JP5442543B2 - Content similarity calculation device and content similarity calculation method - Google Patents

Description

  The present invention relates to an inter-content similarity calculation apparatus and an inter-content similarity calculation method.

  2. Description of the Related Art Conventionally, an inter-content relevance calculation device using collaborative filtering on an EC (Electronic Commerce) site represented by Amazon is known as a device that calculates content related to a certain content. In addition, an inter-content relevance calculation device using content filtering that calculates relevance by comparing contents representing content such as metadata is also known (see Patent Document 1).

JP 2008-134941 A

T. Joachims, "Optimizing Search Engines Using Clickthrough Data", Proc. ACM SIGKDD International Conf. Knowledge Discovery and Data Mining (KDD02), ACM Press, pp. 132-142, 2002.

  In general, collaborative filtering has a problem that a user needs to evaluate each content and a large amount of browsing history or purchasing history. In addition, content filtering is calculated as only relevant content with high relevance, so there is a high possibility that a known one is searched for an area of interest to the user, and there are few new discoveries. There is a problem.

  In a content distribution system based on an open platform such as Android and Windows Mobile, a method for efficiently distributing content that a user potentially desires from a huge number of contents is required as new content is added every day.

  In particular, in the content of a mobile terminal, not only the functional similarity but also the similarity of the situation in which the content is used as the relevance between the contents. For example, the content “weather forecast” and the content “world weather” are considered to have similar functions. On the other hand, it is considered that the content of “route guidance” and the content of “gourmet guide” have similarities in the situation where the content is used. This is because there is a possibility of using “route guidance” in the same time zone and the same area in order to go to the restaurant found in “Gourmet Guide”.

  Therefore, the present invention has an object to calculate the similarity between contents in consideration of the similarity of the usage situation as a means for efficiently distributing the content that the user potentially seeks in the content distribution system. .

The inter-content similarity calculation apparatus of the present invention includes:
An inter-content similarity calculation device that calculates the similarity between contents,
A usage history receiver that receives usage history from the mobile device when the content is activated on the mobile device;
A usage history totaling unit that counts the usage history received by the usage history receiving unit into a preset usage status;
A usage status feature vectorization unit that generates a usage status feature vector for each content from the usage history tabulated by the usage history tabulation unit;
A similarity calculation unit that calculates the similarity between contents based on the feature vector of the usage situation;
It is characterized by having.

The content similarity calculation method according to the present invention includes:
An inter-content similarity calculation method in an inter-content similarity calculation device for calculating the similarity between contents,
Receiving usage history from the mobile device when the content starts on the mobile device,
Totalizing the usage history received by the usage history receiving unit into a preset usage status;
Generating a usage state feature vector for each content from the usage history tabulated by the usage history tabulation unit;
Calculating a similarity between contents based on a feature vector of usage status;
It is characterized by having.

  According to the embodiment of the present invention, it is possible to calculate the similarity between contents in consideration of the similarity of the usage situation.

1 is a configuration diagram of a communication system according to an embodiment of the present invention. Functional block diagram of portable terminal and search server according to an embodiment of the present invention The figure which shows the utilization log | history management table in the portable terminal which concerns on the Example of this invention. The figure which shows the utilization log | history management table in the search server based on the Example of this invention. The figure which shows the metadata management table in the search server which concerns on the Example of this invention. The figure which shows the feature-value management table in the search server based on the Example of this invention. The figure which shows the similarity management table in the search server which concerns on the Example of this invention. The flowchart of the similarity calculation method between contents in the search server which concerns on the Example of this invention. Flow chart showing the procedure for converting the usage status into a feature vector

  In the embodiment of the present invention, an inter-content similarity calculating apparatus that calculates the similarity between contents is used. The inter-content similarity calculation device may be a search server that can communicate with the mobile terminal.

