Analysing the performance of Recommendation System using different similarity metrics

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 10 Issue: 04 | Apr 2023 www.irjet.net p-ISSN: 2395-0072
© 2023, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 560
Analysing the performance of Recommendation System using
different similarity metrics
Kopanathi Sonali1, S. V. G. Reddy2, K. Thammi Reddy3, V. Valli Kumari4
1M. Tech Student, Department of Computer Science and Engineering, GITAM University, Visakhapatnam,
Andhra Pradesh, 530045, India.
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - In today'smodernera ofinformationtechnology,
finding a favourite item in a large dataset has become an
essential issue. So, there is a need for a more effective
recommendation system with better performance. To
achieve this, a Collaborative filtering recommendation
system is proposed in this work. Here, the comparison is
made with various similarity metrics like Pearson
Correlation, Cosine Similarity, Jaccard Coefficient, MSD
(Mean Squared Difference), Sorensen Dice Coefficient and
SVD (Singular Value Decomposition) on the Movielens 100k
dataset. It is observed that the Jaccard Similarity metric,
compared to Pearson correlation and cosine similarity,
produces better outcomes with improved accuracy and less
time complexity.
Key Words: Recommended System, Jaccard Index,
Pearson Correlation, Cosine Similarity, Spearman rank
Correlation, Sorensen Dice coefficient, Prediction,
Recommendations.
1. INTRODUCTION
A type of system known as a recommendation system is
utilized for filtering or sorting information with the purpose
of predicting a user's preference or rating for a specificitem.
These systems are commonly used to provide suggestions
for items such as books, TV shows, movies, music, and apps
that may be of interest to a group of users.
To generate recommendations, the system analyses users'
past interests, which can be gathered either explicitly,
through user ratings of items, or implicitly, by tracking user
behaviour like purchasing history, browsing data, and
downloaded applications. In addition,thesystemmayutilize
information from the user's profile, such as age, gender,
nationality, preferences, and habitsoftheirgroupofusers, to
compare and present personalized recommendations.
Fig.1. Recommendation System
In the depicted scenario, a user's purchase of a camera
triggers a recommendation from the system to purchase a
tripod. The recommendation usestheuser'spreviousactions
and preferences as a basis, which are used to suggest
products that align with their interests. This isanexampleof
how recommendation systems operate.
1.1 Types
In terms of recommendation techniques, there exist various
types including:
Content-based:
This type of recommendation system relies on user reviews,
ratings, and product features to generate recommendations.
To find similar items to those previously liked by the user,
the system calculates the similarity between items based on
their associated features. The system then recommends
items to users who have shown similar preferences.
Recommendations are generated by evaluating similarities
among items and considering the top-most regular items
amongst neighbours.
2Associate Professor, Department of Computer Science and Engineering, GITAM University,
Visakhapatnam, Andhra Pradesh, 530045, India.
3Professor, Department of Computer Science and Engineering, GITAM University,
Visakhapatnam, Andhra Pradesh, 530045, India.
4Professor, Department of CS and SE, College of Engineering (Andhra University),
Visakhapatnam, Andhra Pradesh, 530045, India

Fig.2. Content-Based Filtering
In the given illustration, a user's purchase of juice prompts
the recommendation system to suggest purchasing coke
based on the similarity of the items. Recommendations are
made by taking into account the user's previous behaviour
and preferences, demonstrating how content-based
recommendation systems function.
Collaborative filtering:
Collaborative filtering is a strategy used to suggest items to
users by finding those with similar preferences and
recommending items they have previously preferred. The
system evaluates the similarity of users' preferences by
examining their rating history,alsoreferredtoas"people-to-
people correlation." Thisapproachiscommonlyemployedin
Recommender Systems and is implemented using various
methods.
Neighbourhood methods and item-item approachesaretwo
strategies that focus on the relationships between items or
users in Recommender Systems. The item-item approach
models a user's preference for an item basedontheirratings
of similar items. However, the nearest-neighboursmethodis
more widely used due to its efficiency, simplicity, and
capacity to generate precise and personalized
recommendations, particularly for smaller datasets.
Additionally, a variety of collaborative filtering algorithms
are availabletoaccommodatelargerdatasetswithnumerous
users and a greater number of products than items.
Fig.3. Collaborative Filtering
As shown in the above figure, User A orders Salad and pizza,
while User B likes to order Salad, pizza and coke. Both users
have ordered several times from the samefoodordering app
and are given high ratings for their preferred items. The
collaborative filtering system identified that User B has
similar preferences to User A and suggested some of their
favourite items to User A, who may be more likely to enjoy
those items based on their past behaviour. Collaborative
filtering helps to personalise the recommendations and
make them more relevant andappealingto eachUser.Hence,
this is how the collaborative filtering recommended system
works.
Hybrid recommender systems:
Recommender systems that combine content-based and
collaborative filtering techniques are referred to as hybrid
systems. This approach aims to take advantage of the
strengths of each technique while addressing its limitations.
The hybrid system works by using the strengths of one
technique to overcome the weaknesses of the other.
For instance, collaborative filtering systems face challenges
when recommending new items that users still need to rate.
However, the content-based approach does not face this
limitation since it relies on item features and descriptions,
which are typically readily available. By combining the two
techniques, the hybrid recommender system can overcome
the limitations of each technique and provide moreaccurate
and personalised recommendations to users.
Fig.4. Hybrid Recommender system
As depicted in the preceding image, hybrid recommender
systems utilize the advantages of content-based and
collaborative filtering methods to offer users more precise
and varied recommendations.
1.3 Collaborative Filtering approach
User-to-User Collaborative filtering approach.
The user-to-user Collaborative Filtering approach is widely
used for generating recommendations based on the
preferences of other users who share similar interests with

