JP5153451B2 - Similar item search device, similar item search method, program thereof, and recording medium - Google Patents

Description

  The present invention relates to a similar item search apparatus, a similar item search method, a program thereof, and a recording medium that search a database for items whose attributes are similar to those of a content signal input by a user and output a list of items.

  Search technology (recommendation technology) for recommending items such as music, movies, books, etc. based on user preferences, features of items used by users in the past, and the like is widely used in shopping sites and the like. For the recommendation technology, similar item recommendation method using a method such as collaborative filtering using the user's purchase record or rating data as in Non-Patent Document 1, or each item as in Patent Document 1 A similar item search method that expresses a plurality of attributes (genre, musical instrument used, etc.) by numerical data such as vectors, and searches for item data (such as a title) of an item having a high similarity to the item attribute data input by the user from the database. In addition, there are similar item search methods such as Non-Patent Document 2, which use item content signals (acoustic signals themselves for music) without using purchase records, scoring, and attribute data. There is also a similar item search method that combines search by collaborative filtering as in Non-Patent Document 1, as in Non-Patent Document 3, and search based on content signals as in Non-Patent Document 2.

FIG. 13 is a diagram showing an outline of a similar item search method based on attribute data.
FIG. 13A shows an example of attribute data. When the item is a song, a genre such as classic or jazz, a mood such as sad or fun, an instrument used, etc. can be used as attributes. The attribute value is a numerical value indicating the degree to which the item includes each attribute. In FIG. 13A, the attribute value of the item x is (attribute 1, attribute 2, attribute 3) = (1, 0, 2). This vector is called an item attribute vector. In addition, a space having coordinates of each attribute of an item is called an attribute space.

FIG. 13B shows a method for calculating the similarity between items in the attribute space. For example, when a user selects an item x on a system such as a shopping site and outputs another item similar to the item x as a recommended item, the attribute represented as shown in FIG. Using this, the similarity with each of the items y, z, and w on the database is calculated. FIG. 13B shows an example using a similarity measure based on the Euclidean distance. Note that the attribute space in FIG. 13B is two-dimensional, but is merely an example, and the number of attributes is not limited to this.
FIG. 13C is a list of similar items output based on the similarity calculated in FIG. The system outputs a list in the order of similarity calculated in FIG.

FIG. 14 is a diagram showing an outline of a similar item search method based on a content signal.
FIG. 14A shows an example where the content signal is an acoustic signal such as music. Each item x ′, y ′, z ′, w ′ can be expressed as a vector by extracting features by processing such as frequency analysis. This vector is called an item feature vector. In addition, a space having the extracted features as coordinates is referred to as a feature space.
FIG. 14B shows an example in which two features are extracted from each item x ′, y ′, z ′, w ′ and plotted in the feature space. When another item similar to the item x ′ is output as a recommended item while the user selects the item x ′ on the system, as in the similar item search method based on the attribute data, The similarity between the item x ′ and the items y ′, z ′, and w ′ is calculated. Note that the feature space in FIG. 14B is two-dimensional, but is merely an example, and the number of features is not limited to this.
FIG. 14C is a list of similar items output based on the similarity calculated in FIG. The system outputs a list in the order of similarity calculated in FIG.

If the content signal is long-time series data, etc., one content signal can be divided into multiple segments, and feature extraction can be performed for each segment. In subsequent processing, each segment can be used instead of the content signal. It is. In this case, the average value of the similarities of all the combinations of all the segments of the item x ′ and all the segments of the item y ′ is regarded as the similarity between the items x ′ and y ′, or between the segments having the highest similarity The list shown in FIG. 14C can be created by using a method in which the similarity is regarded as the similarity between items x ′ and y ′.
US Pat. No. 7,003,315 Sarwar et al., “Analysis of Recommendation Algorithm for E-Commerce”, “Proceedings of the 2nd ACM conference on Electronic commerce”, 2000, pp. 158-167 B.Logan., “MUSIC RECOMMENDATION FROM SONG SETS”, “Proceedings of the 5th International Conference on Music Information Retrieval”, 2004, pages 425-428. Qing Li et al., “A Music Recommender Based on Audio Features”, “SIGIR'04”, 2004, pp. 532-533.

