Car Recommendation System Using Customer Reviews

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 10 | Oct 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 983
Car Recommendation System Using Customer Reviews
Saumya Singh1, Soumyadeepta Das2, Ananya Sajwan3, Ishanika Singh4, Ashish Alok5
1,2,3,4,5Vellore Institute of Technology, Tamil Nadu, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Consumers face a major challenge today in
choosing from the many alternatives available to any
product category. With the growing demand and
production of cars, there are thousands of great brands
which make hundreds of new models every year. The need
for proper recommendation of the cars based on customer’s
specific needs is essential and acts integral to both the
manufacturers as well as clients. Recommendation based on
specifications and details, are becoming somewhat
unrealistic and not everyone has the detailed knowledge of
cars. Clients prefer to have a review-based recommendation
from the people who have hands on experience with the
particular product. We can see how recommendation
systems have a surprisingly large impact on the materials
consumers engage with over the course of their daily lives.
Hence, our proposed solution is to provide a system for car
recommendation based on customer reviews, using the
power of ML, NLP and Data Analytics.
Key Words: Machine Learning, Natural Language
Processing, Data Analytics, Data Visualization,
Recommendation, Recommendation System
1. INTRODUCTION
As the world's population grows, so does the value of the
product in the market. Due to global distribution, there is
an increase in global trade leading to different types of
products e.g., buy soap, there are different types
depending on the flavours, aroma, brand (international)
etc. it also applies to cars.
As the number of cars on the international market
increases, so does the information that each individual
gains from an online product. People’s madness with cars
dates back to the Neolithic period, the last part of the
Stone Age when the making of a wheel was made. As time
goes by, technology grows and stands on what we see
today. Today most people are aware of what is happening
around them. As market competition grows, cars with
similar features enter the market. People will be confused
about what to choose. Here the recommendation
algorithm plays a role because it assists the customer or
end user in promoting the right product based on its taste.
This research project is about a web-based program for
vehicles. The existing systems that rely on
recommendation based on specifications and details, are
becoming somewhat unrealistic and not everyone has the
detailed knowledge of cars. Different customers use
different strategies, some are more knowledgeable, and up
to date whereas some need advice, reviews from peers,
and advice. Our goal is to bring the best out of both worlds,
be it specific search, or recommendation, be it seller
claims, or reviews. Existing solutions require the users to
have proper knowledge of specification and their
requirement and, mostly are based on the content type
recommendation. Using the customer reviews, our system
becomes more customer oriented as well as less
knowledge driven but is not confined to that.
The main purpose of this function is to recommend the car
according to the user model and object profile. In this
paper, a proposed algorithm to recommend hybrid-based
user-to-user and interactive filtering techniques is used
based on actual textual data from the end users or clients
of the product, using the ML, NLP and Data Analytics.
The dataset for the project is Consumer Car Reviews
dataset from Edmunds.com (also available on
Kaggle.com), as well as other sources, which contains
lakhs of reviews from multiple brands given by
consumers.
2. LITERATURE REVIEW
1. T. G. Thomas 1, V. Vaidehi , worked on “Vehicle
Recommendation System Design The Web uses a
Hybrid Recommender Algorithm” developed a web
based complimentary program for vehicles. The main
purpose of this function is to recommend the car
according to the user model and object profile. In this
paper, the proposed hybrid recommendation
algorithm is used from user-to-user and interactive
filtering methods aimed at generating vehicle
recommendations. The user model is built with
personality features, click data and browsing history.
The profile of the item is built using various car
attributes. Forty car types are used including 224 car
types in this project.
2. Srivastava, A. Kumar, S. Samee, P. Thokal Vijay, P. S.
Tanesh, worked on “Vehicle Recommendation Method
Using Vengatesan K Machine Learning Algorithm”
studied and found that more than 90% of planned
drivers regularly show that eliminating any driving
pollution can control their chances and adaptability.
Competitive drivers have exerted pressure on the low-

