Page 32 - 2022

Ph.D.
(Engineering & Technology)
COLD START RECOMMENDER SYSTEM USING
TENSOR BASED METHODS.

Ph.D. Scholar : Gondaliya Shital Nikhil
Research Supervisor : Dr. Kiran R. Amin

Regi. No.: 17146051002
Abstract :
In the era of internet, users are facing the problem of data overload. Hence, all internet
activities become unmanageable which cause less efficiency and low productivity. One
solution to data overload problem is personalized recommendations, which help online
users to manage big volumes of varied data on the internet. Many intellectual systems
are there which enable users to locate appropriate items that meet up their desires.
Among all the existing systems, most of the systems are considering only two entities
(user and item) for prediction. Recommendation accuracy can be increased if we
consider a extra information. Since last few years, the multidimensional attributes which
clarify user performance and likings have been increasing interest. In this direction, many
research works are going on it, but these methods do not give good results in case of
multidimensional data. In relation to multidimensional data, Social Tagging System (STS)
has attracted the attention of many researchers. In STS, user can choose tags freely
based on the description provided by frequent users. STS has the two major issues like
cold start and data sparsity. Data sparsity is arising when only few users are interested in
rating the item and cold start is denoting the difficulty in the recommendation for the new
user, where the system is not able to recommend items to users. Both the problems limit
the effectiveness of Collaborative Filtering (CF) methods.

Our research makes an important contribution towards solving the cold start issue and
provides more precise and efficient recommendations with multidimensional data as well
as multi attribute data by proposing three different frameworks.
The first framework is Latent-Neighborhood Higher Order Singular Value Decomposition
(LN-HOSVD) which represents multidimensional relations among user, item, and tag
proposed with the integration of the nearest neighbourhood model and factorization
methods. This approach is recommending items by considering only single attributes
which is one of it's limitations of it. So, to improve the accuracy of the recommendations
we need a system that consider multiple attributes of the item. In the second framework,
we have a proposed Multi-Batch Quasi Newton (MBQN) method with the integration of
Artificial Neural Network (ANN) and Non-Negative Matrix Factorization (NNMF). To
estimate a parameter, we have used the Limited Memory

01