Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

1st IEEE International conference on Innovations in High-Speed Communication and Signal Processing (IEEE-IHCSP) 4-5 March, 2023

Smart Model For Career Guidance Using Hybrid


Deep Learning Technique
1st Vijaya R. Kumbhar 2nd Manisha M. Maddel 3rd Yudhishthir Raut
Computer Science Dept. Operations Dept., Computer Science Dept.
2023 1st International Conference on Innovations in High Speed Communication and Signal Processing (IHCSP) | 979-8-3503-4595-7/23/$31.00 ©2023 IEEE | DOI: 10.1109/IHCSP56702.2023.10127152

Indira College of Commerce and Indira School of Business Studies, D.Y. Patil International University,
Science, Pune,India Pune, India.
Pune, India. manisha.maddel@indiraisbs.edu.com yudhi.raut@dypiu.ac.in
veejeya.kumbhar@gmail.com

Abstract— Career guidance is required mostly for the degree classes that will have the most impact on their chances of
students as they have not clear vision about it. Most students being successful in the future. We've put together a series of
don't know what kind of work they want to do or what degree tests to help the student get off to a good start in their
would be best for them. Considering the user's skills, interests, professional life. The goal of these tests is to find out which
and preferences, the system will suggest different career paths. areas of study are most useful and to point the student in the
The results of this assignment can help students zero in on a
right direction for further growth. Because it is good to
subject area that fits their knowledge and interests the best. The
Smart Career Guidance System using chatbot, is built using improve one's ability to think critically, the site starts with a
recurrent neural network (RNN) and Long Short Term Memory test of one's skills. The first thing a student should do on our
(LSTM) . To proposed Smart Model for Career Guidance Using website is sign up for an account. When a student signs up on
Hybrid Deep Learning Techniques. Our proposed hybrid deep our website, they start getting ready for real work. We give
learning algorithm work on the concepts of RNN and LSTM and them a place to take an aptitude test. When a student signs up
Natural language processing (NLP). for an account on our website, he will be able to choose a
testing level that is right for his level of education. Students
Keywords— Career guidance, machine learning , hybrid Deep will then move on to the main part of the test. After some
learning , NLP , LSTM preliminary questions, the test questions are changed to
reflect what the student is interested in. Jobs in engineering
I. INTRODUCTION and technology, business, law, the humanities, hospitality
There are a lot of different ways that machine learning management, and the social sciences are all on the list of
has been used to make prediction algorithms. Decision trees, possible career paths. These topics will be the basis for the
artificial neural networks, matrix factorization, collaborative questions that will be asked. After the test is over, the teacher
filters, and probabilistic graphical models are all examples of will give the student a detailed analysis of how he or she did,
these methods [6]. Most of these studies think that making a along with their top recommendations. A knowledgeable
prediction is a one-time thing and don't look at how students' person can help a student figure out what kind of job they
learning has changed over the course of their academic want to do. After the expert in the system tells the student
careers. A method called three-mode tensor factorization (on what his results are, the student can make an informed
student, problem, and time) [6] was made to figure out how decision about his future based on the things that really
well a student will do in the field of information technology interest him. Because they are made for a certain place, the
(IT). This method takes the effects of time and order into systems that are already in place don't offer a solution that can
account. When making a model like this one, there are two help a lot of different people. This is because students are
main things that need to be considered: (a) whether or not the looking for answers and come from many different cultures.
student is willing to build a career based on his interests and Some of the solutions that have been made include mining
compassions [3], and (b) whether or not the student has the data that hasn't been found before, making highly complex
right idea of how to improve his skills by taking certification statistical models, and coming up with new ways to use
courses that are based on his interests[3]. For the model to algorithms. Using these methods to figure out how a system
work, both of these questions must be answered in the works will lead to systems that are very complicated and hard
affirmative. Because of this, a Questioner built with this to manage. Even though the internet has a lot of information,
architecture will need to sort the students' individual it can be hard and take a long time to find what you need.
reflections on their individual results. It's no secret that the
level of competition in today's world is always getting II. RELATED WORK
tougher, and it's also no secret that, as the world gets more Kamal, et al[1] The data were put together with the
technical, choosing a career path has become one of the most help of professionals, who made a questionnaire. This
important things for young people to do. Both of these are research was done with the help of 392 people who had just
common knowledge. More than any other question, the finished college. Because there is a lot of variation in the data,
question of what the kids are most interested in comes up the SMOTE oversampling method is used to test the machine
most often in the classroom. Every parent hopes that their learning classifiers[4]. The Authors compared the XGBoost
child will become a scientist or a doctor when they grow up, and Random Forest classifiers' AUC-ROC performance
but no one ever asks them what they are interested in in the scores to find the best career paths. The XGBoost got a score
real world. This programme is helpful for both students and of 0.9952, while the Random Forest classifier got a score of
their families. If a student knows what they are really 0.9963..
interested in, they may be able to steer themselves toward the
327 979-8-3503-4595-7 03/2023 © 2023 IEEE