key concept of open-vehicle integration. This weight
conveys the idea that it is actually improving in terms
of how half of the late respondents had eliminated any
driving distractions that they felt were an open car
that, or in some way, lacked.
3. G. Prabowo, Md. Nasrun, R. A. Nugrahaeni worked on
“Recommendations for the Combined Filtering Program
(CF)”, proposed a program that can help provide
information about vehicles that are in line with user
preferences, i.e., a recommendation system. The
recommendation system requires appropriate
recommendations In this study you will focus on the
problem of recommending a car selection system by
creating a recommendation system using a
collaborative filtering process.
3. DRAWBACKS AND CHALLENGES
Recommendation systems are an important part of
business and e-commerce including the auto industry. The
currently available car recommender systems depend
heavily upon users having proper knowledge of cars and
their specifications.
Most of the systems use a content-based recommender
approach which is based on the user’s history and
suggests items similar to their past purchases. This type of
system is extremely disadvantageous as users could be
first-time buyers and may not have proper knowledge on
car specifications.
These drawbacks have been eliminated with our
application which uses a collaborative filtering approach
i.e., “people to people” approach. Collaborative filtering
methodology is applied to
filter products that might interest a particular user
depending on reactions received by similar users. This
method is more suited to recommend cars as it is less
knowledge-driven and more customer-oriented, keeping
in mind that not all customers are informed on specifics of
cars.
Challenges: One of the challenges faced was cleaning the
user reviews available in the datasets. This was done by
incorporating different methods of natural language
processing like lemmatization, removal of stop words and
punctuations. The datasets were combined onto a single
data frame and cleaned by additionally removing
unnecessary columns.
Another challenge included integrating the machine
learning notebook with the Flask application to create a
fully functional product. The challenge here was to ensure
that the recommendation was fast, accurate and efficient.
This involved multiple testing scenarios and removing
bugs which improved the accuracy and the efficiency of
the recommender system.
4. REQUIREMENTS & PROPOSED SYSTEM
4.1 Software Requirements
• Python (as programming language version>=3.0)
• Python compiler
• Conda or Jupyter environment (used for development)
• Flask
• Web Browser
4.1.1 Python Module Dependencies
• nltk
• genism
• PyLDAVis
• Sklearn
• Pandas
• Numpy
• Pickle
• Json
• re
• Random
• Textblob
• Math
4.2 Proposed System:
The Recommendation system for the car proposed
involves development processes like Data Collection, Data
Pre-processing, Model Design, Model Building,
Recommendation. The system architecture involves
Python as the main programming language, for the
recommendation model.
The architecture involves the dataset collected from
Edmunds Car Reviews dataset, which consists of the
details of the reviewer, the car reviewed, the text review
and the rating given by the customer for the car. The
dataset is pre-processed which involves processes like
stop word removal, lemmatization, removal of common
words, etc. After pre-processing, the processed dataset is
fed to Topic Modelling model using Non-Negative Matrix
factorization after, count vectorization and inverse
document factorization, after this stage the topics are
extracted and distributed based on the dataset and the
reviews on each car. To optimal number of topics in the