  The inter-content similarity calculation device receives a usage history such as time information or position information when content is activated on a mobile terminal from the mobile terminal. The received usage history is aggregated in a preset usage status. For example, the usage history is aggregated for each area, for each usage time zone, or for each length of usage time. A feature vector of usage status is generated for each content from the totaled usage history. The similarity between contents is calculated based on a feature vector of usage status. The usage state feature vector is an n-dimensional vector representing the strength of connection between n usage states and contents set in advance.

  Also, a metadata feature vector may be generated, and the similarity between contents may be calculated based on the usage state feature vector and the metadata feature vector. The feature vector of metadata is an m-dimensional vector representing the strength of connection between m feature words extracted from the content metadata and the content.

  Examples of the present invention will be described in detail below.

<Configuration of communication system>
First, the overall configuration of the communication system according to the present embodiment will be described. As shown in FIG. 1, the communication system according to the embodiment of the present invention includes a mobile terminal 10, a search server 20, a content management server 30, and a mobile communication network 40.

  The search server 20 is an inter-content similarity calculation device that calculates the similarity between contents. The search server 20 receives a content search request from the mobile terminal 10, searches for content highly relevant to the content of the search request, and transmits the search result to the mobile terminal 10 via the mobile communication network 40.

  The content management server 30 is a device that stores and manages content distributed to the mobile terminal 10. The content is a service or information provided from the content management server 30 to the mobile terminal 10 and includes, for example, an application, data, video, music, a combination thereof, and the like.

  With reference to FIG. 2, the portable terminal 10 and the search server 20 will be described in more detail. The mobile terminal 10 includes a usage history transmission unit 101, a usage history storage unit 102, a usage history acquisition unit 103, a current location acquisition unit 104, and a search unit 105.

  The usage history acquisition unit 103 acquires usage history such as time information when the content is executed on the mobile terminal 10. For example, the name of the executed content, the start date / time, and the length of use time are acquired.

  The current location acquisition unit 104 acquires position information when the content is executed. For example, the latitude and longitude of the mobile terminal 10 are acquired.

  The usage history storage unit 102 stores the contents acquired from the usage history acquisition unit 103 and the current location acquisition unit 104.

  The usage history transmission unit 101 transmits usage history such as time information and position information stored in the usage history storage unit 102 to the search server 20.

  The search unit 105 transmits a search request to the search server 20 and receives a search result from the search server 20. The search request is a content ID, and the search result is a list of content highly relevant to the content of the search request.

  The search server 20 includes a usage history reception unit 201, an area information acquisition unit 202, a usage history totaling unit 203, a metadata storage unit 204, a feature word extraction unit 205, a metadata feature vectorization unit 206, a usage A history storage unit 207, a usage situation feature vectorization unit 208, a feature amount storage unit 209, a similarity calculation unit 210, a similarity storage unit 211, and a search unit 212 are configured.

  The usage history receiving unit 201 receives usage history such as time information and position information from the mobile terminal 10.

  The area information acquisition unit 202 acquires peripheral facility information such as a station and a shopping center from the latitude and longitude of the mobile terminal 10 in the usage history. The peripheral facility information may be acquired from a database (not shown) in which the peripheral facility information is associated with the latitude and longitude information in the search server 20. In addition, the peripheral facility information may be acquired from an external server.

  The usage history totaling unit 203 totals the contents acquired from the usage history receiving unit 201 and the area information acquisition unit 202 for each content according to usage conditions set in advance.

  The usage history storage unit 207 stores the contents totaled by the usage history totaling unit 203.

  The usage status feature vectorization unit 208 generates a usage status feature vector that represents the usage status of each content. As a result, the representative usage status of each content is converted into a feature vector.

  The metadata storage unit 204 stores metadata including the name, category, description, etc. of each content.

  The feature word extraction unit 205 extracts a characteristic word representing content from the explanatory text in the metadata.

  The metadata feature vectorization unit 206 generates a metadata feature vector that represents the function of each content based on the extracted feature words.

  The feature quantity storage unit 209 stores the usage status feature vector and the metadata feature vector of each content calculated by the usage status feature vectorization unit 208 and the metadata feature vectorization unit 206.

