This paper proposes an algorithm for making recommendation so that the neutrality toward the viewpoint speci ed by a user is enhanced. This algorithm is useful for avoiding to make decisions based on biased information. Such a problem is pointed out as the lter bubble, which is the in uence in social decisions biased by a personalization technology. To provide such a recommendation, we assume that a user speci es a viewpoint toward which the user want to enforce the neutrality, because recommendation that is neutral from any information is no longer recommendation. Given such a target viewpoint, we implemented information neutral recommendation algorithm by introducing a penalty term to enforce the statistical independence between the target viewpoint and a preference score. We empirically show that our algorithm enhances the independence toward the speci ed viewpoint by and then demonstrate how sets of recommended items are changed.
H.3.3 [INFORMATION SEARCH AND RETRIEVAL]: Information ltering
A recommender system searches for items and information
that would be useful to a user based on the user's behaviors
or the features of candidate items [
During the RecSys panel discussion, panelists made the following assertions about the lter bubble problem. Biased topics would be certainly selected by the in uence of personalization, but at the same time, it would be intrinsically impossible to make recommendations that are absolutely neutral from any viewpoint, and the diversity of provided topics intrinsically has a trade-o relation to the tness of these topics for users' interests or needs. To recommend something, or more generally to select something, one must consider the speci c aspect of a thing and must ignore the other aspects of the thing. The panelists also pointed out that current recommender systems fail to satisfy users' information need that they search for a wide variety of topics in the long term.
To solve this problem, we propose an information neutral recommender system that guarantees the neutrality of recommendation results. As pointed out during the RecSys 2011 panel discussion, because it is impossible to make a recommendation that is absolutely neutral from all viewpoints, we consider neutrality from the viewpoint or information speci ed by a user. For example, users can specify a feature of an item, such as a brand, or a user feature, such as a gender or an age, as a viewpoint. An information neutral recommender system is designed so that these speci ed features will not a ect recommendation results. This system can also be used to avoid the use of information that is restricted by law or regulation. For example, the use of some information is prohibited for the purpose of making recommendation by privacy policies.
We borrowed the idea of fairness-aware mining, which we
proposed earlier [
In this section, we discuss information neutrality in recommendation based on the considerations on the lter bubble problem and the ugly duckling theorem.
We here summarize the lter bubble problem posed by Pariser
and the discussion in the panel about this problem at the
RecSys 2011 conference. The lter bubble problem is a
concern that personalization technologies, including
recommender systems, narrow and bias the topics of information
provided to people while they don't notice these facts [
He furthermore demonstrated a collection of search results from Google2 for the query \Egypt" during the Egyptian uprising in 2011 from various people. Even though such a highly important event was occurring, only sightseeing pages were listed for some users instead of news pages about the Egyptian uprising, due to the in uence of personalization. In this example, he claimed that personalization technology spoiled the opportunity to obtain information that should be commonly shared in our society.
We consider that Pariser's claims can be summarized as follows. The rst point is the problem that users lost opportunities to obtain information about a wide variety of topics. A chance to know things that could make users' lives fruitful was lessened. The second point is the problem that each individual obtains information that is too personalized, and thus the amount of shared information is decreased. Pariser claimed that the loss of sharing information is a serious obstacle for building consensus in our society. He claimed that the loss of the ability to share information is a serious obstacle for building consensus in our society.
RecSys 2011, which is a conference on recommender systems,
held a panel discussion the topic of which was this lter
bubble problem [
The absence of the absolutely neutral viewpoint is pointed out in the above panel. We here more formally discuss this point based on the ugly duckling theorem.
The ugly duckling theorem is a classical theorem in pattern
recognition literature that asserts the impossibility of
classication without weighing certain features or aspects of
objects against the others [
Why is an ugly duckling ugly? As described above, an ugly duckling is as ugly as a normal duckling, if all features and functions are treated equally. The attention to an arbitrary feature such as black feathers makes an ugly duckling ugly. When we classify something, we of necessity pay attention to certain features, aspects, or viewpoints of classi ed objects. Because recommendation is considered as a task to classify items into a relevant class or an irrelevant one, certain features or viewpoints must be inevitably weighed when making recommendation. Consequently, the absolutely neutral recommendation is intrinsically impossible.
We propose a neutral recommendation task other than the absolutely neutral recommendation. Recalling the ugly duckling theorem, we must focus on certain features or viewpoints in classi cation. This fact indicates that it is feasible to make a recommendation that is neutral from a speci c viewpoint instead of all viewpoints. We hence advocate an Information Neutral Recommender System that enhances the neutrality in recommendation from the viewpoint specied by a user. In the case of Pariser's Facebook example, a system enhances the neutrality so that recommended friends are conservative or progressive, but the system is allowed to make biased decisions in terms of the other viewpoints, for example, the birthplace or age of friends.
In this section, we formalize a task of information neutral recommendation and show a solution algorithm for this task.
