JP2011108034A - Web page recommendation method using multiple attributes - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To recommend a Web page by extracting frequent access patterns of combinations of attributes of each Web page, focusing not on LCS of an access pattern of the Web page but on the attributes. <P>SOLUTION: A Web page has a plurality of attribute information for each page. A session is converted to an attribute vector sequence of the attribute information of each Web page from access histories of a plurality of users, and frequent Web page relationships of the attribute vector sequence is extracted. Attribute information of all Web pages are extracted as attribute vectors from an active session of a user who visits a Web site, each of the extracted attribute vectors is compared with a part of the frequent Web page relationships. When there is partial matching in the frequent Web page relationships, a Web page having an attribute vector of the part except the matched frequent Web page relationships is recommended. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a Web page recommendation method for guiding a user who visits a Web site to an appropriate page. In particular, the present invention focuses on a plurality of attributes of a Web page and extracts a combination pattern of the Web page. It relates to a method of making a recommendation.

In recent years, with the increase in the amount of information on websites, the choices of users of websites have increased, and there are many cases where users cannot reach the information they want to provide. For this reason, the importance of Web page recommendation for guiding a user who visits a Web site to an appropriate page is increasing.
For example, when searching for a restaurant on a restaurant information search site or the like, if there is a large amount of restaurant information posted on the site, information on the restaurant that the user who visited the site met his request It's not always easy to get to. For this reason, it is important to recommend a web page of a restaurant that is assumed to be desired by the visiting user as a candidate.

Web page recommendation methods can be broadly classified into a method of directly collecting and analyzing user evaluation information and a method of analyzing an access history remaining on a Web site. The former requires users to evaluate in order to collect information on how to evaluate the accessed Web pages, and requires effort from each user. It depends on each user whether or not it can be done.
On the other hand, the latter is introduced in comparison with the former because it is possible to obtain the same level of information from all users without requiring any special effort for users to analyze patterns in general access history. It can be said that there is little variation in evaluation.

  As a method for analyzing the latter access history, a user session is extracted from the access history, and each Web page accessed during the session is used as an item to apply a mining algorithm to extract a correlation rule and recommend it. A method (see Non-Patent Documents 2 and 3), a method of extracting and recommending frequent access patterns of Web pages, a method of extracting and recommending LCS (Longest Common Subsequence) of access patterns (see Patent Document 1), etc. are proposed. Has been.

The method of extracting the correlation rule has a problem that the access trend of the user cannot be accurately extracted because the access order is not considered. In addition, the method of extracting frequent access patterns of Web pages can only be recommended if they are exactly the same access patterns. Therefore, patterns with sufficient frequency are extracted especially when the number of Web pages increases and the types of generated patterns become enormous. Can not do it. It has been reported that the method of extracting the LCS of the access pattern is more effective than the other methods because it can extract the characteristics of the access order even when the access order does not completely match. (Refer nonpatent literature 6).
However, when the number of Web pages becomes enormous, the number of types of patterns increases even with the LCS extraction method, and patterns with sufficient frequency can be extracted as in the method of using the access pattern as it is. It becomes difficult.

  In the Web page recommendation method that applies the correlation rule mining method to the access history of the Web page described above, a Web page that frequently occurs with the Web page accessed so far is recommended for the Web page in the active session ( Non-patent document 2). However, there is a problem that when a new page is accessed, the recommendation cannot be made, and an improved method has been proposed (see Non-Patent Document 3), but the cost is high because the user needs to provide bookmark information. Furthermore, none of the methods considers the order of Web page access. In a Web site that narrows down Web pages in order as the method according to the present invention targets, it is necessary to consider the order.

In addition, for the purpose of product recommendation at sites selling books, a method using collaborative filtering (see Non-Patent Document 1) and a user preference model are created based on user evaluation history, and recommendations are made based on the model. A method of performing the above has also been proposed (see Non-Patent Document 5). However, in both methods, the user's evaluation needs to be registered in some form, so the cost is still high.
On the other hand, a model that quantifies the content and properties of content and predicts an evaluation value for each user has been proposed (see Non-Patent Document 4). How to quantify the content to apply to actual cases It has become a challenge.