  The similarity calculation unit 210 calculates the usage situation similarity and the metadata similarity based on the contents stored in the feature amount storage unit 209, and further calculates the similarity between the contents.

  The similarity storage unit 211 stores the similarity calculated by the similarity calculation unit 210.

  The search unit 212 receives a search request from the mobile terminal 10 and transmits a content list having a high degree of similarity to the mobile terminal 10.

  In FIG. 2, the mobile terminal 10 transmits the usage history to the search server 20, but the usage history receiving unit 201 of the search server 20 may monitor the mobile terminal 10 and collect the usage history.

  FIG. 3 shows a usage history management table of the usage history storage unit 102. The usage history storage unit 102 stores, for each usage history of content executed on the mobile terminal 10, a content ID that uniquely identifies the content, a start date / time, a latitude, a longitude, and a length of usage time.

  FIG. 4 shows a usage history management table of the usage history storage unit 207. The usage history storage unit 207 stores, for each content ID, the usage status when the content is activated, and the activation count of the content corresponding to the usage status. Here, the usage status is expressed in the form of usage status ID. For example, each component (S01 to S06) of the usage status ID includes the time zone (morning, noon, evening) when the content was used, the length of usage time (within 5 minutes, 30 minutes or more, etc.), area information (station) , Restaurant, school, movie theater, etc.). For each component, 0 (not applicable) and 1 (applicable) are stored. For example, if content ID = C0001 is used at station (S06) at noon (S02), (S01, S02, S03, S04, S05, S06) = (0,1,0,0,0,1) Increase the activation count of the corresponding usage status by one. The content usage history is sorted in descending order of the number of times content is activated (or the number of downloads), and a record in which all components are 0 is inserted at the bottom of the sorted usage history of each content. The usage history may be aggregated based on either time information (utilization time period or length of utilization time) and position information (area information), and may be aggregated based on both time information and location information. May be.

  FIG. 5 shows a metadata management table of the metadata storage unit 204. The metadata storage unit 204 stores content metadata for each content ID. Content may be categorized by category.

  FIG. 6 shows a feature quantity management table of the feature quantity storage unit 209. In the feature quantity storage unit 209, when the number of extracted feature words and usage status ID elements is m and n, respectively, an m-dimensional vector (metadata feature vector) calculated by the metadata feature vectorization unit 206 is stored. ), An n-dimensional vector (usage status feature vector) calculated by the usage status feature vectorization unit 208, and a weight coefficient α for determining the weight of the two feature vectors when measuring the similarity between the contents. ing. For example, the metadata feature vector includes whether or not each of m feature words “route”, “station”, “izakaya”, “weather”, and “time” has appeared, the probability that the feature word will appear, and the feature Indicates the co-occurrence between words. The usage status feature vector represents the usage status of the content converted by SVM (Support Vector Machine) or the like.

  FIG. 7 shows a similarity management table of the similarity storage unit 211. The similarity storage unit 211 stores the cosine similarity (metadata similarity) of the metadata feature vector and the cosine similarity (usage status) of the usage feature vector for the combination of content IDs (content ID1, content ID2). Content) when there is a search request using the content ID 1 content as a search key and the content similarity calculated by weighting the metadata similarity and usage status similarity according to α The number of download requests indicating the number of times ID2 content has actually been downloaded is stored.

<Method for calculating similarity between contents>
Next, a method for calculating the similarity between contents in the search server will be described. FIG. 8 is a flowchart of the inter-content similarity calculation method in the search server 20 according to the embodiment of the present invention.

