A Bayesian User-Controllable Recommender System

Abstract

In this research, we propose a Bayesian User-Controllable Recommender System. Our approach allows the user to control the contextual information, i.e., the user can de ne the content (other users and items) and parameters (users, items, novelty and popularity) used by the recommender to compute predictions. To demonstrate the usefulness of our proposal, we present di erent scenarios where we change the context con guration and discuss the system outputs. The student was supervised by Evandro Costa { Federal University of Alagoas & Joseana Fechine { Federal University of Campina Grande.

MOTIVATION AND RESEARCH CHAL

LENGES

Historically, the main goal of the Recommender Systems (RS) have been increasing the accuracy of the recommendation [ 8 ]. However, the RS's accuracy is not always correlated with a good user experience with the system [ 10 ]. User experience is de ned as the analysis of the human factors captured through user's interaction with the RS, e.g., user satisfaction and user engagement [ 8 ]. A strategy to enhance the user experience in RS is to give the control of the system to the user. For example, Knijnenburg et al. [ 7 ] allow users to order the recommendation list according to di erent attributes in a RS of best practices for energy efciency. Some approaches allow their users to modify the content that will be used by the RS [ 2, 6 ]. Another line of work, using hybrid models, allows their users to modify the weights assigned to each recommender via sliders, e.g., [ 10, 4 ]. However, the e ects caused by giving the control to the users and how they are related to the user characteristics are issues that still need investigation.

In this research, we address these issues by proposing an approach to RS that allows the user to control the system. Our research challenges are: (i) How to de ne the RS elements controllable by the user? (ii) How to design a user interface to capture the user's preferences? (iii) How to dene the computational techniques used in the RS? (iv) How to represent the items and users in the system? and (v) How to validate our approach via user study? Until now, we have de ned a Bayesian user-controllable model and apply it in the scienti c paper recommendation. Our proposal is based on the concept of context, Context-Aware RS make use of contextual information with the aim of improving the recommendation. Contextual information is any information that could be used to impact the accuracy of recommendation, e.g., climate weather information, demographic information [ 1 ]. In our model, users can control the elements (researchers and papers) that will be used as content to generate the recommendations. Furthermore, users can setting up the parameters weights of the RS, such as, users, items, novelty and popularity. To demonstrate the utilization of our approach, we present scenarios where we illustrate the use of the system, i.e., showing the system inputs and discussing the outputs. In summary, we aim to achieve the following contributions for the eld: (i) a new recommendation method to capture user preferences; (ii) a new user interface for RS; and (iii) we intend to bring more evidence on the relationship among the size of the control given to the user, the user experience and the user characteristics. 2.

RELATED WORK

The analysis of the controllability in RS is a relatively recent topic at the same time that is gaining more attention from the academy. Knijnenburg et al. [ 7 ] study the e ects of di erent mechanisms of interaction in a RS for energy saving. They conclude that the best interaction mechanism depends on the user characteristics: more experienced users, who had more domain knowledge had preferences for more controllable interfaces, e.g., hybrid and explicit mechanisms. In contrast, less experienced users opted for interfaces with less controllable, e.g., non-personalized methods.

Bostandjiev et al. [ 2 ] present the TasteWeights, a controllable hybrid RS of music that integrates the content from three sources, Wikipedia, Facebook and Twitter. Their results indicate that using their proposed interface to explain the hybrid process of recommendation increased the user satisfaction. Knijnenburg et al. [ 6 ] also explore the TasteWeights, they use the musics liked by the user in Facebook (items) and the user's Facebook contacts (users). They conclude that the controllability led to a better user experience, however the di erences between the two types of control, items and users, was not statistically signi cant.

Harper et al. [ 4 ] study the user controllability in the context of movie recommendation using the MovieLens system. They allow the users to control the weight of two variables, the popularity of the movie and and the year of movie release. They did not nd an overall optimization of parameters that works for all users, simply, some users prefer to change the con guration of the recommendation list and others not. Also considering the MovieLens system, the work of Ekstrand et al. [ 3 ] allows the user to choose among four algorithms, one to be used. However, the authors found no evidence that would lead a user to choose a particular algorithm.

Considering article recommendation, Parra and Brusilovsky [ 10 ] present the SetFusion, a controllable graphical interface composed by sliders, where the users could assign weights to di erent RS algorithms. In addition, the interface showed a Venn diagram, indicating which algorithm was predominant on the recommendation calculation. Their results indicated that users were engaged and had a better user experience with the controllable interface, however the e ect was signi cant only in the case that the user has gained experience with the basics of the system, i.e., the users who used the controllable interface after the non-controllable.