  In a business that provides services on a website such as the above-mentioned service for presenting restaurant information, it is important to identify and recommend access trends of multiple users. It is useful to consider For this reason, a method of extracting association rules that do not consider the order is not suitable. However, if the frequency of the access pattern is used as it is, it can be recommended only in the same access order, so that what can be recommended is limited. For this reason, the method of extracting and using the LCS from the access sequence is more effective than the method of using the correlation rule or the method of using the frequent access pattern as it is.

  Of the partial sequence of sequence x and the partial sequence of sequence y, what is included in both sequences is called a common subsequence, but the longest common subsequence is the longest common subsequence. ) And abbreviated as “LCS”. For example, the LCS of x = A-F-B-D and y = A-B-C-D is “A-B-D”. By storing the LCS extracted from the list of Web pages of user sessions extracted from the access history, it is possible to extract a pattern through which many users pass except for access that has been detoured along the way. By searching for the LCS corresponding to the first half from the stored LCS in the order of access of the Web page of the active session to be recommended, and showing the second half of the Web page, many access patterns similar to the first half can be obtained. It is possible to recommend a Web page accessed by the user thereafter (see Non-Patent Document 6).

FIG. 1 is a diagram showing a flow of processing in a conventional method for extracting an LCS from a web page access pattern and recommending a web page. An outline of this method will be described below in correspondence with FIG.
[Step 1]
First, a user session is extracted by combining the Web page access history determined to be the same user based on the IP address information and the cookie information included in the user access history. Here, the user session is a sequence of Web page IDs. In FIG. 1, user sessions A-F-B-D are extracted by combining access histories of Web pages having the same IP address (IP1).
[Step 2]
Next, for all the extracted user sessions from user 1 to user n, LCSs included in the two arbitrary sessions are calculated and stored together with the frequency information. In the example of FIG. 1, from the user session sets A-F-B-D and A-B-C-D, its LCS A-B-D is replaced with another session set A-B-C. The LCS A-C-D and B-C-D are obtained from -D and B-A-C-D.
[Step 3]
A recommendation candidate is obtained by comparing the current active session to be recommended with the LCS obtained and accumulated above. In the example of FIG. 1, the user X's active session AB is pattern-matched with the first half of ABDD in the accumulated LCS, and a recommended candidate Web page D is obtained.

さらに、この抽出したＬＣＳを用いて、同一の飲食店サイトに対して２００９年８月２１日からサンプリングした１，０００アクティブセッションに対して推薦候補の抽出を試みたところ、実際に推薦候補を見つけることができたセッションは２６セッションにとどまった。これは、適用した飲食店サイトにおいては、推薦対象となるＷｅｂページの飲食店の数が非常に多く存在し、抽出したＬＣＳの数に対して、ユーザのアクセスしたパターンの数のほうがはるかに大きい為と言える。 In this method, high-precision recommendation is realized by emphasizing access patterns with high appearance frequency (see Non-Patent Document 6). However, when the number of Web pages to be recommended increases, an appropriate Web page cannot be recommended even if the LCS of the Web page access pattern is used. Here, as a result of sampling 1,000 sessions from the access history of an actual restaurant site on November 1, 2008 and extracting LCS, the length and type of LCS shown in Table 1 could be extracted. As can be seen, few LCSs had a sufficient length.

Furthermore, using this extracted LCS, we tried to extract recommendation candidates for 1,000 active sessions sampled from August 21, 2009 for the same restaurant site. Only 26 sessions were possible. This is because, in the applied restaurant site, there are a large number of restaurants on the Web page to be recommended, and the number of patterns accessed by the user is much larger than the number of extracted LCSs. That's right.

