Capturing users' preferences drift is a crucial property of a successful recommender systems. Based on TeRec, proposed by Chen et al., we propose DeMF, which is an online recommender system based on matrix factorisation technique, introduced by Koren et al.. We evaluated performance of both two models on MovieLens-100k and Ciao. The results show that compared with TeRec, DeMF show its superior properties in following aspects:
- Lower mean square error (MAE).
- Lower trend to diverge.
Files are organised in following manner:
|-DeMF
|-mySVD.py
|-sampleInput.py
|-demfNoffline.sh
|-iteration.sh
|-cvxMAE.sh
|-reservoirSize.sh
|-TeRec
|-mySVD.py
|-sampleInput.py
|-demfNoffline.sh
|-iteration.sh
|-cvxMAE.sh
|-reservoirSize.sh
Requirements:
- Python2.7
- NumPy
- SciPy
sh cvxMAE.sh
generates the convex figures.
sh iteration.sh
generates the effects of T.
sh reservoirSize.sh
generates the effects of size of reservoir.
sh demfNoffline.sh
generates time and memory consumption comparison.
Experimental results will be written in following .txt file:
- memoValueImpML.txt:
- memory consumption of users. Offline result is shown on terminal console.
- resultTimeImpML.txt:
- time consumption of users. Offline result is shown on terminal console.
- performance.txt:
- each row has format:
-
$\alpha$ ,$\beta$ , sizeofResv, numEpoches, iteration, pureMFmae, improvedMAE, maeImprovements
-
- each row has format:
- mae_distribution.txt:
- MAE improvements of each user.
Please note that running a new shell file will remove results of previous experiment!
Using MATLAB to generate all plots:
- demfNoffline.m:
- Comparison of memory and time consumption between online and offline
- cvxMAE.m:
- MAE performance with various
$\alpha$ and$\beta$
- MAE performance with various
- maeDisDec.m:
- MAE improvements for each user
- iteration.m:
- MAE various with iteration
- reservoirSize.m:
- MAE various with reservoir size
The memory and time consumption for each user is written in files memoValueImpML.txt
and resultTimeImpML.txt
. To run this some manual actions have to be made since every time the number of rating events could be different. However, this figure is intuitive and simple thus not included a .m file.
Following figures show MAE performance comparisons between DeMF
and TeRec: