Temporal integration based factorization to improve prediction accuracy of collaborative filtering
A recommender system provides users with personalized suggestions for items based on the user’s behaviour history. These systems often use the collaborative filtering (CF) for analysing the users’ preferences for items in the rating matrix. The rating matrix typically contains a high percentage of...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2016
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/69372/1/FSKTM%202016%2040%20IR.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | A recommender system provides users with personalized suggestions for items based on
the user’s behaviour history. These systems often use the collaborative filtering (CF) for analysing the users’ preferences for items in the rating matrix. The rating matrix typically contains a high percentage of unknown rating scores which is called the data sparsity problem. The data sparsity problem has been solved by several approaches such as Bayesian probabilistic, machine learning, genetic algorithm, particle swarm optimization and matrix factorization. The matrix factorization approach through temporal approaches has the accurate performance in addressing the data sparsity problem but still with low accuracy. The existing temporal-based factorization approaches used the long-term preferences and the short-term preferences. The difference between long-term preferences is that it utilizes the whole recorded preferences while the short-term preferences utilizes the recorded preferences within a session (e.g. week, month, season, etc.). However, there are four issues when a factorization approach is adopted which are latent feedback learning, score overfitting, user’s interest drifting and item’s popularity decay over time. This study proposes three approaches which are (i) the Ensemble Divide and Conquer (EDC) which achieved accurate latent feedback learning, (ii) two personalized matrix
factorization (MF) based temporal approaches, namely the LongTemporalMF and
ShortTemporalMF to solve overfitting during the optimization process, user’s interest drifting and item’s popularity decays over time and (iii) TemporalMF++ approach which solved all the issues. The TemporalMF++ approach relies on the k-means algorithm and the bacterial foraging optimization algorithm. The Root Mean Squared Error metric is used to evaluate the prediction accuracy. The factorization approaches such as the Singular Value Decomposition, Baseline, Matrix Factorization and Neighbours based Baseline are used to be compared against the proposed approaches. In addition, the Temporal Dynamics, Short-Term based Latent, Short-Term based Baseline, Long-Term, and Temporal Interaction approaches are used
to benchmark the proposed approaches.
The MovieLens, Epinions, and Netflix Prize are real-world datasets which are used in
the experimental settings. The experimental results show the TemporalMF++ approach
is higher prediction accuracy compared to the approaches of EDC, LongTemporalMF, and ShortTemporalMF. In addition, the TemporalMF++ approach has a prediction
accuracy higher than the benchmark approaches of factorization and temporal. In
summary, the TemporalMF++ approach has a superior effectiveness in improving the accuracy prediction of the CF by learning the temporal behaviour. |
|---|