0% found this document useful (0 votes)
100 views17 pages

Movie Recommendation System

This document discusses several research papers on movie recommendation systems using graph-based and machine learning algorithms. It provides a comparison table summarizing key aspects of different papers, including the algorithms and datasets used. The table shows that papers have used techniques like graph traversal, autoencoders, BERT, sentiment analysis, KNN and its variants to build recommendation systems and evaluate their performance on popular movie rating datasets.

Uploaded by

Mritunjay Jha
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
100 views17 pages

Movie Recommendation System

This document discusses several research papers on movie recommendation systems using graph-based and machine learning algorithms. It provides a comparison table summarizing key aspects of different papers, including the algorithms and datasets used. The table shows that papers have used techniques like graph traversal, autoencoders, BERT, sentiment analysis, KNN and its variants to build recommendation systems and evaluate their performance on popular movie rating datasets.

Uploaded by

Mritunjay Jha
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 17

Term Paper On Movie Recommendation Based on Graph Traversal

Algorithms
Mritunjay Jha 11906046, Rizul Sharma 11910306 Krishna Bhabat 11916423
School of Computer Science and Engineering
Lovely Professional University

Comparison Table:
Graph- Zahra Elsevier Autoencode Relation between
based Zamanza r and Hybrid users based on
hybrid deh recommend their similarities
recommend Darban, ation as nodes in a
ation system Mohamm similarity graph
with ad Hadi and combining
application Valipour them with the
to movie users’ side
recommend information to
ation solve the cold-
start issue.