the target user. It assumes that users who have ratedsimilar
items in the past will likely rate future items similarly as
well.
To generate recommendations,thesystemfirstidentifiesthe
most similar users to the target user based on their past
ratings. Then, it considers the ratings of those similar users
for items that the target user has not yet rated to predict
their potential ratings. Finally, the system generates a list of
top recommendations for the target user based on those
predicted ratings.
In neighbourhood algorithms, the system selects a subset of
users similar to the target user based on a similarity metric
such as cosine similarity. The system then computes a
weighted average of the ratings of those selected users to
generate predictions for the target user. The weights
assigned to each user are typically based on their similarity
to the target user.
Item-to-Item Collaborative filtering approach
The item-to-item Collaborative Filtering approach
recommends items to users based on the ratings that users
have given to specific items. Instead of focusing on the
preferences of other users, this approach analyses the
similarity between the target itemandthecollectionofitems
that the user has already rated.
To generate recommendations, thealgorithmusessimilarity
measures to identify the k most similar items to the target
item. It then computes the similarities and commonalities
between the selected items to generate recommendations
for the target user based on their past ratings.Thisapproach
differs from the user-to-user Collaborative Filtering
algorithm, which focuses on identifying users with similar
preferences to the target user.
Fig.5. Difference between User-to-User and Item-to-Item
As shown in Figure 5, there are two main approaches for
generating recommendations: User-to-User and Item-to-
Item Collaborative Filtering.
In the User-to-User approach, recommendations are made
by identifying users who have similar preferences to the
target user. For example, if Jane and Tim both liked Item 2
and disliked Item 3, it suggests that they may have similar
tastes. Therefore, Item 1 might be a good recommendation
for Tim. However, this approach may not be scalable for
millions of users.
On the other hand, the Item-to-Item approach recommends
items based on their similarity to otheritemsthatusershave
liked. For instance, if Tom and Sandra both liked Item 1 and
Item 4, it suggests that people who liked Item 4 will also like
Item 1, and Item 1 will be recommended to Tim. This
approach is scalable to millions of users and items,makingit
a preferred choice for large-scale recommendation systems.
Hence, the user-to-userapproachcanbepractical forsmaller
datasets but can become computationally expensive when
dealing with large amounts of data. In contrast, the item-to-
item approach is more scalable and can handle larger
datasets by focusing on item similarity.
2. LITERATURE REVIEW
Resnick et al. [1] presented the user-to-user approach to
collaborative filtering, where recommendationsforanactive
user are generated by finding users with similar historical
rating behaviours and using their ratings of items that, the
active User has not yet seen. The similarity between users
can be calculated using various metrics likecosinesimilarity
or Pearson correlation coefficient, which are based on the
users' historical rating behaviours. After identifying similar
users, the weighted average of their ratings can be
calculated, with the weights determinedbyhowsimilareach
User is to the active User.
Breese et al. [2] proposed a prediction problem in
collaborative filteringandconductedanempirical analysisof
prediction algorithms. The algorithm aims to predict the
rating an active user will give an active item. To generate
recommendations, the algorithm relies on historical data of
ratings and associated content of both users and items. This
data is used to make predictions about potential user
preferences and item popularity. The active User is the User
for whom the prediction is made, and the active item is
predicted. By analysing the logged data of user-item
interactions, these algorithms estimate the rating the active
User would give the active item.
In a study by Herlocker et al. [3], the authors explored the
critical decisions involved in evaluating collaborative
filtering recommender systems.Theseincludedtheselection
of user assignments to be evaluated, the types of analysis