JP 2004-152209 A

Linden, G., B. Smith and J. York: Amazon.com recommendations: Item-to-item collaborative filtering ,, IEEE Internet Comput., Vol. 4, No. 1 (2003). Mobasher, B., Dai, H., Luo, T. and Nakagawa, M .: Effective personalization based on association rule discovery from Web usage data, Proc. 3rd Intl. Workshop on Web information and data management, pp. 9-15 (2001). Mobasher, B., Dai, H., Luo, T. and Nakagawa, M .: Using sequential and non-sequential patterns in predictive Web usage mining tasks, Proc.IEEE International Conference on Data Mining (ICDM'02), pp. 669-672 (2002). Hideki Aso, Tomohiro Ono, Yoichi Motomura, Shigeaki Kurokawa, Akihito Sakurai: On the integration of collaborative filtering and attribute-based filtering, IEICE Technical Report NC2006-54 (2006-10) (2006). Shigeaki Kurokawa, Tomohiro Ono, Yoichi Motomura, Hideki Aso, Akito Sakurai: Experimental evaluation of user preference model for movie content recommendation, IEICE Tech. Bulletin NC2004-182 (2005-03) (2004). Rie Yamamoto, Dai Kobayashi, Akihiro Yoshihara, Takashi Kobayashi, Haruo Yokota: Use and Analysis of LCS for Web Page Recommendation Based on Access Log, IPSJ Transactions Database No.SIG11 (TOD34), Vol.48 (2007).

  The present invention has been made in view of the circumstances as described above, and focuses on a plurality of attributes of each Web page, not the LCS of the access pattern of the Web page itself, and a combination of attributes between Web pages. It is an object of the present invention to provide a method for recommending Web pages by extracting a relationship in which patterns frequently appear (hereinafter referred to as “frequent Web page relationship”).

The present invention relates to a web page recommendation method by a programmed computer for analyzing an active session of a user who has visited a web site and recommending a web page to be accessed next to the user. The above object of the present invention is to
The web page has a plurality of attribute information for each page,
In the method, the computer
Extracting user sessions related to the web pages of the plurality of users from access histories of the plurality of users;
Extracting the plurality of attribute information for each Web page and converting the session for each user into an attribute vector sequence of the attribute information for each Web page;
Extracting a frequent web page relationship of the attribute vector sequence among the plurality of users;
With
The attribute information of all the Web pages included in the active session is extracted as an attribute vector from the active session of the user who visited the Web site, and the extracted attribute vector and a part of the relation between the frequent Web pages And recommending a web page having an attribute vector of a portion excluding the matched portion of the frequent web page relationship when the matching frequent web page relationship portion exists. This is achieved by a Web page recommendation method using multiple attributes.

  In addition, the object of the present invention is that the step of extracting the frequent Web page relationship of the attribute vector sequence as a frequent sequence is performed using an LCS algorithm, or the association rule mining using the frequent Web page relationship as a frequent set. It is effectively achieved by a Web page recommendation method using a plurality of attributes characterized by being performed using

According to the method of the present invention, even if there are a large number of Web pages in one Web site, a tendency common to many users is appropriately grasped based on attributes and used for recommendation. can do. There is an effect that the browsing page can be recommended more accurately from the browsing history in the site to the site visitor.
In addition, since an attribute is used, it is possible to target an active session including a new Web page or to recommend a new Web page.

It is a figure for demonstrating the conventional web page recommendation method using LCS of a web page access pattern. It is a figure for demonstrating the web page recommendation method (LCS method) which paid its attention to the multiple attribute which concerns on this invention. It is a flowchart which shows the flow of a process of the method (LCS method) based on this invention. It is a figure for demonstrating the web page recommendation method (association rule mining method) which paid its attention to the multiple attribute which concerns on this invention. It is a flowchart which shows the flow of a process of the method (association rule mining method) based on this invention. It is a figure which shows the relationship between the granularity of an attribute, and LCS extracted.

Hereinafter, a Web page recommendation method from an access history using a plurality of attributes, which is a method according to the present invention, will be described in detail with reference to the drawings.
The Web page recommendation method according to the present invention is performed by analyzing the access history of users remaining on the Web site, and is premised on storing the access history of Web pages (step 0). Since the method of accumulating the access history of the Web page is possible with the existing technology, it will not be described in detail here. This method is executed by a Web server (not shown) that operates a Web site in which a predetermined computer program is installed. Therefore, the following description will be made on the assumption that the access history of the Web page is accumulated.