PROGRESS TO DATE

We de ne a model that gives the RS control to the users and, thus, they will be able to de ne the elements and parameters that will be used as input by the recommendation model. We use the following contextual information: Users { this variable concerns the researchers who are saved in the system, the user can de ne which researchers she wants to include into the context. Thus, to calculate the prediction of the items, the RS will give more importance to the items similar to those included researchers. Items { this attribute represents the items that the user can insert into the context, thus the RS will search for similar articles to compose the recommendation list. Novelty { this variable is related to how much novelty a paper has, e.g., we consider that survey paper have a low level of novelty. Likewise, newer papers receive a higher value than the new older articles. Popularity - this attribute is de ned by how popular a paper is, we consider that the more citations the article has, the more popular it will be and increase the chance of it being recommended.

De nitions { Let U = fu1; :::; ujUjg; jU j > 0 be the set of all users in the systems and let D = fd1; :::; djDjg be the set of all documents in the system. Each document dj 2 D has three attributes represented by a 3-tuple dj = (sj ; yj ; rj ), where sj represents a textual description of the document, yj indicates the year that the document was released, and rj 0 indicates the number of citations of the document. Each user ui 2 U has a set Pi = fpi;1; :::; pi;jPijg; jPij 0, that represents her portfolio. Each item pi;l 2 Pi has two attributes pi;l = (si;l; yi;l), where si;l represents a textual description of the document and yi;l yd indicates the year that the item was included in her portfolio Let T = ft1; :::; tjT jg be the set of terms used to index the documents and to represent user pro les. So, each document dj 2 D is represented by the vector d~j = (wj;1; :::; wj;jT j), where wj;k 0 represents the importance of the term tk 2 T to the document dj . The user ui 2 U has a pro le denoted by the vector u~i = (wi;1; :::; wi;jT j), where wi;k 0 represents the importance of the term tk 2 T to the user ui. Each user ui 2 U has a set Ci = fci;1; :::; ci;jCijg, 0 jCij 10 of contexts, where a context ci;m 2 Ci is de ned by a 7-tuple ci;m = ftitlei;m; Ui;m; Di;m; useri;m; itemi;m; novi;m; popi;mg , where: titlei;m { is a not-empty string that represents the context title; Ui;m U; 0 jUi;mj 5 and Di;m D; 0 jDi;mj 5 { are, respectively, the researchers set and the items set inserted by the user ui into the context ci;m; useri;m, itemi;m, novi;m, popi;m 2 [0; 1] { represent, respectively, the weights assigned by the user to the attributes Users, Items, Novety, Popularity of the context ci;m. To build the user pro les u~i; u~2; :::; u~jUj, we adopt our approach published in [ 9 ]. So, given a user ui, we create the user portfolio Pi using her curriculum vitae, i.e., resume, formation, projects and production. We crawl the user curriculum vitae from the CVLattes (http://lattes.cnpq.br/) and merged it with DBLP publications. To index the items and create the vectors d~1; d~2; :::; dj~Dj, we use the TF-IDF procedure, more details can be found in [ 9 ]. In the rest of this section, when we present the equation sim, we are referring to the cosine similarity between two vectors.

The Bayesian Network { The RS is a Bayesian Network (BN), wherein for each available paper for recommendation is created one BN that calculates its prediction for the given context. This prediction is obtained by the value of the variable P rediction of the BN, and it is in uenced by four variables of the context, U sers, Items, P opularity and N ovelty. Each context variable, receives the in uence of two variables, one corresponding to a user's preference and another related to an item feature. All BN nodes have two states, T rue and F alse, the BN has four types of node: Prediction node { in the BN, only the node Prediction is such, it represents the predictive value of an item to a user in a given context. In other words, the higher the value of P (P rediction) greater will be the predicted value of the item to the user. Parameter nodes { this type of node represents the parameters used in the RS, the nodes of this type are parents of the node Prediction. We de ne four parameters, Users, Items, Popularity and Novelty, so we create a node with the same name for each parameter. Each node of this type has two parents, one representing a user preference and another representing an item feature. User preference nodes { this type of node represents the weight given by the user to a speci c parameter, i.e., it serves as an interface between the user and the model. There are four nodes of this type, UserUsers, UserPopularity, UserNovelty and UserItems. Item feature nodes { such node represents a paper feature. There are four nodes of this type, one for each parameter, ItemUsers, ItemPopularity, ItemNovelty and ItemItems.

