New Recommendation Techniques

for Multi-Criteria Rating Systems

Gediminas Adomavicius YoungOk Kwon

gedas@umn.edu ykwon@csom.umn.edu

Department of Information and Decision Sciences

Carlson School of Management
University of Minnesota

Abstract. While traditional single-rating recommender systems have been successful in a

number of personalization applications, the research area of multi-criteria recommender systems

has been largely untouched. In order to take full advantage of multi-criteria ratings in various

applications, new recommendation techniques are required. In this paper we propose two new

approaches the similarity-based approach and the aggregation function-based approach to

incorporating and leveraging multi-criteria rating information in recommender systems. We also

discuss multiple variations of each proposed approach, and perform empirical analysis of these

approaches using a real-world dataset. Our experimental results show that multi-criteria ratings

can be successfully leveraged to improve recommendation accuracy, as compared to traditional

single-rating recommendation techniques.

Keywords: Personalization, recommender systems, collaborative filtering, multi-criteria ratings,

rating estimation.

1
1. Introduction and Motivation

In order to make good decisions in any situation, it is typically necessary to possess a certain

sufficient amount of information. Technologies enable us to easily obtain more information,

especially on the Internet. For instance, if an individual wants to rent a movie online, there are

numerous choices available. However, too much information can make decision-making

inefficient, leading to information overload. Personalization technologies and recommender

systems help to overcome this problem by providing personalized suggestions regarding which

information is most relevant to users. Most of online shopping sites and many other applications

now use recommender systems. The most popular examples include Netflix, which recommends

movies, and Amazon.com, which recommends books, CDs, and various other products. If users

offer their feedback on purchased or consumed items, the task of recommender systems is to

predict user preferences for the yet unseen items based on users prior feedback and activities

and, subsequently, to recommend the item(s) with the highest estimated relevance to the user.

Recommender systems are usually classified into three categories based on their

approach to recommendation: content-based, collaborative, and hybrid approaches (Balabanovic

& Shoham 1997). Content-based recommender systems recommend items similar to the ones the

user preferred in the past. Collaborative (or collaborative filtering) recommender systems

recommend items that users with similar preferences have liked in the past. Finally, hybrid

approaches combine content-based and collaborative methods, which can be done in many

different ways (Adomavicius & Tuzhilin 2005). Furthermore, recommender systems can also be

classified based on the nature of their algorithmic technique into memory-based and model-based

approaches (Breese et al. 1998). In particular, memory-based techniques usually represent

heuristics that calculate recommendations on the fly based directly on the previous user

2
activities. In contrast, model-based techniques use previous user activities to first learn a

predictive model (typically using some statistical or machine-learning methods), which is then

used to make recommendations.

While a comprehensive survey of recommender systems research literature is beyond the

scope of this paper,1 it is important to note that vast majority of current recommender systems

typically use a single criterion (i.e., a single numerical rating) to represent the utility of an item to

a user in the two-dimensional Users Items space. The recommendation process starts with the

specification of the initial set of ratings that is either explicitly provided by the users or is

implicitly inferred by the system. For example, in case of a movie recommender system, user

John Doe may assign a rating of 11 (out of 13) for movie Vertigo, i.e., set R(John_Doe,

Vertigo) = 11. Once these initial ratings are specified, a recommender system tries to estimate

the rating function R

R: Users Items R0 (1)

for the (user, item) pairs that have not been rated yet. R0 is usually represented by a totally

ordered set (e.g., integers or real numbers within a certain range). Once function R is estimated, a

recommender system can recommend the highest-rated item (or a set of N highest-rated items)

for each user. In summary, one goal of a typical recommender system is to correctly estimate the

ratings of unrated items based on the given ratings; another goal is to find items that maximize

the users utility.

While single-rating recommender systems have been successful in a number of

applications, multi-criteria rating systems are being more and more commonly employed in many

industries. Restaurant guides, such as Zagats Guide, provide three criteria for restaurant ratings

1
A recent survey of recommender systems research literature can be found in (Adomavicius & Tuzhilin 2005).

3
(e.g., food, dcor, and service). Online shopping malls, such as Circuitcity.com and Buy.com,

use multi-criteria ratings for consumer electronics (e.g., display, performance, battery life, and

cost). Note that the aforementioned multi-criteria rating systems are not used in the context of

personalization, i.e., the rating on each criterion is the same for all users (for example, the food

rating for a specific restaurant published by Zagats Guide) and not personalized to each

individual consumer. In order to take full advantage of existing multi-criteria ratings in

personalization applications, new recommendation techniques are required. For example,

recently Yahoo! Movies launched a movie recommendation service that uses multi-criteria

ratings for each movie, which indicates that multi-criteria data provides value both to online

content providers and consumers and may become an important component in different

personalization applications. Therefore, in this paper we propose several new approaches on how

to extend recommendation technologies in order to incorporate and leverage multi-criteria rating

information.

The rest of this paper is organized as follows. In Section 2, we briefly discuss some of

the research related to multi-criteria ratings, including from the recommender systems literature.

In Section 3, we provide some background on a traditional single-criterion collaborative filtering

algorithm, which is used as an example throughout the paper. We then propose new

recommendation techniques for multi-criteria ratings in Section 4. In Section 5, we describe our

empirical analysis based on a real-world dataset. Finally, we conclude our paper in Section 6.

