Group Recommender Systems

Some Experimental Results

Vineet Padmanabhan

, Prabhu Kiran

and Abdul Sattar

Department of Computer & Information Sceinces, University of Hyderabad, Hyderabad, India

Institute for Integrated and Intelligent Systems, Grifﬁth University, Brisbane, Australia

Keywords:

Machine Learning, Recommender Systems.

Abstract:

Recommender Systems (RS) are software applications which aim to support users in their decision making

while interacting with large information spaces. Most recommender systems are designed for recommending

items to individuals. In this paper we provide experimental results related to developing a content-based

group recommender system. To this end we make two important contributions. (1) Implementation of a

group recommender system based on decision-lists as proposed recently in (Padmanabhan et al., 2011)

using MovieLens dataset which is a relatively huge data-set (100,000 ratings from 943 users on 1682 movies)

as compared to the data-set size of 150 used in (Padmanabhan et al., 2011) (2) We use seven variants of

decision-tree measures and built an empirical comparison table to check for precision rate in group

recommendation based on different social-choice theory strategies.

1 INTRODUCTION

Though it is a well known saying that Information

is Wealth people nowadays ﬁnd it difﬁcult to get

useful information because of the huge amount of

data available in the Internet in the form of books, ar-

ticles,movies, music, web sites etc. Therefore, select-

ing an item that is of ones own interest have become

a very difﬁcult task. So we need systems that help

in ﬁltering the content available and suggest only the

data that is of our interest. Such systems are com-

monly known as Personalised Recommender Sys-

tems. Recommender systems have become valuable

resources for users seeking intelligent ways to search

through the enormous volume of information avail-

able to them. Based on what kind of recommendation

techniques are used, personalized recommender sys-

tems are usually classiﬁed into three categories (Ado-

mavicius and Tuzhilin, 2005) (a) Collaborative Filter-

ing (CF) (b) Content based Recommending (CB) and

proposed in the Literature which makes use of the

above techniques as well as other recommendation

strategies like demographic-based (Pazzani, 1999),

utility-based, knowledge-based and association rule-

based techniques (Jananch et al., 2010). In this pa-

per we are mainly concerned with Content-based rec-

ommendation. In content-based recommendation one

tries to recommend items similar to those a given user

has liked in the past, whereas in collaborative recom-

mendation one identiﬁes users whose tastes are simi-

lar to those of the given user and recommends items

they have liked. For instance, a content-based recom-

mendation would be something like Movie X is rec-

ommended because its category is Action and con-

tains the term Bruce Willis, which are features con-

tained in article you rated. A collaborative recom-

mendation would be like Movie X is recommended

because other users similar to you have liked it. For

example, if Bob and Wendy liked the same movies as

you in the past and they both rated Star Wars highly,

you might like it, too.

Most of the previously published studies in rec-

ommender systems focus on the technique of building

personalized/single-user recommender systems and

hence is not suitable for supporting purchasing de-

cisions of a group. Those that have addressed the

problem of group recommender systems (Masthoff,

2003; McCarthy and Anagnost, 1998; O’Connor

et al., 2001) assume that the input of the system is

comprised of items ratings given by individuals and

the group recommendation is obtained by combining

or aggregating (based on some predeﬁned aggrega-

tion strategy) the individual recommendations of the

members in the group. The problem with this ap-

proach is that (1) the ratings are combined without

370

Padmanabhan V., Kiran P. and Sattar A..

Group Recommender Systems - Some Experimental Results.

DOI: 10.5220/0004331003700376

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 370-376

ISBN: 978-989-8565-39-6

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

considering the interaction of group members which

may lead to incorrect recommendations for a group.

(2) It is difﬁcult to specify the additional informa-

tion which may be required from the user to deter-

mine the exact combination/aggregation strategy and

(3) Lot of time will be required even if opinion from

domain experts are sought to guide the combination

process. In this paper we follow our previous work

on group recommender system (Padmanabhan et al.,

2011) wherein the group recommender problem is de-

ﬁned as: Let I = i

, i

, . . . , i

be the set of all

items and U = u

, u

. . . u

be the set of all users.