In [
We formalize an information neutral variant of a predicting ratings task. x 2 f1; : : : ; ng and y 2 f1; : : : ; mg denote a user and an item, respectively. An event (x; y) is a pair of a speci c user x and a speci c item y. Here, s denotes a rating value of y as given by x. We here assume that the domain of ratings is real values, though domain of ratings is commonly a set of discrete values, e.g., f1; : : : ; 5g. These variables are common for an original predicting ratings task.
To treat the information neutrality in recommendation, we additionally introduce a viewpoint variable, v, which indicates a viewpoint neutrality from which is enhanced. This variable is speci ed by a user, and its value depends on an event. Possible examples of a viewpoint variable are a user's gender, which depends on a user part of an event, movie's release year, which depends on an item's part of an event, and a timestamp when a user rates an item, which depends on both elements in an event. In this paper, we restrict the domain of a viewpoint variable to a binary type, 0; 1, but it is easy to extend to a multinomial case. An example consists of an event, (x; y), a rating value for the event, s, and a viewpoint value for the event, v. A training set is a set of N examples, D = f(xi; yi; si; vi)g; i = 1; : : : ; N .
Given a new event, (x; y), and its corresponding viewpoint value, v, a rating prediction function, s^(x; y; v), predicts a rating value of an item y by a user x. While this rating prediction function is estimated in our task setting, a loss function, loss(s ; s^), and a neutrality function, neutral(s^; v), are given as task inputs. A loss function represents the dissimilarity between a true rating value, s , and a predicted rating value, s^. A neutrality function quanti es the degree of the neutrality of a rating value from a viewpoint expressed by a viewpoint variable. Given a training set, D, a goal of an information neutral recommendation (predicting rating case) is to acquire a rating prediction function, s^(x; y; t), so that the expected value of a loss function is as small as possible and the expected value of a neutral function is as large as possible over (x; y; v). We formulate this goal by nding a rating prediction function, s^, so as to minimize the following objective function: loss(s ; s^(x; y; v)) neutral(s^(x; y; v); v); (1) where > 0 is a parameter to balance between the loss and the neutrality.
In this paper, we adopt a latent factor model for predicting
ratings. This latent factor model, which is a kind of a matrix
decomposition model, is de ned as equation (3) in [
R; (3)
X (xi;yi;si)2D where R represents an L2 regularizer for parameters bx, cy, px, and qy, and is a regularization parameter. Once we learned the parameters of a rating prediction function, we can predict a rating value for any event by applying equation (2).
We then extend this model to enhance the information
neutrality. First, we modify the model of equation (2) so as to
depend on the value of a viewpoint variable, v. For each
value of v, 0 and 1, we prepare a parameter set, (v), b(xv),
c(yv), p(xv), and q(yv). One of parameter sets is chosen
according as a value of v, and we get a rating prediction function:
s^(x; y; v) =
(v) + b(xv) + c(yv) + p(xv)q(yv)>:
(4)
We next de ne a neutrality function to quantify the degree
of the information neutrality from a viewpoint variable, v.
In this paper, we borrow an idea from [
I(s^; v) =
Pr[s^; v] log
X Z v2f0;1g v2f0;1g = X
Pr[v]
Z
Pr[s^jv] Pr[s^]
ds^ Pr[s^jv] log
Pr[s^jv] Pr[s^] ds^: (5) The marginalization over v is then replaced with the sample mean over a training set, D, and we get Note that Pr[s^] can be calculated by Pv Pr[s^jv] Pr[v], and we use a sample mass function as Pr[v].
Now, all that we have to do is to compute distribution Pr[s^jv], but this computation is di cult. This is because a value of a function s^ is not probabilistic but rather deterministic depending on x, y, and v; and thus distribution Pr[s^jx; y; v] has a form of collection of Dirac's delta functions, (s^(x; y; v)). Pr[s^jv] can be obtained by marginalizing this distribution over x and y. As a result, Pr[s^jv] also becomes a hyper function like Pr[s^jx; y; v], and it is not easy to manipulate. We therefore introduce a histogram model to represent Pr[s^jv]. Values of predicted ratings, s^, are divided into bins, because sample ratings are generally discrete. The distribution P~r[s^jv] is expressed by a histogram model. By replacing Pr[s^jv] with P~r[s^jv], equation (6) becomes (v)2D s^2Bin 1 X N
X P~r[s^jv] log P~r[s^jv] ;
P~r[s^] (6) (7) where Bin denotes a set of bins of a histogram. Note that because a distribution function, Pr[s^jv], is replaced with a probability mass function, P~r[s^jv], an integration over s^ is replaced with the summation over bins.
L(D) = By substituting equations (4) and (7) into equation (1) and adopting a squared loss function as in the original latent factor case, we obtain an objective function of an information neutral recommendation model:
X
(si s^(xi; yi; vi))2+ I(s^; v)+ R; (8) (xi;yi;si;vi)2D where a regularization term, R, is a sum of L2 regularizers of parameter sets for each value of v. Model parameters, f (v); b(xv); c(yv); p(xv); q(yv)g; v 2 f0; 1g, are estimated so as to minimize this objective function. However, it is very di cult to derive an analytical form of gradients of this objective function, because the histogram transformation used for expressing Pr[s^jv] is too complicated. We therefore adopt the
Powell optimization method, because it can be applied without computing gradients.