  First, in the search server 20, when new content is added to the metadata storage unit 204 (step S11), a feature vector of metadata is generated as advance preparation (step S18). Specifically, the feature word extraction unit 205 extracts words such as nouns from the metadata of each content by morphological analysis. Among the extracted nouns, general nouns having a low relationship with the feature of the content that can be noise are determined and excluded by threshold value processing of TFIDF (term frequency inverse document frequency). For example, if the stored metadata is the content of FIG. 5, words such as “route”, “station”, “izakaya”, “weather”, “time” are extracted as feature words, and “search”, “display” Such general words are expected to be excluded. Using the m feature words extracted here, the metadata of each content is converted into an m-dimensional vector. Here, as a method of determining each component of the vector, a method of substituting 1 (appears) or 0 (does not appear) depending on whether the feature word appears, appearance probability of feature words, and between feature words There is a method of assigning a weighted value in consideration of the co-occurrence degree. The calculated m-dimensional vector is stored in a column of metadata feature vectors in the feature amount management table of the feature amount storage unit 209 shown in FIG.

Then, the similarity calculation unit 210 calculates a cosine distance simA between m-dimensional vectors generated from the metadata for all combinations of contents (step S19). For example, the cosine distance between the vector v _i and the vector v _j is calculated by the following equation (1).

算出されたメタデータ特徴ベクトルのコサイン距離simA（メタデータ類似度）は、図７に示す類似度格納部２１１の類似度管理テーブルにおけるメタデータ類似度の列に記憶される。

The calculated cosine distance simA (metadata similarity) of the metadata feature vector is stored in the metadata similarity column in the similarity management table of the similarity storage unit 211 shown in FIG.

  On the other hand, in the mobile terminal 10, the usage history acquisition unit 103 monitors the activation status of the content, detects the activation or termination event of the content, and activates the content ID, activation date / time, usage time, latitude and longitude of the activated location. Record. Here, as a positioning method, there are a method using a GPS module built in a portable terminal and a method of acquiring position information from a base station. The usage history of the content stored in the usage history storage unit 102 as shown in FIG. 3 is transmitted to the usage history reception unit 201 of the search server 20 at a determined timing (step S12).

  Next, the usage history totaling unit 203 causes the area information acquisition unit 202 to search the usage history received by the usage history receiving unit 201 based on the latitude and longitude information, and around the location where the content is activated Request facility information at Further, the content usage history is totaled for each usage status ID including n elements set in advance, such as the time zone, the length of usage time, and area information (step S13). Here, 1/0 is assigned to each element of the usage status ID depending on whether or not it corresponds. For example, time zone (morning, noon, night) as an element, usage time (within 5 minutes, 5-29 minutes, 30-59 minutes, 60 minutes or more), area information (station, restaurant, movie theater, In the case of 11 convenience stores, the history of using content for 3 minutes in the convenience store in the station in the morning is the usage status ID (1,0,0,1,0,0,0,1,0,0) , 1). Thus, all the histories received from the mobile terminal 10 are given usage status IDs, and when the usage history corresponds to the usage status ID, the number of activations corresponding to the usage status ID is increased by one.

  The totaled usage history is converted into a feature vector (step S14). A specific procedure is shown in FIG. The tabulated usage histories are sorted in descending order of the number of activations for each content ID as shown in FIG. 4 and stored in the usage history storage unit 207 (step S20). Next, a record in which all components are 0 is inserted in the lowest part of each content (step S21). The sorted usage history is converted into one usage status feature vector representing each content by SVM (Support Vector Machine) (see Non-Patent Document 1) and stored in the feature amount storage unit 209 (S22). By sorting in order of the number of activations, the situation where the number of activations is large is reflected more strongly. Since the usage situation feature vector depends on the number of activations, the usage situation feature vector may be dynamically calculated based on the collected usage history.

  Then, the metadata similarity calculation unit 210 calculates a cosine distance simB between n-dimensional vectors generated from the usage status for all combinations of contents (step S15). This cosine distance simB is also calculated in the same manner as in equation (1). The calculated cosine distance simB (metadata similarity) of the usage status feature vector is stored in the usage status similarity column in the similarity management table of the similarity storage unit 211 shown in FIG.

