A framework for quick & easy deployment of customizable RS user studies. The framework is modular, highly extensible and easy to use. It offers out-of-the-box functionality for all phases of the user study, i.e., data loading and presentation, preference elicitation, baseline recommending algorithms as well as result comparison layouts. It also features a fastcompare plugin implementing the most common study flow, i.e., comparing perceived utility of several personalized recommenders.
See the paper for detailed description of existing plugins and their functionality.
- fastcompare allows comparison of 2-3 algorithms on implicit feedback dataset. Is further extensible by providing classes implementing predefined baseclass (for pref. elicitation, algorithms, data loaders).
- utils contains only shared functionality, does not serve as "plugin template".
- layoutshuffling a plugin that we have used for our internal user study. Is kept here for illustrational purposes.
- vae plugin providing MultVAE and StandardVAE algorithm implementations for the fastcompare plugin. Similarly to utils this plugin only contains shared functionality and cannot be used to create new user study. Note that these algorithms are itentionaly separated from fastcompare plugin itself since they have extra dependencies that only have to be installed when this plugin is going to be used.
- empty_template i.e. the minimal working plugin, see below.
If you want to quickly try the framework, feel free to use the links above. If you need to have a local deployment of the framework, details are given in this section:
We provide a Dockerfile for building an image with all the dependencies. To build the docker image, you can use the build_server_container.sh script.
To run the container, you may use the start_server.sh script. The default port is set to 5555, but you can change it by modifying the start_server.sh script and the Dockerfile.
To start user studies (from fastcompare) successfully, you have to provide the datasets that are provided by fastcompare and any other datasets that you plan to support in your newly designed plugins/fastcompare data loader extensions. There is a small limitation, that images or files in Flask has to be in the static directory. It can be static directory inside one of the plugins or of the whole Flask app, and we decided to use the latter, so all images have to be put into static/datasets/X/img where X denotes dataset name ({'ml-latest', 'goodbooks-10k'}). Note that although this may seem a bit restricting, in the end, it is not, because you can always put just symlinks there, that will point to arbitrary directory in your system. The exact images that were used during Demonstration are available in static/datasets/ml-latest/ml_latest_img.zip and static/datasets/goodbooks-10k/goodbooks_img.zip and should be extracted into static/datasets/ml-latest/img and static/datasets/goodbooks-10k/img respectively.
Datasets itself (csv files) can be downloaded from official links: ml-latest, goodbooks-10k and the csv files should be extracted to static/datasets/ml-latest/*.csv and static/datasets/goodbooks-10k/*.csv respectively.
Alternatively you can run the server without the docker container, by just relying on Flask and using flask --debug run or flask run from the server directory. However, in that case, you have to ensure to have all the dependencies installed on your system (you can try to mimick what the Dockerfile does with dependencies). Running the server this way is especially useful during development of new functionality.
The fastcompare plugin itself can be extended with new data loaders, recommendation algorithms, preference elicitation methods, or evaluation metrics. The extension is done by adding new implementation as a class that is subclassing appropriate base classe. The new class can be put either directly into plugins/fastcompare/algo/*.py or into separate plugins, e.g. plugins/newplugin/*.py
Subclass the DataLoaderBase and implement all the abstract methods and properties, those are described in the code comments (you can get inspired by e.g. GoodbooksDataLoader)
Then put your dataset folder under static/datasets
Later you can return the following url_for('static', filename=f'datasets/goodbooks-10k/img/{item_id}.jpg')
alternatively, you can always return remote urls ('http://...') but they are slowing down the system.
If you cannot add a new folder to root static (e.g. your deployment only gives permission to modify /plugins), you can include it directly within your new plugin, and then address the image files using:
pm.emit_assets('pluginName', filename) with from app import pm (again, make sure to check has_app_context even before importing pm!)
Subclass the AlgorithmBase base class and implement it. All abstract methods and properties are described in the code comments.
Subclass the PreferenceElicitationBase and implement it. All abstract methods and properties are described in the code comments.
Subclass the EvaluationMetricBase and implement it. All abstract methods and properties are described in the code comments.
Adding new plugin is done by adding new folder under server/plugins directory. This directory has several requirements (see Flask-PluginKit's documentation) and the plugin itself should expose the following endpoints in order to become a valid plugin:
/create(typically renders some HTML page where user enters parameters) and in the end, it MUST invoke/create-user-studyendpoint/initializethis will be invoked after/create-user-studyhas been called and this is the place where the plugin should perform its initialization. Any long running initialization should be done in the background daemon process, without blocking the request. Once the initialization is done, the plugin is responsible for marking the particular user study asinitialized=Trueandactive=True/joinwill be called when user attempts to join the user study, by following the generated URL link. This is where full control is passed to the plugin and plugin should decide future steps./results(optional) is called upon clicking on "Results" in administration UI. If the implementation is not provided, fallback endpoint from theutilsplugin is used. Being able to hook custom/resultsendpoint is very useful because plugins with specific flow may require very individual and customized evaluation strategies that cannot by captured easily by common interfaces (e.g.EvaluationMetricBase).
We have prepared a minimal working plugin template that you can use as a starting point when developing new plugins.