To de ne the conditional probabilities of the Prediction node, we follow the approach de ned by Zapata and Greer [ 12 ]. Their approach simpli es conditional probabilities definition by simply de ning the in uence of each parent node on the child node. In this work, we admit that all Parameter nodes have a strong in uence on the Prediction node. To de ne the conditional probabilities of the Parameter nodes we use the Noisy-AND strategy with = 0:05 [ 5 ]. Figure 1 presents the BN and its conditional probabilities tables.

Computing the item prediction { Firstly, the user preference nodes are calibrated by the values given by the user on the context. Then, the weights inserted by the user are mapped to the BN as follows: P (U serU sers) = useri;m, P (U serItems) = itemi;m, P (U serN ovelty) = novi;m and P (U serP opularity) = popi;m. This con guration of the BN will be valid for the context ci;m while the user does not perform new changes in their weights. Then, for each available paper for recommendation, we make a copy of the BN and modify the weights of its item feature nodes according to the paper features. Given a document dj 2 D, in the following we present how the item feature nodes, ItemUsers, ItemItems, ItemNovelty and ItemPopularity, are calibrated.

ItemUsers node { Given the users set Ui;m U created by the user ui in the context ci;m and the paper dj 2 D, the weight of this node is obtained by the average of the similarities between the document dj and the users u 2 Ui;m:

Pu2Ui;m sim(d~j;u~) P (ItemU sers) = jUi;mj . ItemItems node { Given the items set Di;m D of the context ci;m created by the user ui, its weight is calculated by the average of the similarities between the document dj and the documents Pd02Di;m sim(d~j;d~0) d0 2 Di;m: P (ItemItems) = jDi;mj . ItemNovelty node { We calculate its weight using the auxiliary BN presented in Figure 2. The weights of the BN also follow the Zapata and Greer's approach [ 12 ], where nodes ItemYear and ItemSurvey have a strong in uence on the node ItemNovelty. We calculate the weight of the node ItemSurvey analyzing the document description sj, if the description contains the terms survey or review, your weight will be P (ItemSurvey) = 0, otherwise P (ItemSurvey) = 1. The weight of the node ItemYear is calculated based on the date of publication yj of the item dj: P (ItemY ear) = 0:9ynow yj , where ynow is the current year. ItemPopularity node { The weight of this node is proportional to the number of citations rj of the document dj. Let R = (r1:::; rjDj) be a list of all citation numbers of the papers dj 2 D, so, we calculate the probability P (ItemP opularity) according to:

P (ItemP opularity) = if rj 2]Q1(R); Q2(R)]; (1) 80 if rj = 0; > > >>0:25 if rj 2]0; Q1(R)]; > <

0:5 >>>0:75 if rj 2]Q2(R); Q3(R)]; > :>1 otherwise, where Q1(R); Q2(R); Q3(R) are, respectively, the 1st, 2nd and 3rd quartiles of the list R.

Scenarios { For each scenario we present the model input and discuss its outputs. Table 1 presents the papers we use in the scenarios to compose the recommendation list. Table 2 presents the models outputs for all scenarios.

Scenario 1 { In this rst scenario, which consists of the most basic case, we suppose that the user started and saved a context without changing the parameters. So, the context ci;m is set up with the following parameters: Ui;m = ?, Di;m = ?, useri;m = 0:5, itemi;m = 0:5, novi;m = 0:5 and popi;m = 0:5. We can see from the Table 2 that the paper d2 has the best balance of P opularity and N ovelty, so, therefore it is on the top of the list.

Scenario 2 { Now, assume that the user has made the following changes in the context. She added the researcher D. Parra, the papers [ 11 ] and [ 4 ], and changed the con guration to give more importance to the U sers variable. Thus, the context will have the following con guration: Ui;m = fD:P arrag, Di;m = f[ 11 ], [ 4 ]g, useri;m = 1:0, itemi;m = 0:3, novi;m = 0:3 and popi;m = 0:3. Thus, rst we calculate the similarities among the papers and the elements that may be used in context, Table 1 displays such similarities. Note that the values shown in Table 1 are illustrative and may vary according to the dataset used. It also presents the probabilities P (ItemU sers) and P (ItemItems). Verifying Table 2, we can see that the recommendation order has been modi ed, this was because the article d3 has a greater similarity with the elements in the context.

Scenario 3 { In this scenario, we assume that the user wants to receive more similar items to the papers she inserted into the context. Thus, the user changes the context to: Ui;m = fD:P arrag, Di;m = f[ 11 ], [ 4 ]g, useri;m = 0:3, itemi;m = 1:0, novi;m = 0:3 and popi;m = 0:3. We can see in Table 2 that the list ordering has changed and the paper d2 is on the top of the list, because it has higher similarity with the items d 2 Di;m.