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 20,2023 at 14:46:50 UTC from IEEE Xplore. Restrictions apply.
1st IEEE International conference on Innovations in High-Speed Communication and Signal Processing (IEEE-IHCSP) 4-5 March, 2023

J. Britto, et al[2] This is a paper that students in their to figure out what the user wants and give the right answer.
last year of school should read. Data Mining and Machine The system's ability to handle human input is based on the
Learning are the two techniques it uses to get such precise way it is made, which is modular. Each part has its own
results. Both the students' academic performance and their knowledge base, way to start up, and set of rules for how to
areas of interest are taken into account. The more information handle user requests. The system's modular design gives it a
it has, the more accurate its predictions will be. Students will lot of flexibility when it comes to taking in new types of data
benefit from this strategy because it will give them or information. Each part has its own knowledge base and
suggestions that fit both their skills and their interests. way to get started, as well as its own set of rules for how to
A. Rohini et al[3] As part of the proposed study, a handle any requests from users. There will be two separate
number of location-based metrics, such as the distance parts to the chatbot module. First, there is the integration of
between nodes, their centrality values, and the residual factors capabilities for handling requests. In the form of a question,
of the parameters, will be evaluated and analysed. In this the user's information will be sent to the conversational agent,
study, they collected data from students. Then, they used K- and with the help of the proposed module, the processing
Means algorithms and the programming language R to predict stages will continue. The platform's module will search
the connections between the nodes and the similarities through the data corpus to find an intent that matches the text
between the users, which leads to the formation of an that the user entered. Remember that the real setting up will
integrated community. happen in the second part of the process. As a direct result,
A. Ghimire et al[4] The results of this study back up there are no obstacles in the way between the two modules
the idea that a person's personality traits can be used to figure and the server. For a module's configuration to be useful, it
out what their favorite area of academic study is. They needs to include model-specific logic (in our system this is
showed that a decision tree model is able to classify things done by the ontology module). The one we chose to use for
with a higher than 90% accuracy rate. Also, they looked at this project was the Naive Bayes Algorithm. We decided to
two different ways to reduce the number of dimensions: one use this method because, in the end, it will tell us how likely
was based on Principal Component Analysis (PCA), and the each category is and which category has the highest chance.
other was based on research from the social sciences on the We have changed our procedure so that it now always gives
Big Five personality traits (OCEAN indices). They used these three outputs, no matter what inputs are used.
methods to cut the number of dimensions in half and make Neural Networks that are strong and effective Artificial
sure that our classifier's performance was not affected in any neural networks are made up of many layers of interconnected
way. artificial neurons. These networks are based on how the
Reddy, K.Sai et al[5] Students can use an automated human brain works. These neurons take in information, get to
method based on ensemble learning to help them decide know its parts, and then work together to make a meaningful
which job path in the software industry is best for them. This result. Neural networks need to be able to look at a huge
research will help them do this. This system takes into amount of data in order to find patterns and trends that are
account not only how well students do in school, but also how hidden in the data. To properly evaluate this programme, we
they are as people and what they are interested in. They used need to find out if the chatbot can simulate human language
three different Machine Learning algorithms on the dataset patterns, keep a natural flow of conversation in response to
and then compared the results to find the best model. Based user input, and meet the user's needs. In light of this, a
on the results, it looks like the random forest method is the modified version of the Turing test might be a good litmus
best one. test. One of the biggest problems with neural networks is that
it can be hard to understand how they work. It is hard to figure
III. PROPOSED METHODLOGY out how a single neuron in a network processed a certain set
of inputs or came to a certain conclusion. Also, neural
Because of how the software is made, both the user networks often need a lot of information to learn well and give
(or student) and the software can talk to each other. The accurate results. Because it's easy, the vast majority of chatbot
system will conclude based on both the data that was environments don't have such a huge amount of data. As a
programmed into it and the AI that it uses. This program is direct result of this, efficient algorithms have been made that
mostly made to help people who aren't sure what careers they can process textual data.
want to go into in the future or what steps they need to take Natural Language Processing (NLP), which is also called
to get there. Because of this, the most important parts of our "machine learning," is an important part of chatbots because it
solution are a way to make suggestions and a way to controls how the bot understands and responds to text input.
summarize information. The first set of questions a user sees The best chatbots would be able to have conversations with
will cover everything about the topic he or she has chosen. users in which the users would have no idea they were talking
After the user answers each of these questions, the domains to a computer. The goal of this software is to learn about the
will be narrowed down until the last set of questions, at which complexities of human language by applying machine
point the user will see a relevant result for their situation. learning techniques to a large database of conversational
data[7]. By using natural language processing, the bot can
Because of this, the exam is broken up into different parts.
understand text data, including grammar, emotion, and
The first part is a single general question meant to get a sense
purpose. Natural language processing [11] gives users a lot of
of how interested the student is in the topic. As you move helpful tools, such as text summarization, word vectorization,
through the levels, the number of questions you must answer topic modelling, point-of-sale (PoS) tagging, n-gram analysis,
will go up. A Suggested The chatbot talks to the user through and sentiment polarity analysis, to name a few. Testing this
whatever platform the user is using. The next step is for it to method is a lot like testing the chatbot's ability to
use the module's natural language processing (NLP) features communicate, since NLP methods are meant to be like human