Movie Raja Scientific Cosine Cosine similarity


Recommend Marappan Publicati similarity algorithm is a
ation , ons and good fit for the
System S.Bhaskar Sentiment movie
Modeling an analysis recommendation
Using system since it is
Machine fast and
Learning accurate. The
SVM classifier
performs
significantly
better than the
NB classifier
when identifying
movie reviews.
Improving CheonSol MDPI BERT and https://www.kaggle.com/ improved graph-
Graph- Lee, IGMC stefanoleone992/rotten-tomatoes- based movie
Based DongHee movies-and-critic-reviews-dataset recommender
Movie Han, system using
Recommend Keejun Cinematic
er System Han, Mun Experience with
Using Yi BERT by building
Cinematic a novel stacking
Experience ensemble
architecture.
Cinematic
Experience (CX)
manifested via
sentiment and
emotion can
contribute to
improving the
recommendation
performance.
Movie Sandipan IEEE Multi- Our approach to
Popularity Sahu, dimensional focused movie
and Target Raghvend KNN, Deep popularity and
Audience ra Kumar, learning, finding out the
Prediction Mohd Fuzzy C target audience
Using the Shafi mean of an upcoming
Content- Pathan, movie is very
Based Jana much unique. we
Recommend Shafi, have used a
er System Yogesh recommendation
Kumar, system to find
Muhamm similar movies
ad Fazal from a given
Ijaz movie and use
similar movies
for forecasting
purposes.
Movie Sonu Springer KNN-Basic, https://www.kaggle.com/datasets/ This paper
Recommend Airen, KNN- grouplens/movielens-20m-dataset proposes
er System Jitendra WithMeans, different
Using K- Agrawal KNN- variations of K-
Nearest WithZScore nearest
Neighbors , KNN- neighbors (KNN)
Variants Baseline algorithm with
different
similarity
measures
namely cosine,
msd, pearson
and pearson
baseline for
Movie
Recommender
System. These
different
variations of
KNN algorithms
have been
implemented for
real data from
MovieLens
dataset and
compared on
accuracy metrics
like fraction of
concordant
Pairs, mean
absolute error,
mean squared
error, root mean
squared error,
precision@k and
recall@k for
Movie
Recommender
System.
A systematic Deepjyoti Springer Content- This review has
review and Roy, Mala based gathered diverse
research Dutta filtering, details like
perspective collaborativ different
on e filtering application fields,
recommend approaches techniques used,
er systems , simulation tools
optimization used, diverse
filtering applications
approaches focused,
performance
metrics, datasets
used, system
features, and
challenges of
different
recommender
systems. Overall,
this paper
provides a
comprehensive
understanding of
the trend of
recommender
systems-related
research and to
provides
researchers with
insight and future
direction on
recommender
systems.
Multimodal Yongheng MDPI multimodal https://movielens.org/ 0.9908 he multimodal
Movie Mu data and movie
Recommend andYun analysis 0.9096 recommendation
ation Wu and deep for test system using
System learning sets of deep learning
Using Deep MovieL could provide
Learning ens better
100 K personalized
and 1 recommendation
M results.
dataset Meanwhile, the
s sparse data
problem was
alleviated to a
certain degree.
We suggest that
the
recommendation
system can be
improved
through the
combination of
the deep-
learning
technology and
the multimodal
data analysis.
A Matrix Dan Liu, Arxiv UISVD++ https://movielens.org/ UISVD++ model
Decompositi Hou-biao model is superior to
on Model Li other baseline
Based on methods in rating
Feature prediction, and
Factors in can efficiently
Movie improve the
Recommend recommendation
ation accuracy and
System address the
problem of data
sparsity. The
model can
extract the
underlying
factors
from the user-
item ratings
matrix and
capture the
implicit
relationship
between the user
and item.
Optimal Deepjyoti Springer water cycle https://www.kaggle.com/code/ HAN of A new technique
hierarchical Roy, Mala algorithm laowingkin/netflix-movie- 89.81 for sentimental
attention Dutta and recommendation/notebook %, analysis-based
network- earthworm MSE of movie
based optimization 0.1,RM recommendation
sentiment algorithm, SE of is devised that
analysis for ROCK 0.430. helps the users
movie algorithm,K to choose the
recommend endall Tau positive
ation distance recommended
measure movie reviews.
an efficient
movie
recommendation
based on the
sentiment
analysis is done
by providing a
suitable positive
recommended
movie reviews to
users.
A Movie ZAKARIA Researc K-nearest https://movielens.org/ the an efficient
Recommend SULIMAN hgate neighbours Apriori hybrid movie
ation ZUBI , ALI (KNN) algorith recommender
System A. algorithm,A m, system has been
Design ELROWA priori dataset designed using
Using YATI , algorithm with the association
Association IBRAHIM subsystem minimu rules mining
Rules SAAD m technique and
Mining and ABU suppor collaborative
Classificatio FANAS t was filter technique.
n 50, ts were
Techniques and a evaluated in
confide term of the
nce important
value degree. The
60% proposed system
improves the
important degree
and gives better
accuracy than
the existing
techniques used.
KNN and Apriori
algorithm
improved the
lists of
userrecommend
ed movies that
are close to their
liking, depending
on which movie
the user selects
the first time
Systematic R IEEE SVD++ Hybrid Movie
analysis of Lavanya, Algorithm Recommendatio
Movie B. Bharati n System using
Recommend sentiment
ation analysis with
System opinion mining, is
through a system that
Sentiment helps the user by
Analysis giving
recommendation
s about the
movies, based
on the user
preference and
ordinal ratings of
the movies. It
removes the
need for the user
to go through a
big list of movies
before deciding
what should be
watched.
A R ICAIS K-means While talking
Comprehen Lavanya, Algorithm, about the
sive Survey Utkarsh HAC computational
