2017 International Conference on Innovations in information Embedded and Communication Systems (ICIIECS)

Recommender Systems: An overview of different

Abstract— This paper presents an overview of the ¿eld of different suggestions. In non-personalized recommendations,
recommender systems and describes the present generation of all the users get same suggestions. RSs are generically
recommendation methods. Recommender systems or classified into the following categories, considering how
recommendation systems (RSs) are a subset of information recommendations are made: Collaborative recommendations,
¿ltering system and are software tools and techniques providing
suggestions to the user according to their need. Many popular E-
Content-based recommendations, and Hybrid approaches. The
commerce sites widely use RSs to recommend news, music, following sections describe the most popular techniques used
research articles, books, and product items. Recommendation now-a- days for building RSs such as Collaborative filtering,
systems use personal, implicit and local information from the Content-based and Hybrid approach.
Internet. This paper attempts to describe various limitations of
recommendation methods and their advantages. II.COLLABORATIVE FILTERING
The term “collaborative filtering (CF)” was coined by
Index Terms— Recommender systems, information retrieval Goldberg et al., in 1992 who proposed that the information
system, recommendations, collaborative ¿ltering, content
filtering process becomes more effective when humans are
¿ltering, hybrid ¿ltering.
involved. [3]. In the collaborative filtering method,
recommendations for each user are generated by making
I.INTRODUCTION comparisons with the liking for one alternative over another of
Recommender systems became an active area of research other users who have qualified the product similarly to the
since the appearance of the ¿rst papers on collaborative active user. CF are based on the idea that people who agree
¿ltering in the mid-1990s. In recent years, many websites are with the evaluation of items in the past are likely to agree
widely using Recommender system. “You may know this,” again in future. CF methods are grouped into two general
“other products you may like” ,”customers who bought this methods neighborhood based and model-based [1].
item also bought,” “you may like this.” Everyone has seen
A. Neighborhood Based Method
these suggestions when browsing the web, be it on Facebook
or Amazon or some other platform. As these Web sites began In neighborhood based (memory-based or heuristic-based)
to develop, a pressing need emerged for providing CF, the user-item ratings stored in the memory are used
recommendations compiled from ¿ltering the whole range of directly to predict classifications for new items. which can be
available options. The drastic growth in data and variety of done in two ways, first user-based and second item-based
information available on the Web and the rapid introduction of recommendation. The User-based approach evaluates the
new E-business services (product buying, product comparison, interest of a user u for an item i by taking into consideration
auctioning, etc.) often overwhelmed users, leading them to the ratings for this item by other users, called as neighbors that
make inappropriate decisions. From the immense variety of have similar rating patterns. The item-based approach predicts
items (services and products) that these Web sites were the rating for an item i of a user u based on the ratings of u for
offering, users were ¿nding it very critical to make the most items similar to i. In item-based approach, the similarity
appropriate choices. between two items is calculated by the ratings provided by
other users of the system in a similar way.
A. Recommender System Background The main advantages of neighborhood-based methods are:-
RSs are basically software tools and techniques for • Simplicity: Relatively simple to implement.
information retrieval and filtering that aims to provide • Justifiability: Provide a brief but comprehensive and
meaningful and effective item recommendations to the active easily understandable justification for the computed
user [2]. These suggestions also called as Recommendations predictions.
relate to various decision-making process such as which news • Efficiency: Do not require costly training phases,
article to read, what products (items) to buy or which song to which need to be carried out at frequent intervals in
listen. “Item” is a general term that is used to denote what the large commercial applications.
RSs recommends to the users. Recommendations are of two
types, personalized and non-personalized. In personalized
recommendations, different users or user groups receive

‹,(((
Authorized licensed use limited to: AMC Engineering College ¿ Karnataka. Downloaded on April 21,2025 at 05:48:58 UTC from IEEE Xplore. Restrictions apply.
2017 International Conference on Innovations in information Embedded and Communication Systems (ICIIECS)