and datasets used, and the methods of measuring prediction
quality. To analyse the performance of different accuracy
metrics on a single domain, the authors reviewed previous
research and conducted their own experiments. They
classified the different accuracy metrics into three
equivalence classes and found that the metrics within each
class were highly correlated. However, they also found that
the metrics across different classes were notcorrelatedwith
each other. The authors emphasized the importance of
considering multiple evaluation techniques to provide a
comprehensive understanding of the system'sperformance.
In a study by Deshpande and Karypis et al. [4] proposed a
method for the model-based recommendation that involves
evaluating similarities between the items based on their
attributes or features. They first identify how to calculate
similarity between pairs of items using various methods,
such as cosine similarity or Pearson correlation coefficient.
Subsequently, they aggregate these similarities to compute
an overall similarity score between a recommended
candidate item and an item the user has already purchased.
By using this method, the system generates personalised
recommendations for active users based on their past
interactions.
Sarwar [5] conducted an analysis of different algorithms
used for generating item-based recommendations. This
analysis included examining techniquesforcomputingitem-
item similarities and strategies for obtaining accurate
predictions. To evaluate the effectiveness of these
algorithms, they were compared to the basic k-nearest
neighbour approach. Based on the results of the study, item-
based algorithms were found to have better performance in
terms of time compared to user-based CF algorithms.
However, the study alsofoundthatuser-basedCF algorithms
provided better quality recommendations. The study
highlights the importance of considering both performance
and recommendation quality when selecting an algorithm
for generating recommendations.
Fkih et al. [6] review and compare similarity measures used
in Collaborative Filtering-based Recommender Systems. It
categorises the most common similarity measures and
provides an overview. The authors perform experiments to
compare these measures using popular datasets and
evaluation metrics. They find that the optimal similarity
measure depends on the dataset and evaluation metric. The
paper also recommendsthe mostsuitablesimilarity measure
for each dataset and evaluation metric. This study is a
valuable resource for researchers and practitioners in
recommender systems who want to select the most
appropriate similarity measures for their CF-basedsystems.
Bell and Koren et al. [9] proposed a scalable approach for
collaborative filtering using jointly derived neighbourhood
interpolation weights. Their approach addressed the
limitations of traditional collaborative filtering algorithms,
such as high computational costs and data sparsity, by
deriving neighbourhood weights using a combination of
user-to-user and item-to-item approaches. The proposed
method was more accurate and scalable than traditional
approaches on benchmark datasets. The paper contributes
to developing collaborative filtering algorithms for large-
scale recommendation systems.
In their study, Saranya et al. [10] compare the performance
of various similarity measures for theCollaborativeFiltering
(CF) technique. Four similarity measures, includingPearson
Correlation Coefficient, Cosine Similarity, Mean Squared
Difference, and Adjusted Cosine Similarity, were evaluated
using the MovieLens dataset. The study aimed to determine
which similarity measure provides the best performance for
the CF technique. The evaluation is based on three metrics:
MAE, RMSE and Precision. The resultshowsthatthePearson
Correlation Coefficient and the Cosine Similarity perform
better than the other measures regarding accuracy and
Precision. The study provides insights into selecting
appropriate similarity measures for CF-based
recommendation systems.
3. Problem Identification & Objectives
3.1 Problem statement
The primary objective is to determine the most efficient
method of computing similarity between users and items in
a dataset. Various similarity measures, including theJaccard
coefficient, cosine similarity and correlation-based
similarity, are usedtocalculatesimilarityandgeneratetop-N
recommendation lists.Thegoal istoidentify whichsimilarity
measure offers the quickest and most efficient output for
recommending items.
3.2 Motivation:
The increasing number of users on online sites such as
Amazon and Netflix have led to a large user-item matrix,
requiring the efficient and quick generation of personalised
recommendations. However, traditional collaborative
filtering systems need help producing high-quality
recommendations for users in the shortest time possible,
especially for large datasets. Therefore, this study aims to
compare the efficiency and advantagesofdifferentsimilarity
measures, including the Jaccardcoefficient,cosinesimilarity,
Pearson correlation, MSD and adjusted cosine similarity, in
generating a top-N recommendation list.
This study aims to identify the most efficient similarity
measure that can generate accurate recommendations even
for large datasets. User-based top-N recommendation
algorithm is also explored as a potential solution to the
challenges faced by traditional collaborative filtering

systems. Ultimately, the recommendeditemstousersshould
satisfy their preferences and interests.
3.3 Objectives
similarity metrics in a recommendation system.
4) To identify the strengths and weaknesses of different
similarity metrics in generating recommendations.
4. PROPOSED SYSTEM METHODOLOGY
The proposed system methodology uses a collaborative
filtering approach. The system will collect and analyse data
from the user’s profile, location, and interests. It will then
determine parameters that can be used to comparethisdata
with other members' data in the database. The system will
search for similarities between users and other members,
such as shared interests or locations. The steps to achieve
the study's objectives:
Fig.6. Flowchart of the Collaborative
Filtering Approach
In Fig.6, it describes the flow of collaborative filtering
approach and its major functionality.
Process:
To build a movie recommender system using user-to-user
collaborative filtering method with the help of different
similarity metrics.
1) Prepare the data: Merge the movie and ratingdata intoa
single data frame, pivot the table to create a matrix of users
and their movie ratings, and fill any missing values with
zeros.
2) Compute user similarities: Calculate the similarity
between each pair of users based on their movie ratings.
3) Find similar users: identifying users who share similar
preferences to the target user by selecting the k users with
the highest similarity scores.
4) Predict ratings: To generate a prediction for each movie
that a user has not yet rated, the algorithm calculates a
weighted average of the ratings given to that movie by the k
most similar users.
5) Generate recommendations: Recommend the top n
movies with the highest predicted ratings to each user.
Fig.7. Process Flow Diagram
The process flow diagram describes the steps involved in a
recommendation system that uses the K Nearest Neighbour
(KNN) approach to find similar users and recommend items
to them. The steps involved are:
1) User Input: The user provides their User ID and some
Facts such as gender, age, and pin code.
2) Data Pre-processing: The raw data is pre-processed
using NumPy and Pandas libraries. This step involves
cleaning and transforming the data intoseparateframesthat
can be used for further analysis.
3) K Nearest Neighbour (KNN): The KNN approach finds
similar users within a community/group. It involves
selecting a value for k (the number of nearest neighbours to
consider) and computing the distance between users.
The main goals of this study are:
1) To enhance the accuracy of the proposed
recommendation system.
2) To reduce the time complexity of the system.
3) To assess and contrast the performance of various

