ optimal stopping algorithms
VI. CONCLUSION
more specifically the 37% rule.
The Optimal Stopping Algorithm has received large
A. 37% Rule: attention by statisticians and applied mathematicians all over
This approach caters to those who don’t have much time the world. One reason for this is the problem’s amazing
despite that they would like to select the best candidate. 37% optimal policy. Under it, the Decision Maker skips the first
rule suggests that we look for at least 37% of the choices n/e candidates and then takes the next applicant whose value
and then think about hiring. Although the original rule is a maximum. What is surprising is that in n/e, the n͢͢͢͢͢ ∞,
doesn’t allow for going back and selecting a candidate who and that in the limit the best overall candidate is selected
was rejected earlier, but we don’t take this consideration with probability 1/e.
into our implementation. We began this paper by presenting Optimal
Stopping Algorithm for Modern Interview Process. The
B. Machine Learning:
problem introduced in the current paper has similar intuitive
Although we have implemented 37% rule to optimize the solution. The balance of time and quality is an important
process, we need a helping hand for the panel or a single problem that we face also in our daily lives. In the modern
interviewer that can suggest efficiently based on data that he interview process it becomes necessary to handle this
has whether he should select or reject this candidate. We situation carefully as the candidate pool can be quite large.
have experimented with different machine learning Future research will focus on the actual
classifiers and found decision tree classifier to be the most implementation of Optimal Stopping Algorithm for Modern

All rights reserved by www.ijsrd.com 1140

 Optimal Stopping Algorithm for Modern Interview Process using Machine Learning
(IJSRD/Vol. 8/Issue 2/2020/263)

Interview Process using Machine Learning in generalized

Optimal Stopping problems. Compared to Classical Optimal
Stopping Algorithm Problem, it seems to us that the solution
mentioned above has more scope.

ACKNOWLEDGEMENT
It gives us great pleasure to present this paper on Optimal
Stopping Algorithm for Modern Interview process Using
Machine Learning. We would like to thank our research
guide and Project coordinator Prof. Jubber S. Nadaf,
Assistant Professor, Department of Computer Engineering,
Keystone School of Engineering, Pune for their constant
encouragement & valuable guidance during the completion
of this research activity.
We are also grateful to Prof. Vilas G. Mankar,
Head of Department, Department of Computer Engineering,
Keystone School of Engineering, Pune for his constant
support and valuable suggestions.

REFERENCES
[1] Brian Christian and Tom Griffiths, “Algorithms to Live
By: The Computer Science of Human Decisions”, ISBN
9781250118363
[2] Zishuo Zhao, “An Asymptotic Analysis on Generalized
Secretary Problem”, Computer Science and Game
Theory (cs.GT), 16 Sep 2019
[3] Thomas S. Ferguson, “Who Solved the Secretary
Problem?”, Statistical Science, Vol. 4, No. 3 (Aug.,
1989), pp. 282-289
[4] Hill, Theodore P. “Knowing When to Stop: How to
Gamble If You Must—the Mathematics of Optimal
Stopping.” American Scientist, vol. 97, no. 2, 2009, pp.
126–133.

All rights reserved by www.ijsrd.com 1141