• Stability: Little influence by the constant addition of predict ratings of users for new items. There are many Model-
items, users, and ratings, which are typically observed based approaches for the task of recommending items. Which
in major E-commercial applications. include techniques such as Bayesian Clustering [4], Latent
Formal Definition of problem: Set of users in the system Semantic Analysis [5], Support Vector Machines [6], and
are denoted by U, the set of items by I, and the set of ratings Singular Value Decomposition [7].
by R in the system, and S is the set of possible values for a Advantages of collaborative filtering technique are listed
rating (e.g. S = [1 through 5] or S =like, dislike). The notation below:
ui denotes the subset of users that have rated an item i. Iu • Implementation of RS using Memory-Based
represents the subset of items rated by user u. Two important Collaborative filtering technique is easy.
problems associated with RS are the best item and top-N • Additional of new data in an incremental manner is
recommendation problems. The first problem consists of easy in Memory-Based Collaborative filtering.
finding, the new item i € I or Iu for a particular user u, in • Improved prediction performance by a Model Based
which user is most likely to be interested. This task of finding Collaborative filtering technique.
such item for the user is modeled as a regression or
classification problem. Where the aim is to find a function f : The disadvantage of collaborative filtering technique:
U × I ĺ S that predicts the rating f(u, i) of a user u for a new • Cold Start Problem: In the case of the new user the
item i. The obtained function then recommends an item k to system either doesn’t know what to recommend or
the uk (active user) for which the predicted rating has the has very poor performance.
highest value: • Scalability: The CF technique generates
k = argmax f(uk , j) recommendations over billions of users and products,
Accuracy is used to evaluate models performance of which require a significantly huge amount of
the recommendation method. Usually, the rating set R is split computational power.
into a training set Rtr used to learn f , and a test set Rte used to • Sparsity: Only small subsets of the items are rated by
evaluate the prediction accuracy. Mean Absolute Error (MAE) the users from the available dataset of items. Hence
and Root Mean Squared Error (RMSE) are two of the most very few ratings are available to generate a
common metrics used to measure accuracy. MAE assess the recommendation which leads in poor performance.
average magnitude of the errors in a generated set of
predictions, without taking into consideration their direction.
RMSE is a quadratic scoring rule that also assesses the Supervised Unsupervised
average magnitude of the error. It’s the square root of the Learning Learning
average of squared differences between prediction and actual
observation. Consider a situation where only the items
purchased by each user are known to the RS, measuring the
rating prediction accuracy is not possible. In such cases, the Classificat Regression Clustering Association
problem of finding the best item is usually modeled into the ion Rule
task of predicting recommendations to an active user uk a list
L(uk) containing N items likely to interest him or her [8, 9].
The quality of this method can be measured by splitting the
items of I into a set Itr, used to learn L, and a test set Ite. Sequence Market Recomm
Consider T(u) is the small subset of test items that a user u Predictio Basket ender
found relevant. So, the performance of the method is then Analysis System
computed using the measures of precision and recall:
1
Precision ( L ) = ¦ L ( u ) ∩ T ( u ) / L ( u ) One or few Outputs Multiple Outputs
u
1
Recall ( L ) = ¦ L ( u ) ∩ T ( u ) / L ( u ) Fig 1. RSs in Machine Learning
u
B. Model-Based Recommendation Methods
Neighborhood based systems, which uses the stored The above figure clearly shows that recommender system
ratings directly into the prediction whereas model based comes under the Unsupervised Learning of the Machine
approaches use these ratings to acquire knowledge and learn a Learning paradigm. In unsupervised learning the data
predictive model. The idea used in model based systems is to available to us is not labeled so the hidden associations and
imitate and model the interactions of user-item with factors cluster formation is revealed by performing unsupervised
representing the hidden characteristics of the users and the learning using algorithms that comes under that paradigm.
items in the system, like the preference category of users and Different approaches to recommendations are nothing but
the category class of items. Then the model is trained using algorithms that comes under sub category Association rule
the available data (train dataset) and trained model is used to mining.

Authorized licensed use limited to: AMC Engineering College ¿ Karnataka. Downloaded on April 21,2025 at 05:48:58 UTC from IEEE Xplore. Restrictions apply.
2017 International Conference on Innovations in information Embedded and Communication Systems (ICIIECS)

