Proceedings of the workshop "Intelligent Educational Systems on the World Wide Web",
8th World Conference of the AIED Society, Kobe, Japan, 18-22 August 1997

Toward Learning from Surfing


Toyohiro Nomoto, Noriyuki Matsuda, Tsukasa Hirashima, Jun'ichi Toyoda
ISIR, Osaka University
8-1 Mihogaoka, IBARAKI, OSAKA 567, JAPAN
Phone : +81-6-879-8426
Fax : +81-6-879-8428
nomoto@ai.sanken.osaka-u.ac.jp
http://www.ai.sanken.osaka-u.ac.jp/thome2/nomoto/

Abstract. Browsing is one of the most popular ways to gather information in hypertext, for example, digital libraries or WWW. The surfing has a learning effect as its secondary one, although the purpose of the surfing is not learning. It is difficult to control the learning process from the educational viewpoint. However, the amount of knowledge gathered during browsing cannot be disregarded. Therefore, the way to support learning from browsing is an important issue in learning systems.
Usually, the browsing is carried out under a clear task. However, people often browse WWW or CD-ROM encyclopedia without a clear task. Such Browsing is called "surfing." In surfing, educational interruption may be allowed, because the purpose of surfing is only to surf. During surfing, user's interests often shift depending on the collected information, it is not haphazard one (so, we often call surfing "context-sensitive browsing"). As a first step to realize the support facilities for learning from surfing, we propose a filtering method to support surfing. In surfing, a user is often provided too many choices to reflect the collected information. In such situation, because user's cognitive load is heavy, learning from surfing is inefficient. In the filtering method, the shifting interests are modeled and the choices are ordered following the interests. The user may afford to learn from the collected information, because the load to find adequate information is reduced.
In this paper, we propose a filtering method for surfing. We have implemented it for CD-ROM encyclopedia. The results of experimental evaluation are reported in this paper.

Keywords: Filtering, Learning, Context, Surfing, Hypertext

1. Introduction

Browsing is one of the most popular ways to gather information in database with hypertext structure, for example, digital libraries or WWW. Although the purpose of the browsing isn't learning, a user can get much knowledge from the collected information. Because this learning is a secondary effect of the browsing and isn't organized, it is difficult to control the learning from the educational viewpoint. However, the amount of knowledge acquired from browsing cannot be disregarded. Therefore, the way to support learning from browsing is an important issue in learning systems. Usually, the browsing is carried out under a clear task. In such browsing, to complete the task should take precedence to learn. To support learning from the collected information may interrupt to complete the task. Therefore, the browsing with a clear task isn't an adequate target of the support. Every browsing, however, doesn't always have a clear task. People often browse WWW or encyclopedia in CD-ROM format without clear tasks. Such browsing is called "surfing." Because the purpose of surfing is only to surf, educational interruption may be allowed. Our research objective is to support learning from surfing.

Although user's interests often shift depending on the collected information, it's not haphazard one (so, we often call surfing "context-sensitive browsing"). In this paper, as the first step to realize the support facilities for learning from surfing, we propose a filtering method to support surfing. In surfing, a user is often provided too many choices to reflect the collected information. In such situation, because user's cognitive load is heavy, learning from surfing is inefficient. In the filtering method, the shifting interests in the surfing are modeled and the choices are ordered following the interests. The user can easily find the choices which are nearer to his/her interests. Because the load to find adequate information is reduced, the user may afford to learn from the collected information. In this paper, we propose a model of the shift of user's interests. Then, we describe a method to filter information based on the model of the interests. We have implemented the filtering method for encyclopedia in CD-ROM format. The results of experimental evaluation are reported in this paper.

2. Context-Sensitive Filtering

2.1 Support of the Context-Sensitive Browsing

There are two issues to be understood in support of the context-sensitive browsing: one is the author links and the other is the excess of choices. In most hypertext, a user has to follow links designed by the author. These links are made based on the author's viewpoint for the content of the hypertext. When the user has the same viewpoint as the author, the user can comfortably browse useful information on the hypertext. However, a user isn't always able to follow the author's viewpoint. Especially when the user browses with his/her shifting interests, a wide variety of viewpoints should be available. It is impossible for the author to completely cover all aspects of users interests.