979-8-3503-4595-7 03/2023 © 2023 IEEE 328

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 20,2023 at 14:46:50 UTC from IEEE Xplore. Restrictions apply.
1st IEEE International conference on Innovations in High-Speed Communication and Signal Processing (IEEE-IHCSP) 4-5 March, 2023

conversation. There is no well-known standard or set of in a movie into groups based on what kind of thing they are.
criteria that can be used to measure how well NLP algorithms A normal neural network would have to remember what had
work. One of the most common and helpful ways to test NLP already happened in the movie before it could correctly
algorithms is to measure and analyse how happy users are with predict what would happen next. Recurrent neural networks
them[14]. It might be easy to do data analysis and text analysis were made so that a solution to this problem could be found.
on user satisfaction ratings, but it's not a good idea to do so They are a type of network that has connections that go in a
because it doesn't lead to any useful information being found. circle and can store information forever. The picture on the
Since natural language processing often tries to act like how right shows a part of a neural network that is labelled with the
people talk, the Turing test can be used to judge it [5]. The
letter A. This part of the network oversees processing the
Turing test is a way to figure out how well a computer
information shown by xt and putting out the result shown by
programme imitates human language and how people talk to
each other [1]. Users could also be asked to tell the difference ht. One piece of information can be passed from one level of
between talking to a real person and talking to a chatbot that a network to the next with the help of a loop. Because of this,
is powered by natural language processing. So that the testing recurrent neural networks tend to make people feel like there's
can be as thorough as possible, it is important for the user to something mysterious going on. But if you think about it
give the chatbot inputs that are hard or likely to cause more, you'll see that they're not all that different from a
mistakes. A good chatbot not only learns what it needs to normal neural network.
know to give the right answers, but it also learns what it doesn't
need to know to do its job Since this strategy is based on the
chatbot's ability to learn from experience, it may be possible
to come up with a way to figure out how well the input was
represented. If the chatbot thinks the information is not
important, it will stop giving correct answers. Another way to
prove this claim is to check the algorithm's ability to recognise
correct sentence structure. To put it another way, the algorithm
should be tested by giving it data with different levels of
grammar, spelling, and punctuation mistakes and by changing
the order of the words in the data.
IV. PROPOSED ALGORITHM
1) Prepare a dataset of questions and possible answers
that are used in career counselling, guidance.
2) Bifurcate it in two parts training and testing as 80% Fig 2: Working of RNN and LSTM for career guidance
and 20% respectively.
3) Build a model using training dataset of career They are built with neural networks, which are the most
counselling. natural way to handle this kind of data. RNNs are used to
4) Use the combined method of techniques of NLP, solve a wide range of problems now a days, such as speech
RNN and LSTM. recognition, language modelling, and so on. A lot of things
5) Test it using testing data. are on the list. Andrej Karpathy has already written a great
blog post about this topic called "The Unreasonable
6) Validate the new and existing algorithms by finding
Effectiveness of Recurrent Neural Networks." In it, he talks
suitable metrics evaluation. about the amazing results that can be achieved with RNNs, so
I won't go over his analysis again here. Instead, I'll tell you
what I think about the subject. On the other hand, they are
interesting on their own. The use of "LSTMs," a specialised
type of recurrent neural network, is crucial to these findings
because, compared to the standard version, it performs a
number of tasks much better. Almost all of the interesting
things that can be done with recurrent neural networks can
only be done by using them. The main focus of this study will
be on long-term and short-term memories. This is done by
letting the gate layers keep an eye on what's going on inside
the cell.
Fig. 1: Approach for Career Guidance Using Hybrid deep Learning 𝑓𝑓𝑡𝑡 = 𝜎𝜎(𝑊𝑊𝑓𝑓 . [𝐶𝐶𝑡𝑡−1 , ℎ𝑡𝑡−1 , 𝑥𝑥𝑡𝑡 ] + 𝑏𝑏𝑓𝑓 )
Techniques
𝑖𝑖𝑡𝑡 = 𝜎𝜎(𝑊𝑊𝑖𝑖 . [𝐶𝐶𝑡𝑡−1 , ℎ𝑡𝑡−1 , 𝑥𝑥𝑡𝑡 ] + 𝑏𝑏𝑖𝑖 )
It's not true that every second, the human mind starts 𝑜𝑜𝑡𝑡 = 𝜎𝜎(𝑊𝑊𝑜𝑜 . [𝐶𝐶𝑡𝑡−1 , ℎ𝑡𝑡−1 , 𝑥𝑥𝑡𝑡 ] + 𝑏𝑏𝑜𝑜 )
over from scratch. In this essay, each word's meaning builds
on the meaning of the word before it. It's not as easy as just In a second possible step, the input gate and the
throwing away everything and starting over. Your mental forget gate are linked together to make a new circuit[18]. We
fortitude is admirable. Most neural networks that are used don't make these decisions separately; instead, we talk about
today have this very big problem. Let's say, for the sake of them and agree on what should be in the discussion and what
this example, that you want to put all of the things that happen should be left out. The only time we remember we forgot

329 979-8-3503-4595-7 03/2023 © 2023 IEEE

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 20,2023 at 14:46:50 UTC from IEEE Xplore. Restrictions apply.
1st IEEE International conference on Innovations in High-Speed Communication and Signal Processing (IEEE-IHCSP) 4-5 March, 2023

