Flex Recs: Expressing and combining Flexible recommendations Georgia Koutrika Bercovitz. Hector Garcia-Molina Computer Science Depa Stanford University, Stanford, California, USA outr O, hector/@cs. stanford. edu ABSTRACT Recommendation systems have become very popular but most rec- bedded in the system code, not expressed declaratively. From ommendation methods are 'hard-wired into the system making ex the designer viewpoint, this fact makes it hard to modify the perimentation with and implementation of new recommendation algorithm, or to experiment with different approaches radigms cumbersome. In this paper, we propose FlexRecs,a No Flexibility: The recommendations provided are fixed. End framework that decouples the definition of a recommendation pro. users are given few choices. For example, a user may be unable cess from its execution and supports flexible recommendations over to request recommendations for movies that could be jointly structured data. In FlexRecs, a recommendation approach can be seen by her and her friend, or that her recommendations be defined declaratively as a high-level parameterized workflow com- based on what people in her age group are watching. Users rising traditional relational operators and new operators that gen may expect diverse recommendations in different contexts erate or combine recommendations. We describe a prototype flex Limited world model: Recommendation approaches deal with ible recommendation engine that realizes the proposed framework two types of entities, users and items(e. g, movies), represented and we present example workflows and experimental results that as sets of ratings or features. Providing recommendations using show its potential for capturing multiple, existing or novel, recom- richer data representations is not straightforward. For example mendations easily and having a flexible recommendation system a user may want recommendations for courses from users with that combines extensibility with reasonable performance. similar grades and similar ratings Categories and Subject Descriptors <. In this paper, we propose FlexRecs, a framework that allows flex- nendations to be easily defined, customiz H.3.3 (Information Storage and Retrieval Information Search cessed over structured data. FlexRecs(a)decouples the definition and Retrieval-Search process of a recommendation process from execution, (b)declaratively de- fines a recommendation process as a high-level workflow and (c) General terms enables generating any recommendations with the same engine A given recommendation approach can be expressed as a high- Algorithms, Languages, Performa level workflow, which may contain traditional relational operators such as select, project and join, plus new recommendation opera- Keywords tors that generate or combine recommendations. A workflow handle data(including recommendations) in relational form. flexible recommendations, recommendation operators, recommen- a designer can easily create multiple, customizable workflows dation queries for content-based, collaborative, as well as novel recommendation paradigms. The end user can select from them, depending on her 1. INTRODUCTION information needs. This selection is done through a gul, which Recommendation systems provide advice on movies also allows the user to enter parameters for workflows in order travel, leisure activities, and many other topics, and I urate and personalized recommendations. For in- ery popular in systems, such as Google News [10]. Al stance, the user may specify that her recommendations be based and MovieLens 19]. Since the appearance of the first on what people in her age group are watching. This choice gets dation systems [12, 22, 261, many recommendation approaches ranslated into a select condition, which is passed to the appropri- have been proposed both by the industry and academia. However ate workflow. This functionality is essentially similar to advanced most recommendation systems have a number of limitations searches: a designer builds a set of parameterized SQL queries End users can execute these queries choosing parameter values to receive different results through an advanced search interface ble recommendation en realizes the proposed framework. The system allows executie is granted without fee provided that copies are workflow over a conventional DBMs by"compiling" it into a se not made or di rofit or commercial advantage and that copies quence of SQL calls. The recommendation operators may call upon bear this notice and the full citation on the first page. To copy otherwise, to functions in a library that implement common tasks for generating publish, to post on servers or to redistribute to lists, requires prior specifi recommendations, such as computing the Jaccard or Pearson sim of two sets of objects, dictions. When possible, library functions are compiled into theFlexRecs: Expressing and Combining Flexible Recommendations Georgia Koutrika, Benjamin Bercovitz, Hector Garcia-Molina Computer Science Department, Stanford University, Stanford, California, USA {koutrika, berco, hector}@cs.stanford.edu ABSTRACT Recommendation systems have become very popular but most rec￾ommendation methods are ‘hard-wired’ into the system making ex￾perimentation with and implementation of new recommendation paradigms cumbersome. In this paper, we propose FlexRecs, a framework that decouples the definition of a recommendation pro￾cess from its execution and supports flexible recommendations over structured data. In FlexRecs, a recommendation approach can be defined declaratively as a high-level parameterized workflow com￾prising traditional relational operators and new operators that gen￾erate or combine recommendations. We describe a prototype flex￾ible recommendation engine that realizes the proposed framework and we present example workflows and experimental results that show its potential for capturing multiple, existing or novel, recom￾mendations easily and having a flexible recommendation system that combines extensibility with reasonable performance. Categories and Subject Descriptors H.3.3 [Information Storage and Retrieval]: Information Search and Retrieval—Search Process General Terms Algorithms, Languages, Performance Keywords flexible recommendations, recommendation operators, recommen￾dation queries 1. INTRODUCTION Recommendation systems provide advice on movies, products, travel, leisure activities, and many other topics, and have become very popular in systems, such as Google News [10], Amazon [17] and MovieLens [19]. Since the appearance of the first recommen￾dation systems [12, 22, 26], many recommendation approaches have been proposed both by the industry and academia. However, most recommendation systems have a number of limitations: Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SIGMOD’09, June 29–July 2, 2009, Providence, Rhode Island, USA. Copyright 2009 ACM 978-1-60558-551-2/09/06 ...$5.00. • Hard Wired: The recommendation algorithm is typically em￾bedded in the system code, not expressed declaratively. From the designer viewpoint, this fact makes it hard to modify the algorithm, or to experiment with different approaches. • No Flexibility: The recommendations provided are fixed. End users are given few choices. For example, a user may be unable to request recommendations for movies that could be jointly seen by her and her friend, or that her recommendations be based on what people in her age group are watching. Users may expect diverse recommendations in different contexts. • Limited world model: Recommendation approaches deal with two types of entities, users and items (e.g., movies), represented as sets of ratings or features. Providing recommendations using richer data representations is not straightforward. For example, a user may want recommendations for courses from users with similar grades and similar ratings. In this paper, we propose FlexRecs, a framework that allows flex￾ible recommendations to be easily defined, customized, and pro￾cessed over structured data. FlexRecs (a) decouples the definition of a recommendation process from execution, (b) declaratively de- fines a recommendation process as a high-level workflow and (c) enables generating any recommendations with the same engine. A given recommendation approach can be expressed as a high￾level workflow, which may contain traditional relational operators such as select, project and join, plus new recommendation opera￾tors that generate or combine recommendations. A workflow can handle data (including recommendations) in relational form. A designer can easily create multiple, customizable workflows for content-based, collaborative, as well as novel recommendation paradigms. The end user can select from them, depending on her information needs. This selection is done through a GUI, which also allows the user to enter parameters for workflows in order to get more accurate and personalized recommendations. For in￾stance, the user may specify that her recommendations be based on what people in her age group are watching. This choice gets translated into a select condition, which is passed to the appropri￾ate workflow. This functionality is essentially similar to advanced searches: a designer builds a set of parameterized SQL queries. End users can execute these queries choosing parameter values to receive different results through an advanced search interface. We describe a prototype flexible recommendation engine that realizes the proposed framework. The system allows executing a workflow over a conventional DBMS by “compiling” it into a se￾quence of SQL calls. The recommendation operators may call upon functions in a library that implement common tasks for generating recommendations, such as computing the Jaccard or Pearson simi￾larity of two sets of objects, or perform more fancy statistical pre￾dictions. When possible, library functions are compiled into the