  However, since the similar item search method using collaborative filtering as described in Non-Patent Document 1 uses user purchase records and rating data for search, users who do not have such data cannot search for items. There is a problem. Moreover, since the similarity item search method based on attribute data like the above-mentioned patent document 1 calculates a similarity by the attribute data created in advance, there is a problem that an item having no attribute data cannot be recommended. is there. The similar item search method combining collaborative filtering and a method based on a content signal as described in Non-Patent Document 3 described above solves these problems, but items that do not have rating data by the user are rated data. There is a problem in that it is not possible to compare on the same scale when comparing with items having.

  The present invention has been made in view of the above points, and its object is to make it possible to compare and recommend items that do not have attribute data created in advance on the same scale as items that have attribute data. It is an object to provide an item search apparatus, a similar item search method, a program thereof, and a recording medium.

The present invention has been made in order to solve the above-described problems. An attribute important component estimation database and an item attribute vector representing an item attribute indicated by a learning content signal are used as axes representing attribute important components. Attribute / attribute important component conversion means for projecting and generating an item attribute important component vector having a dimensionality lower than that of the item attribute vector, and generating a learning feature vector having the features of the learning content signal as elements Using the feature extraction means, the item attribute important component vector p and the learning feature vector f, the attribute important component estimation parameter C used to estimate the attribute important component satisfying the relational expression represented by p = Cf + d. And d are obtained, and the attribute important component estimation parameter is stored in the attribute important component estimation database. Meter feature learning means, item database in which item and item attribute important component vector of the item are registered, and second feature extraction that extracts an item feature vector having the feature of the content signal as an element from the received content signal Means for substituting the attribute feature component estimation parameters C and d stored in the attribute feature component estimation database in the relational expression by the content signal that has received the input. Attribute important component estimation means for generating an item attribute important component vector of an item to be generated, an item attribute important component vector of an item indicated by the content signal that has received the input, and at least one item attribute important stored in the item database Ingredients Calculating a similarity between Le a similar item retrieval apparatus characterized by having a list creation means for outputting to create a list based on the value of the similarity.

Further, the present invention comprises a database registration unit for registering an item and an item attribute important component vector of the item in the item database, and the database registration unit, when the item is associated with an item attribute vector, When an item attribute important component vector obtained by projecting an item attribute vector onto an axis representing an attribute important component is registered in the item database, and the database registration unit is not associated with the item attribute vector The item attribute important component vector calculated by substituting the item feature vector f of the item and the attribute important component estimation parameters C and d stored in the attribute important component estimation database into the relational expression, Register in item database Characterized in that that.

Also, the attribute / attribute important component conversion means of the present invention is characterized in that, in an attribute space that is a space having coordinates of each attribute of an item, an axis with a large variance having a certain direction is an axis representing an attribute important component. And

  Further, the list creating means of the present invention inputs not only the content signal but also attribute data indicating the attribute, and using the attribute important component vector calculated from the attribute data, at least stored in the item database The similarity with one item attribute important component vector is calculated.

Further, the present invention is characterized in that in the second feature extracting means, the input content signal is an acoustic signal.

  Further, the present invention is characterized in that in the feature extraction means, each element of the output item feature vector is an acoustic power spectrum at each frequency that is equally spaced on the logarithmic frequency axis.

Further, the present invention, the attribute / attributes important component conversion means, the item attributes vector representing the attributes of the item indicated by the learning content signal, projected on the axis representing the attributes important component, than the item attribute vector generates a low item attributes important component vector of dimensionality, first feature extraction means, the characteristics of the learning content signal to generate a training feature vector whose elements, parameter learning unit, the item attributes important component vector P, and the learning feature vector f, and the attribute important component estimation parameters C and d used for estimating the attribute important component satisfying the relational expression p = Cf + d are obtained, and the attribute important accumulating parameter for component estimation attribute key component estimation database, the second feature extraction means, content signal, the input of which is accepted The characteristics of the content signal to extract the items feature vector whose elements from the attribute important component estimation means, the item feature vector is f, the item feature vector and the attribute are stored in the key component estimation database attributes By substituting the important component estimation parameters C and d into the relational expression , an item attribute important component vector of the item indicated by the content signal that has received the input is generated, and the list creation means receives the input The similarity between the item attribute important component vector of the item indicated by the content signal and at least one item attribute important component vector stored in the item database in which the item data and the item attribute important component vector of the item data are registered. Calculate the similarity A similar items search method characterized by creating and outputting a list based on the value.