Of the methods focusing on the plurality of attributes of the Web page according to the present invention, the flow of processing for extracting the relationship between frequent Web pages using the LCS algorithm will be described below in association with FIGS.
First, a user session is extracted from a user's access history in the same manner as in the conventional LCS extraction method for each Web page (step 1).
Next, the sequence of the Web pages included in the user session is converted into a sequence of the attributes of each Web page expressed as a vector (hereinafter referred to as “attribute vector”) (step 2). In the example of FIG. 2, page A has an attribute (α1, β2, γ1), page B has an attribute (α3, β1, γ2), and page D has an attribute (α3, β3, γ2). F has the attribute (α3, β2, γ1), and the user session A-F-BD is (α1, β2, γ1)-(α3, β2, γ1)-(α3, β1, γ2 )-(Α3, β3, γ2).
In addition, as a method of extracting the attribute which it has from a Web page, for example, it carries out by keyword search from the description content of a page, or embeds an attribute with meta information etc. at the stage of creating a page in advance. Conceivable.

  When all the elements of the attribute vectors are equal to the attribute vector sequence obtained above, it is determined that they are the same, and the LCS is extracted and stored (step 3). In the example of FIG. 2, a set of attribute vectors corresponding to user sessions A-F-B-D of user 1 and A-B-C-D of user 2 is (α1, β2, The LCS of the attribute vector γ1) − (α3, β1, γ2) − (α3, β3, γ2) is extracted.

  On the other hand, the Web page accessed by the user X is also converted into an attribute vector for the active session targeted for page recommendation for the user X (step 4). In the example of FIG. 2, it is assumed that page P has an attribute vector of (α1, β2, γ1) and page Q has (α3, β1, γ2).

The attribute vector sequence of the active session and the first half of the LCS of the attribute vector are compared, and pattern matching is performed (step 5), and the attribute vector of the second half of the LCS of the matching attribute vector is used as a recommendation candidate. In the example of FIG. 2, (α3, β3, γ2) is an attribute vector of recommendation candidates. Therefore, a Web page having the same attribute vector as the recommendation candidate attribute vector is recommended (step 6). In the example of FIG. 2, page R has an attribute vector of (α3, β3, γ2) and is recommended. In the method according to the present invention, the use of attribute vectors increases the possibility that a long LCS can be achieved by determining that a plurality of attributes match even if they are not the same Web page.
Further, in the recommendation using the LCS, it is possible to recommend even a Web page not included in the access history by using the attribute vector. In the example of FIG. 2, recommendation is possible even for pages that are not included in the access history, such as the web pages P and Q of the active session and the web page R to be recommended.
It should be noted that this is a specific method of page recommendation. If a user who viewed a page called P and a page called Q is to recommend a page called R next by the method according to the present invention, he / she sees it now. It is conceivable to include information about the R page in the Q page (for example, placing a clickable thumbnail of the R page). This can be realized by going from the P page to the Q page (clicking forward) to go to a page that includes the R page in Q. In addition, if cookies are used, information on the R page can be dynamically embedded in the Q page, so that a more flexible response is possible.

On the other hand, in the conventional recommendation by the LCS for each Web page shown in FIG. 1, both the Web page appearing in the active session and the Web page to be recommended must be included in the access history.
In general, it can be assumed that each Web page has a plurality of attributes. If only one attribute is used, recommendation is biased to that attribute, and there is a possibility that the user cannot be guided to an appropriate page.
Therefore, the method according to the present invention uses a plurality of attribute vectors. As the plurality of attributes, what kind of attribute is used and what granularity is used is important.

  Next, in the method focusing on a plurality of attributes of the web page according to the present invention, when the frequent web page relationship is considered as a frequent set having no order, the frequent web page relationship is extracted by association rule mining. The flow of processing will be described below in association with FIGS.

First, a user session is extracted from a user's access history in the same manner as the conventional frequent item set extraction method for each Web page (step 11).
Next, a set of Web pages included in the user session is converted into a set of attributes (hereinafter referred to as “attribute vectors”) that represent the attributes of each Web page (step 12). In the example of FIG. 4, page A has the attribute (α1, β2, γ1), page B has the attribute (α3, β1, γ2), and page D has the attribute (α3, β3, γ2). F has the attribute (α3, β2, γ1), and the user session A-FBD is represented by [(α1, β2, γ1), (α3, β2, γ1), (α3, β1, It is shown that it has been converted into a set of attribute vectors γ2), (α3, β3, γ2)].