IG is the set of items that are not rated by G (sub-

set of U). Goal is to ﬁnd ratings for items in IG or

to ﬁnd whether the items in IG are recommendable to

the given group of people (G). Other related works in

the area of group recommender systems and how they

differ from our model is summarized below.

In (Chen et al., 2008) a group recommendation

approach based on collaborative ﬁltering and genetic

programming is proposed. In this work the authors

assume that the input data already contains ’items’

ratings given by individuals and then use genetic al-

gorithm to get ’items’ group ratings. In our ap-

proach both individual as well as group ratings are

learned. The individual ratings are learned by the rule

learner and the group ratings using social choice the-

ory strategies. Moreover we use content based ap-

proach whereas in (Chen et al., 2008) the approach

is that of collaborative ﬁltering and hence suffers

from cold-start, ﬁrst-rater and popularity bias prob-

lems. (Yu et al., 2006) make use of content based

approach and outlines a method to merge individual

user proﬁles to get common user proﬁle. Here the

merging is based on individual user preferences on

features (e.g. genre, actor and keyword about a pro-

gram) whereas we combine individual user ratings on

whole programs rather than features. The obvious dis-

advantage of this approach is that it increases the time

and effort required to do the recommendation. (Tubio

et al., 2008) also uses content-based approach but the

focus is more on developing an ontology language

like OWL through which digital TV programs can

be described and then to relate them through their se-

mantic characteristics. There are no experimental re-

sults to show how this can be acheived. In (Masthoff,

2004) no mention is made on how to get user pro-

ﬁles. Social choice theory strategies are mentioned

but again how to include a learning component to

make use of those strategies is not shown. (de Cam-

pos et al., 2007) proposes a group recommender sys-

tem based on Bayesian Network. They do not discuss

about how groups are formed but sums up by saying

that a group is a new entity where recommendations

are made by considering the particular recommenda-

tions of its members in some way. In our case we are

more interested in combining individual user models

to adapt to groups such as how humans select a se-

quence of television items to suit a group of viewers.

2 VSW METHOD

Vineet et.al., (Padmanabhan et al., 2011) proposed

a movie group recommender System based on

Decision List Rule Learner (Rivest, 1987; Cohen,

1995; Quinlan, 1996) and social choice theory strate-

gies (Masthoff, 2003) (here we refer their approach as

VSW Method). They used a Data set of 150 movies

where each is a collection of 12 attribute-value pair.

Movies are rated on a 5-scale 0, 1, 2, 3, 4 [Bad, Av-

erage, Above-average, Good, Excellent]. The VSW

approach is shown in Figure 1 and is based on the

following 5 steps (1) content based recommendation

technique (2) merging recommendation list strategy

(3) social choice strategies to get group recommenda-

tion list (4) RTL strategy to get group recommenda-

tion list and (5) Decision List Rule Learner (DLRL).

Figure 1: Recommender system based on VSW method.

The working of vsw method as shown in Figure

1 is as follows: initially, the system has no idea to

recommend any programs except if we add any ex-

ternal rules. For few weeks, it will collect ratings for

each and every program the user has watched. These

are the training examples to the learning algorithm.

From these training examples the learning algorithm

learns the set of rules that cover all training examples.

This process will be done for every user i.e., for each

user, we get a separate rule base (User proﬁle). Now

we know that any TV guide contains information re-

garding TV programs i.e., Day, Date, Time, Channel

etc. By using the rule base of each user we generate

predicted ratings of programs in the TV guide which

GroupRecommenderSystems-SomeExperimentalResults

371

is shown as TV Guide + Classiﬁcation in Figure 1.

These programs with predicted ratings are nothing but

recommendation list for each individual user. Finally,

social choice theory strategies are used to get a com-

bined (group) recommendation list.