The present invention is also a program for causing a computer to function as the above-described similar item search device.

  The present invention is also a computer-readable recording medium that records the above-described program.

  According to the present invention, attributes can be estimated from content, and attribute data can be assigned to items that do not have attribute data. Since the similarity between items is calculated using the assigned attribute data, it is possible to compare / recommend an item that does not have attribute data using the same scale as an item that has attribute data. Further, by extracting the important component of the attribute, it is possible to shorten the processing time and improve the accuracy.

(First embodiment)
Hereinafter, a first embodiment of the present invention will be described in detail with reference to the drawings. The first embodiment is an example in which a similar item is searched by inputting a content signal of an item and estimating an item attribute from an item feature vector representing the feature of the content signal.
In the present embodiment, an item is a concept that identifies an object, and indicates, for example, a song, a movie, a book, or the like. The item indicates a content signal that represents the content of the item, an attribute that represents characteristics unique to the content signal, such as an genre to which the item belongs or an impression generated when a person views or listens to the content signal, and unique information of the item Represented by item data. The relationship between an item, item data, and a content signal is, for example, when an item is a song, a person can associate the contents of the song by knowing the title (item data) of the song, By listening to the sound signal (content signal), that is, the content of the music, the title of the music can be associated.
The item data is represented by, for example, item title information, ID numbers, specifications, and the like.

  FIG. 1 is a schematic block diagram showing a configuration of a similar item search apparatus 1 according to the first embodiment of the present invention. Hereinafter, an example in which the content signal of an item is a music acoustic signal (hereinafter referred to as an item acoustic signal) will be described. The configuration of the apparatus is the same for other content signals such as video.

  The similar item search device 1 includes a recommendation list creation unit 110, a parameter learning unit (parameter learning unit) 120, an attribute estimation parameter storage unit (attribute estimation database) 130, and a database registration unit (attribute registration unit) 140. , A music database (item database) 150. The recommendation list creation unit 110 includes a feature extraction unit (feature extraction unit) 111, an attribute estimation unit (attribute estimation unit) 112, and a list creation unit (list creation unit) 113. The parameter learning unit 120 includes a feature extraction unit 121 and a parameter learning unit 122. The attribute estimation parameter storage unit 130 stores attribute estimation parameters used by the attribute estimation units 112 and 143. The database registration unit 140 includes an attribute registration unit 141, a feature extraction unit 142, and an attribute estimation unit 143. The music database 150 stores item attribute data (attribute data).

  In the similar item search device 1, the feature extraction unit 111 extracts the feature of the input item acoustic signal. The attribute estimation unit 112 estimates the attribute of the item using the feature extracted by the feature extraction unit 111 and the attribute estimation parameter stored in the attribute estimation parameter storage unit 130. The list creation unit 113 compares the attribute estimated by the attribute estimation unit with a plurality of item attribute data stored in the music database 150, and the item data indicates item data stored in the music database 150. A list arranged in the order of similarity between items and items indicated by the input item acoustic signal is created and output. The number of items to be output can be, for example, all songs whose similarity exceeds a preset threshold, the number set in advance, or the like.

First, the parameter learning unit 120 that registers the attribute estimation parameter in the attribute estimation parameter storage unit 130 will be described.
The parameter learning unit 120 calculates attribute estimation parameters used by the attribute estimation unit 112 of the recommendation list creation unit 110 and the attribute estimation unit 143 of the database registration unit 140. The parameter learning unit 120 receives learning music data including both item attribute data and item acoustic signals (learning content signals) as inputs. The feature extraction unit 121 reads an item acoustic signal, extracts features from the read item acoustic signal, and generates an item feature vector f (learning feature vector).

