الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 113 |  |  |
| | | from | 113 | | |
Abstract
Today's Recommender system is a relatively new area of research in machine learning. Also,
it's an essential issue for E-commerce business for on-line websites today to increase sales of
products to users. It helps users to find their interests more efficiently according to his favorite
movies, songs, and products using rated history items or matching profiles with other users.
There are four main ways that recommender systems produce a list of recommendations for a
user: Content-based filtering, Collaborative filtering, Demographic filtering, and Hybrid
filtering. We do a survey about these system and list advantage and disadvantages.
In this thesis, we build a new system that has three models to solve some of the collaborative
filtering challenges. Our new system will solve the Cold start problem, Scalability of the
approach, Recommending the items in the Longtail, Sparse, Missing data, Accuracy of the
prediction, Novelty, and diversity of recommendation.
In this system we need to solve sparsity issues using association rule then using SVD to
reduce the data size and delete unused data to recommendation system to achieve high
performance then we use three models to detect the emotions using face, colors and arousal
map and using new recommendation system to recommend items to users with high accuracy
and high performance. Also, we change the user’s mood from negative to positive through
tracking status in the system.
The first model of the system is using association rule with clustering based to solve the
Longtail challenge when recommending items, accuracy, and sparsity challenges of
recommended items. We use association rule to count every song listen per transaction to find
the number of played times of the same song without a rating so we will find a hidden pattern
and use number of played counts as explicit data rated from users to this songs and compute
similarities by cosine vector similarity to make a prediction list and clustering data items and
users based on the similarity values.
The second model is using k-means clustering and SVD (singular value decomposition we use
SVD to reduce the size of the cluster to solve the scalability challenge.
The third model, we use emotion to make a hybrid model "Hybrid emotion-based music
recommendation," generating playlist which suits and matches your mood of listening to
music using three methodologies to detect the user's emotions. We use face, colors, and
arousal map and merge the output of these steps and detect the accurate user's emotion. This
model will track the user's emotion in the system, and if the user was in a bad or negative
mood, we try to change his mood to positive or pleasant.
Our experiments applied to last.FM music data sets with the user's implicit interaction records
and explicit rating records of items named by hybrid feedback, movie lens datasets.
Name of Candidate: Marwa Hussien Mohamed Hassan
Degree: PHD
Title of Thesis: Enhanced Recommendation System based on a Collaborative Filtering Technique
Supervisors:
1. Assoc. Prof. Dr. Mohamed Helmy Khafagy 2- Assoc. Prof. Dr. Mohamed Hassan Ibrahim
Approval: --- / ----- / 2020 Department: Information Systems
The summary not more than 500 words
Our experimental results show that the new proposed system has better performance and more
accuracy compared to basic collaborative filtering techniques when data are spare using
precision, recall, and F-metrics evaluations and recommend novel items to users.
The first model using association rule and clustering-based has more accuracy in
recommendation we enhance in Precision by 37 %, recall by 7 %, and F-measure by 19%.
The second model using k-means and SVD has more accuracy. We enhance precision by
36%, recall by 5%, and F-measure by 17%.
We try more experiments to detect the user's emotions mood firstly we merge the three
emotions datasets, and then we try to use our first recommendation model with face emotions
only, colors, arousal map, and our hybrid model. Hybrid system is best in accuracy while
selecting the positive mood. Hybrid has accuracy by 11 % than arousal ,13% than colors and
15% at face emotion. Surprise mood enhanced by 18% than using arousal mood
recommendation, by 23% than using colors and by 20% than using face emotions only. Happy
mood enhanced by 14% than using arousal mood recommendation, by 23% than using colors
and by 20% than using face emotions only.