We implemented our information neutral recommender system in the previous section and applied it to a benchmark data set.
We used a Movielens 100k data set [
We tested the following two types of viewpoint variable. The
rst type of variable, Year, represents whether a movie's
release year is newer than 1990, which depends on an item
part of an event. In [
We used the implementation of the Powell method in the
SciPy package [
Experimental results are shown in Figure 1. Figure 1(a)
shows the changes of prediction errors measured by a mean
absolute error (MAE) index. The smaller value of this index
indicates better prediction accuracy. Figure 1(b) shows the
changes of the mutual information between predicted
ratings and viewpoint values. The smaller mutual information
indicates a higher level of neutrality. Mutual information is
normalized into the range [0; 1] by the method of employing
the geometrical mean in [
MAE was 0.90, when o ering a mean score, 3:74, for all users and all items. In Figure 1(a), MAEs were better than this baseline, which is perfectly neutral from all viewpoints.
Furthermore, the increase of MAEs as the neutrality parameter, , was not so serious. Turning to the Figure 1(b), this demonstrates that the neutrality is enhanced as the neutrality parameter, , increases from both viewpoints, Year and Gender. By drawing attention to the fact that the Y-axis is logarithmic, we can conclude that an information neutrality term is highly e ective. In summary, our information neutral recommender system successfully enhanced the neutrality without seriously sacri cing the prediction accuracy.
Figure 2 shows the changes of mean predicted scores. In both gures, the X-axes represent parameter values of , and the Y-axes represent mean predicted scores for each case of using di erent viewpoint value. Figure 1(a) shows mean predicted scores when a viewpoint variable is Year.
Dashed and dotted lines show the results under the condition a viewpoint variable is \before 1990" and \after 1991", respectively. Figure 1(b) shows mean predicted scores when a viewpoint variable is Gender. Dashed and dotted lines show the results obtained by setting a viewpoint to \male" and \female", respectively.
We rst discuss a case that a viewpoint variable is Year. According to Figure 1(b), neutrality was drastically improved in the interval that is between 0 and 10. By observing the corresponding interval in Figure 2, two lines that were obtained for di erent viewpoints became close each other.
This means that prediction scores become less a ected by a viewpoint value, and this corresponds the improvement of neutrality. After this range, the decrease of NMI became smaller in Figure 1(b), and the lines in the corresponding interval in Figure 2 were nearly parallel. This indicated that the di erence between two score sequences less changes, and the improvement in neutrality did too. We move on to a Gender case. By comparing the changes of NMI between Year and Gender cases in Figure 1(b), the decrease of NMI in a Gender case was much smaller than that of a Year case. This phenomenon could be con rmed by the fact that two lines were nearly parallel in Figure 2(b). This is probably because the score di erences in a Gender case are much smaller than those in a Year at the point = 0, and there is less margin for improvement. Further investigation will be required in this point.
To enhance the neutrality, we borrowed an idea from our
previous work [
These techniques might be useful for detecting biases in recommendation.
Another type of fairness-aware mining technique focuses on
classi cation designed so that the in uence of sensitive
information to classi cation results is reduced [
Information neutrality can be considered as diversity in
recommendation in a broad sense. McNee et al. pointed out
the importance of factors other than prediction accuracy,
including diversity, in recommendation [
cation is a technique for enhancing the diversity in a
recommendation list [
There are several reports about the in uence of
recommendations on the diversity of items accepted by users. Celma
et al. reported that recommender systems have a
popularity bias such that popular items have a tendency to be
recommended more and more frequently [
Because information neutral recommenders can be used to
avoid the exploitation of private information, these
techniques are related to privacy-preserving data mining [
In this paper, we proposed an information neutral recommender system that enhanced neutrality from a viewpoint speci ed by a user. This system is useful for alleviating the lter bubble problem, which is a concern that personalization technologies narrow users' experience. We then developed an information neutral recommendation algorithm by introducing a regularization term that quanti es neutrality by mutual information between a predicted rating and a viewpoint variable expressing a user's viewpoint. We nally demonstrated that neutrality could be enhanced without sacri cing prediction accuracy by our algorithm. The most serious issue of our current algorithm is scalability. This is mainly due to the di culty in deriving the analytical form of gradients of an objective function. We plan to develop another objective function whose gradients can be derived analytically. The degree of statistical independence is currently quanti ed by mutual information. We want to test other indexes, such as kurtosis, which are used for independent component analysis. We will develop an information neutral version of other recommendation models, such as pLSI/LDA or nearest neighbor models.
We would like to thank for providing a data set for the Grouplens research lab. This work is supported by MEXT/JSPS KAKENHI Grant Number 16700157, 21500154, 22500142, 23240043, and 24500194, and JST PRESTO 09152492.