A solution to this problem is to link each node through indexes [3, 5]. In this method, first, each node is assigned a listing of indexes. When two nodes have the same index, they are linked through that index. This method of connecting a node to every node, which can be associated by indexes included in the node, provides a dense hypertext structure in response to various usages from several viewpoints. Such a hypertext system allows a user to browse freely following her/his interests.

There is a trade-off to this advantage. An excess of choices can occur [3, 5]. When a user is provided several choices at each browsing node, the user is often at a loss and finds it difficult which to choose. This is a typical problem of information filtering [4]. Links designed by the author are filtered from the author's viewpoint, so the quality and number of links are guaranteed to a certain extent. Adaptive filtering (ordering) is one of the methods to help a user to browse a hypertext, where the nodes are linked through indexes, without difficulty [6].

So we propose a method to filter the links which connect nodes through indexes, based on the user's browsing history [9]. We assume that the history reflect shift in user's interests. Even when a user browses, following changeable interests without a clear task, the browsing is not a haphazard one. It is reasonable to assume that the useršs current interests are reflected in the content and order of nodes in the browsing history. The filtering method models user's current interests from the user's browsing history and puts the next choices in order of the nearness to his/her interests. We call the filtering method "context-sensitive filtering". Here, "context" is used to emphasize that the sequence of nodes in browsing history should be considered.

2.2 Model of User's Interests

Model of user's interests depends on the contents of each node and their order in the browsing history. Currently, a list of indexes for each node is used as the description of the contents. When a user visits and accepts a node, two assumptions are made (I) that the user is interested in the indexes for the node, (II) that the user is more interested in an index included in several nodes and in more recently browsed nodes. Based on these assumptions, the model of interests is described as a set of pairs: an index and its current weight (Index weight : Iw). The weight means the user's interest for the index. As the sum total of indexes included in a node, this method calculates the weight of nodes (Node weight : Nw) which are choices for the next browsing nodes, and then, judges that the heaviest node is the nearest node for the user's current interest.

To calculate the weight of index, Iw(i, k), the weight of node, Nw(k), should be defined. Nw(k) has to satisfy three conditions:

(Condition-1) the values of Nw(k) have to increase monotonically. This is because the weight of the latter node should be higher than the weight of the former node. This is a basic condition to reduce the weights of older nodes.
(Condition-2) the weight of an index should change following increments of n and shouldn't diverge, even when n becomes a large number. This condition is necessary to be context-sensitive because the user's interests should be influenced and changed by newly collected information.
(Condition-3) the computational quantity which is required to renew the weight of an index, should be finite and independent of n. This condition is necessary to be practical. In this method, the weight of kth node should be decreased when one more node is added in the browsing history. This means that every node weight in the history should be recalculated following increments of n. If the recalculations are simply carried out, the recalculations require numerous computational quantities and this method may be not practical.
We use a geometric progression sequence which satisfies these three conditions:

Nw(k) = r^(n-k) (0 < r < 1) (1)

where r corresponds to the ratio of decrease of user's interests. When n increases one, the weight of index term i changes depending on whether the nth node include index i or not. The weight of an index can be renewed by a recurrence formula. The verifications are omitted in this paper due to limitations of space.

3. Experimental Evaluation by Real Users

3.1 Browser for Experiment

To evaluate the context-sensitive filtering, we developed a browser which allows a user to browse nodes in CD-ROM encyclopedia as hypertext, where the nodes are linked through indexes. In the encyclopedia, there are 18,892 nodes which were given indexes beforehand. In the indexes, 183,011 words are used. From two to 2,270 nodes (the mean is seven nodes) are connected through one index.