something is when we try to switch out one thing for another. where the number of jobs is growing quickly, many people
We only enter a new value if we can't remember what the old could benefit from some kind of career advice or help. This
one was. system has both a chatbot and an AI-powered engine that
works well and gives accurate results. The engine's career
𝐶𝐶𝑡𝑡 = 𝑓𝑓𝑡𝑡 ∗ 𝐶𝐶𝑡𝑡−1 + (1 − 𝑓𝑓𝑡𝑡 ) ∗ 𝐶𝐶̌𝑡𝑡 ….[11] suggestions are based on data from top experts in the field.
This means that it can use the knowledge of these people. This
The Gated Recurrent Unit, which is sometimes method is helpful for anyone who wants to find direction in
called the GRU, is a more extreme version of the LSTM [11]. their career. The results of this study show that putting AI into
It's the same as having a single "update gate" instead of two career counselling systems is likely to be very helpful for both
separate "forget gates" and "input gates." It changes both the the global community now and in the future.
hidden state and the cell state by mixing them together.
Traditional LSTM models are harder to understand and use
than the model that came out of this. REFERENCES
𝑍𝑍𝑡𝑡 = 𝜎𝜎(𝑊𝑊𝑧𝑧 . [ℎ𝑡𝑡−1 , 𝑥𝑥𝑡𝑡 ]) ……[5]
𝑟𝑟𝑡𝑡 = 𝜎𝜎(𝑊𝑊𝑟𝑟 . [ℎ𝑡𝑡−1 , 𝑥𝑥𝑡𝑡 ]) ……[5] [1] Kamal, B. Naushad, H. Rafiq and S. Tahzeeb, "Smart Career Guidance
̌ℎ𝑡𝑡 = 𝑡𝑡𝑡𝑡𝑡𝑡ℎ(𝑊𝑊. [𝑟𝑟𝑡𝑡 ∗ ℎ𝑡𝑡−1 , 𝑥𝑥𝑡𝑡 ]) ……[5] System," 2021 4th International Conference on Computing &
Information Sciences (ICCIS), 2021, pp. 1-7, doi:
ℎ𝑡𝑡 = (1 − 𝑍𝑍𝑡𝑡 ) ∗ ℎ𝑡𝑡−1 + 𝑍𝑍𝑡𝑡 ∗ ℎ̌𝑡𝑡 ) ……[5] 10.1109/ICCIS54243.2021.9676408.
[2] J. Britto, S. Prabhu, A. Gawali and Y. Jadhav, "A Machine Learning
Even so, those don't cover all the most well-known LSTM Based Approach for Recommending Courses at Graduate Level," 2019
variations. Also, there are a lot of others, like depth-gated International Conference on Smart Systems and Inventive Technology
recurrent neural networks. Clockwork RNNs are another way (ICSSIT), 2019, pp. 117-121, doi:
to solve the problem of how to deal with these kinds of 10.1109/ICSSIT46314.2019.8987815.
dependencies. We need to figure out which of these two [3] A. Rohini and T. Sudalaimuthu, "An Analysis of Structural Properties
options is better. What could be important about the in Social Network Links in Educational Data Mining," 2019
International Conference on contemporary Computing and Informatics
differences between these two versions? when they compare
(IC3I), 2019, pp. 295-299, doi: 10.1109/IC3I46837.2019.9055654.
common variants, they usually find that they are mostly the
[4] A. Ghimire, T. Dorsch and J. Edwards, "Introspection with Data:
same. Researchers tried out different RNN designs and found Recommendation of Academic Majors Based on Personality Traits,"
that some of them did some tasks better than LSTMs. 2022 Intermountain Engineering, Technology and Computing (IETC),
2022, pp. 1-6, doi: 10.1109/IETC54973.2022.9796766.
[5] Reddy, K.Sai Vivek and Reddy, M.Akhila and Kaur, Veerpal and Kaur,
V. RESULT ANALYSIS Gagandeep, Career Guidance System Using Ensemble Learning (July