on Movie Singh, Algorithm, time of a movie
Recommend Vibhor recommendation
ation Tyagi system, few
Systems models have
also been
prepared which
keep the
timeliness and
scalability of the
system in mind.
For ex. The
authors of paper
[18] developed a
simple but
effective real
time movie
recommendation
system. Firstly,
they partitioned
users by
evaluating their
user profile
attributes and
using a simple
but highly
efficient
algorithm.
Secondly, a
virtual opinion
leader is chosen
for every cluster
who is set to
represent the
opinions of that
particular cluster
on behalf of all
other users in
that cluster
Similarity Raghaven ICICCS CF There are
Based dra C K; Algorithm, various
Collaborativ K.C. RS approaches to
e Filtering Srikantaia Algorithm build these
Model for h recommendation
Movie s few of them are
Recommend Content-Based
ation (CBF), Filtering,
Systems Demographic,
Knowledge-
Based,
Collaborative
Filtering (CF)
and Hybrid
approaches. One
of the most
widely used
techniques is
Collaborative
Filtering which
filters the
information for a
target user
based on the
details of users
having similar
interests. The CF
algorithms are
classified as
user-based and
item-based.
Enhanced Tharun ICOSEC Clustering KG is a
Movie Thomas Algorithm, knowledge base
Recommend Ajith, Ajay Page Rank that integrates
ation using Krishnan Algorithm data and
Knowledge C V, summarises
Graph and Nandakis relationships
Particle hore J using a graph-
Filtering structured data
model. First
introduced and
popularized by
Google in the
year 2012 under
the term
”knowledge
map”. It offers a
framework and a
common
interface for all
data, allowing
the construction
of smart
multilateral
relations
between
databases and
the use of
exploratory
searches when
working with
large amounts of
data
Movie Christ ICITech Collaborativ the collaborative
Recommend Zefanya e filtering filtering
ation Omega;H Algorithm, recommendation
System endry SVD system is made
using Algorithm using the SVD
Weighted (matrix
Average factorization)
Approach algorithm. The
SVD algorithm
will form a
pattern where
the matrix of
ratings is
factored into a
matrix product
representing the
item's latent
factor. This
algorithm
establishes a
recommendation
system that
allows us to "fill
in the gaps" in
the rating matrix
and predict each
user's rating for
each item in the
dataset to
minimize errors
when predicting
existing ratings.
The SVD
algorithm can
also measure
RMSE and MAE.
Movie Harsh ICIRCA- K-means Recommender
Recommend Katter, 2021 Algorithm, K systems can be
ation Nishita nearest utilized in many
System Goel, Algorithm, contexts to
using Naina KNN generate
Cosine Gupta Algorithm recommendation
Similarity s to the users
with that might
Sentiment interest them.
Analysis Recommendatio
n systems were
first developed
by Tapestry
project in
1992[5]. The
Tapestry project
(first
commercialreco
mmender
system)
introduced the
term
“collaborative
filtering” Existing
movie
recommendation
systems are
mostly built using
the content
based and
collaborative
filtering approach
Factors that Mariam ICCCNT K-means, The ultimate goal
has Effect Akter, CNN, RNN, of this paper is
on Early Nishat SVD how to
Movie Sultana,M recommend a
Recommend d. Fahad movie at a better
ation: A Hossain approach.
Systematic Hopefully, By
Literature looking at our
Review work new
researchers will
be able to get an
idea of the
previous work
and create an
idea of what they
will do in the
future.
Personalize Mariam ICCCNT K-means, The system that
d Movie Akter, RNN mostly values
Hybrid Nishat user choice in
Recommend Sultana,M predicting or
ation Model d. Fahad filtering
Based on Hossain preferences from
GRU big data is called
a recommender
system.
Recommender
systems can
predict whether a
particular user
would prefer an
item or not based
on the user's
profile. The
motives of this
recommendation
system can be
summed up as
“Help the users
find relevant
items”. The main
operationalizatio
n of this goal is
focused on the
capability to
predict the user’s
preferences
numerically for
the items that are
not watched by
the users or to
give them a list
of items based
on the prediction
Analysis of Shavak ICCMC Boosting music
Intelligent Chauhan, Algorithm, recommendation
movie Rajdeep SVM system is
recommend Mangrola Algorithm constructed by
er system physiological
from facial sensors. Here
expression emotion is
recognized
through
physiological
sensors by
galvanic skin
response (GSR)
and
photoplethysmog
raphy (PPG) [4].
This data is then
used for the
recommendation
purpose of the
music. It can be
seen that it used
real-time data of
the body and
mind and then
convert it to
music which
improves the
recommender
system over the
traditional
recommender
system.
Integrating Qiming IAEAC K-GBDT Based on the
KNN and Du; Nan Algorithm, idea of matrix
Gradient Li; XGBoost, decomposition,
Boosting Shudan CatBoost the hidden
Decision Yang; Algorithm vectors of user
Tree for Daozhu items are mined
Recommend Sun; to improve
ation Wenfu Liu prediction
accuracy. The
classic
collaborative
filtering algorithm
has obvious
head effects and
weak
generalization
ability. To solve
this problem, the
algorithms of
matrix
factorization
family such as
SVD(Singular
Value
Decomposition)
[4], NMF (Non-
negative Matrix
Factorization)[5]
have been
proposed one
after another. By
mining potential
features of users
and items from
rating data, the
accuracy of
rating prediction
has been greatly
improved.
Movie Akansh IRJET Collaborativ proposed
recommend Surendra e filtering, algorithm uses
ation system n, Aditya Content textual metadata
using Kumar based of the movies like
machine Yadav, filtering plot, cast, genre,
learning Aditya release year and
Kumar other production
information to
analyse them
and recommend
the most similar
ones.