Learning algorithm plays major role in content

based recommendation approach. It is used to learn

user proﬁles. Our learning algorithm (DLRL) as

mentioned in Figure 1 is a decision list rule learner

based on RIPPER (Cohen, 1995) and FOIL (Quin-

lan, 1996) rule learners. It is a multi-class rule learner

wherein there are ﬁve classes : bad, average, above

average, good, excellent. Initially, all training ex-

amples are divided into two sets: training data and

prune data. Training data is used to learn the set

of rules. Prune data is used to prune the rules to

avoid over-ﬁtting. FOIL Information gain is given

as FOIL Gain(L, R) = t(log

(

) − log

(

))

where L is the candidate literal to add to rule R, p

the number of positive bindings of R, n

is the number

of negative bindings of R, p

is the number of positive

bindings of R + L, n

is the number of negative bind-

ings of R + L, t is the number of positive bindings of

R also covered by R + L. The formula used to prune

the rule is deﬁned as v =

(p−n)

(p+n)

where p is the number

of positive examples covered by the rule in prune data

set and n is the number of negative examples covered

by the rule in the prune data set. Pruning criteria is

deleting the ﬁnal sequence of conditions that maxi-

mizes v. The different steps involved in our learning

algorithm is shown in Algorithm 1.

Algorithm 1: Learning Algorithm used in VSW.

Input: Train Data,Prune Data

Output: set of rules

Step 1: foreach class do

Find the number of training examples for that class;

Take the class with maximum number of examples, make that as

Default class;

Step 2: Take an empty RuleSet;

while No class has left do

take the next smallest class;

Consider training examples for that class as positive,

remaining as negative;

while All positive examples covered do

Take empty Rule;

Add conjuncts to rule as soon as it improves FOIL

Information gain;

prune the rule by deleting any ﬁnal sequence of

conditions;

Mark covered positive examples by this rule;

Add this rule to RuleSet;

Step 3:Add Default Rule to RuleSet;

Return RuleSet;

Table 1: Example to demonstrate Social choice Strategies.

Tv-Programs A B C D E F G H I J

John 10 4 3 6 10 9 6 8 10 8

Adam 1 9 8 9 7 9 6 9 3 8

Mary 10 5 2 7 9 8 5 6 7 6

In VSW, mention is made about four social

choice theory strategies as out- lined in (Masthoff,

2003). The main idea behind social choice the-

ory is (also called as group decision making) decid-

ing what is best for a group given the opinions of

individuals. The strategies used in the implemen-

tation are (1) Utilitarian Strategy (Hogg and

Jennings, 1999): In this strategy, instead of using

ranking information, utility values are used. This can

be done in multiple ways, additive or multiplicative.

For example, the utility values for the 10 programs in

Table 1. will be 21, 18, 13, 22, 26, 26, 17, 23, 20,

22 respectively (column-wise addition). The TV pro-

gram E and F are having highest utility values. So

either E or F is the recommended program. (2) Least

misery strategy (O’Connor et al., 2001): In

this strategy, the item with large minimum individ-

ual rating will be recommended. The idea behind this

strategy is that a group is as happy as its least happy

member. For Example, the group rating for the 10 TV

programs based on Least Misery Strategy will be 1, 4,

2, 6, 7, 8, 5, 6, 3, 6 respectively. From the above group

ratings, F has the highest rating. So TV program F

is recommended by Least misery strategy. (3) Most

pleasure strategy (Masthoff, 2004): Making

new list with the maximum of individual ratings. For

Example, from Table 1 the group rating for 10 TV

programs based on this strategy will be 10, 9, 8, 9,

10, 9, 6, 9, 10, 8 respectively. From the above group

ratings, A, E, I are having the highest rating values.

So either A, E or I will be the recommended TV

program. (4) Average without Misery strategy

(McCarthy and Anagnost, 1998): In this strat-

egy, a new list of ratings is made with the average of

the individual ratings, but without items that score be-

low a certain threshold for individuals. The item with

maximum value will be recommended. For example,

from Table 1. considering a threshold of 4 the average

values for 10 TV programs will be -, 18, -, 22, 26, 26,

17, 23, -, 22 respectively. The TV program E and F

are having highest utility values. So either E or F is

the recommended program.