dataset to avoid overfitting and high variance decisions,
we check the coherence values vs number of topics, the
number with the highest coherence value is taken as the
number of topics. Upon deciding the number of optimal
topics, the topics and the output is fed to the Latent
Dirichlet Allocation (LDA) model, to score each record and
find out the underlying relationship of a particular query
with the review of the car, as a result of which a specific
topic number is be allocated to each car, which is used for
recommendation based on whatever query the user
provides.
The application allows both quantitative searches based
on car type like SUV, sedan and also qualitative search like
how a user describes their requirements to an expert, who
in return uses their expertise to recommend their cars.
This application and recommendation model provides the
capability for both.
Stages:
1. Data Collection
2. Data Pre-processing
3. Model Building:
a. Topic Modelling using NMF
b. Selecting optimal number of topics
c. LDA model training
d. Assign topic scores to cars and mapping back to
dataset
4. Recommendation
Fig -1: Architecture Diagram
5. MODULES DESCRIPTION
a) Dataset: The dataset used for the project is Consumer
Car Reviews dataset from Edmunds.com (also available on
Kaggle.com), which contains lakhs of reviews from
multiple brands given by consumers. Contains datasets
pertaining to different companies, which could be used
separately for each company or combined to be used as
whole general database.
The usability of the dataset is marked 7.1-7.5, which is
decent, and means it has low null and ill formatted data.
Hence, as the source provides a huge data, along with
proper columns and fields and has high usability score, we
decided to work on this dataset. Consumer Car Reviews
Dataset contains a huge database of car reviews from
varying brands.
Contains 7 columns namely:
• Key,
• Review Date,
• Author Name,
• Vehicle Title,
• Review Title,
• Review, Rating.
The models will be trained based on the Date, Vehicle
Name, Review Title, mainly the Review and Rating.
b) Data Pre-processing: The datasets are combined and
loaded onto a single data frame and then we only consider
data after 2010 due to relevancy. The dataset is cleaned by
removing unnecessary columns and then we are left with
the main data – Vehicle, Review and Rating columns.
Data Pre-processing techniques like:
• Stop Word Removal
• Lemmatization
• Removal of common basic words
One important inference during pre-processing stage was,
cars with higher rating had reviews with lesser text review
length. Even after cleaning and pre-processing the dataset,
it was holding true, and is used for model building.
Fig -2: Preprocessing Inference (ratings inversely
proportional to review length)

c) Topic Modelling: The processed data is used for
training the topic modelling model, topic modelling is
done using the NonNegative Matrix Factorization (NMF)
model from sklearn, which takes the input matrix and
outputs two matrices
• The W factor contains the document membership
weights relative to each of the k topics. Each row
corresponds to a single document, and each column
correspond to a topic.
• The H factor contains the term weights relative to each of
the k topics. In this case, each row corresponds to a topic,
and each column corresponds to a
unique term in the corpus vocabulary.
Topics Distributed and top words for each distributed
topic is shown.
Fig -3: Topics from topic modelling, with top words for
each topic
After this stage, optimal number of topics to be keep for
actual recommendation is found out from all the topics
distributed. This is done to keep the recommendations
from overlapping too much due to redundant topics. This
is done by plotting coherence values against the number of
topics.
Fig -4: coherence values vs. number of topics
From the plot the number of optimal topics is determined
to be 8.
d) Latent Dirichlet Allocation (LDA) model fitting: This
helps us find the underlying relationships in the
distributed topics so that every text pertaining to each car
can be scored and each car can be assigned a topic. The
topic having highest value for a particular car is assigned.
And based of thi LDA model the users query is compared
and the top topics for the query are scored and from that
topic the top-rated cars are recommended.
The below plot shows the distribution of topics and the
most salient terms and its frequency distribution.
Fig -5: LDA topic distributions and word distributions for
each topic
Word Count of each topic’s top words from the LDA model
and their Importance in that topic (Fig. 6 & Fig. 7):
Fig -6
Fig -7: Visualization showing distribution of topics
assigned

Fig -8: Visualization of topic distributed (8 optimal topics
spatial distribution)
e) Recommendation: After the LDA model assigns each
record with a specific topic, the users query is scored with
the LDA model to find out the relations defined in the user
query and find those topics which reflect the query. Then
from the top topics discover the top-rated cars are
recommended back to the users. For specific types of cars
such as SUV and sedan or hatchback, we extracted the
types of cars from their model names and appended the
categories for class wise recommendation. Allowing a
holistic recommendation ability based on quantitative
features like category, type, etc as well as qualitative based
on the users’ abstract requirements.
6. RESULTS & DISCUSSIONS
Pre-processing inference:
From the visualization, it can be inferred that the ratings
increase with decrease in review length, for the
population.
Fig -9: Preprocessing Inference
(ratings inversely proportional to review length)
The distribution shows the topics are evenly distributed
and explains a different section, showing an efficient
model.
Fig -10: LDA topic distributions and word distributions for
each topic
Post topic modelling: choosing optimal number of topics,
from the plot the number of optimal topics is determined
to be 8.
Fig -11: Coherence Values vs. No. of Topics
Topic Distributions, results: Intertopic distances show
that the topics are properly spaced out and each topic
explains a different category of automobile and features.
Fig -12: Visualization of topic distributed (8 optimal topics
spatial distribution)

