@@ -21,7 +21,7 @@ It can be clearly stated that the \textit{texisting baselines} have been \textit
\subsection{Experiment Realization}
As the \textit{Netflix-Prize} has shown, \textit{research} and \textit{validation} is \textit{complex} even for very \textit{simple methods}. Not only during the \textit{Netflix-Prize} was intensive work done on researching \textit{existing} and \textit{new reliable methods}. The \textit{MovieLens10M-dataset} was used just as often. With their \textit{experiment} the authors \textit{doubt} that the \textit{baselines} of \textit{MovieLens10M} are \textit{inadequate} for the evaluation of new methods. To test their hypothesis, the authors transferred all the findings from the \textit{Netflix-Prize} to the existing baselines of \textit{MovieLens10M}.
Before actually conducting the experiment, the authors took a closer look at the given baselines. In the process, they noticed some \textit{systematic overlaps}. These can be taken from \textit{table} below.
\input{overlaps}
...
...
@@ -36,6 +36,12 @@ As a \textit{first intermediate result} of the preparation it can be stated that
In addition, it can be stated that learning using the \textit{bayesian approach} is better than learning using \textit{SGD}. Even if the results could be different due to more efficient setups, it is still surprising that \textit{SGD} is worse than the \textit{bayesian approach}, although the \textit{exact opposite} was reported for \textit{MovieLens10M}. For example, \textit{figure}\ref{fig:reported_results} shows that the \textit{bayesian approach BPMF} achieved an \textit{RMSE} of \textit{0.8187} while the \textit{SGD approach Biased MF} performed better with \textit{0.803}. The fact that the \textit{bayesian approach} outperforms \textit{SGD} has already been reported and validated by \citet{Rendle13}, \citet{Rus08} for the \textit{Netflix-Prize-dataset}. Looking more closely at \textit{figures}\ref{fig:reported_results} and \ref{fig:battle}, the \textit{bayesian approach} scores better than the reported \textit{BPMF} and \textit{Biased MF} for each \textit{dimensional embedding}. Moreover, it even beats all reported baselines and new methods. Building on this, the authors have gone into the detailed examination of the methods and baselines.
\subsubsection{Experiment Implementation}
For the actual execution of the experiment, the \textit{authors} used the knowledge they had gained from the \textit{preparations}. They noticed already for the two \textit{simple matrix-factorization models SGD-MF} and \textit{Bayesian MF}, which were trained with an \textit{embedding} of \textit{512 dimensions} and over \textit{128 epochs}, that they performed extremely well. Thus \textit{SGD-MF} achieved an \textit{RMSE} of \textit{0.7720}. This result alone was better than: \textit{RSVD (0.8256)}, \textit{Biased MF (0.803)}, \textit{LLORMA (0.7815)}, \textit{Autorec (0.782)}, \textit{WEMAREC (0.7769)} and \textit{I-CFN++ (0.7754)}. In addition, \textit{Bayesian MF} with an \textit{RMSE} of \textit{0.7653} not only beat the \textit{reported baseline BPMF (0.8197)}. It also beat the \textit{best algorithm MRMA (0.7634)}.
As the \textit{Netflix-Prize} showed, the use of \textit{implicit data} such as \textit{time} or \textit{dependencies} between \textit{users} or \textit{items} could \textit{immensely improve existing models}. In addition to the two \textit{simple matrix factorizations}, \textit{table}\ref{table:models} shows the \textit{extensions} of the \textit{authors} regarding the \textit{bayesian approach}.
\input{model_table}
As it turned out that the \textit{bayesian approach} gave more promising results, the given models were trained with it. For this purpose, the \textit{dimensional embedding} as well as the \textit{number of sampling steps} for the models were examined again. Again the \textit{gaussian normal distribution} was used for \textit{initialization} as indicated in \textit{section}\ref{sec:experiment_preparation} . \textit{Figure} XY shows the corresponding results.
\textit{Matrix-Factorization}&\textit{u}, \textit{i}& Simple \textit{matrix-factorization} similar to \textit{biased matrix-factorization} and \textit{RSVD}. \\\hline
\textit{timeSVD}&\textit{u}, \textit{i}, \textit{t}& Based on the \textit{matrix- factorization}, \textit{time dependencies} are taken into account. \\\hline
\textit{SVD++}&\textit{u}, \textit{i}, $\mathcal{I}_u$& Based on the \textit{matrix-factorization}, the \textit{items}$\mathcal{I}_u$ that a \textit{user} has \textit{viewed} are included. \\\hline
\textit{timeSVD++}&\textit{u}, \textit{i}, \textit{t}, $\mathcal{I}_u$& Combination of \textit{SVD++} and \textit{timeSVD}. \\\hline
\textit{timeSVD++ flipped}&\textit{u}, \textit{i}, \textit{t}, $\mathcal{I}_u$, $\mathcal{U}_i$& Extension of \textit{timeSVD++} whereby all other \textit{users}$\mathcal{U}_i$ who have seen a certain \textit{item} are also taken into account. \\\hline
\end{tabular}%
}
\caption{\textit{Models} and their \textit{features} created and used by the \textit{authors}.}