2. Related Work

Multi-criteria problems have been studied extensively in operation research and decision science

fields. The majority of engineering problems are essentially multi-criteria optimization problems

(Statnikov & Matusov 1995). For example, when an airplane is being designed, its reliability,

4
longevity, efficiency, cost, and the combination of other utilization factors need to be considered.

Typical methods to solve the multi-criteria optimization problems include: finding Pareto

optimal solutions; optimizing the most important criterion and converting other criteria to

constraints; consecutively optimizing one criterion at a time, converting an optimal solution to

constraints, and repeating the process for other criteria.

The decision science field treats organizational decision making as a multi-criteria

problem, i.e., it considers various points of view, such as financial, human resources-related, and

environmental aspects in making a decision (Figueria et al. 2005). The objective of multi-criteria

decision analysis is to assist a decision maker in choosing the best alternative when multiple

criteria conflict and compete with each other. Most commonly used decision aiding methods,

such as outranking methods and the analytical hierarchy process, are based on multi-criteria

aggregation procedures. Outranking methods determine which alternatives are preferred to

others by systematically comparing possible alternatives on each criterion. The analytical

hierarchy process structures multi-criteria into a hierarchy and calculates the score of each

criterion as a weighted sum of its sub-criteria.

Similarly, in marketing research literature, buying a product also can be regarded as a

multi-criteria decision problem. For example, when we purchase a car, we consider its multiple

attributes, such as price, brand, and color. The conjoint model is most commonly used technique

for solving multi-criteria problems in this field (Green et al. 2001). This model determines the

importance weights of product attributes and the values of the attributes. The customers

preference for the product then can be calculated as a linear combination of weights and values.

Multi-criteria information is also used in certain electronic market mechanisms, such as

multi-attribute auctions (Bichler 2000). Multi-attribute auctions are typically used in

5
procurement settings and enable auction participants negotiate not only on price, but also on

other attributes of a deal (e.g., quality level, style, delivery date). It has been demonstrated that

multi-attribute auctions have several advantages over their single-attribute (i.e., price-only)

counterparts, including the improvements in the overall utility and suitability for various

application domains (Bichler 2000).

However, the multi-criteria problems addressed in above-mentioned fields typically are

not intended for personalization and recommendation settings. These problems find the

solutions or items that are optimal in general (i.e., optimal with respect to all users), and

differences in individual user preferences are not explicitly considered. Recently, multi-criteria

rating problems have started receiving attention in recommender systems research and are

regarded as one of the important issues for the next generation of recommender systems

(Adomavicius & Tuzhilin 2005). In recommender systems literature, the roots of multi-criteria

ratings could be traced to the approaches that started incorporating content-based features into

collaborative recommendation techniques. This allowed the recommender systems to identify

favorite content attributes (e.g., comedy movies) based on the content analysis of the

previously rated items, and then also to recommend items to a user based not only on the ratings

of similar users, but also based on these favorite content attributes (Balabanovic & Shoham

1997). However, the users were able to submit just a single rating for each item, and could not

specify their individualized feedback about a specific movie component/aspect (such as movies

visual effects).

In addition, Ricci et al. (2002) developed a recommender system for personalizing travel

using case-based reasoning techniques. The recommendations are performed by ranking and

aggregating elementary items (locations, activities, services) based on the users preferences and

6
a repository of past travels. While these techniques do not use multi-criteria ratings per se, the

recommendation process does take into account multiple criteria, and the optimization is

performed over a multidimensional solution space.

Furthermore, there has been some research on providing recommendation filtering

capabilities based on item content information. For example, Schafer (2005) implements a meta-

recommendation system that allows users to indicate the preference for each content attribute

(e.g., movie genre, MPAA rating, or film length) and rate the importance of these attributes. For

example, users can indicate that they want only comedy movies, and that it is the most

important condition for recommendations the users requirements will filter the potential

recommendations towards what the users really want. Note, however, that this does not

represent a multi-criteria rating environment, since the users are specifying general filtering

requirements for all movies (such as specifying the preferred value and weight for movie genre

attribute). Similarly, Lee et al. (2002) also obtain the importance weights of content attributes

directly from the user. They use each attributes rank to compare the items, but the value or rank

of each attribute is assumed to be the same for all users. In contrast to Schafer (2005) and Lee et

al. (2002), in multi-criteria rating environment users would be able to specify subjective ratings

for various components of individual items (e.g., to rate visual effects component for the Star

Wars movie), which could then be leveraged for prediction and personalization purposes.

In summary, while there have been several different approaches discussed in

personalization literature that are somewhat related to the issue of incorporating and leveraging

multi-criteria ratings in recommender systems, it would be fair to say that this issue is largely

unexplored. For this reason, in this paper we focus on new recommendation techniques for

multi-criteria rating systems.

7
3. Background: Traditional Single-Rating Similarity-Based Collaborative

Filtering Approach

Before proceeding with the discussion on new recommendation techniques for multi-criteria

rating settings, we briefly describe one of the traditional and commonly used single-rating

collaborative recommendation techniques, which we will use as an example throughout the

paper.

