Expanding User Interests by Recommending Innovative Blog Entries

1. Introduction

Consumer-generated media (CGM) such as blogs and social networking services (SNSs) are becoming more popular for publishing and discussing shared interests among users. In some CGM services like last.fm [1], a music SNS, users disclose their interests not only actively by writing entries in their blogs but also passively through listening to music. Information-sharing systems of these types could enable users to expand their interests by browsing collected blog entries or listening histories published by other users. Online recommendations made by analyzing these published user interests are essential for service providers to increase the sales of their contents. We can automatically create user profiles about their interests from listening histories by checking how many times a user listens to the same songs. By applying collaborative filtering (CF) [2]–[4] to measure the similarity between user profiles, we can recommend several songs to users comparatively easily. However, CF is apt to recommend information resources that are the same as the concepts in the user's profile. Furthermore, there is a major problem called the sparsity problem: this occurs when there is not enough data in the user profile to measure the similarity among users.

In our research [5], we define an innovation as a new concept that is likely to be interesting to the user even though it is not included in his or her user profile. We try to expand the scope of user interests significantly by recommending innovative information. In particular, we apply innovation detection to blogs (web logs) because they have become a popular publishing format and targets for searching for information that could expand user interests.

To achieve this purpose, we first construct a user profile automatically as a user-interest ontology, which is a class hierarchy of user interests with interest weights. We present a method of automatically extracting an interest ontology with an interest weight assigned to each class and instance. Bloggers are apt to describe their interests about topics in several service domains freely. Thus, we use blog entries to specify user interests by introducing a template ontology, which is a domain ontology of one service. We classify user entries according to a template ontology and remove classification mistakes by using class characteristics and the continuity of descriptions about user interests using a top-down approach. We also provide a user interface to update the interest ontology using a bottom-up approach.

Next, we present a method for measuring the similarity of interest ontologies considering the degree of interest agreement for each class and instance. Then, we present Semantic-Driven CF by applying our measurement of similarities between user-interest ontologies to the previous CF with Pearson correlation coefficients [2]–[4]. Thus, we detect innovative blog entries for each user by analyzing other users' ontologies that have a high degree of similarity to that of the user. We apply our techniques to help users create blog communities by browsing innovative blog entries that include information that is unknown to users and has a high probability of being interesting. We executed offline experiments based on a large number of blog entries (1,600,000 entries of 55,000 users) on an actual blog portal Doblog [6], which is one of the biggest blog portals in Japan. In December 2006, it had about 55,000 users. For these experiments, we used a music template ontology with 114 classes and 4300 instances. We confirmed that our automatic ontology extraction and innovation detection have potential for creating user-oriented blog communities that correspond to user interests.

The specific contributions of this article are as follows.

1. We impose a class taxonomy of instances of user interests in user profiles. Thus, we improve the accuracy of recommendations, especially for users with a small number of instances in their user profiles, by considering the similarities between instances and user-interest ontologies. This is important because many content providers have to handle a lot of users with a small number of entries in their profiles, such as light users who have just begun to use the service. We confirmed the effectiveness of our technique by using extracted results of interest ontologies to compare Semantic-Driven CF with the previous CF.

2. The evaluation was done in two steps. The first step was an offline evaluation for extracting interest ontologies and comparing the previous CF with our Semantic-Driven CF. The second step was an online evaluation for analyzing user reactions to recommen-

dations based on innovation detection. (We ran an experimental service called DoblogMusic for Doblog users from August to December 2006.) Most previous studies evaluated their recommendation techniques by using only offline experimental results. However, it is very important to analyze user reactions to

recommendations to check whether the recommendation techniques are actually effective. By analyzing the change in the frequency of user accesses to innovative instances of our recommendations, we confirmed the effectiveness of our ontology extraction and innovation detection. Through an evaluation of the frequency of comments between users who got to know each other through our online recommendations, we also found that recommendations based on innovation detection facilitated the creation of new communities.

2. Interest-ontology extraction

Here, we explain the algorithm for generating an interest ontology by analyzing the distribution of user interests, as shown in Fig. 1.

Fig. 1. Procedure for generating interest ontologies.

(1) First, we make index files for all blog entries collected through the ping server. Here, we assume that each collected blog entry has a unique user ID.