4) Utility Matrix: A utility matrix is created after applying
the KNN approach. Thismatrixdefinestheaveragerating the
user gives each other.
5) User Similarity: User similarity is calculated using the
utility matrix and Pearson correlation. This step involves
calculating the similarity between users to determine how
similar their preferences are.
6) Recommendation:Thesystemusestheutilitymatrixand
suggests items to the user based on the ratings of user’s
similarity. The recommended items have received high
ratings from similar users but have yet to be interacted with
by the user.
Hence, the process flow diagram describes a
recommendation system that uses collaborative filtering to
recommend items to users based on their similarity to other
users within a community/group.
Properties of similarity metrics
1) Non-negativity: The similarity between any two objects
must be non-negative.
S (X, Y) ≥ 0 for all X and Y.
2) Symmetry: The similarity between two objects should be
symmetric.
S (X, Y) = S (Y, X) for all X and Y.
3) Reflexivity: The similarity between an object and itself
should be maximum.
S (X, X) = 1 for all X.
4) Triangle inequality: The similarity between two objects x
and z should be less than or equal to the sum of their
similarities with a third object Y.
S (X, Z) ≤ S (X, Y) + S (Y, Z) for all X, Y, and Z.
5) Range normalisation: Similarity measures should be
normalised to a certain range, often [0,1], to be compared
across different datasets or applications.
Similarity Metrics
Pearson Correlation:
Pearson Correlation Coefficient (PCC) isusedtomeasurethe
linear relationship between the two variables. Inthecontext
of recommendation systems, these two variables are the
ratings two users give to the same items. The PCC value
ranges between (-1, +1). Where,
-1 = perfectly negative correlation.
0 = no correlation.
+1 = perfect positive correlation.
A positive value represents a positive correlation or a
relationship in which two variables move in the same
direction. It is used when
(1) Linear relationship,
(2) Both variables are quantitative,
(3) Normally distributed
(4) They Have no outliers.
Advantages of Pearson correlation coefficient
1) It is an accurate method of computing the correlation
between two variables.
2) The coefficient helps to determine the degree and
strength of the correlation between the variables.
3) It is standardised,allowingfordirectcomparison between
different datasets and variables.
4) It is robust to small amounts of noise or outliers in the
data.
5) It can identify both positive and negative correlations
between variables.
Disadvantages of Pearson correlation coefficient
1) Pearson's correlation coefficient (PCC) is unsuitable for
testing attributive research hypotheses involving only one
variable. It is because PCC is a bivariatestatistical model that
analyses the relationship between two variables.
2) It cannot determine the nonlinear relationships between
variables.
3) It does not distinguish between dependent and
independent variables.
4) PCC is sensitive to outliers in the data, which can
significantly affect the calculated correlation value.
Cosine similarity:
Cosine similarity is a popular similarity measure utilized in
Collaborative Filtering to evaluate the similarity between

two sets of rating vectors, which could be the rating vectors
of two users or two items. It calculates the cosine of the
angle between the two vectors in a multi-dimensional space.
By comparing the cosine similarity between two rating
vectors, the Collaborative Filtering approach can identify
users who share similar tastes or items that exhibit similar
rating patterns. The cosine similarity score is bounded
between -1 and 1, where a score of 1 represents identical
vectors, -1 signifies diametrically opposed vectors, and 0
indicates orthogonal or independent vectors.
Advantages of cosine similarity:
1) Cosine similarity is computationally efficient and does
not require a lot of memory, which makes it suitable for
large datasets.
2) It is scale-invariant, which means that it is not affected
by the magnitude of the ratings, only their directions.
This property makes it useful for handling sparse data
and data with missing values.
3) It is widely used in many applications, including text
mining, image analysis, and recommendation systems.
Disadvantages of cosine similarity:
1) It does not take into account the magnitude of the
ratings, only their directions.
2) This can lead to inaccuracies if some users tend to rate
items much higher or lower than others.
3) It assumes that the ratings are distributed uniformly
across all dimensions, which may not be the case in
some datasets.
4) It is sensitive to outliers, which can have a significant
impact on the similarity scores if they are not handled
properly.
Adjusted cosine similarity:
It is a modification of the cosine similarity measure that
addresses its limitations. In Collaborativefiltering,similarity
between users is computed based on the rating matrix. In
contrast, in item-to-item Collaborative filtering, the
similarity is computed based on the columns. However, the
standard cosine similarity measure used in item-to-item
Collaborative filtering does not consider user rating
behaviour differences.
Advantages of Adjusted-cosine similarity
1) Overcomes the drawback of cosine-based similarity
2) It subtracts the user average from each co-rated pair.
3) It considers the differences in rating scales across users.
4) It effectively handles sparse data, where many entries in
the user-item matrix are missing.
Spearman Rank Correlation
Spearman Rank Correlation is a similarity measure that
computes similarity based on rankings rather than ratings,
thus eliminating the need for ratingnormalization.However,
this method is not suitable for incomplete orderings, even if
the ratings are comparable.
Mean Squared Difference
The Mean Squared Difference (MSD) approachconsiders the
absolute ratings by calculating the mean of the squared
differences between ratings instead of the traditional
approach of considering the total standard deviations. This
mean is used to determine the similarity between two
vectors. A smaller value of the mean squared difference
indicates a higher similarity between the vectors.
Jaccard Similarity Coefficient:
The Jaccard coefficient is a metric used to quantify the
similarity and dissimilarity between two sample sets. It is
computed by dividing the size of the intersection of the sets
by the size of their union. When both sets are empty, the
coefficient is 1. The coefficient ranges from 0 to 1, where 0
implies that there is no overlap between the sets, and 1
indicates that the sets are identical.