TABLE I. TABLE TYPE STYLES constructs a model of users interest which generates
Paper Title Approach to Conclusion recommendations [10]. For example, if a user likes a web page
recommendation with the words “Camera,” “Lens” and “Tripod” the Content-
Recommender CB, Neighborhood-based, The hybrid robust based filtering will recommend pages related to the electronics
Systems Handbook. CF, Multi-Criteria filtering methods are world. The item description and a profile of the user’s
Springer. Recommender, Robust CF. preferable over unique
approaches. orientation are important in Content-based filtering. CB filters
A Case Based CB, CF, Case-based CBR a new approach try to generate recommendations based on similarity count.
Recommendation reasoning (CBR). in hybrid filtering is The top, matching items are recommended by comparing more
Approach for considered as the than one candidate items with items that are already rated by
Market Basket preferable method for
Data, IEEE recommendation
the user.
intelligent systems. assuming transactions Advantages of Content-based filtering:-
as the case. • Content-based recommender system provides user
Music CB, Low-level Descriptors, Optimization of both independence through exclusive ratings which are
Recommendation Correlation Analysis, the feature vector and
System. Feature Vector classification algorithm used by the user to build their profile.
Effectiveness, is essential. • Provides transparency to the active user by providing
Collaborative filtering. an explanation of working of CB filter.
Modeling CF,CB, Neighborhood- Designed new CF
relationships at aware factorization methods based on
• CB filter is good to recommend items that are not yet
multiple scales to models that try hard to rated or viewed by any user. This will be
improve the minimize quadratic advantageous for a new user.
accuracy of large errors, and
recommender demonstrated high
systems. performance on a Limitations of Content-based filtering :-
large, real-world • In CB Filters, it is very difficult to generate
dataset. Hence they characteristics of an item.
find CF to be more
preferable.
• CB Filters suffer from an over-specialization problem
Item-based top-N CF-based top-N Conditional because it advocates the same types of items.
recommendation recommender Systems: probability-based item • It is more difficult to get feedback from users in CBF
algorithms. user-based, model-based. similarity scheme and because users normally do not sort the items (as in
Item-based top-N higher-order item-
recommendation based models provide CF) and therefore, it is not possible to determine if the
algorithms. reasonably accurate recommendation is correct.
recommendation and
are better than
traditional user-based IV.HYBRID APPROACH TO RECOMMENDATIONS
CF techniques.
Implemented scheme A wide and diverse variety of techniques are proposed for
is independent of the generating recommendations which include collaborative,
size of the user–item
content based, knowledge based and other techniques. These
matrix.
kNN versus SVM User profiling, kNN is dominant on methods are blended in hybrid recommenders to improve
in the collaborative collaborative filtering, datasets with relatively performance. Collaborative filtering and Content-based
filtering Support Vector Machine, low sparsity. On filtering approaches are extensively used in information
framework. k-Nearest Neighbor datasets with high to filtering application. Commonly, collaborative filtering is
extremely high level of
sparsity. integrated with other techniques to nullify the ramp-up
In case of extreme problem. Hybrid approaches can be put into action in various
sparsity, it is best to ways:
use a model-based
approach, such as
• Individual implementation of collaborative and
SVM classifier or content-based methods and aggregation of their
SVM regression predictions to generate recommendations.
• Integration of some pro characteristics from content-
based methods into a collaborative approach,
III.CONTENT-BASED FILTERING • Integration of some pro characteristics from
Content-based filtering, sometimes also referred to as collaborative approach methods into a content-based
cognitive filtering, recommends the items based on a approach,
comparison between the content of the items and a user profile • A generic consolidative model that is the assimilation
data. The content of each item is represented as a set of of both content-based and collaborative
descriptors, terms or feature vector. This descriptor can be a characteristics.
genre for movie item or can be frequent terms that occur in a
document for the document as an item. The content based filter A. Weighted
analyzes set of descriptors feed to it as input for a particular In the weighted hybrid recommendation score or weight of
item which is previously rated by the user. This filter then a recommended item is calculated from the results of all

Authorized licensed use limited to: AMC Engineering College ¿ Karnataka. Downloaded on April 21,2025 at 05:48:58 UTC from IEEE Xplore. Restrictions apply.
2017 International Conference on Innovations in information Embedded and Communication Systems (ICIIECS)