Specifically, based on the recommender systems classification schemes mentioned in

Section 1, lets consider the memory-based collaborative filtering technique that estimates R(u, i)

the rating that user u would give to item i by computing the weighted average of all known

ratings R(u', i), where user u' is similar to u. Two popular ways to compute this weighted

average are (Breese et al. 1998):

Weighted sum approach, i.e.,

R (u, i ) = z
u N ( u )
sim(u , u ) R(u, i ) ; (2)

Adjusted weighted sum approach, i.e.,

R (u, i ) = R(u ) + z
u N ( u )
( )
sim(u, u) R(u, i ) R(u) . (3)

Here the value of rating R(u', i) is weighted by the similarity of user u' to user u the more

similar the two users are, the more weight R(u', i) will have in the computation of rating R(u, i).

Furthermore, multiplier z serves as a normalizing factor and is usually set to

z =1 u N ( u )
sim(u , u ) , R(u ) represents the average rating of user u, and N(u) represents the

set of users that are similar to user u. The size of set N(u) can range anywhere from 1 to all users

in the dataset. Limiting the neighborhood size to some specific number (e.g., 3) will determine

how many similar users will be used in the computation of rating R(u, i).

8
 Furthermore, there are several ways to compute similarity sim(u, u') between two users,

including cosine-based and correlation-based computations (Breese et al. 1998). We will use

the cosine-based similarity in this paper, since it is arguably the most commonly used technique

for determining how similar two users are in memory-based collaborative filtering algorithms.

Assuming I(u, u') represents the set of all items rated by both users u and u', the cosine-based

similarity can be calculated as follows:

sim(u , u ) = R(u, i ) R(u, i ) R(u, i ) 2 R(u , i ) 2 (4)

iI ( u ,u) iI ( u ,u ) iI ( u ,u )

In addition, because of the inherent symmetry between users and items in the traditional

memory-based collaborative filtering setting, this approach can be either user-based or item-

based, depending on whether we want to calculate the similarity between users or items.

Equations (2) and (3) represent the user-based approach, but they can be straightforwardly

rewritten for the item-based approach. For example, the item-based adjusted weighted sum can

be calculated as follows (Sarwar et al. 2001):

R (u , i ) = R(i ) + z
i N ( i )
(
sim(i, i) R(u, i) R(i) ) (5)

and z, R (i ) , sim(i, i'), and N(i) are analogous to their user-based counterparts.

In the rest of the paper, unless explicitly stated otherwise, by traditional collaborative

filtering approach we will refer to the user-based adjusted weighted sum approach (3) that uses

the cosine-based similarity function (4).

Finally, recommender systems typically recommend the items with the highest predicted

rating to the user. In other words, recommenders often are not concerned about predicting the

ratings of all items as accurately as possible, but rather about accurately predicting the highest-

rated items, since users in real-world personalization applications are usually interested in

9
looking only at few highest-ranked item recommendations. Therefore, it is useful to evaluate the

recommender system performance based on items that get the top N highest scores for each user,

assuming, of course, that the values of the top N ratings are high enough to merit an actual

recommendation. This is the evaluation approach that we adopt in this paper, as will be

discussed in Section 5.

4. Extending Recommender Systems to Incorporate Multi-Criteria Ratings

In addition to the overall rating, multi-criteria ratings provide additional information about user

preferences regarding several important aspects/components of an item. Leveraging this

additional information in recommender systems should be beneficial, since it can potentially

increase the accuracy of the recommendations. Therefore, new techniques are needed in order to

effectively incorporate the multi-criteria rating information into the recommendation process.

The goal of multi-criteria recommender systems is to find items that maximize each

users utility, just as in the single-rating recommender systems. Therefore, one of the important

goals of recommendation systems is to be able to predict the overall rating of each item for each

user, because the system ultimately needs to compare the items based on their overall ratings and

recommend the best items to the users. The difference between single-rating and multi-criteria

rating systems is that the latter have more information about the users and items, which can be

effectively used in the recommendation process. More formally, the general form of a rating

function in a multi-criteria recommender system is:

R: Users Items R0R1Rk (6)

where R0 is the set of possible overall rating values, and Ri represents the possible rating values

for each individual criterion i (i = 1, , k), typically on some numeric scale (e.g., from 1 to 13).

10
 In the remainder of this section we propose two new recommendation approaches and

present several different variations of each. The first approach is designed to extend the

traditional single-criteria memory-based collaborative filtering algorithm, while the second

approach is not restricted to any specific algorithm. In other words, any existing single-criteria

recommendation algorithm (i.e., content-based, collaborative, or hybrid) can be used in

conjunction with this approach. And, as mentioned earlier, throughout the paper we will use one

of the common user-based collaborative filtering techniques for illustration purposes.

4.1 Similarity-Based Approach to Extending Standard Collaborative Filtering Techniques

Consider a movie recommendation application, where users provide the recommender system

with a single rating (between 1 and 13) for each movie they have seen. Moreover, suppose that

this recommender system is using a traditional user-based collaborative filtering approach for

rating prediction, as described in Section 3. In this case, according to Equation (3), any rating

that user u would give to yet unseen movie i would be estimated based on how users u' that are

similar to target user u rated movie i, i.e., unknown rating R(u, i) is calculated based on ratings

R(u', i). Therefore, the more accurately the system determines who the true peers (or nearest

neighbors) of u are, the more accurate the rating prediction should be. The traditional (two-

dimensional) collaborative filtering calculates the similarity between users u and u' based on how

similar their ratings are for the movies they both have seen.