  By the way, as shown in FIG. 7, the similarity storage unit 211 stores the number of times content ID2 has been downloaded (the number of download requests) when a related search is performed using the content with content ID1 as a search key. The weighting coefficient α for calculating the similarity between contents is calculated by the following equation (2) according to the number of download requests (step S16).

ただし、request(i)は、コンテンツiを検索キーとした検索リクエストがあった際に他のコンテンツのダウンロードがリクエストされたダウンロードリクエスト総数であり、request(i|simA>simB)は、ダウンロードリクエスト総数のうち、simA>simBを満たす組み合わせの数である。なお、α_iの初期値を0.5としているが、これはダウンロードリクエストが閾値未満である場合には、simAとsimBとを同等に扱うことを意味している。このように、１つのリクエストによってα_iが大きく変化しないようにリクエスト総数が閾値以上になるまではα_i=0.5とする。α_iの初期値は0.5以外の値でもよい。

However, request (i) is the total number of download requests that were requested to download other content when there was a search request with content i as the search key, and request (i | simA> simB) was the total number of download requests. Of these, the number of combinations satisfying simA> simB. Note that the initial value of α _i is 0.5, which means that simA and simB are treated equally when the download request is less than the threshold. In this way, α _i = 0.5 is set until the total number of requests is equal to or greater than the threshold value so that α _i does not change significantly by one request. The initial value of α _i may be a value other than 0.5.

  In the example of the content C0001 shown in FIG. 7, the contents of C0002, C0003, and C0004 are downloaded 50 times, 20 times, and 30 times, respectively. Of the total number 100 of download requests from the related content search for the content C0001, the number of requests satisfying simA> simB is 20, so α = 0.2, which is stored as the weighting factor α of the feature amount storage unit 209. α is a value calculated for each content, and the closer to 0, the closer the usage status is returned as a search result. Thus, α may be dynamically changed according to the number of download requests.

Next, for all combinations of content, the similarity calculation unit 210 calculates the inter-content similarity sim _{i, j} between the content i and the content j from simA, simB, α using the following equation (3). .

算出されたコンテンツ間類似度は、類似度格納部２１１に記憶される。図７ではC0001のコンテンツ間類似度は、α=0.5として算出された値である。上記のようにダウンロードリクエスト数に応じてα=0.2に更新された場合、C0001からC0002、C0003、C0004への各類似度は0.82、0.46、0.68となりC0002、C0004との結びつきが強くなる。つまり、C0001からの関連度検索は状況類似依存傾向となる。

The calculated similarity between contents is stored in the similarity storage unit 211. In FIG. 7, the similarity between contents of C0001 is a value calculated as α = 0.5. As described above, when α = 0.2 is updated according to the number of download requests, the similarities from C0001 to C0002, C0003, and C0004 are 0.82, 0.46, and 0.68, and the connection with C0002 and C0004 becomes strong. That is, the relevance degree search from C0001 has a situation similarity dependency tendency.

  When receiving the request from the mobile terminal 10, the search unit 212 of the search server 20 delivers a content list in which the content is rearranged in the descending order of similarity stored in the similarity storage unit 212 with respect to the received content ID. To do.

<Effect of Example>
According to the embodiment of the present invention, the similarity between contents can be calculated taking into account the similarity of the usage situation. For example, for a favorite content explicitly specified by the user, not only a function but also a content used in a similar usage situation can be presented, and a new discovery can be provided. In addition, it becomes possible to reach content that meets the user's potential purpose with few procedures, and the use of resources such as client terminals and network equipment can be reduced. In addition, whether the related content depending on the metadata or the usage status is preferred for each content is reflected in the search result, so that there is an effect of suppressing the appearance of inappropriate content.

  For example, although functions such as news reader and weather forecast, media player and book reader are different, content that is likely to be used in similar usage situations will be associated, and there is no clear image but search for useful content Search that matches the purpose.

  Furthermore, by converting the usage status and metadata into feature vectors, the user informs the server of the known content as a search key, so that the server resembles typical usage status and metadata for the content. A list of contents can be distributed.

