Putting Users in Control of their Recommendations F. Maxwell Harper, Funing Xu, Harmanpreet Kaur, Kyle Condiff, Shuo Chang, Loren Terveen GroupLens Research University of Minnesota Minneapolis, MN, USA {harper, xuxx0572, kaurx090, cond0155, schang, terveen}@cs.umn.edu ABSTRACT fect their recommendations1; the system is a “black box”. The essence of a recommender system is that it can recom- Therefore, while users may want a recommender to behave mend items personalized to the preferences of an individual di↵erently (“my recommendations are too obscure”, or “my user. But typically users are given no explicit control over recommendations show too many expensive things”), they this personalization, and are instead left guessing about how are not given any means to tell the system to behave di↵er- their actions a↵ect the resulting recommendations. We hy- ently. pothesize that any recommender algorithm will better fit In this research, we explore the idea of giving users con- some users’ expectations than others, leaving opportunities trol over their recommender. What if the user could tell the for improvement. To address this challenge, we study a rec- recommender to de-emphasize obscure content, or to priori- ommender that puts some control in the hands of users. tize a↵ordable content? We hypothesize that such a system Specifically, we build and evaluate a system that incorpo- could leave users feeling more satisfied with their recommen- rates user-tuned popularity and recency modifiers, allowing dations and more in control of the process. users to express concepts like “show more popular items”. To this end, we evaluate a system that gives users control We find that users who are given these controls evaluate over a single variable that is external to (or subsumed by) the the resulting recommendations much more positively. Fur- recommendation algorithm. By examining a single variable, ther, we find that users diverge in their preferred settings, we allow users to control the recommender through simple confirming the importance of giving control to users. actions like “show less expensive items”. Our recommendation method uses a linear weighted com- bination of a personalization variable (the output of a rec- Categories and Subject Descriptors ommender algorithm such as item-item collaborative filter- H.5.3 [Group and Organization Interfaces]: Computer- ing) and a blending variable (a non-personalized content at- supported cooperative work; H.1.2 [User/Machine Sys- tribute such as price); we give users control over the weights. tems]: Human factors; H.3.3 [Information Search and This approach has a number of advantages. It can be incor- Retrieval]: Information filtering porated into any existing item-ranking recommender sys- tem by simply re-ranking result lists. It is computationally cheap, both to recommend items, and to change the recom- Keywords mender’s weights. It is simple to understand, both from a recommender systems; collaborative filtering; social com- user perspective and from a data analytics perspective. puting; user control; personalization; user study; simulation We structure this work around the following research ques- study; MovieLens tions: RQ1: Do users like having control over their recom- 1. INTRODUCTION • mendations? Recommender systems are usually not user-configurable. Recommendation algorithms personalize their responses for RQ2: Given control, how di↵erent are users’ tuned rec- • users by measuring behaviors (ratings, clicks, purchases, sur- ommendations from their original recommendations? vey questions, etc.), sometimes matching these with con- tent attributes (popularity, price, author, etc.) Typically, RQ3: Do users converge to a common “best tuning” • users are given no explanation of how their behaviors af- setting? We focus our analysis on two attributes available in most 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 recommender systems: item popularity, and item age. Item for profit or commercial advantage and that copies bear this notice and the full cita- popularity — the number of users who have interacted with tion on the first page. Copyrights for components of this work owned by others than the item — may be operationalized by measuring data such ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re- as clicks, ratings, and mentions. Item age may be oper- publish, to post on servers or to redistribute to lists, requires prior specific permission ationalized by the time when the item was added to the and/or a fee. Request permissions from [email protected]. system, or when the item was created/released. RecSys’15, September 16–20, 2015, Vienna, Austria. c 2015 ACM. ISBN 978-1-4503-3692-5/15/09 ...$15.00. 