Figure 1 illustrates this estimation process with a simple example. Assume, that we have

five users u1, , u5 and five movies i1, , i5. Furthermore, lets suppose that the recommender

system needs to estimate how much the target user u1 would like movie i5 and, as indicated in

Figure 1, that all other ratings of different users to different movies are known. Then, the

traditional collaborative filtering approach finds the users that are closest to u1 and that have seen

11
movie i5. In this case, u2 and u3 seem to be perfect matches for user u1, since all of them rated

the common movies exactly the same (see Figure 1). Since both u2 and u3 rate movie i5 as 9, the

value of target rating R(u1, i5) will be predicted as 9.

Item Item Item Item Item

i1 i2 i3 i4 i5
Target user Rating to be
User predicted
5 7 5 7 ?
u1
User
5 7 5 7 9
Users most u2
similar to the User Ratings to be used
5 7 5 7 9
target user u3 in prediction
User
6 6 6 6 5
u4
User
6 6 6 6 5
u5

Figure 1. Collaborative filtering in a single-criteria setting.

Now lets consider the same scenario as above, but in a multi-criteria setting.

Specifically, lets assume that we have the same five users u1, , u5 and five movies i1, , i5.

Also, rating R(u1, i5) is unknown and needs to be predicted, and, as indicated in Figure 2, all

other overall ratings of different users to different movies are known and are exactly the same as

before (in Figure 1). In addition, lets assume that each user is also asked to provide the

feedback about the movie on four specific criteria: story, acting, direction, and visuals2, and that

the overall rating in this case is a simple average of the four individual criteria ratings.

Following the idea behind the standard collaborative filtering approach, in order to

predict R(u1, i5) the recommender system should find the users that are closest to u1 and that have

seen movie i5. However, because of all the additional information that is available in the form of

multi-criteria ratings, one can clearly see that users u2 and u3 are quite different in their tastes and

preferences from user u1, even though their overall ratings for each movie match perfectly. In

2
As is done on some movie review websites, such as Yahoo! Movies (http://movies.yahoo.com).

12
particular, the movie aspects that u1 hated (story and acting) were really liked by u2 and u3 and

vice versa. However, in recommender systems that are based on single-criteria ratings, this

information would be hidden within the aggregate overall rating, which may lead to inaccurate

insights about the true similarity between user preferences (as in this example). Users u4 and u5

seem to be much better matches for user u1 in this example, since not only their overall ratings

are similar, but their preferences for different movie aspects were very similar as well (see Figure

2). Since both u4 and u5 rate movie i5 as 5, the value of target rating R(u1, i5) would be predicted

as 5, which is a very different outcome from the one obtained in a single-criteria rating scenario.

Item Item Item Item Item

i1 i2 i3 i4 i5
Target user Rating to be
User predicted
52,2,8,8 75,5,9,9 52,2,8,8 75,5,9,9 ?
u1
User
58,8,2,2 79,9,5,5 58,8,2,2 79,9,5,5 9
Users most u2
similar to the User Ratings to be used
58,8,2,2 79,9,5,5 58,8,2,2 79,9,5,5 9
target user u3 in prediction
User
63,3,9,9 64,4,8,8 63,3,9,9 64,4,8,8 5
u4
User
63,3,9,9 64,4,8,8 63,3,9,9 64,4,8,8 5
u5

Figure 2. Collaborative filtering in a multi-criteria setting.

In summary, while the overall rating that a user gives to an item provides the information

regarding how much the user liked the item, multi-criteria ratings provide some insights

regarding why the user liked the item as much as she did. Therefore, having multi-criteria ratings

provides the possibility to estimate the similarity between two users more accurately.

Based on this idea, we propose to extend the standard collaborative filtering algorithm to

include multi-criteria ratings. Specifically, we propose several different ways to include multi-

criteria rating information in the calculation of the similarity between two different users sim(u,

u') or two different items sim(i, i'). Then, given the newly calculated similarity, the rating

13
prediction can be done using the weighted sum or adjusted weighted sum in the same way as

with a standard collaborative filtering algorithm, i.e., using Equations (2), (3), or (5). Below we

describe two different approaches to leverage multi-criteria ratings in the similarity computation.

Aggregating traditional similarities that are based on each individual rating

This approach can use any standard similarity metric, such as cosine-based (4), and calculates the

similarity between users (or items) based on each individual criteria. Lets assume that each

rating that user u gives to item i consists of an overall rating r0, and k multi-criteria ratings r1,

, rk, i.e.,

R(u, i) = (r0, r1, , rk). (7)

Then, k+1 different similarity estimations can be obtained by using some standard metric to

measure similarity between users u and u': sim0(u, u') represents the similarity between u and u'

based on the overall rating; sim1(u, u') similarity based on the first criteria rating; sim2(u, u')

similarity based on the second criteria rating; and so on. The overall similarity then can be

computed by aggregating the individual similarities in several ways:

[Average similarity] By averaging all individual similarities, i.e.,

1
simavg (u, u) = simi (u, u) ,
k
(8)
k +1 i =0

[Worst-case similarity] By using the smallest of similarities, i.e.,

simmin (u , u ) = min simi (u , u ) . (9)

i = 0,..., k

Calculating similarity using multidimensional distance metrics

In multi-criteria rating scenario, each rating R(u, i) = (r0, r1, , rk) represents a point in the k+1-

dimensional space. Therefore, one natural approach to compute similarity between different

users is to use multidimensional distance metrics. Such metrics are easy to understand and

14
straightforward to implement. Note that the metrics of distance and similarity are inversely

related: the smaller the distance between two users, the higher the similarity. We calculate the

similarity between two users in three steps.