We successfully developed a flask application that that
textual query as input, analyses it, runs the topic modelling
and linear discriminant analysis model, and recommends
top 10 cars accordingly. The recommender system uses
not only quantitative inputs but also qualitative inputs
which includes reviews from other customers.
Another feature that is used is that in the flask app itself, if
the type of automobile is specified such as sedan or
hatchback, then the query filters out those cars from the
start giving better results
Fig -13: The web application interface recommending top
10 cars for SUV type, entered by user
Fig -14: Recommending top 10 cars for “stylish and
affordable” text query, entered by user
7. CONCLUSION
Generally, the car recommendation systems are content
based, but we are trying to implement collaborative
recommendation system. This will help us recommend
cars using textual queries as well, which are based on the
reviews given by actual customers. The existing systems
are mostly either search on basis of specification or
content-based recommendation.
The proposed system could be act both as specific search
tool or recommender as well as an expert car advisor or a
community-based suggestion, but in the digital form.
As the global market rises and demand for new products
in the Indian economy is leading to the arrival of new
models. All foreign car manufacturers see the Indian
market as their growth point in their share of the global
automotive economy. As the world progresses to the
climax of a new era, recommendations become an
inevitable reality. Almost all technological and non-
technical items in modern hands raise their hands in
compliments. The main fact that the recommendations are
extremely focused on the new technology is because of its
accuracy, precision, and reliability.
The recommendation provides a personal preference for
user needs. In the proposed method, which is a
combination of user-to-user and object to a collaboratively
based object to recommend filtering an algorithm that
works well for suggesting. The biggest problem with car
databases is that they are dynamic data because it is
difficult to predict the car model that will be released in
their product. In addition, the performance of the
proposed system can be improved by using a real-time
network that allows you to build websites and access
session information. This research activity can be
expanded as information-based complimentary programs
using a variety of information presentations. Expert
recommendations using a professional program can also
be considered using knowledge bases.
REFERENCES
[1] Dataset sources, Edmunds.com / Kaggle.com,
(https://www.kaggle.com/ankkur13/edmundsconsu
mer-car-ratings-and-reviews)
[2] Q. Zhang, J. Wang and K. Fan, "Research on passenger
car recommendation based on comments mining of
Internet," 2017 12th IEEE Conference on Industrial
Electronics and Applications (ICIEA), 2017, pp. 122-
127, doi: 10.1109/ICIEA.2017.8282826.
[3] G. Prabowol, M. Nasrun and R. A. Nugrahaeni,
"Recommendations for Car Selection System Using

Item-Based Collaborative Filtering (CF)," 2019 IEEE
International Conference on Signals and Systems
(ICSigSys), 2019, pp. 116-119, doi:
10.1109/ICSIGSYS.2019.8811083.
[4] Shrey Talati, Anukrity, Priyanka Salian and Anam
Hussain. Article: Recommendation System for
Automobile Purchasing: A Survey. IJCA Proceedings
on National Conference on Advancements in
Computer & Information Technology NCACIT
2016(6):23-27, May 2016.
[5] Vengatesan K, A. Srivastava, A. Kumar, S. Samee, P. T.
Vijay, P. S. Tanesh, “A Novel Approach of Car
Recommendation Using Machine Learning Algorithm”,
(https://www.ijstr.org/final-print/jun2020/A-Novel-
Approach Of-Car-Recommendation-Using-Machine-
Learning-Algorithm.pdf)

Car Recommendation System Using Customer Reviews

More Related Content

Similar to Car Recommendation System Using Customer Reviews (20)

More from IRJET Journal (20)

Recently uploaded (20)

Car Recommendation System Using Customer Reviews