In collaborative filtering recommender systems, users cannot get involved in the choice of their peer group. It leaves users defenseless against various spamming or “shilling” attacks. Other social Web-based systems, however, allow users to self-select trustworthy peers and build a network of trust. We argue that users self-defined networks of trust could be valuable to increase the quality of recommendation in CF systems. To prove the feasibility of this idea we examined how similar are interests of users connected by a self-defined relationship in a social Web system, CiteuLike. Interest similarity was measured by similarity of items and meta-data they share. Our study shows that users connected by a network of trust exhibit significantly higher similarity on items and meta-data than non-connected users. This similarity is highest for directly connected users and decreases with the increase of distance between users.
Recommender systems powered by collaborative filtering (CF)
technologies become a feature of our life. Such popular systems
as Amazon.com, Netflix, Last.fm, and Google News use
Collaborative filtering to recommend us products to buy, movies
to watch, music to listen and news to read. The power of this
technology is based on a relatively simple idea: starting with a
target user’s rating, find a peer cohort (neighborhood) of users
who have similar interests and recommend items favored by this
cohort to the target user. As such, the choice of cohort is an
essential part in CF recommendations and is usually determined
by automatically calculating rating similarities between the target
user and other users. In a typical CF system, this peer cohort (a
group of users selected as the basis for CF) is unknown to target
users. Moreover, the target users cannot add trustworthy users to
their cohort group nor exclude suspicious users from the group.
The success of social linking and bookmaking systems that allow
users to build their networks of trust, stresses a fact forgotten by
modern CF systems: the source of the recommendation is an
important criterion for judging the quality of recommendations [
The popularity of CF technology, revealed some problems. CF
appeared to be not well-protected against malicious users who try
to harm the system or to make a profit by gamming the system.
For example, by copying the whole user profile, a malicious user
is perceived by the system to be a perfect peer user and the
products added by him are therefore recommended to the target
user [
Several research teams attempted to exploit trust between users to
resolve some of the cited problems of CF technology. Massa and
Avesani’s study [
To prove the feasibility of trust network as a source of information for reliable recommendation, several research teams started with checking the main assumption: do users linked by self-defined networks of trust have similar interests.
Singla and Richardson (2008) found the positive correlation of
frequency and time of instant messaging between users with
search interests [
As a source of data for our study we selected CiteuLike, a social Web system for sharing bibliographic references. To pick up initial set of users, we visited this site randomly in September and October of 2008. Users who posted new articles at the time of visit were picked. The information collected for each user included the bibliography (article title, list of authors, journal name, publication year, etc.) and the watchlists (connected users). After collecting a group of initial users, we collected data of their trusted connections. Table 1 shows the descriptive statistics. In collaborative tagging systems explicit connections between users are of special nature. In some sense, they bear more “trust” than the connections between friends in social networking systems. In CiteuLike, users can directly connect to other users who have interesting bibliography by adding them ‘watch list.’ Then the system displays the whole bibliography of watched users.
In this paper, we interpret user’s act of connecting to other users (by adding this user to the watch list) as a sign that she likes the focus and trust the quality of the added user’s references and wants to have direct access to them continuously in future. Thus, watching in CiteuLike could be considered as evidence that connected users are trustworthy to the original user in terms of information collection.
We distinguish two kinds of trusted connections – unidirectional and reciprocal. The act of adding another user to the ‘watch list’ is unidirectional (which is different from social networking systems). If user A added user B to her network, it does not imply that user B will be added to A’s network necessarily. The users in A’s network decide independently whether to add A to their networks. For example, user B may not have A in his network and we call the relationship between A and B as ‘unidirectional’. Another user C in A’s network may add A to his network as well. We call this relationship as ‘reciprocal’ (Figure 1).
In this study we tested how similar the information shared by
people in trusted network is. Specifically, we counted the number
of shared information items (academic papers) and meta-data. In
CiteuLike context, authors and journals (or conferences) is a good
example of meta-data.. In our study, however, we considered
authors only since it is more reliable and easy to track. Following
Ziegler and Golbeck [
For the information similarity in trusted network, the following hypotheses were assessed: H1. Users connected by direct or indirect relationships of trust have more similar information items and meta-data than a non-connected pairs. H2. Users in reciprocal relations have more similar information item and meta-data than users in unidirectional relations.