Learning Xusong IEEE CBF, CF, n MovieLens10M a we consider


and Fusing Chen, RNN efficient
Multiple Dong representations
User Liu,Zhiwei of four aspects of
Interest Xiong,Zhe user interest:
Representati ng-Jun first, we use
ons for Zha latent
Micro-Video representation,
and Movie i.e. user
Recommend embedding, to
ations profile the overall
interest; second,
we propose item-
level
representation,
which is learned
from and
integrates the
features of a
user’s historical
items; third, we
investigate
neighbor-
assisted
representation,
i.e. using
neighboring
users’
information to
characterize user
interest
collaboratively;
fourth, we
propose
category-level
representation,
which is learned
from the
categorical
attributes of a
user’s historical
items
Analysis of Muppana IEEE Collaborativ Movie-Lens-100K paper examines
Movie Mahesh e filtering the
Recommend Reddy; R. recommendation
ation Sujithra s that are
Systems; Kanmani; obtained with
with and B. and without
without Surendira considering the
considering n movies that have
the low never got an
rated above-average
movies rating, where
average rating is
defined here as
the mid-value
between 0 and
maximum rating
used.The
technique used
is “collaborative
filtering” and the
similarity
measure used is
the “Pearson
correlation
coefficient”
Movie Nimish IEEE SVM,KNN,T Sentiment
Recommend Kapoor,S ext mining analysis system
ation aurav is built based on
System Vishal, the TMDB
Using NLP dataset and used
Tools SVM to predict
positive negative
sentiment from
the user's movie
review and then
movies are rated
based on the
sentiment
analysis and
scoring of the
reviews
Machine M. IEEE KNN,collab If a movie is
Learning Chenna orative rated excessive
Model for Keshava, filtering by means of a
Movie S. algorithm consumer who
Recommend Srinivasul also watched the
ation u,P. movie you are
System Narendra watching now,
Reddy,B. it's miles possibly
Dinesh to show up inside
Naik the
recommendation
s.The films with
the highest
overall scores
are in all
likelihood to be
enjoyed by way
of nearly
everyone. The
algorithm which
does all these
features is called
CineMatch. For
personal users, it
also learns from
the conduct of
the person to
higher expect a
movie the
consumer is
anticipated to be
fascinated in.
Here we have to
increase our
CineMatch
algorithm 10% by
using
fashionable
collaborative
filtering
techniques.
Movie F. IRJET content- introduced a
recommend Furtado, based and model combining
ation system A. Singh collaborativ both content-
using e approach based and
machine collaborative
learning approach. It will
give
progressively
explicit outcomes
compared to
different systems
that are based
on content-based
approach.
Content-based
recommendation
systems are
constrained to
people, these
systems don’t
prescribe things
out of the box,
thus limiting your