(2) Second, we classify all collected blog entries into a template ontology. This is knowledge about the content taxonomy such as the hierarchy of genres of music artists; it is constructed by the designers of service providers to manage and promote their contents. For example, consider the template ontology in Fig. 1. Here, the instances are the names of musical bands (e.g., Stone Roses), while classes are the names of musical genres (e.g., Madchester, a term coined to describe the alternative music scene in Manchester, UK, in the late 1980s to early 1990s). We classify a blog entry into the instance "Happy Mondays" of class "Madchester" when the character string "Happy Mondays" is included in the body of a blog entry. However, our algorithm mistakenly classifies blog entries about agricultural farms into the instance "Farm" of class "Madchester" because the only meaning it knows for farm is the name of a British band. To filter out mistakes caused by words with multiple meanings, we make use of characteristics such as class relationships in ontologies and the durability of user interests in a blog.

1. Adjacent classes have similar characteristics. Instances of those classes also have similar characteristics.

2. User interests continue for a certain period and describe an interest lasting two or more days.

(3) Then, we measure the number of users interested in each instance of Ce, which is one of the end classes in the template ontology. We calculate the number of users interested in class Ce by obtaining the number of users interested in all instances in Ce and in class Ce. Thus, the distribution of interested users in the domain can be measured by recurrently counting the number of users from Ce to the root class Cr.

(4) Next, by extracting only the classification results for a certain user ID from all classification results, we can extract an interest ontology for this user ID. In Fig. 1, we can extract an interest ontology of user A when the blog entries of this user describe instances of "Stone Temple Pilots", "New Order", and "Farm".

3. Interest-weight-based similarity measurement

We now explain our similarity measurement using Fig. 2. First, we define the terminology. We denote the interest ontology of users A and B by O_A and O_B. We use two topologies: T₁ is composed of a class and subclass relationship and T₂ is composed of a class and instance relationship. Furthermore, we define common classes and common instances of both ontologies as Ci and Ii, respectively. In particular, we define a common class set that formalizes topology T₁ as C(T₁) and a common class set that formalizes topology T₂ as C(T₂). For example, in Fig. 2, C(T₁) has common classes a1 and b2, and C(T₂) has common classes b2, b3, and c4. We also denote the degree of interest agreement of common instance Ii as I(Ii), that of common class Ci as I(Ci), and that of common topology created by common class Ci as It(Ci).

Fig. 2. Measuring similarity based on the degree of interest agreement.

Maedche and Staab [7] calculated the similarity between ontologies considering the degree of similarity between class topologies T₁. We extend this by additionally taking the following ideas from the viewpoint of creating user-interest-based communities.

1. We evaluate the degree of interest agreement between Ci s and Ii s as a smaller value of interest weight. This idea is for filtering users who only enumerate a lot of instances in an entry and for creating a community among users who have similar or larger interest weight values from the viewpoint of each user.

2. We treat topologies T₁ and T₂ separately because we consider that T₁ reflects the width and depth of a user's interests while T₂ reflects the objects in which users are interested.

3. We achieve low computational complexity by generating the class schema of user-interest ontologies according to that of template ontologies. For ontology mapping, it is important to use a large-scale dataset of the blog community such as that used in our experiments described in Sections 4 and 5.

The procedure of our method is given below.

(1) We analyze classes common to O_A and O_B and extract common classes that belong to C(T₁) and C(T₂).

(2) If common class Ci has common instance Ii between ontologies, we assign the smaller value of the interest weight of common instances Ii to I(Ii). For example, I(a) is 2.

(3) Similarly, we assign the smaller value of the interest weight of common class Ci to I(Ci). For example, I(b1) is 3.

(4) We define product sets of subclasses of Ci, which are common to a class set, as N(Ci), and we define the set union of subclasses of Ci among Ci∈C(T₁) as U(Ci).

For example, N(a1) = {b1, b2} and U(a1) = {b1, b2, b3}. Then, we express It(Ci) as .
For example, I_t(a1) is given by (3+18+0)/3 = 7. Thus, we obtain degree of interest agreement S(T₁) of C(T₁) as Σ_Ci∈C(T1)I_t(Ci). In Fig. 2, S(T₁) = (3+18+0)/3+(9+3)/2.