First, we have to be able to calculate the distance between two users ratings for the same

item, i.e., d rating ( R(u , i ), R(u , i ) ) , where R (u , i ) = (r0 , r1 , , rk ) and R(u , i ) = (r0, r1, , rk) . For

this purpose, any of the standard multidimensional distance metrics can be used:

ri ri ;
k
Manhattan distance: i =0
(10)

2
ri ri ;
k
Euclidean distance: i =0
(11)

Chebyshev (or maximal value) distance: max ri ri . (12)

i = 0,, k

Second, the overall distance between two users u and u' is simply:

1
d user (u , u ) =
I (u , u )

iI ( u ,u )
d rating ( R(u , i ), R(u , i ) ) (13)

where I(u, u') denotes the set of items that both u and u' have rated. In other words, the overall

distance between two users u and u' is the average distance between their ratings for all their

common items.

Finally, because the collaborative filtering techniques operate with the metric of user

similarity (and not user distance), and the distance and similarity are inversely related, we use the

simple transformation between the two metrics:

1
sim(u, u ') = . (14)
1 + d user (u, u ')

Note that this definition of similarity has desired range properties, i.e., the similarity will

approach 0 as the distance between two users becomes larger, and it will be 1 if the distance is

zero (users are identical).

15
 In summary, both of the approaches presented in this section change only the similarity

function in the traditional collaborative filtering technique in order to reflect multi-criteria rating

information, which should result in a more accurate identification of similar users and,

consequently, in better recommendation quality. We provide some empirical results in Section 5.

4.2 Aggregation Function Based Approach

Approaches to integrate multi-criteria rating information into recommender systems discussed in

the previous section apply primarily to the similarity-based recommenders, such as traditional

collaborative filtering techniques. In contrast, in this section we present a different approach that

is not limited to any specific recommendation algorithm. The intuition behind this approach

comes from the assumption that multi-criteria ratings represent users preferences for the

different important components of an item (e.g., story, acting, direction, and visuals aspects in the

case of movie recommender systems). Thus, the overall rating of an item is not just another

rating that is independent of others, but rather serves as some aggregation function f of the

multi-criteria ratings of this item, i.e.,

r0 = f (r1 , , rk ) . (15)

In other words, this approach assumes that the overall rating has a certain relationship with multi-

criteria ratings. For instance, in a movie recommendation application, the story criteria rating

may have a very high priority, i.e., the movies with high story ratings are well liked overall by

some users, regardless of other criteria ratings. Therefore, if the story rating of the movie is

predicted to be high, the overall rating of the movie must also be predicted as high in order to be

accurate.

The proposed approach to rating estimation consists of the following three steps, as

illustrated in Figure 3. First, we decompose k-dimensional multi-criteria rating space into k

16
single-rating recommendation problems and use any traditional single-criteria recommendation

technique to estimate ratings for each individual criterion. Second, we use statistical or machine

learning techniques to estimate aggregation function f based on the known ratings. And third,

using the multi-criteria ratings estimated in step 1 and function f estimated in step 2, we directly

calculate the predicted overall rating. Below we discuss each of these steps in more detail.

Known ratings
R (u , i ) = ( r0 , r1 , , rk )

(1) Predict k multi-criteria ratings using (2) Learn aggregation function f using
any traditional recommendation technique statistical or machine learning techniques
Given: ri (for each i = 1, , k) Given: ( r0 , r1 , , rk )
Compute: ri Estimate: f such that r0 = f ( r1 , , rk )

(3) Predict an overall rating

Given: ( r1, , rk) , f
Compute: r0 based on r0 = f ( r1, , rk)

Figure 3. Overview of the aggregation-function-based approach.

Step 1: Predicting multi-criteria ratings

As mentioned earlier, we decompose k-dimensional multi-criteria rating space into a set of k

single-rating recommendation problems, where each problem can be represented with a

traditional UsersItems matrix (like the one in Figure 1) and addresses the rating prediction for

one of the individual criteria. In other words, instead of the multi-criteria recommendation

problem R:UsersItems R0R1Rk we are dealing with k single-rating recommendation

problems R:UsersItems Ri (where i = 1, , k). This approach provides a lot of flexibility,

since (unlike with similarity-based approaches mentioned in previous section) any existing

17
single-criteria recommendation technique (e.g., collaborative, content-based, or hybrid) can be

used to estimate unknown ratings for individual criteria.