worse. There are literally millions of options available to the


Abstract— Today, it's a regular issue to propose media user. Without a support that filters these options, the work
content. Traditional algorithms based on collaborative becomes quite challenging.
filtering necessitate a current dataset of users and their Systems that make recommendations look into the user's
preferences, which is challenging to assemble for large
databases of things. A disadvantage of the content-based preferences, attitude, or current environment. They
approach is the difficult computation of item similarity. In this produce a precise recommendation that is appropriate for
essay, we suggest a method of recommendation that focuses on the specific user based on the analysis [1]. The methods
how user preferences for movies naturally vary over time. used to produce recommendations vary. Content-based and
Utilizing a graph representation, we tested out modified graph collaborative recommendation systems are the two main
algorithms. We provide a graph-based representation of the
data about movies and a mechanism that leverages this model
groups. For each category, we should search the whole
to make recommendations. We suggest four techniques for entity base to choose the best thing to suggest. This is a
finding recommendations based on starting nodes, whose matrix of users and items in a collaborative method; in a
selection is depending on the user's current preferences. These content-based approach, it is a matrix of items and their
algorithms were put into practice, and they were tested similarity. However, there is no guarantee that the
experimentally on actual people. recommendation is accurate enough for the user and that
Keywords:- the estimation is accurate. Many recommendation engines
attempt to match the user's context and preferences with the
I. INTRODUCTION extracted knowledge base to suggest items. As a result,
There is a lot of media content produced every day in the recommender systems may experience performance
modern world, including films, TV series, programs, etc. problems that prevent real-time use.
Choosing the proper material can be challenging and In this post, we'll design and assess a new recommendation
intimidating for the user when there are so many mechanism for recommending things with a variety of
possibilities available. With regard to television, the user qualities, like movies. Instead of using the aforementioned
has access to dozens of channels at any given time. He strategies, we chose to develop a solution that makes use of
misses something he might love on another station if he graph structure. We tested many graph algorithms to compare
choose the incorrect one. With more and more people using them and show their advantages and disadvantages. There has
televisions or videos-on-demand, the issue just becomes
been testing of graph-based recommender systems, and the
findings are encouraging [2]. We compare graph algorithms
and their changes as part of our contribution. In a web
service named Televido, we implemented our technique for
multimedia recommendation, allowing us to test out different
algorithms on actual consumers.

II. RELATED WORK

Due to the information deluge we all experience on a


daily basis, recommender systems have become highly
popular. There are methods like content-based or
collaborative filtering [1]. By employing graph
representation to expedite the process of finding the most
suited objects, both approaches could be improved.
Since we use only one type of connection between
people and things, homogenous bipartite graph
representation is a need for collaborative filtering. Mele et
al.'s [3] intriguing method used graph representations of
users' direct relationships. These connections represent the
users' behaviour. When user u found a particular item
before user v, a connection between the two users is
added. This collaborative filtering's temporal modification
is comparable to a well-known model of PageRank[4].

In case of content-based approaches we commonly use


text optimized databases to quickly compute the similarity
among items [5]. However, a graph could be one way to
hold information on items. Especially if we use
known metadata (categories, keywords, attributes) as the
content to calculate similarity. A graph could be used as
alternative to store similarity relations among items. This
actually means that we need more connection types
thus working with
heterogeneous graphs. This is typical for multidimensional
recommender systems as presented by Lee et al. [6]. movie 1 actor
person 3 person 4 person 5

Another option is to combine both collaborative filtering actor


producer

and content-based approach as proposed by Huang et al.


genre director
actor

[2]. In this work they proposed recommender system for genre


movie 2
director

producer
digital library using Hopfield net algorithm. This work person 2
genre 1