(5) We also define an instance set of Ci in ontology O_A as I_A(Ci), and we define an instance set of Ci in ontology O_B as I_B(Ci) among Ci∈C(T₂). Then, we express It(Ci) as .
For example, I_t(C3) is given by ((2+0+3+0)/4) = 5/4. Thus, we assign the degree of interest agreement S(T₂) of C(T₂) as Σ_Ci∈C(T2)I_t(Ci). In Fig. 2, S(T₂) = 2/1+5/4+0.

(6) By using evaluation function f(X) corresponding to the relative degree of importance of a topology, we finally assign the similarity score between ontologies S_o(AB) as S(T₁)+f(S(T₂)).

4. Results of offline experiment

Here, we present the results of offline experiments and simulation studies that show the performance of interest ontology extraction and innovative blog-entry detection. We also compared the predictions of Semantic-Driven CF and the previous CF.

We evaluated the performance of our methods using the large-scale blog portal Doblog. We also used the template ontology of the music domain (Fig. 1), which was created by referring to public information on goo music [8], a web portal containing music artist genre information. Our experimental template ontology contained 114 classes as genres and 4300 artists as instances. Each class and instance could have two or more name attributes. For example, the instance "R.E.M." has the name attributes of "R.E.M." and "REM". In total, we gave 7600 name attributes to 4300 instances.

To evaluate the accuracy, we defined correct answers as blog entries that have descriptions of classified classes or instances and evaluated the generated interest ontology by using precision and recall in the classified results. In this paper, precision means the proportion of correct answers in classified results and recall means that of correct answers in all blog entries. When the recall was high, extracted interest ontologies covered user interests better. However, when the precision was lower, created interest ontologies included classification mistakes, and innovation detection for the user was unreliable. Thus, it is essential to achieve high precision. In the evaluation, we applied filtering algorithms to instances consisting of only one word such as "police", because we considered a single word to have a high possibility of having several meanings. To generate index files of blog entries, we used Namazu [9].

5. Conclusion

We presented Semantic-Driven CF for expanding user interests significantly by measuring the similarity between user-interest ontologies. By using class characteristics of instances in a user profile, we can detect information that is innovative for users and improve the recommendation results, especially for users with few entries in their profiles. Comparing the effectiveness of Semantic-Driven CF with the previous CF, we found that Semantic-Driven CF achieved good prediction performance for users with few entries in their profiles. Through an online evaluation, we confirmed the effectiveness of innovation detection by checking the access frequency of users that clicked on innovative artists.

Further work is required to study how users control their interest ontologies by providing feedback about their collective knowledge to update class hierarchies of template ontologies constructed by expert designers of ontologies. In addition, we need to evaluate our techniques by comparing them with template ontologies of other domains such as those of fashion or movies.

References

[1]	http://www.last.fm/
[2]	B. M. Sarwar, G. Karypis, J. Konstan, and J. Riedl, "Application of dimensionality reduction in recommender systems—a case study, In ACM WebKDD Workshop, 2000.
[3]	J. S. Breese, D. Heckerman, and C. Kadie, "Empirical Analysis of Predictive Algorithms for Collaborative Filtering, Proc. of the Fourteenth Annual Conference on Uncertainty in Artificial Intelligence, pp. 43–52, 1998.
[4]	P. Resnick, N. Iacovou, M. Suchak, P. Bergstorm, and J. Riedl, "GroupLens: An Open Architecture for Collaborative Filtering of Netnews, Proc. of ACM 1994 Conference on Computer Supported Cooperative Work, Chapel Hill, North Carolina, ACM, pp. 175–186, 1994.
[5]	M. Nakatsuji, Y. Miyoshi, and Y. Otsuka, "Innovation Detection Based on User-Interest Ontology of Blog Community, International Semantic Web Conference (ISWC2006), pp. 515–528, 2006 (http://iswc2006.semanticweb.org/items/Nakatsuji2006dp.pdf).
[6]	http://www.doblog.com (in Japanese).
[7]	A. Maedche and S. Staab, "Measuring Similarity between Ontologies, Proc. of the European Conference on Knowledge Acquisition and Management, EKAW-2002. Madrid, Spain, LNCS/LNAI 2473, Springer, pp. 251–263, 2002.
[8]	http://music.goo.ne.jp/ (in Japanese).
[9]	http://www.namazu.org/index.html.en

↑ TOP