Introduction
Svdfeature is a toolkit developed by Apex Data & Knowledge Management Lab in the KDD CUP11 contest. Its purpose is to effectively solve the feature-based matrix decomposition. The new model can be implemented only by defining new features. This feature-based setting allows us to include a lot of information in the model, so that the model is more time-keeping. With this toolkit, it is easy to integrate other information into the model, such as time dynamics, domain relationships, and hierarchical information. In addition to scoring predictions, you can also implement pairwise ranking tasks. Model
The svdfeature model is defined as follows:
The input contains three features <α,β,γ>, which are user characteristics, item characteristics and global features.
The final version of the model is:
The usual choices for active functions and loss functions are as follows (ignoring the following identities):
Detailed information can be consulted svdfeature-manual. input Format
The input format is a sparse feature format similar to the SVM format. For an input sample, we need to specify three feature,<α,β,γ> and forecast targets. The format is as follows:
The ID and value here correspond to feature IDs and eigenvalues of non-0 items. The signature file first specifies the target of the prediction, then the global, user, and item eigenvectors the number of non-0 items. And then the sparse feature format to enumerate the non-0 global, user and item characteristics. For example, if we use the basic matrix decomposition model, user 0 gives item 10 a rating of 5:
5 0 1 1 0:1 10:1
The <0,1,1> here represents 0 global features, 11 user features, and one item feature. 0:1 represents a user trait, and 10:1 represents an item feature.
The rest of the details are self-checking svdfeature-manual.
Files in the Svdfeature Toolkit:
* Solvers:all The customization of Svdfeature solvers, not included in the basic package
* Tools:the Auxiliary tools that can is used for experiment
* Demo:the examples that can-help-get started on the toolkit operation
I'm using the Ubuntu14.04. Compiling environment requires g++4.6 and above, as to how to install g++, Baidu.
Copy the svdfeature-1.2.2.tar.gz to Ubuntu and unzip it.
Enter "make" into the main directory and tools directory to compile.
After compiling, go to the demo directory:
Svdfeature provides 5 examples, namely: BASICMF, Binaryclassification, Implicitfeedback, Neighborhoodmodel and Pairwiserank.
Enter a separate directory containing a run.sh file that runs "./run.sh" to complete the training and testing phase.
If there are run-ml100k.sh files in the directory, you can use Movielen data, the steps are: Download ml-100k data, put Ua.base and ua.test into the directory, run "run-ml100k.sh".
Normal running process:
Run End:
Forecast results:
Run with Movielen:
The results of the run are saved in Pred.txt:
The other few examples are not shown.