3.2 Value of r

The value of common ratio r in Equation 1 corresponds to the decreasing ratio of user's interests. In the case of r=1, the weight of every node in the browsing history is one. It means the oldest browsed node has the same influence to the user's current interests as the most recently browsed node. Following this, the weights of the candidate nodes are decided by a set of the browsed nodes, independent of their sequence in the history. On the other hand, in the case of r=0, only the current node has weight one, but the weight of all other nodes in the browsing history is zero. Here, the weights of the candidate nodes are decided only by the current node independent of the browsing history. We defined our method in the range of 0 In this paper, we experimented the cases r=0.0, r=0.6, and r=1.0.

3.3 Procedure of the Experiment

To evaluate the method, we compared the following three methods:

(Ordering Method-1: OM-1) Our ordering method in the case of r=0.6 in Equation 1: This ordering method is the context-sensitive filtering proposed in this paper. The candidate nodes are ordered in response to the user's browsing context. (Ordering Method-2: OM-2) In the case of r=0 in Equation 1: this method orders the candidate nodes by considering only the current node. This means the weight of the candidate nodes is decided statically independent of the browsing history. (Ordering Method-3: OM-3) The order of the raw results of retrieval.

First, a user is provided a browser which uses one ordering method among the three. The user is asked to browse freely an encyclopedia. After the user has browsed several nodes, he/she is presented a list of candidate nodes in random order. The user is, then, asked to read each of them and to judge whether or not each node is acceptable as the next node. The results are ordered again by the three methods. Each of the three methods is, then, evaluated whether or not they can put the accepted candidate nodes higher than the rejected nodes in their lists. This procedure is described in more detail in the next section.

In this experiment, when the user has already browsed more than five nodes and the number of candidate nodes is 30 to 60, he/she is asked to judge the candidate nodes. Each user is asked to browse three times. Every time the ordering method is different; the user, however, isn't informed of the differences. Twelve users participated in this experiment and they were divided into two units. Each unit was formed of six users who used the three methods in different sequences. This was done to balance the order of the use of the three methods. All of the users were graduate and undergraduate students who regularly use e-mail and browsers of WWW.

3.4 Procedure of Evaluation

The scoring procedure of the three ordering methods is explained with Table I. Assuming that five candidate nodes (CNs) are presented to a user in the order in the first column, he/she, then, accepted two nodes as the next browse node, CN-B and CN-D. A check means the accepted node; a blank means the rejected nodes. These nodes are ordered again by each of the three ordering methods. The orders are then, scored. As an example, the score of the fourth column is explained. Each accepted candidate node is given a score dependent on its rank. If the accepted node is ranked higher in the list, it is given a higher score. When the number of candidate nodes is N, the score of each accepted node is calculated as (N-j+1), where j is the ranking of the candidate node. In the fourth column, because CN-D is placed first (j=1) and the total number of candidate nodes is five (N=5), its value is five. Similarly, the value of CN-B is three. The score of an ordering method is the sum total of the scores of the accepted nodes. The score of the ordering method used in the fourth column is eight.

Table I. An example of scoring procedures of the three ordering methods.
User's
evaluations		 Ranking
(j)
 Score (N=5)
(N-j+1)
 Score of
an ordering method
CN-A
1
5 : (N)
 CN-D * 5
CN-B *
2
4 : (N-1)
 CN-A
CN-C
3
3 : (N-2)
 CN-B * 3
CN-D *
4
2 : (N-3)
 CN-C
CN-E
5
1 : (N-4)
 CN-D
Total Score
8

These scores are described by number values. The measure of the value, however, is different depending on each browsing context and the regularity of intervals isn't guaranteed. The score differences are dealt with in a ranking scale. Here, two of the three conditions are examined by a two-sided sign test. The results are explained in the next section.