(Padmanabhan et al., 2011) suggested that a single

strategy alone would not be sufﬁcient to get the most

accurate result as far as group recommendation is con-

cerned. To address this problem a combined strategy

was put forward that considers three factors: (1) Least

group member happy (like least misery strategy) (2)

Most group member happy (like most pleasure strat-

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

372

egy) and (3) Total group happy (like Utilitarian strat-

egy) and named the strategy as RTL (Repeat Total plus

Least group happiness strategy). The strategy can be

explained as follows: Let G be a group consisting of

N users and I be the set of instances. Remove in-

stances with a user rating ”0”. If all instances in I

have user rating 0 then continue. For each instance

in I, calculate the sum of least happiness and total

happiness. Recommend the instance with maximum

value. If we have maximum value for multiple in-

stances, remove other instances from I and remove the

minimum values (least happiness) from instances in I

and apply the same above process for the new set of

instances repeatedly. Hence the name Repeat Total

plus Least group happiness strategy. For ex-

ample, let us take ﬁve users and two TV programs

with ratings, I

={1, 2, 1, 1, 4}, I

= {1, 2, 2, 2, 2}.

=group happiness + least happiness = 9 + 1 = 10.

Similarly C

= 9 + 1 = 10, here maximum value =10.

Therefore, I

will be {2, 4} and I

will be {2, 2, 2, 2}.

Now, C

= 6 + 2 = 8 and C

= 8 + 2 =10. Here maxi-

mum value is 10 for C

. Therefore instance 2 will be

recommended.

3 EXPERIMENTAL RESULTS

USING VSW METHOD WITH

MovieLens DATASET

The VSW method (Vineet et.al. [6]) was imple-

mented on a movie data set which had around 150

entries. The main reason for outlining this work is

that we tried to evaluate the performance of VSW

method by using a real-world movie data set like that

of MovieLens data which has more than one hun-

dred thousand entries. We also looked into other as-

pects of the VSW method like trying to use a data

structure like decision tree instead of Decision list

and re-evaluated the performance. Experiments are

performed with a real data set(MovieLens) that has

been used as benchmark in prior works. MovieLens

data sets were collected by the GroupLens Research

Project at the University of Minnesota. This data set

consists of (a) 100,000 ratings from 943 users on 1682

movies (b) The range of rating is between 1 (bad) and

5 (excellent) (c) Each user has rated at least 20 movies

(d) Movies are classiﬁed according to their 19 genres

and these genres are the features of the movies in our

data set (e) The 19 genres are: Unknown, Action, Ad-

venture, Animation, Childrens, Comedy, Crime, Doc-

umentary, Drama, Fantasy, Film-Noir, Horror, Musi-

cal, Mystery, Romance, Sci-Fi, Thriller, War, Western

(f) 1 under a particular genre indicates that movie is of

that genre and 0 indicates it is not (g) Movies can be

in several genres at once. The metric that we used to

calculate the performance of our group recommender

system (VSW-GRS) based on the data given above is:

value =

∑

t=1

∑

j=1

max

precision =

value

where k is the number of test instances, m is the size

of group, r

is the rating of user j on test instance

t and max

is the maximum rating given by user j.

The performance of our group recommender system

using the above metrics and FOIL information gain

(FOIL-Gain) as mentioned earlier, with groups of size

10, 20, 30, 40, 50, 75 and 100 is shown in Figure 2.

Utility, MPS, LMS, and RTL stands for the respec-

tive social startegies used. It is imperative to mention

here that in (Padmanabhan et al., 2011) the implement

ion of a movie Group Recommender using a Decision

List Rule Learner was done with a data set of size

150. In the current set of experiments with VSW we

have used the Movielens data set which is in tune of

around 100000. It should be clear from Figure 2 that

a decision list based recommender system with FOIL-

GAIN and using RTL and utilitarian social startegies

gives similar precision.

Figure 2: Decision-List with FOIL-gain.