Scenario 4 { Now, the user is concerned with the popularity of the recommended papers and change the context to the following con guration: Ui;m = fD:P arrag, Di;m = f[ 11 ], [ 4 ]g, useri;m = 0:3, itemi;m = 0:3, novi;m = 0:3 and popi;m = 1:0. Analyzing Table 2, we verify that the papers are ordered by the citation number, then the paper with more citations, d1, is on the top of the list now.

Scenario 5 { In this last scenario, the user wants more recency papers, i.e., paper with novelty, thus she modi es the context to: Ui;m = fD:P arrag, Di;m = f[ 11 ], [ 4 ]g, useri;m = 0:3, itemi;m = 0:3, novi;m = 1:0 and popi;m = 0:3. In this con guration, the paper d2 assumes the top of . the list, because it is a relatively new paper and it is not a survey, and it presents a good balance among variables.

FUTURE WORK

For future work, we aim two steps, the rst one is to de ne a controllable interface for the user to manage her contexts. Basically, the interface must provides the following features to the users: (i) create a context; (ii) set up a context (iii) save a context; (iv) delete a context; (v) duplicate a context; and (vi) provide feedback on the recommendation. In the second step, we aim to perform a user study in laboratory. Thus, we must de ne the experiment planning, i.e., de ne the procedure, subjects, variables, statistical tests, etc.

References

[1] G. Adomavicius and A. Tuzhilin . Recommender Systems Handbook, chapter Context-Aware Recommender Systems , pages 217 { 253 . Springer US , Boston, MA, 2011 .

[2] S. Bostandjiev , J. O'Donovan , and T. Ho llerer. Tasteweights: A visual interactive hybrid recommender system . In Proceedings of the Sixth ACM Conference on Recommender Systems, RecSys '12 , pages 35 { 42 , New York, NY, USA, 2012 . ACM.

[3] M. D. Ekstrand , D. Kluver , F. M. Harper , and J. A. Konstan . Letting users choose recommender algorithms: An experimental study . In Proceedings of the 9th ACM Conference on Recommender Systems, RecSys '15 , pages 11 { 18 , New York, NY, USA, 2015 . ACM.

[4] F. M. Harper , F. Xu , H. Kaur , K. Condi , S. Chang , and L. Terveen . Putting users in control of their recommendations . In Proceedings of the 9th ACM Conference on Recommender Systems, RecSys '15 , pages 3 { 10 , New York, NY, USA, 2015 . ACM.

[5] M. Henrion . Some Practical Issues in Constructing Belief Networks . In L. N. Kanal, T. S. Levitt , and J. F . Lemmer, editors, Uncertainty in Arti cial Intelligence 3 , pages 161 { 173 . North-Holland, 1989 .

[6] B. P. Knijnenburg , S. Bostandjiev , J. O'Donovan, and A. Kobsa . Inspectability and control in social recommenders . In Proceedings of the Sixth ACM Conference on Recommender Systems, RecSys '12 , pages 43 { 50 , New York, NY, USA, 2012 . ACM.

[7] B. P. Knijnenburg , N. J. Reijmer , and M. C. Willemsen . Each to his own: How di erent users call for di erent interaction methods in recommender systems . In Proceedings of the Fifth ACM Conference on Recommender Systems, RecSys '11 , pages 141 { 148 , New York, NY, USA, 2011 . ACM.

[8] J. A. Konstan and J. Riedl . Recommender systems: From algorithms to user experience. User Modeling and User-Adapted Interaction , 22 ( 1-2 ): 101 { 123 , Apr . 2012 .

[9] J. Magalha ~es, C. Souza , E. Costa, and J. Fechine . Recommending scienti c papers: Investigating the user curriculum . In The Twenty-Eighth International Flairs Conference, 2015 .

[10] D. Parra and P. Brusilovsky . User-controllable personalization: A case study with setfusion . International Journal of Human-Computer Studies , 78 : 43 { 67 , 2015 .

[11] N. Tintarev and J. Mastho . Evaluating the e ectiveness of explanations for recommender systems . User Modeling and User-Adapted Interaction , 22 ( 4-5 ): 399 { 439 , Oct . 2012 .

[12] J.-D. Zapata-Rivera and J. E. Greer . Interacting with inspectable bayesian student models . International Journal of Arti cial Intelligence in Education , 14 ( 2 ): 127 { 163 , 2004 .