Where, J = Jaccard Similarity
A = Set 1
B = Set 2
Advantages of the Jaccard Index:
1) Measures the similarity between two asymmetric binary
vectors or sets.
2) This similarity measure is beneficial when duplicates
are not necessary.
Disadvantages of the Jaccard Index:
The Jaccard index has a significant disadvantage, mainly
when applied to large datasets, as the data size strongly
affects the index. In such cases, a slight change in the union
can significantly impact the index while keeping the
intersection the same.
Sorensen–Dice coefficient:
The Sorensen-Dice coefficient isa similaritymeasureused to
compare sample sets. It is also referred to as the Sorensen-
Dice index, Dice's coefficient, or Dice similarity coefficient.
This coefficient is closely related to the Jaccard index and is
computed by dividing twice the intersection of two sets by
the sum of their sizes.
The Sorensen-Dice coefficient computes a value between 0
and 1, where 0 represents no overlap between two sets and
1 indicates complete overlap. Unlike the Jaccard index, the
Sorensen-Dice index is more easily understood as the
percentage of overlap between the two sets. This index is
typically used to measure the similarity of two samples,
particularly in the case of discrete data.
Where
DSC = Dice Similarity Coefficient
|A|, |B| = The num of elements in each set.
|A ∩ B| = The Common num of elements
Disadvantages of Sorensen–Dice coefficient:
1) It weights each item differently based on the size of the
relevant set instead of treating them equally.
2) It does not satisfy the triangle inequality. It is considered
as a semi-metric version of the Jaccard Index.
4.3 Difficulty of the User-to-User Collaborative
Filtering Algorithms
User-to-User collaborative filtering algorithmshavebecome
popular in many domains but have several things that could
be improved that make them more challenging to use
effectively. These include sparsity in the data, scalability
issues, lack of diversity in recommendations, the cold start
problem, and privacy concerns. While thesealgorithmshave
been successful, there is still a need to address these
challenges to improve their accuracy and effectiveness.
4.4 Proposed Improvement
An approach has been proposed to improve the challenges
faced by user-to-user collaborative filtering algorithms. It
analyses the user-item matrix using different similarity
metrics to identify similarities and relationships among
different products. The Jaccard index similarity metric
provides more accurate recommendations to users in less
time.
Similarity measures like Pearson's correlation coefficient
and cosine similarity are limited in their ability to make
recommendations and are susceptible to the cold-start
problem. The proposed approach uses co-related and non-
related ratings to improve performance and enhance the
quality of recommendations.
5. Implementation
This segment explains implementing a movie
recommendation systemusingPythonprogramming withK-

Nearest Neighbour. Implementing the system involves
several sub-sections:
Dataset: The MovieLens 100k dataset is a popular dataset
used as a standard benchmark in the field of recommender
systems. It comprises 100,000 ratings of 1,682 movies,
provided by 943 users. The ratings are scored on a scale of 1
to 5, with 5 being the highest. Demographic data about the
users, including their age, gender, occupation, and zip code,
is also included in the dataset, as well as details about the
movies, such as the movie title, release year, and genre.
Data Cleaning: Before building the model, the data must be
pre-processed to clean and transform it intoa usableformat.
It involves handling missing data, removing duplicates, and
transforming it into a matrix form suitable for analysis.
Model Analysis: Once the data has been cleaned and pre-
processed, it is necessary to analyse it to gain insights and
determine the appropriate model. It involves exploring the
relationships between variables, identifying trends, and
selecting relevant features.
Model Building: In this case, the K-Nearest Neighbour
algorithm predicts movie ratings based on similar users'
preferences. The model is trained on the MovieLens 100k
dataset, and its performance is evaluated using various
metrics such as RMSE and MAE.
Finally, the model is used to make predict ratings for non-
rated items and the results are displayed to the user in the
form of recommended movies.
5.1 Technology
Python: It is a dynamically typed programming language
used for web development, software, automation, data
analysis, and visualisation. It is easy to learn and used by
non-programmers for everyday tasks.
NumPy: NumPy is a Python library that enablesthecreation
and manipulation of multi-dimensional arraysandmatrices,
with a rich set of high-level mathematical functions to
facilitate tasks such as linear algebra, Fourier transforms,
and array manipulation.
Pandas: A Python library for working with data sets,
including analysis, cleaning, exploration, and manipulation.
It provides fast and expressive data structures.
Matplotlib: Matplotlib is a Python library that facilitatesthe
creation of visualizations in various forms, including static,
animated, and interactive graphics.
Scikit-learn: A free machine-learning library with tools for
statistical modelling and machine learning.
Surprise: A Python library for building and evaluating
recommender systems. It supports collaborative filtering
and matrix factorisation techniques,parallel processing,and
hyperparameter tuning.
Google Collaboratory: Google Collaboratory is a cloud-
based Python Integrated Development Environment (IDE)
launched in 2017 by Google, which provides data scientists
with a platform to develop machine learning and deep
learning models. It offers cloud storage capabilities to store
and share notebooks, datasets, and other files with
collaborators.
6. Results and Discussions
6.1 Discussion
Example 1: Sample Data (5*5)
The sample dataset consists of ratings for fiveitemsgiven by
five users where Item 5 is not rated by user Alice.
Item 1 Item 2 Item 3 Item 4 Item 5
Alice 5 3 4 4 ?
User 1 3 1 2 3 3
User 2 4 3 4 3 5
User 3 3 3 1 5 4
User 4 1 5 5 2 1
Tab.1 User rating for 5 Items
Tab.1 represents the user ratings for five items, where Alice
did not provide a rating for Item 5. Thisstudyaimstopredict
the rating for the non-rated item using the "User-Based
Collaborative Filtering Approach".Itusesdifferentsimilarity
metrics, such as Pearson Correlation Coefficient, Jaccard
Similarity Coefficient, and Sorensen Dice Coefficient, to
determine user similarity. Based on the similarity score, the
non-rated item rating is predicted.
Similarity comparison
Users
Pearson
Correlation
Coefficient
Jaccard
Similarity
Coefficient
Sorensen Dice
Coefficient
Sim (Alice,
User1)
0.85 0.2 0.25
Sim (Alice,
User2)
0.70 0.6 0.5