FIG. 2 is a diagram illustrating a concept of a method for calculating an item feature vector from an item acoustic signal. Here, an item feature vector calculation method will be described using an example in which a band filter bank including N band filters is used to calculate an item feature vector. In this case, the feature extraction unit 121 includes a band filter bank and a feature vector calculation unit. FIG. 2A shows an example of the input item acoustic signal. When the item acoustic signals (x ₁ , x ₂ ,...) Represented in FIG. 2 (A) are input to the band filter bank, the same sampling frequency as that of the item acoustic signal in FIG. N sample strings are output. The band-pass filter bank sets a frame (signal interval) of length W for each of the output N sample strings, and calculates the average value of the signal strength in the frame or the logarithm of the average value of the signal strength. . The band-pass filter bank skips the frame by a predetermined number of samples L, and calculates the average value of the signal strength in the frame or the logarithm of the average value of the signal strength until the end of the item acoustic signal is reached. The N number sequences thus obtained (for example, y ₁₁ , y ₁₂ ,... In the case of the first number sequence) are referred to as frame feature sequences. FIG. 2B shows a frame feature sequence. The feature vector calculation unit takes an average of each of the N frame feature sequences output by the band filter bank, and outputs an item feature vector f = (f ₁ , f ₂ ,... F _N ). In addition, the feature vector calculation unit sets the frame feature string output by the band filter bank to a norm of N values of each frame (for example, y ₁₁ , y ₂₁ ,..., Y _{N1 in the} first frame) as 1 or the like. You may normalize so that.

  The center frequency of the band-pass filter bank uses an arrangement (acoustic power spectrum) by a geometric series as shown in Equation (1) using the characteristics of the scale so that the center frequency is equally spaced on the logarithmic frequency axis. It is effective.

However, F _C (k) represents the center frequency of the k-th band filter, and F ₀ represents the lowest center frequency.
R represents a ratio between adjacent center frequencies, and a value exceeding 1, for example, 2 ^1/12 is set. For example, if an item's acoustic signal is composed of multiple sections and there is a large change in acoustic characteristics, for example, the item's acoustic signal is divided into multiple segments every 30 seconds, and each segment is the same as one item's acoustic signal. It is also possible to calculate item feature vectors.

  The parameter learning unit 122 receives the item feature vector f output from the feature extraction unit 121 and the item attribute data of the corresponding music data for learning. The parameter learning unit 122 generates an item attribute vector t based on the item attribute data, and calculates an attribute estimation parameter. As the item attribute vector conversion formula (attribute data calculation formula), for example, formula (2) can be used.

Here, t is an item attribute vector of n _t dimension and f is an item feature vector of n _f dimension. The matrix A is a matrix of n _t × n _f , and the vector b is a vector of n _t × 1. The i-th row components of A and b in Expression (2) are a plurality of item acoustic signals and item attribute data using the i-th element of the item attribute vector t as an objective variable and each element of the item feature vector f as an explanatory variable. It can be calculated by applying linear regression analysis to the combination. The calculated attribute estimation parameter is stored in the attribute estimation parameter storage unit 130.

Next, the database registration unit 140 that registers item attribute vectors and item data in the music database 140 used by the list creation unit 114 of the recommendation list creation unit 110 will be described.
The feature extraction unit 142 is the same as the feature extraction unit 121 of the parameter learning unit 120, receives an item acoustic signal, and outputs an item feature vector f. The attribute estimation unit 143 uses the item feature vector f generated by the feature extraction unit 142 and the attribute estimation parameter stored in the attribute estimation parameter storage unit 130 to calculate the item attribute vector t.