To perform the experiment, a script is written in 14, 2022). Proceedings of the Advancement in Electronics &
Python on Anaconda platform and Precision and Recall. The Communication Engineering 2022, Available at SSRN:
Smart Career Guidance System using chatbot, is built using https://ssrn.com/abstract=4157249 or
RNN and LSTM, is then evaluates using standard metrics. http://dx.doi.org/10.2139/ssrn.4157249
Our proposed hybrid deep learning algorithm worked on the [6] S. Vignesh, C. Shivani Priyanka, H. Shree Manju and K. Mythili, "An
Intelligent Career Guidance System using Machine Learning," 2021
concepts of RNN and LSTM and NLP. model. This research
7th International Conference on Advanced Computing and
concludes that better understanding of the career concepts
Communication Systems (ICACCS), 2021, pp. 987-990, doi:
makes students to take right decision about their career. In 10.1109/ICACCS51430.2021.9441978.
this research we included all such keywords and model built [7] P. Udagama, C. Wijayanama and M. Vithanapathirana, "An innovation
around such keywords using chatbot gives optimal result in Career Guidance in Higher education: Effectiveness and
compared to existing other models. Sustainability of Institutionalization of Service Learning in the
University of Colombo," 2019 From Innovation to Impact (FITI), 2019,
TABLE I. COMPARATIVE ANALYSIS OF EXISTING AND pp. 1-5, doi: 10.1109/FITI49428.2019.9037627.
PROPOSED METHOD [8] L. Acheson and R. Rybarczyk, "Integrating career development into
F-1 computer science undergraduate curriculum," 2016 11th International
Model Accuracy Precision Recall
Score Conference on Computer Science & Education (ICCSE), 2016, pp.
DT 0.9167 0.8765 0.4444 0.9453 177-181, doi: 10.1109/ICCSE.2016.7581576.
[9] J. R. D. Atienza, R. M. Hernandez, R. L. Castillo, N. M. De Jesus and
RF 0.9248 0.9855 0.9851 9.6775 L. J. E. Buenas, "A Deep Neural Network in a Web-based Career Track
RNN 0.9474 0.7654 0.9773 0.9799 Recommender System for Lower Secondary Education," 2022 2nd
Asian Conference on Innovation in Technology (ASIANCON), 2022,
LSTM 0.9733 0.8765 0.9351 0.9845
pp. 1-6, doi: 10.1109/ASIANCON55314.2022.9908965.
Proposed [10] A. B. Chowdhury, B. Tan, S. Garg and R. Karri, "Robust Deep
0.9933 0.7654 0.9345 0.9844
Method Learning for IC Test Problems," in IEEE Transactions on Computer-
Aided Design of Integrated Circuits and Systems, vol. 41, no. 1, pp.
183-195, Jan. 2022, doi: 10.1109/TCAD.2021.3054808.
VI. CONCLUSION [11] L. Zhu, W. Yang, B. Chen, F. Lu and S. Wang, "Enlightening Low-
It is interesting to think about how a network of career Light Images With Dynamic Guidance for Context Enrichment," in
counsellors could work. Since students have access to the IEEE Transactions on Circuits and Systems for Video Technology, vol.
internet, they will be able to use this type system to choose a 32, no. 8, pp. 5068-5079, Aug. 2022, doi:
line of work that fits their skills the best. In today's world, 10.1109/TCSVT.2022.3146731.