3.1 Comparison of Decision Tree

Measures in VSW-GRS

In the previous section we outlined the implementa-

tion results of VSW-GRS which in turn makes use

of the Decision-List data structure and FOIL-Gain

information measure. In this section we make use of

different decision-tree (Mingers, 1989) selection

measures along with each of the social choice

strategies and RTL and compare the precision

rate in making group recommendations based on

the MovieLens dataset. Our implementation results

GroupRecommenderSystems-SomeExperimentalResults

373

Table 2: Results show that RTL is out-performing than other strategies.

Measure Utility MPS LMS RTL

(ADD)

chi-square 0.935000 0.872500 0.860000 0.950000

Info-gain 0.913607 0.875000 0.915000 0.913607

G-stat 0.927672 0.910000 0.910000 0.927672

Gain Ratio 0.865000 0.796667 0.840000 0.865000

GINI 0.910000 0.820000 0.820000 0.910000

Marshalls 0.959999 0.790000 0.790000 0.959999

show that RTL out-performs all other Social choice

strategies. The different selection measures used in

the implementation are (1) Quinlans Information

measure(IM), (2) The chi-square statistic

(3) The G statistic(G) (4) GINI index of

Diversity(GINI) (5) Gain-Ratio Measure(GR)

(6) FOIL Gain and (7) Marshalls Correction.

Since these decision tree measures are well known

in the Machine learning community we do not

feel the necessity of explaining each one. The

implementation details are as follows:

Training and test sets are formed by dividing the

entire data set into 80% − 20% sets respectively. A

Model is built on the training set and we evaluated its

performance on the test set using the metric given in

the previous section. Results of this analysis with size

of group 10 is shown in Table 2. The data is plotted

using a matrix. When the data is represented in the

matrix format, the X and Y coordinates are the index

of row and column. The Z coordinate value ranges of

matrix cells. The performance of above metrics with

size of groups

{

10, 20, 30, 40, 50, 75, 100

}

opposite to

every selection measure is depicted in ﬁgures (Figure

4, Figure 5, Figure 8, Figure 7, Figure 6, and Figure

9). Here the X and Y coordinates are the size of a

group and precision. The precision is evaluated for

every social choice strategy (MPS, LMS, Utility) and

RTL. Talking about the results as they stand, the Mar-

shall correction (Figure 9) and the G-statistic (Fig-

ure 8) is marginally best and the gain-ratio (Figure 5)

with probability not the best. In fact, the results show

that accuracy is not improved signiﬁcantly by using

a measure at all. However, the choice of measure

does signiﬁcantly inﬂuence the size of unpruned trees.

Randomly selecting attributes produces trees roughly

twice as large as those produced with an informed

measure. Between the measures, the gain ratio gener-

ates the smallest trees, whereas chi-square produces

the largest, which will affect the performance of the

Recommendation system.

Figure 3: Precision with group size=10.

Figure 4: Precision with Chi-Square.

4 CONCLUSIONS

One important conclusion of this paper is that the em-

pirical results show that the predictive accuracy of

Group recommendation is not affected by the selec-

tion measures used in decision tree construction. A

second important conclusion is that the RTL method

is giving better results than any other social choice

strategies irrespective of measures used in the con-

struction of decision trees. In fact, the results show

that accuracy is not improved signiﬁcantly by using

a measure at all. Selecting attributes entirely ran-

domly produces trees that are as accurate as those

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

374

Figure 5: Precision with Gain-ratio.

Figure 6: Precision with Gini.

produced using a particular measure. In the case of

decision-list based VSW we saw that both RTL and

Utilitarian strategy were having the same precision.

Moreover we have used MovieLens data set for ex-

perimental evaluation which is a relatively huge data

set and is a benchmark dataset from an industrial per-

spective too. There are several ways in which our

method could be extended. One aspect which we

have not looked into is how requirements( Adityaet

al.,2011 (Parameswaran et al., 2011)) affect recom-

mendations. For instance, in a University environ-

ment, for a student to graduate the student needs to

satisfy a bunch of requirements like take 2 courses

from a, b, c, d, but b and c together don’t count.

Also we have not looked into the problem of pre-

requisites (Parameswaran et al., 2010a) wherein when

we make recommendations we need to make sure that

we recommend a package of items such that the pre-

requisites are present in the package itself like the

course linear algebra needs to be taken before cal-

culus. There is also some recent work on how se-

quence mining (Parameswaran et al., 2010b) can be

used to form an aggregated recommendation and en-

vironment which we have not looked into.