FIG. 3 is a diagram illustrating a concept of a method for calculating an item attribute vector from an item feature vector. FIG. 3A is an example in which items X and Y are plotted on a three-dimensional feature space represented by features a, b, and c. The n _f- dimensional item feature vectors representing the positions of the items X and Y are respectively represented as f _X and f _Y. In this example, n _f = 3. FIG. 3B is a plot of items X and Y on a three-dimensional attribute space. Here, the feature space and the attribute space have the same number of dimensions, but they need not be the same. Item X, respectively _{n t} dimensions of the item attribute vector representing a position in the attribute space of Y _t X, and _{t Y.} In this example, n _t = 3. FIG. 3C shows a calculation formula for converting from the feature space to the attribute space. The item feature vector is f, the item attribute vector is t, and a conversion formula t = T (f) is used to calculate f _X to t _X and f _Y to t _Y as estimated values.

  When the vector is converted using the equation (2) as the conversion equation, the matrix A and the vector b are stored in the attribute estimation parameter storage unit 130 (FIG. 1). Then, the attribute estimation unit 143 calculates the item attribute vector t by using the item feature vector f generated by the feature extraction unit 142 and the attribute estimation parameters A and b stored in the attribute estimation parameter storage unit 130. To do.

  The database registration unit 140 accepts both an item having no item attribute data and an item having item attribute data as input. In the case of an item having no item attribute data, the feature extraction unit 142 receives an item acoustic signal and outputs an item feature vector f. The attribute estimation unit 143 receives the item feature vector f output from the feature extraction unit 142 and outputs an item attribute vector t. The item attribute vector output by the attribute estimation unit 143 is stored in the music database 150. In the case of an item having item attribute data, the item attribute data is input to the attribute registration unit 141, and the attribute registration unit 141 creates an item attribute vector from the input item attribute data, and the item of the item indicated by the item attribute vector It is registered in the music database 150 in association with the data.

  Here, the user searches for an item similar to the item indicated by the item acoustic signal input using the similar item search device 1. In the search, the item acoustic signal that the user inputs as a search question is input to the recommendation list creation unit 110. The recommendation list creation unit 110 receives the input item sound signal. This item acoustic signal corresponds to the item x in FIGS. 14A and 14B, and represents the content signal of the item selected by the user in a system such as a shopping site. The recommendation list creation unit 110 extracts an item having a high similarity to the item indicated by the item acoustic signal input from the items indicated by the item attribute vector stored in the music database 150, and information such as the title ( Item data) is output as a music list. The number of songs to be output can be, for example, all songs whose similarity exceeds a preset threshold value, a preset number, and the like.

FIG. 4 is a flowchart showing the process of the recommendation list creation unit 110.
With the above process, the attribute estimation parameter is stored in the attribute estimation parameter storage unit 130, and the following process is started with at least one item attribute data stored in the music database 150.
First, the feature extraction unit 111 reads an item acoustic signal (content signal) (step S1). Next, the feature extraction unit 111 extracts features from the read item acoustic signal and generates an item feature vector f (step S2). Next, the attribute estimating unit 112 uses the item feature vector f by the feature extraction unit 111 to generate a parameter for the attribute estimation attribute estimation parameter storage unit 130 stores, calculates the item attribute vector t _in ( Step S3).

Next, the list creation unit 113 searches the song database 150 for item attribute vectors for which no read completion flag is set. (Step S4). If the item attribute vector read completion flag is not set in the music database 150, the list creation unit 113 reads the item attribute vector t _d corresponding the music database 150, a read completion flag in the read item attribute vector t _d Stand up (step S5). List creation unit 113 calculates the item attribute vector _{t d} read from the music database 150, the similarity between the item attribute vector _{t in} the attribute estimating unit 112 is calculated (step S6).

As the similarity S (t _in , t _d ) between the item attribute vectors t _in and t _d , an inner product on the attribute space, a cosine, a Gaussian radial basis function such as Equation (3), or the like can be used. .

  Here, r is a parameter set in advance. The list creation unit 113 performs the processing from step S4 onward until all item attribute vectors are read from the music database 150. When all the item attribute vectors are read, the list creation unit 113 outputs a list of item data of items with high similarity on the music database 150 (step S7). When the item sound signal is divided into a plurality of short-time segments in the feature extraction unit 111, the similarity is calculated for each segment, and the item data of the item to which the segment having a high similarity belongs is output as a list.