979-8-3503-4595-7 03/2023 © 2023 IEEE 330

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 20,2023 at 14:46:50 UTC from IEEE Xplore. Restrictions apply.
1st IEEE International conference on Innovations in High-Speed Communication and Signal Processing (IEEE-IHCSP) 4-5 March, 2023

[12] Yasrab, Robail et al. “PhenomNet: Bridging Phenotype-Genotype Gap: Steering System," in IEEE Transactions on Intelligent Vehicles, vol. 6,
A CNN-LSTM Based Automatic Plant Root Anatomization System.” no. 4, pp. 622-633, Dec. 2021, doi: 10.1109/TIV.2020.3044180.
bioRxiv (2020): n. pag. [15] W. Yang, S. Wang, Y. Fang, Y. Wang and J. Liu, "Band
[13] X. Guo, C. Yang, Y. Liu and Y. Yuan, "Learn to Threshold: Representation-Based Semi-Supervised Low-Light Image
ThresholdNet With Confidence-Guided Manifold Mixup for Polyp Enhancement: Bridging the Gap Between Signal Fidelity and
Segmentation," in IEEE Transactions on Medical Imaging, vol. 40, no. Perceptual Quality," in IEEE Transactions on Image Processing, vol.
4, pp. 1134-1146, April 2021, doi: 10.1109/TMI.2020.3046843. 30, pp. 3461-3473, 2021, doi: 10.1109/TIP.2021.3062184.
[14] Z. Yan, K. Yang, Z. Wang, B. Yang, T. Kaizuka and K. Nakano, [16] J. Carlson, "Hands-on learning opportunities at the Ocean Technology
"Intention-Based Lane Changing and Lane Keeping Haptic Guidance Foundation," Proceedings of OCEANS 2005 MTS/IEEE, 2005, pp.
2279-2281 Vol. 3, doi: 10.1109/OCEANS.2005.1640105.

331 979-8-3503-4595-7 03/2023 © 2023 IEEE

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 20,2023 at 14:46:50 UTC from IEEE Xplore. Restrictions apply.

You might also like