Sim (Alice,
User3)
0.00 0.5 0.5
Sim (Alice,
User4)
- 0.79 0.2 0.25
Tab.2 Users Similarity Comparison by using different
similarity metrics
Tab.2 compares the similarity scores between Alice and the
other users in the data using three different similarity
metrics: Pearson Correlation Coefficient, Jaccard Similarity
Coefficient, and Sorensen Dice Coefficient. The table shows
that the similarity scores betweenAliceandUser1andUser2
are relatively high, indicating a strongersimilarity.However,
the Jaccard Similarity Coefficient and Sorensen Dice
Coefficient give higher scores for Alice's similarity with
User2 and User3. Based on these similarity scores, the KNN
algorithm will predict the rating for item 5.
Fig.8 Shows the bar chart comparing user similarity
scores using different similarity metrics.
In fig.8, the x-axis shows the users, while the y-axis shows 3
the similarity score. Three different similarity metrics are
used to calculate the similarity between Alice and other
users. The bar chart shows that user 1 has the highest
similarity score using the Pearson Correlation coefficient.
Whereas, User 2 has the highest similarity score using
Jaccard Similarity Coefficient. User 3 and user 4 have low
similarity scores with Alice using Pearson Correlation
Coefficient but have average similarity scores using Jaccard
Similarity Coefficient and Sorensen Dice Coefficient.
Pearson
Correlation
Coefficient
Jaccard
Similarity
Coefficient
Sorensen Dice
Coefficient
Pred(Alice,
Item5)
2.25 3.91 3.67
Tab.3 Comparison of Prediction Value using different
similarity metrics.
Tab. 3 shows that the predicted rating value for Alice and
Item 5 is highest using the Jaccard Similarity Coefficient,
followed by Sorensen Dice Coefficient, and then Pearson
Correlation Coefficient.
Fig.9 Shows the bar chart comparing the predicted value
for non-rated items.
In Fig.9, a comparison of the predicted rating values for the
non-rated item (Alice, Item 5) is presented using three
different similarity metrics. The Jaccard Similarity
Coefficient is observed to yield the highest predicted rating
value for (Alice, Item 5), with a value of 3.91.
Fig.10 Shows the comparison of execution time for
different similarity metrics.

Fig.10 displays a bar chart where the x-axis shows the
similarity metrics, and the y-axis shows time. The chart
indicates that the Jaccard Similarity Coefficient takes less
time than the Pearson Correlation Coefficient.
Example 2: Movielens 100k Dataset
The MovieLens 100K dataset is widely used as a standard
reference forevaluatingcollaborativefilteringrecommender
systems. It comprises of ratings given by users to movies on
a 5-point scale, ranging from 1 to 5. The dataset contains
100,000 ratings and is made up of 943 unique users and
1682 unique movies. Additionally, each user in the dataset
has rated a minimum of 20 movies.
When evaluating a filtering technique, statistical accuracy
metrics are employed to measure accuracy. These metrics
directly compare the predicted ratings with the actual user
rating. Commonly used statistical accuracy metrics include
Mean Absolute Error (MAE), Root Mean Square Error
(RMSE), and Correlation.
MAE is the most popular and widely used. It measures the
deviation of the recommendation from the user's specific
value.
Root Mean Square Error (RMSE)metric givesmore weightto
larger deviations in prediction and is used to evaluate the
accuracy of recommendation engines. The lower the RMSE
value, the higher the accuracy of the prediction.
6.2 Results :
Comparision Table
Similarity Metric RMSE MAE
Pearson Correlation 0.9586 0.751
Cosine similarity 0.9645 0.7631
Jaccard Coefficient 0.8656 0.7276
MSD 0.9569 0.7548
SVD 0.9508 0.7492
Tab.4 Comparison of RMSE and MAE values for different
similarity metrics
Tab.4 compares the RMSE (Root Mean Squared Error) and
MAE (Mean Absolute Error) values for different similarity
metrics. The metrics compared are Pearson Correlation,
Cosine Similarity, Jaccard Coefficient, MSD (Mean Squared
Difference) and SVD (Singular Value Decomposition).
The highest values for both RMSE and MAE are observed for
Cosine Similarity, followed by Pearson CorrelationandMSD.
The lowest values for both metrics are observed for Jaccard
Coefficient and SVD. Hence, the metrics with higher RMSE
and MAE values indicate a larger prediction error, while
lower values indicate better accuracy.
Fig.11 Shows the bar chart comparing RMSE and MAE
values of different similarity metrics.
Fig.11 displays a bar chart showing Cosine Similarity's
highest values for RMSE and MAE. The Jaccard Coefficient
has the lowest RMSE and MAE values, indicating thatitisthe
most accurate metric for predicting user-item ratings.
However, the differences between the metrics are relatively
small.
Some of the Advantages of the Jaccard index:
The Jaccard index has several advantages, including
1) It can make the recommendation system more reliable by
providing accurate similarity scores betweenitemsorusers,
which can then be used to make accurate predictions.
2) It increases efficiency by reducing the number of
comparisons to calculate the similarity scores because the
Jaccard index only considers the intersection and union of
two sets, which can be computed efficiently.
3) Execution is faster because it involves simple set
operations, which can be computed quickly.
4) It gives accurate results when used appropriately,
especially for sparse data sets with missing ratings.