Step 2: Learning the aggregation function

The goal of this step is to estimate relationship f between the overall rating and the underlying

multi-criteria ratings of items, such that r0 = f (r1 , , rk ) . We are already able to predict the

individual multi-criteria ratings (see Step 1 above), but one of the important goals of

recommendation systems is to be able predict the overall rating of each item for each user, which

can be helpful in different situations. For example, having the overall rating for each item

enables the recommender system to rank all items for each user in terms of their predicted utility

(i.e., overall rating) and recommend only the most relevant items. In contrast, to determine the

most relevant items without the presence of the overall rating, the recommender system would

have to deal with a much more complex multi-criteria optimization problem (Statnikov &

Matusov 1995). Thus, finding the aggregation function is crucial for recommender systems, and

there are several ways in which this function could be obtained:

Domain expertise. Based on her prior experience and knowledge of the domain, the

domain expert may suggest the appropriate aggregation function. For example, it may

be the case that the overall rating is a simple average of the underlying multi-criteria

ratings for each item, i.e., r0 = (r1 + + rk ) / k .

Statistical techniques, including various linear and non-linear regression analysis

techniques. For example, in the case of linear regression, the aggregation function for

the overall rating would be a linear combination of the multi-criteria ratings, i.e.,

r0 = w1r1 + + wk rk + c , where weight wi associated with criterion i can be interpreted as

18
 the importance of this criterion in determining the overall rating. The weights wi (i = 1,

, k) and constant c are estimated based on the set of known ratings.

Machine learning techniques. Various sophisticated computational learning techniques

can also be used for this purpose, e.g., artificial neural networks (Mitchell 1997).

Besides the ability to use different learning techniques, the aggregation function can also be of

three different scopes: total, user-based, or item-based. In particular, f is a total aggregation

function if it is used to predict all unknown ratings, e.g., if the criteria weights wi in a regression-

based function mentioned above are consistent for all users and items. However, depending on

the domain specifics, it may be useful to consider user-based or item-based aggregation

functions in some applications. For example, in a movie recommender system, user u may have

a much larger weight on the story component that is consistent for all movies, whereas user u'

may have a significant weight on the visuals component. In this case, it would be

advantageous for user u to have her own user-based aggregation function fu, which would be

learned exclusively from the known ratings of user u (as opposed to all known ratings) using the

aforementioned techniques. Similarly, with the item-based aggregation function fi we would

assume that each item i will have its own aggregation function that is consistent for all the

ratings involving this item.

Finally, note that a variety of different techniques are available for testing the fitness or

accuracy of the predicted aggregation function(s). For example, in the case of linear regression,

one can estimate the predictive power using its R2 value. Or, more generally, one could use

standard n-fold cross validation techniques to estimate the predictive accuracy of the aggregation

function (Mitchell 1997). Therefore, we have the ability to restrict the use of user-based (or

item-based) aggregation functions only to the ones that exhibit sufficient predictive performance,

19
e.g., whose accuracy is greater than some pre-specified threshold. The remaining users (or

items) could use other techniques, e.g., the total aggregation function. As with every data-driven

computational learning technique, there will be application domains where this approach will

work well (i.e., domains where users/items exhibit consistent preferences on each criterion) and

domains where other techniques will be more advantageous.

Step 3: Predicting overall ratings

Finally, as mentioned earlier, we compute each unknown overall rating r0 directly by using the

multi-criteria ratings estimated in step 1 and function f estimated in step 2: r0 = f (r1, , rk) .

4.3 Other Benefits of Multi-Criteria Ratings in Recommender Systems

Up to this point, we have focused on how new techniques can potentially improve the estimation

of overall ratings by leveraging multi-criteria rating information. In addition to this

enhancement, the usage of multi-criteria ratings in recommender systems can provide other

benefits to their users.

In particular, most recommender systems are inflexible in customizing recommendations

according to user-specific requests. In other words, recommendations typically are fixed for all

users (e.g., provide 5 most relevant items to each user), and cannot be adjusted by the users on

the fly. There have been interesting attempts to provide recommendation filtering capabilities

based on some item content information (see, for example, Schafer 2005); however, while

undoubtedly useful, this filtering is typically done on the user-specified information that is fixed

to an item and, therefore, same to all the users. For example, in a movie recommender system

the users may be able to narrow their movie recommendations based on the movie genre, MPAA

rating, film length, etc. (Schafer 2005). However, in a multi-criteria recommender system

(similar to the one shown in Figure 2), a certain user may want to request only exceptionally

20
good story movies, where the story component of a movie is completely subjective to each

user and, as mentioned earlier, is estimated individually for each user. Multi-criteria rating

information would allow the recommender systems to respond to users individual dynamic

needs (e.g., expressed as filtering thresholds on individual criteria) in a more personalized

manner and adjust the recommendations accordingly.

5. Experimental Results

To evaluate the proposed approaches, we have collected a set of user-submitted movie ratings

from Yahoo! Movies website (movies.yahoo.com) for several hundred randomly chosen movies

from the last decade. When a user submits movie ratings to Yahoo! Movies, in addition to the

overall rating, she is asked to provide four criteria information for each movie: story, acting,

direction, and visuals. All ratings have 13 possible values and are based on a standard grading

scale from A+ to F; for the analysis purposes we changed them to numerical values from 13 to 1.

In the data preprocessing stage, we invoked two constraints on the dataset in order to ensure that

the dataset is not extremely sparse and has enough data for rating prediction: (a) there should be

at least 10 movie ratings per user and (b) at least 10 user ratings per movie.

As a result, we ended up with a dataset that includes 155 users, 50 movies, and has 2,216

known ratings in total (28.6% of ratings are known). Each user has rated 14.3 movies on

average, and the average number of common movies between two users is 5.2. Each movie has

been rated on average by 44.3 users, and the average number of common users between two

movies is 13.6. The average rating on each criterion is approximately 9 (or B).