To test whether users connected by direct or distant links of trust share more information than non-connected pairs (H1), we compared both absolute numbers of shared information items and their normalized numbers (inlink, outlink, and overall powers) using one-way ANOVA test.
First, we explored the number of shared items and meta-data. Table 2 shows mean numbers of shared items and meta-data for direct and distant relationship on contrast to a non-related pair of users (which we can interpret as infinite distance). At average, direct pairs share the largest number of items and meta-data. The numbers are decreasing with the increase of distance in the network of trust achieving its minimum for non-connected pairs. This is the evidence that users connected in a network of trust do have significantly more similar interests than non-connected users. We can also consider it as an evidence of information propagation along a network of trust, although impressive similarity on the meta-data level (which are hard to propagate!) hints that interest similarity may play a more important role than propagation in the observed phenomenon. As Table 2 shows, reciprocal relationships exhibit the same pattern, also with significant differences between columns in the number of shared information items and meta-data. Second, we explored differences between relative similarity measures – fractions of shared items and meta-data for unidirectional relationship (Table 3) and reciprocal relationships (Table 4). In both cases, same pattern can be observed for relative similarity measures: directly related users have the largest fraction of shared items and meta-data and the fractions decrease with the increase of the distance between users and reach the minimal level for not connected users (infinite distance). In addition to demonstrating a clear connection between item and meta-data level similarity and user closeness in a network of trust, the data shown above allows to make interesting observations. First, as we expected, between-user similarity on the level of meta-data is much larger than similarity on the level of items for both systems. For example, the inlink power similarity of items in direct relation is 2.01% while inlink power similarity of meta-data in the same direct relation was 5.33%. Second, both absolute and relative similarities are pair-wise larger for reciprocal than for unidirectional connections for all distance levels. This difference is most pronounced in relative form reaching its highest level for direct reciprocal relations (6.79% for items and 13.01% for metadata). Next section examines the difference between reciprocal and unidirectional connections in details and checks its significance.
To compare the differences of information sharing pattern between unidirectional and reciprocal relations, we started with comparing the number of shared information items and meta-data, doing it now separately for several distances of relations. In all three distances but meta-data of 2-hop connection, the numbers of shared information items and meta-data in reciprocal relations were significantly larger than in unidirectional relations. In case of meta-data of 2-hop relation, there was no significant difference. Secondly, we checked the significance of observed differences in relative information item similarity between reciprocal and unidirectional relations (Table 6). For direct and 1-hop relationship, the differences appeared to be significant, i.e., users connected by a direct or 1-hop distanced reciprocal relation shared significantly larger fractions of information items than users connected by unidirectional relation. For 2-hops relations the observed difference appeared to be non-significant for one out of three relative similarity measures. On the final step we compared relative information meta-data similarity for reciprocal and unidirectional relations (Table 7). The relative source similarity was significantly higher for users connected by direct and 1-hop reciprocal relation than for users connected by unidirectional relations of the same distance. Two out of three relative similarities were significantly larger for reciprocal relations.
To prove the feasibility of users’ self-defined relations of trust as the bases of recommendation, we examined how similar interests of users connected by a self-defined relation of trust are. Using
CiteuLike datasets, we found that user connected by a self-defined relation of trust have more common information items and metadata than user pairs with no connection. The similarity was largest for direct connections and decreased with the increase of distance between users in the network of trust. Users involved in a reciprocal relationship exhibited significantly larger similarity than users in a unidirectional relationship on all levels. Moreover, similarity on the level of meta-data (authors) was larger than similarity on the level of individual items (references). While the results of our study support the idea of using networks of trust in CF systems, they still do not answer the question how to use this information to improve the quality of recommendation. In out future studies we plan to address this issue. As the first step, we will investigate the impact of trusted networks on recommendation quality using our CiteuLike data set. We will also explore how information propagates within trusted networks and investigate the influence of information authorities who play a leading role in disseminating the information. In later studies, we plan to expand our target domains by adding different data sets.