The Jaccard similarity coefficient performs well in the user-
based collaborative filtering approach because it findsusers
with similar preferences. It can help to predict accurate
ratings for non-rated items.
In the item-to-item collaborative filtering approach, the
choice of similarity metric may not significantly impact the
predicted values because the focus is on finding similar
items rather than users. However, the Jaccard index can still
be a useful metric in this context because it can help identify
items that are frequently rated together, which can be used
to make recommendations based on users' ratings of other
items.
6.1 Comparison
Stitini et al. [31] The paper"InvestigatingDifferentSimilarity
Metrics Used in Various Recommender Systems Types &
Scenario Cases" explores the performance of different
similarity metrics. The authors conduct an empirical study
on four types of recommender systems: user-to-user, item-
to-item, content based, and hybrid. They compare the
performance of eight similarity metrics: Pearson,Spearman,
Cosine, Jaccard, Adjusted cosine, Euclidean, Manhattan and
Mean Squared Distance (MSD).
The study finds that the performance of the similarity
metrics varies for different types of recommender systems
and scenario cases. For user-to-user and item-to-item
recommender systems,Cosineand Jaccardsimilaritymetrics
perform better than the other three metrics. For content-
based recommender systems, Pearson correlation and
Cosine similarity perform the best. In hybrid recommender
systems, the choice of similarity metric depends on the
specific scenario case.
Earlier studies on recommender systems compared various
similarity metrics and found that the cosine and Jaccard
metrics performed better than others in user-based
recommender systems. This current paper builds upon that
previous work and concludes that the model's performance
is optimal when using the Jaccard coefficient similarity
metric. Based on RMSE (0.8656) and MAE (0.7276) values,
the Jaccard metric is shown to predict recommendations
more accurately than all other similarity metrics.
7. Conclusion & Future Scope
7.1 Conclusion:
Recommender/Recommendation systems are becoming an
essential tool for E-commerce on theweb. Theyhelpusers to
find items they like and increase sales for businesses. With
the massive user data volume, recommender systems need
new technologies to improve scalability.
This study presented and evaluated various similarity
metrics like Pearson Correlation, Cosine Similarity, Jaccard
Coefficient, MSD (Mean Squared Difference), Sorensen Dice
Coefficient and SVD (Singular Value Decomposition) for
User-to-user collaborative filtering recommender systems.
The results show that the Jaccard index can perform well for
large data sets while producing high-quality
recommendations.
Hence, the Jaccard index is useful when the data is sparse
and missing values are common. Its advantages include
improved reliability, increased efficiency, faster execution,
and accurate results.
7.2 Future Scope:
1) Enhance the security measures to prevent fake ratings or
user manipulation.
2) Enhance the evaluation approach for scenarios where
there are no ratings available.
 Adopt proactive recommendation systems.
 Utilize privacy-preserving recommendation
systems.
3) Implement a Deep Neural Networks recommendation
system that provides dynamic results/recommendations.
8. REFERENCES
[1] Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., and
Riedl, J. (1994). GroupLens: An Open Architecture for
Collaborative Filtering of Netnews. In Proceedings of CSCW
'94, Chapel Hill, NC.
[2] Breese, J.S., Heckerman, D., Kadie, C. (1998). Empirical
analysis of predictive algorithms for collaborative filtering.
Proceedings of the 14th Conference on Uncertainty in
Artificial Intelligence.
[3] Herlocker, J., Konstan, J.A., Terveen, L., Riedl, J. (2004).
Evaluating collaborative filtering recommender systems.
ACM Transactions on Information Systems (TOIS) 22.
[4] Deshpande, M., and Karypis, G. (2004). Item-based top-n
recommendation algorithms. ACM Transactions on
Information System (TOIS), 22(1),143–177.
[5] Sarwar, B., Karypis, G., Konstan, J., and Riedl, J. (2001).
Item-based collaborative filtering recommendation
algorithms. In Proceedings of the 10th international
conference on World Wide Web (pp. 285-295).
[6] Fkih, F. (2021). Similarity measures for Collaborative
Filtering-based Recommender Systems: Review and
experimental comparison. Journal of King Saud University-