  If all the elements of the attribute vectors are equal to the set of attribute vectors obtained above, they are determined to be the same, and a frequent set is extracted using correlation rule mining and stored (step 13). In the example of FIG. 4, from a set of attribute vector sequences corresponding to user sessions A-F-B-D of user 1 and A-B-C-D of user 2, [(α1, β2 , Γ1), (α3, β1, γ2), (α3, β3, γ2)] are extracted frequently.

  On the other hand, the Web page accessed by the user X is also converted into an attribute vector for the active session targeted for page recommendation for the user X (step 14). In the example of FIG. 4, it is assumed that page P has an attribute vector (α1, β2, γ1) and page Q has (α3, β1, γ2).

A set of attribute vectors of the active session is compared with a subset of the frequent set of attribute vectors, pattern matching is performed (step 15), and attribute vectors not included in the subset of the frequent set of matching attribute vectors are set as recommended candidates. To do. In the example of FIG. 4, (α3, β3, γ2) is the attribute vector of the recommendation candidate. Therefore, a Web page having the same attribute vector as the recommendation candidate attribute vector is recommended (step 16). In the example of FIG. 4, page R has an attribute vector of (α3, β3, γ2) and is recommended. In the method according to the present invention, it is possible to extract a frequent set having a large number of elements by using attribute vectors and determining that the attributes are the same if a plurality of attributes match even if they are not the same Web page. Get higher.
Furthermore, even in the recommendation using the frequent set, it is possible to recommend a Web page that is not included in the access history by using the attribute vector. In the example of FIG. 4, recommendation is possible even for pages that are not included in the access history, such as the web pages P and Q of the active session and the web page R to be recommended.

  Hereinafter, an example in which the above-described frequent web page relationship extraction is performed using the LCS algorithm will be described. First, consider the types of attributes, and then consider specific examples of the granularity.

[Attribute type]
Considering possible attribute candidates of each recommended Web page on a Web site where Web page recommendation is expected to be useful, the following can be listed.
(1) Restaurant web page: business type, average budget, area, number of reviews, private rooms, etc. (2) Ryokan and hotel web pages: Hotel / Ryokan type, accommodation, area, hot springs, etc. (3) Web pages of condominiums and rental properties: Rent, size, area, new / used, etc. (4) Web pages such as news and articles: Article types, article titles, dates, etc. (5) Products such as books and CDs Web page: Genre, author, price, year of publication, size, etc. (6) Web page of music on music download site, etc .: Genre, singer name, price, release year, etc.

Based on the observation of such attribute examples, the attributes are classified according to the following types.
[category]
“Business category (classification of Western, Japanese, Izakaya, etc.)” in restaurant recommendation and “Genre” in book / CD recommendation are representative, and mainly qualitative data. In many cases, this category is classified according to the search specification of the recommended Web site.
[range]
Average quantitative data for restaurant recommendation, accommodation fee for hotel recommendation, rent for apartment recommendation, etc. are representative quantitative data. In many cases, the user desires a certain range such as a lower limit, an upper limit, or both.
[distance]
It is an area for restaurant recommendation, an area for hotel recommendation, and an area for apartment recommendation. In addition to simple positional information, it may take the form of “minute from the station”.
[Evaluation]
The popularity ranking and the access number ranking in each recommendation. If too much emphasis is given, there is a possibility that the recommendation is biased to a specific Web page, and attention should be paid to handling.
[Additional information]
Information that only certain users feel merit about having, such as “Private rooms, parking availability, smoking / non-smoking classification” in restaurant recommendations, and “Separate bathroom” in apartment recommendations is there.
Of the attribute types roughly classified above, since “evaluation” and “additional information” need to be handled, in this embodiment, three attributes “category”, “range”, and “distance” are used. Subject to evaluation.

[Attribute granularity]
Each attribute of the Web page has a classification granularity (unit of subdivision). For example, the attribute of [Category] in the restaurant Web page corresponds to the business type of the restaurant, but it ranges from a large classification such as Japanese food, Western food, and Chinese to a fine classification such as kaiseki, cooking, sushi, and country food. [Distance] also has a wide range of areas in Tokyo, such as Shinjuku and Shibuya, to a slightly narrower level such as around Shinjuku West Exit / Tochomae, Shinjuku Sanchome / Shinjuku Gyoen, Shibuya Dogenzaka / Shinsen.
The average budget for meals, which is [Range], can be rounded to a granularity from 100 yen units to 500 yen units and 1000 yen units.