suggests to use graph algorithms to solve the problem of director


genre

movie 3
recommending items.
Research on media content recommendations frequently
person 1 genre 2
has trouble determining how similar different products are
to one another. (in content-based). Authors frequently
Figure 1. Example of the data model represented in graph structure.
suggest changes to collaborative filtering. Neural networks
were employed by Biancalana et al. [7] to train the model
which ought to be suggested. This data model structure also
just once. As a result, their data's multidimensionality was
makes it possible for us to easily identify entities that are
actually reduced, yet they were unable to offer timely
comparable to other entities. This resemblance is implicitly
recommendations without retraining the algorithm.
stored in the connections between nodes.
Chen et al.'s work on the use of graph algorithms is also
interesting[8]. In the video search, they employed a IV. RECOMMENDATION ALGORITHMS
tripartite network of users, objects, and inquiries. In terms We use graph algorithms for our recommendation
of graph-based suggesting, this approach is extremely systems, which start at the beginning of the graph and go
similar to spreading activation, which shows promise. outward. Users' interests are represented by the initial
Bogers et al. [9] also offered a similar strategy. They nodes. A user's favourite movie or genre could serve as the
employed additional contextual characteristics in their starting node. These initial nodes are either directly chosen
work. (genre, language, director, actor, etc.). They by the user as a query or implicitly chosen based on both
employed a probabilistic approach that makes use of a the user model and the user's input using conventional
matrix containing the intensity of transitions between items techniques. In reality, connecting user models to nodes is
as the user browses through them. not necessary. We only need the list of items that the user
will find relevant, and we can utilise it as a query as we go.
III. REPRESENTATION OF MOVIES AND METADATA
The algorithms search the graph for the nodes that are
We created a network structure that represents the data—
closest to all beginning nodes, and these nodes are
information about media content—to produce precise
subsequently returned as suggestions. Finding the closest
suggestions. Our suggested structure is based on the RDF1
nodes can be approached from a variety of angles,
triplet structure and was influenced by how people actually
particularly in extremely complex graphs. For this reason,
speak or write about movies in everyday life.
we developed and tested four different algorithms.
The primary components are collectively referred to as
These settings allow the recommendation algorithms to
"movies," "TV shows," and "TV programmes." These are
be fairly flexible. For instance, the same algorithm can be
the things that should be suggested. People (actors,
used with different parameters to only propose TV shows
directors) and genres are examples of other entities. The
that will air tomorrow night or movies that are presently
graph's nodes serve as these representations. The
playing in theatres.
relationships between the entities are represented by the
edges between nodes, which operate as the at-tributes of
A. Union Colors Algorithm
the movies. Since the final network only has one
component (connected graph), all movies must be The Union Colors algorithm is based on the basic
indirectly connected to one another. The attributes make Breadth-first search (BFS) graph algorithm. The algorithm
sure that the model is connected and sophisticated by using works in the following way:
real data. Every movie, for instance, is linked to several 1) Mark each initial node with a different color (naturally,
different genres. We assume the graph is one component as colors are represented by numbers).
there are roughly 40 genres and 165 thousand nodes in 2) Perform a simultaneous BFS from each initial node:
total. Figure 1 depicts a sample of the data model graph a) Enqueue all initial nodes.
structure mentioned earlier. b) Dequeue a node and visit it.
When developing the concept, we presupposed that viewers c) Add all neighbors of the visited which are yet
also select the actors or directors of the films they watch. to be visited node into the queue.
Our method makes it straightforward to identify this pattern d) Repeat from step b)
and identify the entities because of the model's structure.
3) When a node is visited:
C1: recommended movie 1 C1: recommended movie 2 C1-C2: recommended movie 3 d) Try to add all neighbors of the visited node into
the queue.
i) If the node is not in the queue, add it and
remember, that it was added by the visited
node.
C1: person 1
ii) If the node is already in the queue, do not
C1-C2:
C1: genre 1
person 2
add it but remember that it has also been
added by the visited node.
e) Repeat from step b)
3) Continue until the required number of nodes is
colored by every color, return them.
C1: initial movie 1 C2: initial movie 2
In Figure 3 the visualization of the algorithm is illustrated.
The numbers in parentheses represent the colors of each
node after one iteration of the algorithm.
Figure 2. Visualization of Union Colors algorithm.
C. Energy Spreading Algorithm
The Energy Spreading algorithm is based on Spreading
a) If it is yet to be colored, color it with the color activation, which is a method for searching associative
of the initial node networks, neural networks, or semantic networks. Similar
b) If it is already colored (with a different color), to previous algorithms, it is a variation on the simultaneous
merge the two colors into one - remember that BFS algorithm. The logic of the algorithm is following:
one color equals the other 1) Set the energy of each initial node to some constant
4) Keep merging colors until the last two colors are value.
merged and only one color remains 2) Perform a simultaneous BFS from each initial node
5) Return the current and the next required number of (as described in the previous section).
nodes in the queue as the result 3) When a node is visited its energy increases by value
E, E = En , where Ep is the energy of the parent node
p
To merge the colors we can make use of the extremely fast
Disjoint-set data structure, using which two colors can be which enqueued the visited node and n is the number
merged in nearly constant time [10]. We actually do not need of nodes the parent node enqueued.
to traverse the whole graph. We only need to merge initial 4) A node’s energy can increase multiple times, but it
colors which count reflects the number of initial nodes. only spreads it when it receives energy for the first
Visualization of this algorithm is shown in Figure 2. time.
Nodes which have the final color are going to be recom- 5) Continue until the required number of nodes is
mended. These are the nodes which come out as output visited from each initial node.
of this algorithm. It actually means that nodes which are 6) Order the nodes by their energies and return the re-
colored with final color earlier are going to be quired number of the nodes, the more energy it has
recommended as more relevant. C1 and C2 are the colors the higher it is.
of initial nodes and the algorithm is stopped when it
reaches the node ”recommended movie 3”. movie 3 (-)(-)(-) movie 4 (-)(-)(-) recommended movie 1
(-)(-)(2,1)

