Table 2. Difference of Shared Items and Metadata 5.2 Information Similarity between a Group Members of the Members of the and the Members Same grou Different Grour In above section, we examined the information similarity among group members and found that each member had a tiny little Items H Numbers Mann-Whitney U= 123.0,P<.001 portion of common information with other members. Will they Relative 001% share enough information with group? In this section, we Mann-WhitneyU=139.5.p<.001 investigate whether the information sharing pattern of the group Absolute and the members is different with the one of group =93,p<00 Metadata Before the computation of similarity, we compared the co Relative 070% 0.16% of groups and the collections of the members. As you collections Mann-Whitney U=15.5.p<.001 the difference, groups' collection with 445. 89 items on average is As the next step, we compared the similarity in two kinds of tags gnificantly larger than the members' personal collections with 188.24 items on average(Wilcoxon Z=-3143, p<. 001).When macro-level tags and micro-level tags. The results of these tags several group members contribute to organize information in the were comparable to the results of items and metadata. The nembers of the same group shared significantly larger macro- similar suggestion can be made in the result of a correlation test. When we calculated the correlation between the number of group Table 3. Difference of Shared Macro-Tags and Micro-Tags members and the size of group collection, there was significantly Members of the Members of the positive correlation(r=0. 22, P<. 001)meaning that the more Different Group members a group has, the more items the group's collection Absolute contains. The group collection may be constituted evenly by the Micro- Numbers Mann-Whitney U=-123.0, p<.001 group members, not by only one or two leading members. We examined this even contribution of the members to the group Relative collection later Measures Mann-Whitney U=-139. 7, P<.001 Absolute We also tested whether the number of group of which each user is Macro- Numbers lann-Whiney U=-93 p<. 001 art correlates with the size of their personal collections. There tags Relative 032% are significantly positive correlations(r=0.18, P<. 001)even Measures Mann-Whitmey U=-15.5, p<.001 though the correlation is relatively small. That is to say, the more Although the same group members shared significantly larger i ups a user participated in, the more information he collected in al collection. It seems that being a member of a group may be helpful to gather useful information. So as to investigate common group in all explored levels, the amount is trivial (i.e. this idea further, we examined the similarities between a group 0.29%o of items. 0.83% of metadata and 0.86 of Said differently, even though each member of the groups have As the similarity test, we computed four different similarity values sufficient amount of ation in their personal collection(M= 251.53 on Table 1), he/she shared a very little information with the absolute numbers of common information the member his/her group members. We considered that this result may be fractions, group fractions and Jaccard coefficients for item, his group collection, the poster information is rather invisible items and 45.73% metadata)was overlapped with the collection of selectively retrieved or sorted by the poster information. Another their groups. Out of 3, 528 group memberships, 997 users did not possible reason of this little overlap is that each member desired have any common information item with their group. On the other hand. users whose item collection was 100%o matched with their very specific information and failed find the right one from other oup's collection were 873 users. The members whose personal members collection. Even though the poster information is not collection was at least 50% overlapped with their group shown very clearly, users are able to see the poster's personal collection were more than 40% of all the users in data set(using collection, when they clicked the poster name. It means that users item similarity, 40.70% of users and using metadata similarity, had a chance to refer what the poster had and to copy interesting 44.20% of users ). This is the interesting finding. People are much re similar to their groups but not their group members even whether this little overlap was caused by the interface problem or though they participate in the same group. Specifically, rather pared the groups collection and the members'collection. The system displays the than information items per se, the similarity of semantic level group collection in the same format of members personal, hence information such as metadata and macro tags is higher. we it is intuitive for users to navigate and refer to it. If there is lar interpret these results as groups or communities are good s to get interesting information but due to the T erlap with the group collection. the small fraction of similarit interface, users may be unable other members'collection etween members may be due to the interface. Otherwise, since Figure 6 and Figure 7 display the distribution of absolute numbers nembers are seeking too detailed information and develop their wn strategy of finding information, being a member of a group of overlapped information. Both figures show that many users ay be just a fruitless attempt to find information. We will check whether it is about inform overlap with their groups regardless have large informatic nation items, metadata or tags. ut the overlap between groups and the members collection and whether group members are sufficiently similar to their groupTable 2. Difference of Shared Items and Metadata Members of the Same Group Members of the Different Groups Items Absolute Numbers .75 .02 Mann-Whitney U = 123.0, p < .001 Relative Measures 0.28% 0.01% Mann-Whitney U = 139.5, p < .001 Metadata Absolute Numbers 7.78 2.77 Mann-Whitney U = 9.3, p < .001 Relative Measures 0.70% 0.16% Mann-Whitney U = 15.5, p < .001 As the next step, we compared the similarity in two kinds of