Figure 7: Precision with Info-gain.

Figure 8: Precision with Gstatistic.

Figure 9: Precision with Marshall’s Correction.

REFERENCES

Adomavicius, G. and Tuzhilin, A. (2005). Toward the next

generation of recommender systems: a survey of the

state-of-the-art and possible extensions. Knowledge

and Data Engineering, 17(6):734 – 749.

Chen, Y. L., Cheng, L.-C., and Chuang, C. (2008). A group

recommendation system with consideration of inter-

GroupRecommenderSystems-SomeExperimentalResults

375

actions among group members. Expert Syst. Appl.,

34(3):2082–2090.

Cohen, W. W. (1995). Fast effective rule induction. In

ICML, pages 115–123.

de Campos, L. M., Luna, J. M., Huete, J. F., and Morales,

M. A. (2007). Group recommending: A methodolog-

ical approach based on bayesian networks. In ICDE

Workshops, pages 835–844.

Hogg, L. and Jennings, N. R. (1999). Variable sociability in

agent-based decision making. In ATAL ’99, Orlando,

USA, July 15-17, 1999, pages 305–318.

Jananch, D., Zanker, M., Felfernig, A., and Friedrich, G.

(2010). Recommender Systems - An Introduction.

Cambridge University Press, Cambridge.

Masthoff, J. (2003). Modeling the multiple people that are

me. In 9th international conference on User modeling,

UM’03, pages 258–262, Heidelberg. Springer.

Masthoff, J. (2004). Group modeling: Selecting a sequence

of television items to suit a group of viewers. User

Model. User-Adapt. Interact., 14(1):37–85.

McCarthy, J. and Anagnost, T. D. (1998). Musicfx: An

arbiter of group preferences for computer aupported

collaborative workouts. In CSCW, pages 363–372.

Mingers, J. (1989). An empirical comparison of selection

measures for decision-tree induction. Mach. Learn.,

3(4):319–342.

O’Connor, M., Cosley, D., Konstan, J. A., and Riedl, J.

(2001). Polylens: A recommender system for groups

of user. In ECSCW, pages 199–218.

Padmanabhan, V., Seemala, S. K., and Bhukya, W. N.

(2011). A rule based approach to group recommender

systems. In MIWAI, pages 26–37.

Parameswaran, A., Venetis, P., and Garcia-Molina, H.

(2011). Recommendation systems with complex con-

straints: A course recommendation perspective. ACM

Trans. Inf. Syst., 29(4):20:1–20:33.

Parameswaran, A. G., Garcia-Molina, H., and Ullman, J. D.

(2010a). Evaluating, combining and generalizing rec-

ommendations with prerequisites. In CIKM ’10, pages

919–928.

Parameswaran, A. G., Koutrika, G., Bercovitz, B., and

Garcia-Molina, H. (2010b). Recsplorer: recommen-

dation algorithms based on precedence mining. In

SIGMOD ’10, pages 87–98, New York, NY, USA.

ACM.

Pazzani, M. J. (1999). A framework for collaborative,

content-based and demographic ﬁltering. Artif. Intell.

Rev., 13(5-6):393–408.

Quinlan, J. R. (1996). Improved use of continuous attributes

in c4.5. Journal of Artiﬁcial Intelligence Research,

4:77–90.

Rivest, R. L. (1987). Learning decision lists. Machine

Learning, 2(3):229–246.

Tubio, R., Sotelo, R., Blanco, Y., Lopez, M., Gil, A., Pazos,

J., and Ramos, M. (2008). A tv-anytime metadata ap-

proach to tv program recommendation for groups. In

Consumer Electronics, ISCE, pages 1–3.

Yu, Z., Zhou, X., Hao, Y., and Gu, J. (2006). Tv program

recommendation for multiple viewers based on user

proﬁle merging. User Model. User-Adapt. Interact.,

16(1):63–82.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

376