available recommendation techniques implemented in the intelligent systems, the correctness and validity of the
system. Recommended components which have different information that is served to the user were of concern. The
scores are combined statistically. Additive aggregation is memory based approach was unable to provide significant
implemented to get normalized scores. accuracy, but it can be used in simpler scenarios. Model based
approach was developed which suffered from the problem of
B. Switching
cold start to improve the quality of collaborative filters. To
In this approach, out available recommendation deal with the problem of cold start, Hybrid Recommender
components at the disposal, the system chooses one Systems as a result of extensive research were designed.
particular component and applies the picked one which best Hybrid systems provide a notable enhancement in accuracy,
suits the purpose. The system has a criterion function to precision, and recall matrices.
switch between recommendation techniques.
REFERENCES
C. Mixed
[1] Recommender Systems Handbook. Francesco Ricci,Lior
Different recommendation approaches provide different Rokach, Bracha Shapira, Paul B. Kantor. Springer 2010.
recommendations that will be introduced together. This [2] A Case-Based Recommendation Approach for Market Basket
hybrid system is based on merging and presenting multiple Data Anna Gatzioura and Miquel Snchez-Marr IEEE
rated lists into a single rated list. This approach avoids the INTELLIGENT SYSTEMS 2015.
“new item” start-up problem. [3] Using collaborative ¿ltering to weave an information Tapestry
D. Goldberg, D. Nichols, B. M. Oki, and D.
D. Feature combination Terry,Communications of the ACM, vol. 35, no. 12, pp. 6170,
1992.
The hybrid system is divided or separated into two parts: [4] Breese, J.S., Heckerman, D., Kadie, C.: Empirical analysis of
contributing and actual recommender. These two co-exist in predictive algorithms for collaborative ¿ltering. In: Proc. of the
the system. The actual recommender depends upon the 14th Annual Conf. on Uncertainty in Arti¿cial Intelligence, pp.
4352. Morgan Kaufmann (1998).
results or data output of the contributing recommender.
[5] Hofmann, Collaborative ¿ltering via Gaussian probabilistic
Feature combination hybrid technique lets the system latent semantic analysis. In: SIGIR 03: Proc. of the 26th Annual
consider collaborative recommender output data without Int. ACM SIGIR Conf. on Research and Development in
depending on it exclusively, which decreases the sensitivity Information Retrieval, pp. 259266. ACM, New York, NY, USA
of the system. (2003).
[6] Grcar, M., Fortuna, B., Mladenic, D., Grobelnik, M.: k-NN
E. Feature Augmentations versus SVM in the collaborative ¿ltering framework. Data
Science and Classi¿cation pp. 251260 (2006).
Feature Augmentation hybrid is similar to the feature
[7] Bell, R., Koren, Y., Volinsky, C.: Modeling relationships at
combination hybrids, but the difference is contributor gives multiple scales to improve the accuracy of large recommender
interesting characteristic. Feature Augmentation hybrid is systems. In: KDD 07: Proc. of the 13th ACM SIGKDD Int.
more elastic than feature combination method. Feature Conf. on Knowledge Discovery and Data Mining, pp. 95104.
Augmentation found to make a significant contribution to ACM, New York, NY, USA (2007).
the quality value of recommendations. [8] Deshpande, M., Karypis, G.: Item-based top-N recommendation
algorithms. ACM Transaction on Information Systems 22(1),
V.CONCLUSION 143177 (2004).
Recommender systems have become popular. RSs are widely [9] Last.fm: Music recommendation service (2009).
http://www.last.fm
utilized and deployed in a variety of areas. RSs provide
recommendations based on users interaction with the system. [10]Mladenic, D.: Text-learning and Related Intelligent Agents: A
Survey. IEEE Intelligent Systems14(4), 44–54 (1999).
Popular approaches to recommendations are content-based, [11]Karim, J., 2014, Hybrid Systems for Personalized
collaborative and Hybrid. An overview of all these approaches Recommendations,
shows that they have both advantages and disadvantages. CF Research Challenges in Information Science (RCIS), 2014 IEEE
system does not need content information about items and Eighth
International Conference, May 2014
users to be machine recognizable. CF approach utilizes only
[12]G. Linden, B. Smith and J. York, “Amazon.com
ratings and there is no need of any additional information recommendations item-to-item collaborative filtering.” In IEEE
about users or items. But the principal disadvantage is that CF Internet Computing, Volume 7,Issue 1, January 2003, pp.76-80,
systems cannot generate recommendations if there are no 2003
ratings available. In the starting phase of development of such

Authorized licensed use limited to: AMC Engineering College ¿ Karnataka. Downloaded on April 21,2025 at 05:48:58 UTC from IEEE Xplore. Restrictions apply.