B. Mixing Colors Algorithm


The Mixing Colors algorithm is similar to the previously
described Union Colors algorithm. These two algorithms
differ in the way they deal with the representation and
person 2
the meeting of colors. (-)(1)(1) genre 1
The algorithm works as follows: (-)(1)(1) movie 2 (-)(-)(-)

1) Mark each initial node as visited by a different color.


2) Perform a simultaneous BFS from each initial node:
a) Enqueue all initial nodes. initial movie 1 initial person 1
(1)(1)(1,2) (2)(2,1)(2,1)
b) Dequeue a node and visit it.
c) When visiting a node, mark it with the colors
of all the nodes which enqueued it. Figure 3. Visualization of the Mixing Colors algorithm.
recommended movie 1 recommended movie 2 recommended movie 3 movie 4 movie3 (4)(5)(9) recommended movie 1
(energy = 5,33) (energy = 5,33) (energy =13,33) (4)(5)(9) (3)(2)(5)

2 2
3,33
3,33 3,33 10
2 4
2 person 2 (-)(-)(x) 4 4 2 movie 2
(3)(2)(5)

genre 1
(energy = 10) person 2 4 2
2
(energy = 10)
person 1 initial
(energy= 4) initial person 1 (1)(0)(x)
movie 1 4genre 1
(4)(5)(9) (0)(1)(x) 1
4 4 6
6
4
Figure 5. Visualization of Modified Dijkstra’s algorithm.
initial movie 1 Initial movie 2
(energy = 12) (energy = 12)

Figure 4. Visualization of Energy Spreading algorithm.

Visualisation of the algorithm is illustrated in Figure 4.