In this embodiment, the business type is classified into two types, “large business type” and “small business type”, and the area is classified into three types, “large area”, “medium area”, and “small area”. The state of the restaurant site as of August 2009 is classified into 12 categories as “large business category” and 127 categories as “small business category”. Similarly, “large area” is classified into 181 classifications, and “small area” is classified into 704 classifications.
Considering the relationship between the extracted LCSs when the attribute granularity is changed, generally, as shown in FIG. 6, the longer the granularity, the longer the obtained LCS and the more types of LCS.
In this comparison, the smallest granularity is in the case of Web page units that do not focus on attributes.

Ｗｅｂページ単位で抽出したＬＣＳに関しても比較のために示すが、どの属性を用いたとしても、得られるＬＣＳの種類は増加し、ＬＣＳ長も長くなっていることが分かる。
また、上で解析したように、粒度が粗いほどＬＣＳの種類が増え、長さが長くなっていることも分かる。ＬＣＳが長くなればなるほど、アクティブセッションと共通の属性を含む可能性が高くなるため、推薦できるアクティブセッションの割合も高くなり、推薦すべき属性を包含する可能性も高くなる。
また、得られるＬＣＳが多くなればなるほど、推薦すべき属性を包含する可能性も高くなるが、属性が粗くなるため、推薦すべきではないＷｅｂページを含む確率も上がる。これらは情報検索の分野における、適合率と再現率の関係と同じと考える。属性を粗くしすぎると、再現率は上昇するが、適合率は下がる可能性がある。つまり、両者はトレードオフの関係にあると言え、最も良い粒度の属性を調整する必要がある。 Here, for business data (category), area (distance), and average budget (range) for data obtained by sampling 1,000 sessions from the access history of the above-mentioned actual restaurant site on November 1, 2008, Table 2 shows the number of LCS, the number of types, the average LCS length, and the longest LCS length obtained when the particle size was changed.

The LCS extracted in units of Web pages is also shown for comparison, but it can be seen that the type of LCS obtained is increased and the LCS length is increased regardless of which attribute is used.
Also, as analyzed above, it can be seen that the coarser the particle size, the more types of LCS and the longer the length. The longer the LCS, the higher the possibility of including an attribute common to the active session. Therefore, the proportion of active sessions that can be recommended increases and the possibility of including the attribute to be recommended increases.
In addition, as the obtained LCS increases, the possibility of including recommended attributes increases. However, since the attributes become coarse, the probability of including Web pages that should not be recommended increases. These are considered to be the same as the relationship between precision and recall in the field of information retrieval. If the attribute is too rough, the recall rate will increase, but the precision rate may decrease. In other words, it can be said that they are in a trade-off relationship, and it is necessary to adjust the attribute with the best granularity.

  Thus, when extracting LCS using the some attribute of a web page, the number and length of LCS extracted also change with the combination of an attribute. Moreover, even if attention is paid to one attribute, it varies depending on the granularity to be classified. For this reason, in order to perform an appropriate Web page recommendation, it is important to use what attributes of the Web page at what granularity and how to combine them. In the embodiment of the present invention, the method according to the present invention is applied to the access history of an actual restaurant site, and compared with the conventional extraction of LCS in units of Web pages, and the selection of attributes and the granularity The effect was investigated. The results are described below.

  Web page recommendation based on a conventional method for extracting LCS in units of Web pages and a method for converting a user session into a multi-attribute vector according to the present invention is applied to the actual access history of the restaurant site. Then, an evaluation experiment was carried out to demonstrate the superiority of the method according to the present invention over the conventional recommendation for each Web page using another access history at different times as a test session. The results are as follows.