  Furthermore, by dynamically changing α, if content with a high usage status similarity is downloaded with respect to the search key, the usage status similarity weight is increased, and content with a high metadata similarity is downloaded. If so, the weight of metadata similarity can be increased. For this reason, it is possible to deal with both contents having characteristics of usage and contents having functions.

  Furthermore, by sorting the usage history in descending order of the number of activations and generating a usage condition feature vector by SVM, it is possible to calculate a typical usage condition for content that strongly reflects the usage condition with a large number of activations.

  For convenience of explanation, the inter-content similarity calculation apparatus according to the embodiment of the present invention has been described using a functional block diagram. However, the inter-content similarity calculation apparatus of the present invention may be hardware, software, or their It may be realized in combination. For example, each function unit of the content similarity calculation device may be realized by software and may be realized in a computer. In addition, two or more embodiments and each component of the embodiments may be used in combination as necessary.

  As mentioned above, although the Example of this invention was described, this invention is not limited to said Example, A various change and application are possible within a claim.

DESCRIPTION OF SYMBOLS 10 Mobile terminal 101 Usage history transmission part 102 Usage history storage part 103 Usage history acquisition part 104 Present location acquisition part 105 Search part 20 Search server 201 Usage history reception part 202 Area information acquisition part 203 Usage history total part 204 Metadata storage part 205 Features Word extraction unit 206 Metadata feature vectorization unit 207 Usage history storage unit 208 Usage situation feature vectorization unit 209 Feature amount storage unit 210 Similarity calculation unit 211 Similarity degree storage unit 212 Search unit

Claims (8)

An inter-content similarity calculation device that calculates the similarity between contents,
A usage history receiver that receives usage history from the mobile device when the content is activated on the mobile device;
A usage history totaling unit that counts the usage history received by the usage history receiving unit into a preset usage status;
A usage status feature vectorization unit that generates a usage status feature vector for each content from the usage history tabulated by the usage history tabulation unit;
A similarity calculation unit that calculates the similarity between contents based on the feature vector of the usage situation;
An inter-content similarity calculation device. A metadata feature vectorization unit that generates a metadata feature vector for each content based on words included in the content metadata;
The similarity calculating unit calculates the similarity of the usage situation from the feature vector of the usage situation, calculates the similarity of the metadata from the feature vector of the metadata, and calculates the similarity of the usage situation and the similarity of the metadata. The content similarity calculation apparatus according to claim 1, wherein the similarity between contents is calculated by weighting.   The similarity calculation unit calculates a weighting coefficient used for weighting the similarity of usage status and the similarity of metadata based on the number of times content is downloaded based on the similarity between content. 2. The similarity calculation apparatus between contents described in 2.   The inter-content similarity calculation apparatus according to any one of claims 1 to 3, wherein the usage status feature vectorization unit generates a feature vector by rearranging usage statuses in descending order of activation counts for each content.   The usage history totaling unit counts the usage history received by the usage history receiving unit for each usage time zone or length of usage time when the content is activated on the mobile terminal. The content similarity calculation device according to claim 1. From the usage history received by the usage history receiving unit, further comprising an area information acquisition unit for acquiring an area when the content is activated on the mobile terminal,
The inter-content similarity calculation apparatus according to any one of claims 1 to 5, wherein the usage history totaling unit totals the usage history received by the usage history receiving unit for each area.   The content search unit according to any one of claims 1 to 6, further comprising a content search unit that generates a list in which content is rearranged in descending order of similarity when a content search request is received from a mobile terminal. Similarity calculation device. An inter-content similarity calculation method in an inter-content similarity calculation device for calculating the similarity between contents,
Receiving usage history from the mobile device when the content starts on the mobile device,
Totalizing the usage history received by the usage history receiving unit into a preset usage status;
Generating a usage state feature vector for each content from the usage history tabulated by the usage history tabulation unit;
Calculating a similarity between contents based on a feature vector of usage status;
The content similarity calculation method which has.

Images