1 DOI: http://dx.doi.org/10.1145/2792838.2800179. Some users will try to reverse engineer the algorithm [12] To deepen our understanding of the impact of using pop- gorithms. Research on Tasteweights [2], an interactive rec- ularity and age in item recommendations, we conduct an of- ommender system that allows users to directly manipulate fline simulation study and an online user study in MovieLens many attributes, found that users were more satisfied when (http://movielens.org), a movie recommendation web site they were given control. We extend this idea by evaluat- with several thousand active monthly users. MovieLens ing a recommendation method that works on top of existing members rate movies on a 5 star scale (with half-star in- recommendation methods, and that allows simple user ma- crements). They expect that rating more movies will help nipulation. the system deliver better recommendations. We make several research contributions. We describe a 3. RECOMMENDATION METHOD functional, computationally efficient recommendation method To build a top-N recommendation list for each user, we that can add user-configurability into any existing item rec- build a personalized blending score (s ) for each item, then ommendation framework. Through a simulation study, we u,i sort by that score. The mechanism for computing these develop several insights concerning popularity and age — scores is kept simple so that we can learn from the results, two widely-available, easily understood variables for user and flexible so that we can experiment with di↵erent blend- manipulation. We determine, via a user study, that users ing strategies. who are given control over their recommender evaluate the Equation 1 shows the general form of our method — a resulting recommendations more positively. weighted linear combination of input variables. ru,i is the 1 n recommender’s score of item i for user u; fi ,...,fi are nu- 0 n 2. RELATED WORK meric representations of item features, and wu,...wu are There has been substantial research e↵ort on optimiz- weights that determine the relative importance of the fea- ing recommender algorithms for system-determined outcome tures. measures such as prediction accuracy, click-through rate, 0 1 1 n n su,i = w ru,i + w f + ... + w f (1) purchase dollars, and return visits. Some of the best-known u · u · i u · i approaches are based on nearest-neighbor (e.g., item-item We operationalize this general method as shown in Equa- collaborative filtering [19]) or matrix factorization approaches tion 2. (e.g., SVD++ [14]). 0 1 2 su,i = w predu,i + w popi + w agei (2) Recently, there has been an emphasis on learning to rank u · u · u · [15], a family of machine learning techniques that are trained The specific variables used are: on a ranking-oriented loss function. There are several related approaches to this idea that seek to maximize ranking out- predu,i: a prediction of the rating that user u will as- • comes while remaining computationally feasible (e.g., [6, 4]). sign item i, as generated by an item-item collaborative Learning to rank has the potential for optimizing recommen- filtering algorithm. We use Lenskit [10], configured to dation lists by examining many sources of behavioral input, use cosine similarity scoring and a neighborhood size including implicit feedback such as clicks and purchases [18]. of 20. We clamp the output values to the range of 0.5-5 Algorithms will continue to improve in their ability to opti- to match the values used in the live system. mize system-determined outcomes. We build on this fact by pop : the number of times item i was rated in the past seeking a method that lets users decide their own outcomes. i • year.2 Our recommendation method can extend any algorithm that scores or ranks items. agei: 50 minus the number of years since item i was In this work, we develop a hybrid recommender — an • released, clamped to the range of 0-50. We subtract ensemble of multiple recommender algorithms. Hybrid rec- from 50 so that newer items have higher values, which ommenders have been widely applied to improve prediction is more consistent with typical user preferences. and recommendation accuracy [3, 13, 23, 21]. Weighted en- Weights w0,w1,w2 may take any value; we experiment sembles — linear combinations of multiple recommendation { } algorithms — are one of the most popular methods in this with weights in the range of 0-1. In addition, we examine only one of w1,w2 at a time, which we accomplish by area [3]. For example, [1] describes introducing a simple lin- { } ear ensemble model in Netflix, incorporating predicted rat- setting the other variable’s weight to zero. ings with many other features. We build on this literature by We scale predu,i, popi,andagei to the range of 0-1 by allowing the user to vary the weights, and by conducting a computing their percentile rank before combining them, in user study to measure user perceptions of di↵erent weights. order to make their weights comparable.