4. Results and Discussion

We asked twelve users to browse the hypertext with the browser shown in Figure 3 three times and evaluate a list of the candidate nodes in each time. A total of thirty-six sets of scores was gathered. The results examined by the two-sided sign test are shown in Tables II, III and IV. The numbers in the second column are the numbers of times when the score of one method is larger than the other. The cases of equivalent scores are omitted. Based on these results, it is significant that OM-1, the context-sensitive filtering proposed in this paper, is better than OM-2 and OM-3 in ordering the candidate nodes which a user wants to browse upward in the list.

In addition, we implemented one more ordering method which is in the case of r=1: OM-4. This method gives equivalent weight to every node independent of the time order in the history, in contrast to OM-1, which reduces the weight of nodes following the sequence in the history. By using the browsing histories collected in the above experiment, the scores of OM-4 for each browsing history were calculated. The scores were then compared with OM-1 in the same way. The result is shown in Table V. It is significant that OM-1 is better than OM-4.

Table II. OM-1 vs OM-2
(* OM: Ordeirng Method) OM-1 OM-2 two-sided sign test
The numbers of times when
the score of one method is
larger than the other
26
8
p = 0.0021

Table III. OM-1 vs OM-3
(* OM: Ordeirng Method) OM-1 OM-3 two-sided sign test
The numbers of times when
the score of one method is
larger than the other
31
3
p = 7.0 e-07

Table IV. OM-2 vs OM-3
(* OM: Ordeirng Method) OM-2 OM-3 two-sided sign test
The numbers of times when
the score of one method is
larger than the other
30
2
p = 2.3 e-07

Table V. OM-1 vs OM-4
(* OM: Ordeirng Method) OM-1 OM-4 two-sided sign test
The numbers of times when
the score of one method is
larger than the other
24
8
p = 0.0070

The OM-1, -2 and -4 use the browsing history in their own way to infer user's interests. It is significant that OM-1, -2, and -4 are better than OM-3, which uses raw results of retrieval. This fact suggests that the browsing history is a useful source to infer the user's interests. In OM-2 (r = 0), only the current node has weight one but the weight of all other nodes in the browsing history is zero. It means the weights of the candidate nodes can be decided statically independent of the browsing history. On the other hand, in OM-4 (r = 1), the weight of every node in the browsing history is one. It means the weights of the candidate nodes are decided by a set of the browsed nodes independent of their sequence in the history. The main characteristics of OM-1 (r = 0.6) are (1) reducing the weights of the nodes to deal with their sequence in the history, and (2) cumulating their weights to deal with the history. It is significant that OM-1 is better than both OM-2 and -4. This fact suggests that the two characteristics are suitable to support the browsing.

The above, however, are only statistical results. In not all cases, OM-1 is better than other methods. When the user's interests change frequently, OM-2 may be better. In contrast, when the user's interests are stable in long term, OM-4 may be better. During the interests shift gradually depending on the local context, OM-1 has an advantage over the other methods. OM-1, overall, is significantly better than the other methods because browsing behaviors, which are suitable to be supported by OM-1, are more frequent than others. This type of browsing is "context-sensitive browsing."

In the experiment, the users were asked to browse freely. In addition, they also browsed on a large scale and a dense hypertext. The hypertext included about twenty thousand nodes, each node with about four hundred links to other nodes ( each node has sixty-eight indexes and each index is shared by seven nodes on the average). Furthermore, the hypertext was an encyclopedia, so the content of each node had high independency. These conditions allowed the users to browse following their changeable interests. When these conditions aren't satisfied, OM-1 may be not useful. The context-sensitive browsing, or surfing, however, becomes one of the most important behaviors of information gathering. Hypertexts are becoming larger and denser rapidly. Diffusion of search engines in WWW also enables a user to browse following his/her changeable interests. The browsing behavior, which OM-1 supports effectively, can steadily becomes widespread.

5. Related Works

Many researchers are developing support facilities (or agents) for browsing. Several promising results have been reported [5, 10]. The browsing task in these investigations is limited, though, for two reasons: (1) the purpose should be specified before a user begins to browse, (for example, looking for a particular software module [7] or technical papers [1]), and (2) the user's interests should be consistent during the browsing. Under these limitations, the browsing task has many repeated behaviors done by users. Assuming that these repeated behaviors include many recurrent patterns of browsing, machine learning is a promising method to predict user's behaviors [11]. Most research has adopted the machine learning approach, where the systems monitor the user's behaviors and find regularities and recurrent patterns.