(1) Data to be evaluated As data to be evaluated, an access history for a request dated November 1, 2008 to the restaurant site described above was used. A user session was created by connecting the access histories using the cookie information included in the access history. Considering that user sessions that access only a small number of Web pages cannot be used for recommendation, among the created user sessions, 40, 312 sessions with 3 or more items accessed during the session were targeted. Since the main purpose is to obtain a granularity of good attributes for recommendation, and the execution time is limited, 1,000 sessions are randomly extracted from the above-mentioned sessions, and the brute force of the sessions is accessed and accessed. LCS was extracted on the page, and the session was converted to business category, average budget, and area attributes to extract LCS. Using the extracted LCS, recommendation was performed using the WRAPL-FL method described in Non-Patent Document 6. In this method, when recommending a certain active session, the LCS is first extracted, and then the Web page common to the extracted LCS and the active session is extracted, and the last of the common part is extracted from the LCS. Remove until.
For example, assuming that the active session is A-B and the LCS is A-C-B-A-D, the A-C-B-A part is removed from the LCS, and D as a recommendation candidate is obtained. In this way, scores corresponding to the appearance frequency are added to the Web pages remaining after removal, and when the score addition with all the LCSs is completed, the Web page with the highest score is set as a recommended candidate. Is. For evaluation, 1,000 sessions accessed on August 21, 2009 are set as a test set, the previous two accesses in the user session are set as active sessions, and Web pages actually accessed thereafter are a set of correct pages. Treated as.

(2) Evaluation and discussion on experimental results The experimental results are evaluated using the precision (Precision), recall (Recall), and F value (F-measure) defined below.
[Definition formula]
Precision = | Recom∩Eval | / | Recom | (∩: intersection)
Recall ＝ ｜ Recom∩Eval ｜ ／ ｜ Eval ｜
F-measure = 2 × Precision × Recall / (Precision + Recall)
Here, Recom and Eval represent a set of recommended pages derived from the target active session and a set of correct pages actually accessed following the target active session, respectively. The precision (Precision) represents the ratio of the number of correct pages to the recommended number of pages, and the recall (Recall) represents the ratio of the correct pages to the number of pages in the evaluation set. The F value is the harmonic average of precision and recall.

前述したように、推薦に対して良い属性の粒度を求めることを主眼として、実験回数を増やすためにＬＣＳ作成のためのセッションをランダムサンプリングとし、テストセットのセッションも少ないものを使ったことから、再現率、適合率とも高くはならなかったが、属性の粒度を変えたことによる違いは出た。
表３から分かるように、Ｗｅｂページ単位でのＬＣＳによって推薦を行うと、再現率・適合率共に最も低くなる。これは、ユーザセッションから求めたアクセスパターンに対して、アクティブセッションのユーザのアクセスパターンの方が多すぎた為と思われる。 In the method of extracting an LCS in units of Web pages and the method according to the present invention, two types of granularity of large and small regarding business conditions, three types of granularity of large, medium, and small regarding areas, an average budget of 100 yen unit, 500 yen unit, 1000 yen Table 3 shows the results of the precision, recall, and F value of the LCS extracted by changing the unit particle size.

As mentioned above, since the main purpose is to obtain a granularity of good attributes for recommendation, the session for creating the LCS is random sampling in order to increase the number of experiments, and the test set has few sessions. Both recall and precision did not increase, but there were differences due to changing the granularity of attributes.
As can be seen from Table 3, when the recommendation is performed by LCS in units of Web pages, both the recall rate and the matching rate are the lowest. This seems to be because there were too many access patterns for the users in the active session compared to the access patterns obtained from the user session.

On the other hand, with respect to the method of making a recommendation after converting to the attributes of business category, area, and average budget, the recall rate and the relevance rate were improved as compared to the recommendation for the Web page. This is considered to be because the use of a plurality of attributes has made it possible to deal with an access pattern that could not be dealt with by recommendation in Web page units. It also shows that it is possible to deal with a case where there is a new Web page at a different time.
In addition, for the method using a small business type, a small area, and an average budget in units of 100 yen, although the relevance rate was higher than that recommended in Web page units, there was almost no change in the recall rate.
This seems to indicate that the difference was difficult to be obtained by sampling and that there was not much difference in the granularity of the Web page. However, both the precision and recall increased as the grain size was further increased. This indicates that the influence of the granularity appears.
If the granularity is further reduced, the matching rate starts to decrease. However, as described above, it seems that Web pages that should not be recommended are included because the granularity is excessively coarsened. However, in the range of evaluation here, the F value increases even if the particle size is increased.