The numbers on the arrows and their width represent
the amount of energy which is being received by the re-
ceiving nodes. In the case shown in the figure, the movie
number 3 is the first in the list of recommended nodes.
D. Modified Dijkstra’s Algorithm
The Modified Dijkstra’s Algorithm is, as the name sug-
gests, based on the well known Dijkstra’s algorithms for
finding the shortest path in a graph. We implement a vari-
ation on this algorithm, which works for multiple initial
nodes.
1) Run the Dijkstra’s shortest path algorithm from each
node to (some constant) maximal depth - calculate
the shortest path from the starting node to each
visited node.
a) If the algorithm is ran from the first initial node,
put all visited nodes with the value of the
shortest path as their total value into a results
set.
b) If the algorithm is ran from other initial nodes,
check if the node is in the results set. If yes, add
the shortest path value to its total value.
2) Check if each node in the results set was visited from
all initial nodes.
3) Order the nodes in the results set by their total values
of the shortest paths.
4) Return the required number of nodes from the
ordered results set.
The advantage of this algorithm is that it can be used
on a graph with weighted edges. For our recommendation
system it means that, for example, we can set all the edges
which connect genres to movies to be longer than the edges
which connect people to movies.
The logic of the algorithm is illustrated in Figure 5. The
first two numbers in parentheses refer to the distance of
each node to the initial nodes. The last value is their sum or
x, which represents that the node was not visited from both
initial nodes.
V. CONCLUSIONS of the 2nd Challenge on Context-Aware Movie
In this term paper, we provide a strategy for Recommendation. ACM, 2011, pp. 5–10.
recommending media content that is based on graphs. Four [8] B. Chen, J. Wang, Q. Huang, and T. Mei, “Personalized
different recommendation algorithms were proposed video recommendation through tripartite graph propagation,”
together with a graph-based data architecture. Although the in Proc. of the 20th ACM int. conf. on Multimedia - MM ’12.
examination reveals encouraging findings, it also highlights New York, New York, USA: ACM Press, 2012, p. 1133.
the need for additional study. Four alternative methods [9] T. Bogers, “Movie recommendation using random walks
were observed, which contributed to an improvement, over the contextual graph,” in Proc. of the 2nd Workshop on
particularly in the suggestion precision. Context-Aware Recommender Systems, 2010.
The experiment demonstrated that we can put less
[10] M. Fredman and M. Saks, “The cell probe complexity of
emphasis on enhancing the Modified Dijkstra's algorithm dynamic data structures,” in Proc. of the twenty-first annual
and perhaps also the Union Colors algorithm. Additionally, ACM symp. on Theory of computing, ser. STOC ’89. New
adding additional kinds of relationships, such keywords, York, NY, USA: ACM, 1989, pp. 345–354.
can help the data model.
Performance is the key benefit of the graph
representation for the item recommendation task. The
implemented techniques appear to be quick enough to
function in practise.

time. Our suggestion is pertinent for the recommendation


of a large number of items and users with diverse interests.
Our method can be used in other fields as well. The
suggested traversal methods may be utilised in fields where
it is necessary to represent entities with a range of relations
and properties. For instance, recommendations for books,
music, and the news can all be found online.
REFERENCES
[1] F. Ricci, L. Rokach, and B. Shapira, “Introduction to rec-
ommender systems handbook,” in Recommender Systems
Handbook. Springer US, 2011, pp. 1–35.

[2] Z. Huang, W. Chung, T.-H. Ong, and H. Chen, “A graph-


based recommender system for digital library,” in Proc. of
the 2nd ACM/IEEE-CS joint conf. on Digital libraries, ser.
JCDL ’02. NY, USA: ACM, 2002, pp. 65–73.

[3] I. Mele, F. Bonchi, and A. Gionis, “The early-adopter graph


and its application to web-page recommendation,” in
Pro. of the 21st ACM int. conf. on Information and
knowledge management - CIKM ’12. NY, USA: ACM,
2012, p. 1682.

[4] S. D. Kamvar, T. H. Haveliwala, C. D. Manning, and G. H.


Golub, “Extrapolation methods for accelerating pagerank
computations,” in Proc. of the 12th int. conf. on World
Wide Web, ser. WWW ’03. NY, USA: ACM, 2003, pp.
261–270.

[5] M. Bielikova, M. Kompan, and D. Zelenik, “Effective hi-


erarchical vector-based news representation for
personalized recommendation.” Comput. Sci. Inf. Syst., vol.
9, no. 1, pp. 303–322, 2012.

[6] S. Lee, “A generic graph-based multidimensional


recommen- dation framework and its implementations,” in
Proc. of the 21st int. conf. on World Wide Web - WWW ’12
Companion. New York, New York, USA: ACM Press,
2012, p. 161.

[7] C. Biancalana, F. Gasparetti, A. Micarelli, A. Miola, and


G. Sansonetti, “Context-aware movie recommendation
based on signal processing and machine learning,” in Proc.

You might also like