Furthermore, in order to obtain reliable results with a relatively small amount of data, we

use a standard 10-fold cross validation technique (Mitchell 1997), where we randomly divide the

dataset into 10 disjoint subsets. We use nine-tenths of the data for training, and the remaining

21
one-tenth for testing rating prediction, and then repeat this process 10 times (each time with a

different test dataset) and perform the evaluation on all predicted ratings.

Numerous metrics for evaluating the performance of recommender systems have been

proposed and used in the research literature (Herlocker et al. 2004), including the statistical

accuracy metrics (e.g., mean absolute error and root mean squared error) as well as decision-

support measures that determine how well the recommendation algorithm can predict high-

relevance items (i.e., items that would be rated highly by the user). Examples of decision-

support metrics include precision (the percentage of truly high ratings among those that were

predicted to be high by the recommender system), recall (the percentage of correctly predicted

high ratings among all the ratings known to be high), and F-measure, which is a harmonic

mean of precision and recall (Herlocker et al. 2004).

In this paper, we have focused on the popular variation of the above-mentioned precision

metric, i.e., precision-in-top-N, which represents the percentage of truly high overall ratings

among those that were predicted to be N most relevant items for each user. This metric was

chosen because of its practicality, since many users in real-life personalization and

recommendation applications are typically interested in looking only at few highest-ranked item

recommendations.

Because precision-related metrics measure the frequency with which a recommender

system makes correct decisions about whether an item that is predicted as highly-ranked is

truly highly-ranked, we needed to define what highly-ranked means in our application. In

other words, every rating had to be defined on a binary scale, i.e., as highly-ranked or non-

highly-ranked. Since Yahoo! Movies rating scale (from A+ to F) was not binary, we translated

the overall movie ratings into a binary scale by treating the ratings greater than 10.5 (A+, A, A-)

22
as highly-ranked and ratings less than 10.5 as non-highly-ranked. The threshold of 10.5 was

chosen with the assumption that the users would really want to focus on the recommendations

about movies that are most relevant to them (i.e., movies they would rate as A+, A, A-), and

therefore the correctness of recommendations for such movies is most desirable.

Also note that, in our dataset, the percentage of the highly-ranked ratings (i.e., overall

ratings above 10.5) was 35.6%, which means that it would be possible to obtain the precision of

35.6% simply by recommending items at random. Any recommender system that does not

achieve 35.6% precision would be worse than a random guess and, therefore, essentially useless.

In order to illustrate the performance of the proposed multi-criteria recommendation

techniques on real-life data, we performed the empirical analysis of the following five approaches

using the above-mentioned movie data (as summarized in Table 1):

standard CF a traditional single-rating user-based CF approach, which uses adjusted

weighted sum and the cosine similarity metric, as described in (3) and (4). This approach

is used as a baseline to illustrate the performance of multi-criteria recommendation

approaches, as compared to a single-rating recommender system.

Two similarity-based techniques (as described in Section 4.1) implemented with the

traditional user-based CF approach:

o cos-min an example of a technique that aggregates traditional cosine-based

similarities for each individual rating.

o Chebyshev an example of a technique that uses Chebyshev multidimensional

distance metric.

Two aggregation-function-based techniques (as described in Section 4.2), where

individual multi-criteria ratings are estimated using the traditional user-based CF

23
 approach:

o total-reg an example of total aggregation function that is based on linear

regression.

o movie-reg95 an example of item-based aggregation function that is generated

separately for each movie and restricted only for the movies that have the best

regression fit (i.e., R2 95%).

Note that we use the standard user-based collaborative filtering approach as an integral

part of every technique in order to minimize the non-essential differences between the techniques

as much as possible and, thus, to maximize the possibility that any differences in performance

between the standard CF and multi-criteria recommender systems are due to the newly

introduced multi-criteria rating information.

Precision in Precision in Precision in

Recommendation Approach: user-based CF
top 3 (%) top 5 (%) top 7 (%)
standard CF 70.7 68.7 69.0
cos-min 70.7 68.8 69.1
Neighborhood size: ALL users Chebyshev 74.5 70.3 70.5
total-reg 71.5 70.9 70.4
movie-reg95 71.8 74.0*** 75.3***
standard CF 64.9 64.9 66.3
cos-min 67.1 67.1 67.8
Neighborhood size: 3 users Chebyshev 66.2 65.5 64.6
total-reg 65.2 66.6 66.5
movie-reg95 69.0*** 70.7*** 72.2***

Table 1. Main experimental results of several recommendation approaches.

Furthermore, for the sake of completeness, we provide results for different CF

neighborhood sizes (the neighborhood of all users vs. the neighborhood of the 3 most similar

users) and for different precision-in-top-N levels (N = 3, 5, and 7). The results are summarized in

Table 1. The shaded cells represent the performance of the baseline CF approach. Note that

nearly every multi-criteria technique performed either better or at least as well as the baseline

24
technique. The precision figures in regular font represent 0%1% improvement over the baseline

approach. The boldface precision figures represent 1%4% improvement over the baseline

approach, and the boldface precision figures marked with *** represent >4% improvement over

the baseline approach. For further comparison, we have also calculated the precision-in-top-N

metric for a simple popularity-based recommendation approach, where each user is

recommended N movies (N = 3, 5, and 7) that are most liked by all other users, based on the

average rating for each movie. The results show that the precision-in-top-N for this simple

approach is: 61.3% (top 3), 53.3% (top 5), and 46.4% (top 7), which performs better than a

random guess approach mentioned earlier but not as well as collaborative filtering techniques.