In the context-sensitive browsing, however, a user begins to browse with a wide variety of purposes. When this occurs, the interests often change. Such browsing is often called "surfing." Balabanovic and Shoham reported that the machine learning approach is often useless to support surfing [2]. Lieberman suggested that even when a user states her/his interests before browsing, the interests decay over time [8]. In the context-sensitive browsing, a user browses information depending upon the local context. Our method which uses the local context for filtering is suitable to deal with the context-sensitive browsing. Besides, our method was able to be evaluated by real users. Few other investigations were evaluated by real users. The reason for this may be that the context-sensitive browsing is natural browsing for them.

6. Conclusion Remarks

As a first step to realize "Toward Learning from Surfing", we have introduced a context-sensitive filtering method for context-sensitive browsing. In the context-sensitive browsing, user's interests gradually shift depending on the browsing context. The main characteristic of context-sensitive filtering is the model of user's shifting interests. Our filtering method can order candidate nodes according to the model. We have implemented a browser with this filtering method as an encyclopedia in CD-ROM format. The effectiveness of this method was confirmed through an experiment where real users browse nodes freely, following their interests.

References

[1] Armstrong, R., Freitag, D., Joachims, T., Mitchell, T.: 1995, 'WebWatcher: A Learning Apprentice for the World Wide Web'. On-line Working Notes of AAAI Symposium on Information Gathering from Distributed, Heterogeneous Environments (available at http://WWW.isi.edu/sims/knoblock/sss95/info-gathering.html).
[2] Balabanovic, M., Shoham, Y., : 1995, 'Learning Information Retrieval Agents: Experiments with Automated Web Browsing'. On-line Working Notes of AAAI Spring Symposium on Information Gathering from Distributed, Heterogeneous Environments, (available at http: //WWW.isi.edu/sims/knoblock/sss95/info-gathering.html).
[3] Beaumont, I., Brusilovsky, P.: 1995, 'Adaptive Educational Hypermedia: From Ideas to Real Systems'. Proc. of ED-MEDIA'95 - World Conference on Educational Multimedia and Hypermedia, pp.93-98.
[4] Belkin, N.J., Croft, W.B.: 1992, 'Information Filtering and Information Retrieval: Two Sides of the Same Coin?'. CACM, Vol.35, No.12, pp.29-38.
[5] Brusilovsky, P.:1996, 'Methods and Techniques of Adaptive Hypermedia'. User Modeling and User-Adaptive Interaction, Vol.6, pp.87-129.
[6] Brusilovsky, P., Pesin, Leonid.: 1994, 'ISIS-Tutor: An Intelligent Learning Environment for CDS/ISIS Users'. Proc. of CLCE'94. Joensuu, Finland.
[7] Holte, R.C. & Drummond, C.: 1994, 'A Learning Apprentice for Browsing'. Proc. of AAAI Spring Symposium on Software Agents.
[8] Lieberman, H.: 1995, 'Letizia: An Agent That Assists Web Browsing'. Proc. of IJCAI95, pp.924-929.
[9] Kaplan, C., Fenwick, J., Chen, J.: 1993, 'Adaptive Hypertext Navigation Based on User Goals and Context'. User Modeling and User-Adapted Interaction, Vol.3, No.3, pp.193-220.
[10] Knoblock, C. & Levy, A.(Eds.): 1995, 'On-line Working Notes of the AAAI Spring Symposium Series on Information Gathering from Distributed, Heterogeneous Environments'. (http://www.isi.edu/sims/knoblock/sss95 /proceedings.html).
[11] Maes, P.: 1994, 'Agents that Reduce Work and Information Overload'. CACM, Vol.37, No.7, pp.31-40.