In addition, instead of the average budget, an experiment using the attributes of the number of seats along with the attributes of the business category and area was also performed (here, detailed results are omitted). The relevance rate / recall rate for the recommendation using LCS based on the number of seats could not show a sufficient advantage over the Web page unit even if the granularity of business category and area attribute was changed. This seems to be because the number of seats does not significantly affect the actual store selection.
From the above, it can be said that the selection of the attribute of the Web page to be recommended and the granularity of the selected attribute greatly affect the recommendation. In the above experiment, attributes of category, range, and distance type are considered. However, depending on the target Web page, the attributes discussed here do not always exist. If the selection of attributes and the granularity can be automatically adjusted, the recommendation accuracy should be further improved.

In addition, the Web pages recommended from the obtained attribute vectors were not particularly narrowed down this time, but in actual recommendation, if a large number of Web pages are displayed as recommendation candidates, the usability is reduced, so that some ranking It is necessary to narrow down by attaching.
For example, a method of considering a distance between attributes for a plurality of attributes or assigning a priority to a target attribute can be considered.
In addition, in the above experiment, the average budget was divided into several ranges, but for the user, even if there was a designation such as “the budget is 5,000 yen or less”, it should be “5,000 yen. I think there are few requests such as Thus, the range parameters can be clustered in advance, and the method according to the present invention can be applied in accordance with the clusters, so that more efficient recommendation can be performed.
Furthermore, in addition to the attributes as described above, it is also possible to make a recommendation using text data or data extracted from an image held by a Web page. For example, feature words such as “organic vegetables” and “home-like atmosphere” included in text data can be an important factor when a user searches a store.

Claims (8)

In a web page recommendation method by a programmed computer for analyzing an active session of a user who has visited a website and recommending the next web page to be accessed to the user,
The web page has a plurality of attribute information for each page,
In the method, the computer
Extracting user sessions related to the web pages of the plurality of users from access histories of the plurality of users;
Extracting the plurality of attribute information for each Web page and converting the session for each user into an attribute vector sequence of the attribute information for each Web page;
Extracting a frequent web page relationship of the attribute vector sequence among the plurality of users;
With
The attribute information of all the Web pages included in the active session is extracted as an attribute vector from the active session of the user who visited the Web site, and the extracted attribute vector and a part of the relation between the frequent Web pages And recommending a web page having an attribute vector of a portion excluding the matched portion of the frequent web page relationship when the matching frequent web page relationship portion exists. A web page recommendation method using a plurality of attributes.   2. The method for recommending a web page using multiple attributes according to claim 1, wherein the step of extracting the frequent web page relationship of the attribute vector sequence as a frequent sequence is performed using an LCS algorithm.   The Web page recommendation method using multiple attributes according to claim 1, wherein the step of extracting the frequent Web page relationship of the attribute vector sequence as a frequent set is performed using an association rule mining method.   The said recommendation is performed using the data extracted from the text data and image which the said Web page has in addition to the said attribute information as said attribute information. Web page recommendation method using multiple attributes.   The method for recommending a Web page using a plurality of attributes according to claim 1, further comprising a step of ranking the plurality of candidates and narrowing down candidates when there are a plurality of Web page candidates to be recommended. .   6. The Web page using multiple attributes according to claim 1, wherein recommendation information is dynamically embedded in a page accessed by the user using the cookie information of the user. Recommendation method. A computer program for analyzing an active session of a user who has visited a web site and causing the computer to recommend a web page to be accessed next to the user, wherein the web page is page by page Has multiple attribute information,
The computer program is stored in the computer.
Extracting user sessions related to the web pages of the plurality of users from access histories of the plurality of users;
Extracting the plurality of attribute information for each Web page and converting the session for each user into an attribute vector sequence of the attribute information for each Web page;
Extracting a frequent web page relationship of the attribute vector sequence among the plurality of users;
The attribute information of all the Web pages included in the active session is extracted as an attribute vector from the active session of the user who visited the Web site, and the extracted attribute vector and a part of the relation between the frequent Web pages And recommending a web page having an attribute vector of a portion excluding the matched portion of the frequent web page relationship when the matching frequent web page relationship portion exists. A computer program that is executed.   A Web server for providing the Web site, wherein the server is installed with the computer program according to claim 7.

Images