tags – macro-level tags and micro-level tags. The results of these tags were comparable to the results of items and metadata. The members of the same group shared significantly larger macro￾level and micro-level tags than the random pairs. Table 3. Difference of Shared Macro-Tags and Micro-Tags Members of the Same Group Members of the Different Groups Micro￾tags Absolute Numbers .37 .00 Mann-Whitney U = -123.0, p < .001 Relative Measures 0.07% 0.00% Mann-Whitney U = -139.7, p < .001 Macro￾tags Absolute Numbers 3.98 .77 Mann-Whitney U = -9.3, p < .001 Relative Measures 2.0% 0.32% Mann-Whitney U = -15.5, p < .001 Although the same group members shared significantly larger amount of information than the random pairs who were not in any common group in all explored levels, the amount is trivial (i.e. 0.29% of items, 0.83% of metadata and 0.86% of macro-tags). Said differently, even though each member of the groups have sufficient amount of information in their personal collection (M = 251.53 on Table 1), he/she shared a very little information with his/her group members. We considered that this result may be related to the Citeulike interface. When a user posted an article to his group collection, the poster information is rather invisible since the font is small and the items in group collection cannot be selectively retrieved or sorted by the poster information. Another possible reason of this little overlap is that each member desired very specific information and failed find the right one from other members’ collection. Even though the poster information is not shown very clearly, users are able to see the poster’s personal collection, when they clicked the poster name. It means that users had a chance to refer what the poster had and to copy interesting items to their own collection. In what follows, in order to check whether this little overlap was caused by the interface problem or members’ very specific needs, we compared the groups’ collection and the members’ collection. The system displays the group collection in the same format of members’ personal, hence it is intuitive for users to navigate and refer to it. If there is large overlap with the group collection, the small fraction of similarity between members may be due to the interface. Otherwise, since members are seeking too detailed information and develop their own strategy of finding information, being a member of a group may be just a fruitless attempt to find information. We will check out the overlap between groups and the members’ collection and whether group members are sufficiently similar to their group. 5.2 Information Similarity between a Group and the Members In above section, we examined the information similarity among group members and found that each member had a tiny little portion of common information with other members. Will they share enough information with group? In this section, we investigate whether the information sharing pattern of the group and the members is different with the one of group members. Before the computation of similarity, we compared the collections of groups and the collections of the members. As you can reckon the difference, groups’ collection with 445.89 items on average is significantly larger than the members’ personal collections with 188.24 items on average (Wilcoxon Z = -31.43, p < .001). When several group members contribute to organize information in the group collection, this asymmetric proportion is natural. The similar suggestion can be made in the result of a correlation test. When we calculated the correlation between the number of group members and the size of group collection, there was significantly positive correlation (r = 0.22, p < .001) meaning that the more members a group has, the more items the group’s collection contains. The group collection may be constituted evenly by the group members, not by only one or two leading members. We examined this even contribution of the members to the group collection later. We also tested whether the number of group of which each user is part correlates with the size of their personal collections. There are significantly positive correlations (r = 0.18, p < .001) even though the correlation is relatively small. That is to say, the more groups a user participated in, the more information he collected in his personal collection. It seems that being a member of a group may be helpful to gather useful information. So as to investigate this idea further, we examined the similarities between a group and the individual member. As the similarity test, we computed four different similarity values – the absolute numbers of common information, the member fractions, group fractions and Jaccard coefficients – for item, metadata, micro-tag and macro-tag levels. As shown in Table 4, nearly half of group members’ personal collection (42.16% of items and 45.73% metadata) was overlapped with the collection of their groups. Out of 3,528 group memberships, 997 users did not have any common information item with their group. On the other hand, users whose item collection was 100% matched with their group’s collection were 873 users. The members whose personal collection was at least 50% overlapped with their group’s collection were more than 40% of all the users in data set (using item similarity, 40.70% of users and using metadata similarity, 44.20% of users). This is the interesting finding. People are much more similar to their groups but not their group members even though they participate in the same group. Specifically, rather than information items per se, the similarity of semantic level information such as metadata and macro tags is higher. We can interpret these results as groups or communities are good source to get interesting information but due to the inappropriate interface, users may be unable to see other members’ collection. Figure 6 and Figure 7 display the distribution of absolute numbers of overlapped information. Both figures show that many users have large information overlap with their groups regardless whether it is about information items, metadata or tags