Recommender systems are an important service utilized in a wide variety of applications, but they rarely explain the processes and data used to generate the nal recommendation. Furthermore, no convenient software resource exists to help developers create recommender systems that use a variety of strategies, such as knowledge-driven recommendation processes. We present a Python framework to support the development of recommender systems with an emphasis on using data sources containing rich semantics and providing explanations for each step involved in producing the recommendation. We illustrate this through an example food recommendation system developed using the framework. Submission Type: Poster Resource Link: https://tetherless-world.github.io/FREx
Recommender systems, which aim to predict preferences and suggest suitable
items to users, have become a standard tool in many application domains.
Various recommendation techniques have been developed, and in recent years there
has been an increasing amount of attention given to factors like the explainability
of recommender systems [
While there are many software tools available to implement algorithms for
computing various recommendation scores [
In this paper, we present preliminary work for FREx (a Framework for Recommendations with Explanations), which provides a framework for developing knowledge-driven recommender systems that integrate explainability into the end-to-end process of the recommender system. Our framework is intended to best serve applications that utilize a source of well-structured semantics, such as an ontology, to provide domain models of the various items, users, and other entities involved in the application. Additionally, we apply our framework to implement an application for healthy food recommendation, to demonstrate its utility and to potentially serve as pedagogical material. 2
FREx is a Python framework aimed at enabling e ective prototyping of knowledgedriven recommendation systems with a focus on integrating explanations. FREx provides extensible classes that can be developed and integrated into a cohesive system for recommending items and providing explanations. At the core of FREx is a pipeline architecture for processing a series of steps to provide recommendations. This general architecture is shown in Figure 1.
FREx's pipeline for recommendations is constructed of PipelineStages. In a recommender system, three key stages that are typically carried out are to (1) generate, (2) score, and (3) lter candidate items to recommend. In FREx, the skeleton of each of these three stages is implemented { as the CandidateGenerator, CandidateScorer, and CandidateFilterer classes { and handles the creation or modi cation candidates as well as attaching explanations associated with the stage. Developers can extend these base classes to provide domain-speci c functionality and supply customized explanations.
As candidates pass through the pipeline, they retain information about the
context in which they were produced, the recommendation scores applied by
scoring stages, and the explanations applied by all stages. Precisely what constitutes
\context" (for which various de nitions exist [
In order to make meaningful use of the data in a domain speci c, knowledgedriven recommendation system, it is important to model the semantics of objects in the domain. FREx has been developed with the intention of enabling the use of Resource Description Framework (RDF) data with rich semantics to model domain objects and their properties. To facilitate this, Python dataclasses for the application are implemented that re ect the data models of the underlying ontology, supporting more object-oriented program design. 3
Using FREx, we have developed an exemplar application, FoodRec, for recommending healthy foods to users. FoodRec incorporates knowledge into the recommendation process by (1) considering the classi cations of ingredients and (2) utilizing guidelines to determine what is considered a \healthy" recipe. In the recommender, we leverage semantic information that is available through FoodOn's class hierarchy to infer additional knowledge about preferences and restrictions. For example, if a user is prohibited from eating pork, we can identify subclasses of pork in FoodOn to infer further restricted ingredients, like \bacon" or \ham." Similarly, we can infer more general preferences based on speci c ingredients that are included in a user's favorite recipes (e.g., a user likes recipes that use bacon, so we can infer a preference for pork in general).
The FoodRec use case utilizes the FoodKG [
In its current state, FREx is intended to be highly exible in terms of the kinds of data used as input and the methods used to produce recommendations. Our example use case, FoodRec, demonstrates how RDF data can be queried and converted into objects, which in turn can be passed through a pipeline to perform knowledge-driven recommendations. FREx can be similarly utilized in new domains by constructing the appropriate domain object models and querying services. New pipelines to perform di erent recommendation strategies can also easily be developed and incorporated.
A key limitation of our current iteration of FREx is the need for developers to de ne the domain object models for their application, as well as the need to manually craft explanation text. If the recommendation application is developed using an ontology, domain objects may be easily modeled, but it may be di cult to determine what pieces of information from the ontology are relevant to the application. Additionally, while our framework focuses on explanations about the recommendation stages, such explanations are not produced automatically. While it is possible to extend the Explanations class to apply some strategies like templating systems (e.g., \This item received a score of fitem scoreg because its fproperty nameg was below the threshold of fthreshold valueg"), such templating systems would once again be fairly speci c to the application or domain and not be widely applicable to certain scoring or ltering techniques. 5
FREx provides a high level structure aimed at aiding developers to prototype recommendation systems that integrate explanations throughout the recommendation process. A exible pipeline system is introduced to handle the common components of generating, ltering, and scoring items to recommend. We demonstrate the use of FREx for a food recommendation use case in the FoodRec application.
Moving forward, we aim to integrate FREx with other tools to better leverage semantics provided by ontologies. Our goal here is to incorporate tools that can automatically generate Python classes from the domain ontology to more easily support FREx's object oriented work ow. We additionally hope to expand our work to provide a greater level of automation in providing explanations. This will help to further reduce the burden on developers to manually supply FREx with explanation text, making prototyping of such applications even easier. Lastly, we plan to perform an evaluation of using FREx for developing recommender systems. This evaluation would mostly focus on the framework's usability, especially in terms of more e ciently creating knowledge-driven recommendation pipelines. 6
This work is partially supported by IBM Research AI through the AI Horizons Network.