Among other notable results:

For user-based CF, precision-in-top-1 measures (as opposed to top-3, top-5, and top-7

measures in Table 1) for various neighborhood sizes were dominated by similarity-based

techniques (such as Chebyshev and cos-min), which typically outperformed both the

baseline approach and the aggregation function-based approaches by 2%-6%.

We also tried movie-based CF (as opposed to user-based CF, as in Table 1), for which

total-reg performed the best of all the techniques for various neighborhood sizes and

typically outperformed the baseline approach by 1%-5%.

Combining similarity-based and aggregation-function-based multi-criteria

recommendation techniques can sometimes improve the predictive performance, which is

generally consistent with similar findings in recommender systems literature about the

advantages of combining different types of recommender systems (e.g., it has been

widely reported that combining content-based and collaborative systems may improve the

recommendation accuracy).

25
As with most recommender systems and, more generally, computational learning techniques, the

performance of a specific technique is highly domain-dependent. In other words, its performance

depends significantly on the characteristics of the underlying data. Thus, while we expect the

proposed techniques to do well in a variety of different application domains, multi-criteria

recommendation techniques cannot be expected to have an advantage over traditional single-

criterion techniques in all domains where multi-criteria information exists, especially in the ones

where multi-criteria ratings do not carry meaningful information or where is no inherent

relationship between the overall rating and multi-criteria ratings for the users or items.

6. Conclusions

While single-rating recommender systems have been successful in a number of personalization

applications, multi-criteria rating systems are getting to be commonly deployed in many

industries. However, in order to take full advantage of existing multi-criteria ratings in

personalization applications, new recommendation techniques are required. In this paper, we

propose two new recommendation approaches the similarity-based approach and the

aggregation-function-based approach to incorporating and leveraging multi-criteria rating

information. Our experimental results on a real-world dataset confirm that, when available,

multi-criteria ratings can be successfully leveraged to improve recommendation accuracy. We

expect that the proposed approaches will be useful in other application domains as well, where

they will be able to predict overall ratings more accurately by utilizing the available multi-criteria

rating information.

The area of recommender systems has made significant progress over the last few years;

many techniques have been proposed and many systems have been developed. However,

modern recommender systems still require further significant improvements in order to provide

26
better recommendations and be viable in more complex personalization applications; the ability

to leverage multi-criteria rating information constitutes one such improvement. We believe that

this paper is just the first step in studying multi-criteria recommender systems and that

significant additional work is needed to further explore this issue.

Acknowledgments

The research reported in this paper was supported in part by the National Science Foundation

grant IIS-0546443.

References

G. Adomavicius and A. Tuzhilin, Toward the Next Generation of Recommender Systems: A

Survey of the State-of-the-Art and Possible Extensions, IEEE Transactions on

Knowledge and Data Engineering, vol. 17, no. 6, pp. 734-749, Jun. 2005.

M. Balabanovic and Y. Shoham, Fab: Content-Based, Collaborative Recommendation,

Communications of the ACM, vol. 40, no. 3, pp. 66-72, 1997.

M. Bichler, An experimental analysis of multi-attribute auctions, Decision Support Systems,

vol. 29, no. 10, pp. 249-268, 2000.

J. S. Breese, D. Heckerman, and C. Kadie, Empirical Analysis of Predictive Algorithms for

Collaborative Filtering, Proc. 14th Conf. on Uncertainty in Artificial Intelligence, July

1998.

J. Figueria, S. Greco, and M.Ehrgott, Multiple Criteria Decision Analysis: State of the Art

Surveys, Springer, 2005.

P. E. Green, A. M. Krieger, and Y.Wind, Thirty years of conjoint analysis: Reflections and

Prospects, Interfaces, vol. 31, no. 3, pp. 56-73, 2001.

27
J. L. Herlocker, J. A. Konstan, L. G. Terveen, and J. T. Riedl, Evaluating Collaborative

Filtering Recommender Systems, ACM Transactions on Information Systems, vol. 22,

no. 1, pp. 5-53, 2004

W. Lee, C. Liu, and C. Lu, Intelligent agent-based systems for personalized recommendations

in Internet commerce, Expert Systems with Applications, vol. 22, no. 4, pp. 275-184,

May 2002.

T. Mitchell, Machine Learning, McGraw-Hill, 1997.

F. Ricci and H. Werthner, Case-Based Querying for Travel Planning Recommendation,

Information Technology and Tourism, vol. 4, nos. 34, pp. 215226, 2002.

B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, Item-Based Collaborative Filtering

Recommendation Algorithms, Proc. 10th Intl WWW Conf., 2001.

J. B. Schafer, DynamicLens: A Dynamic User-Interface for a Meta-Recommendation systems,

Beyond personalization 2005: A workshop on the next stage of recommender systems

research at the ACM Intelligent User Interfaces Conf., Jan. 2005.

R. B. Statnikov and J.B. Matusov, Multicriteria Optimization and Engineering, Chapman &

Hall, 1995.

28