Computer and Information Sciences, 33(4), 431-449.
https://www.sciencedirect.com/science/article/pii/S13191
57821002652?via%3Dihub
[7] D. Goldberg, D. Nichols, B. M. Oki, and D. Terry, “Using
collaborative filtering to weave an information tapestry,”
Communications of ACM, vol. 35, no. 12, pp. 61–70, 1992.
[8] Manos Papagelis and Dimitris Plexousakis “Qualitative
analysis of user-based and item-basedpredictionalgorithms
for recommendation agents” Engineering Applications of
Artificial Intelligence 18 (2005) 781–789
www.elsevier.com/locate/engappai
[9] Bell, R., & Koren, Y. (2007). Scalable Collaborative
Filtering with Jointly Derived Neighbourhood Interpolation
Weights. IEEE International Conference on Data Mining
(ICDM’07), pp. 43–52.
[10] Saranya, K. G., Sudha Sadasivam, G., & Chandralekha, M.
(2016). Performance Comparison of Different Similarity
Measures for Collaborative Filtering Technique. Indian
Journal of Science and Technology, 9(29), 1-7. DOI:
10.17485/ijst/2016/v9i29/91060
[11] A Case-Based Recommendation Approach for Market
Basket Data Anna Gatzioura and Miquel Snchez-Marr IEEE
INTELLIGENT SYSTEMS 2015.
[12] Recommender Systems: An overview of different
approaches to recommendationsKunal Shah,AkshayKumar
Salunke, Saurabh Dongare, Kisandas Antala SIT, Lonavala
India 2017
[13] Recommendation analysis on Item-based and User-
based Collaborative Filtering Garima Gupta, Rahul Katarya,
India
[14] Using collaborative filtering to weave an information
Tapestry D. Goldberg, D. Nichols, B. M. Oki, and D. Terry,
Communications of the ACM, vol. 35, no. 12, pp. 6170,1992
[15] Recommendersystems,Handbook,FrancescoRicci,Lior
Rokach, Bracha Shapira, Paul B. Kantor. Springer 2010.
[16] Zhao, Zhi-Dan, and Ming-Sheng Shang. “User-based
collaborative filtering recommendation algorithms on
Hadoop.” In 2010 Third International Conference on
Knowledge Discovery and Data Mining, pp. 478-481. IEEE,
2010
[17] P. W. Yau and A. Tomlinson, “Towards Privacy in a
Context Aware Social Network Based Recommendation
System,” Privacy, Security, Risk and Trust (PASSAT) and
2011 IEEE Third International Conference on Social
Computing (SocialCom), 2011 IEEE Third International
Conference on, Boston, MA,2011, pp. 862-865.
Doi:10.1109/PASSAT/SocialCom.2011.87
[18] Gao, Min, Zhongfu Wu, and Feng Jiang. “User rank for
item-based collaborative filtering recommendation.”
Information Processing Letters 111, no. 9 (2011): 440-446.
[19] Grcar, M., Fortuna, B., Mladenic, D., Grobelnik, M.: k-NN
versus SVM in the collaborative 3604iltering framework.
Data Science and Classification pp. 251260 (2006).
[20] Hofmann, Collaborative filtering via Gaussian
probabilistic latent semantic analysis. In: SIGIR 03: Proc.Of
the 26th Annual Int. ACM SIGIR Conf. On Research and
Development in Information Retrieval, pp. 259266. ACM,
New York, NY, USA (2003).
[21] Bell, R., Koren, Y., Volinsky, C.: Modelingrelationshipsat
multiple scales to improve the accuracy of large
recommender systems. In: KDD 07: Proc. Of the 13th ACM
SIGKDD Int. Conf. On Knowledge DiscoveryandData Mining,
pp. 95104. ACM, New York, NY, USA (2007)
[22] Wikipedia link
https://en.wikipedia.org/wiki/Collaborative_filtering
[23] Recommender Systems – The Textbook | Charu
C.Aggarwal | Springer. Springer. 2016. ISBN
9783319296579.
[24] “A Study of Hybrid Recommendation Algorithm Based
On User” Junrui Yang1, Cai Yang2, Xiaowei Hu3 2016 8th
International Conference on Intelligent Human-Machine
Systems and Cybernetics
[25] Gomez-Uribe, Carlos A.; Hunt,Neil (28December2015).
“The Netflix Recommender System”. ACM Transactions on
Management Information Systems. 6(4): 1–19.
doi:10.1145/2843948
[26] A Study of Hybrid Recommendation Algorithm Based
On User Xian University of Science and Technology Xian,
China
[27] Crowe, N. (n.d.). Absolute bounds on set intersection
and union sizes. Retrieved March 8, 2023, from
https://faculty.nps.edu/ncrowe/intersect2.htm
[28] Recommender Systems in E-Commerce J. Ben Schafer,
Joseph Konstan, John Riedl GroupLens Research Project
DepartmentofComputerScienceandEngineeringUniversity
of Minnesota Minneapolis, MN 55455 1-612-625-4002
[29] Rajeev Kumar, Guru Basava, Felicita Furtado, "An
Efficient Content, Collaborative – Based and Hybrid
Approach for Movie Recommendation Engine" Published in
International Journal of Trend in Scientific Research and
Development (ijtsrd), ISSN: 2456- 6470, Volume-4 | Issue-3,
April 2020, pp.894-904, URL:
www.ijtsrd.com/papers/ijtsrd30737

[30] Maddali Surendra Prasad Babu, and Boddu Raja Sarath
Kumar. An Implementation of the User-based Collaborative
Filtering Algorithm / (IJCSIT) International Journal of
Computer Science and Information Technologies, Vol. 2 (3),
2011, 1283-1286
[31] Stitini, O., Kaloun, S., & Bencharef, O. (2022).
Investigating Different Similarity Metrics Used in Various
Recommender Systems Types: Scenario Cases. The
International Archives of the Photogrammetry, Remote
Sensing and Spatial Information Sciences, XLVIII-4/W3-
2022, 327-334. doi:10.5194/isprs-archives-XLVIII-4-W3-
2022-327-2022.
9. AUTHOR’S PROFILES
Kopanathi Sonali currently
pursuing M.Tech- Data Science
(CSE) from GITAM Institute of
Technology, GITAM (Deemed to be
University), Visakhapatnam. Her
areas of research work aremachine
learning and deep learning. Her
areas of interest are Recommended
Systems, Neural Networks anddata
science.
Dr. S.V.G. REDDY completed M-Tech
(CST) from Andhra University and
has obtained a PhD in Computer
Science and Engineering fromJNTU
Kakinada. He is an Associate
Professor, the Department of CSE,
GIT, GITAM University. His area of
research work is data mining,
machine learning and deep neural
networks. He has guided various B.
Tech and M. Tech projects and has
publications in several journals.His
areas of interest aredrugdiscovery,
computer vision, brain-computer
interface, climate change and waste
management.
K. Thammi Reddy completed his
PhD (computer science and
engineering) during 2008 from
JNTU Hyderabad. He is working as
Professor, Department of CSE, GIT
and serving as Director, IQAC,
GITAM University. His area of
research work is data mining,
machine learning and cloud
computing with Hadoop. He has
guided various B. Tech, M. Tech
projects and had publications in
several reputed journals and
conferences.
V. Valli Kumari completed her PhD
(CSSE) from Andhra University
during 2006. She is working as
professor, Department of CSSE,
college of engineering, Andhra
University. Her area of interest is
Software engineering, Network
Security and Cryptography, Privacy
issues in Data Mining and Web
Technologies. She has guided
various B.Tech,M.Techprojectsand
had publications in several reputed
journals and conferences. She
received best researcherawardand
other various awards inthefieldsof
teaching and research. She has
undergone and completed various
research projects.

Analysing the performance of Recommendation System using different similarity metrics

More Related Content

Similar to Analysing the performance of Recommendation System using different similarity metrics (20)

More from IRJET Journal (20)

Recently uploaded (20)

Analysing the performance of Recommendation System using different similarity metrics