  Thus, according to the first embodiment, an item attribute vector can be given to an item that does not have item attribute data. Thereby, the item which does not have item attribute data can be compared and recommended using the same scale as the item which has item attribute data.

(Second Embodiment)
Hereinafter, the second embodiment of the present invention will be described in detail. The second embodiment is an example in which a lower-order item attribute important component is extracted from an item attribute vector and a similar item is searched.

FIG. 5 is a schematic block diagram showing the configuration of the similar item search device 1 according to the second embodiment of the present invention.
Hereinafter, as in the first embodiment, an example in which the content signal of an item is a music acoustic signal (hereinafter referred to as an item acoustic signal) will be described.
The similar item search device 1 according to the second embodiment includes an attribute estimation unit 112, a parameter learning unit 122, an attribute estimation parameter storage unit 130, an attribute registration unit 141, and an attribute estimation in the similar item search device 1 according to the first embodiment. The unit 143 includes an attribute important component estimation unit (attribute important component estimation means) 212, an attribute important component estimation parameter learning unit 222, an attribute important component estimation parameter storage unit (attribute important component estimation database) 230, and an attribute important component. The registration unit (attribute important component registration unit) 241 and the attribute important component estimation unit 243 are changed, the parameter learning unit 120 includes the attribute / attribute important component conversion unit 223, and the database registration unit 140 includes the attribute / attribute important component conversion unit. 244 is provided.
The other processing units are the same as those in the first embodiment, and will be described using the same reference numerals.

  The parameter learning unit 120 calculates an attribute important component estimation parameter used by the attribute important component estimation unit 212 of the recommendation list creation unit 110 and the attribute important component estimation unit 243 of the database registration unit 140. As in the first embodiment, the parameter learning unit 120 receives music data for learning including both item attribute data and item acoustic signals as input. The feature extraction unit 121 receives an item acoustic signal and outputs an item feature vector f. Since the feature extraction unit 121 is the same as the feature extraction unit 121 according to the first embodiment, the description thereof is omitted. The attribute / attribute important component conversion unit 223 receives item attribute data. The attribute / attribute important component conversion unit 223 generates an item attribute vector t based on the item attribute data, and calculates an item attribute important component vector p from the generated item attribute vector t.

FIG. 6 is a diagram illustrating a concept of a method for calculating an item attribute important component vector from an item attribute vector. FIG. 6A shows an example in which the item attribute vector t is plotted on a three-dimensional attribute space. The item attribute important component is an important component in the search calculated from the item attributes. For example, the axis with a large variance having a direction in the attribute space (two thick arrows in FIG. 6A). It can be calculated by projecting to FIG. 6B shows an example in which an item in the attribute space is projected onto a two-dimensional attribute important component space. The direction of the axis for calculating the item attribute important component vector p is obtained in advance by a method such as principal component analysis using a number of items having item attribute data as learning data. When using principal component analysis, conversion formula from n _t dimensions of the item attribute vector t to n _{p-dimensional} item attributes important component vector p (attribute important component data calculation formula) is expressed by equation (4) The

However, the matrix E is an n _p × _nt matrix. The vector g is an n _p × 1 vector. The parameter of equation (4) is stored in the attribute / attribute important component conversion unit 223. Then, the attribute / attribute important component conversion unit 223 calculates the item attribute important component vector p by using the item attribute vector t calculated based on the item attribute data and the parameters E and g.

  The attribute important component estimation parameter learning unit 222 receives the item feature vector f output from the feature extraction unit 121 and the item attribute important component vector p output from the attribute / attribute important component conversion unit 223, and estimates the item attribute important component. Calculate the parameters. As the item attribute important component vector conversion formula, for example, Formula (5) can be used.

Here, p is an _np- dimensional item attribute important component vector, and f is an _nf- dimensional item feature vector. The matrix C is an n _p × n _f matrix, and the vector d is an n _p × 1 vector. The i-th row component of C and d in Equation (5) can be calculated by applying linear regression analysis using the i-th element of p as an objective variable and each element of f as an explanatory variable. The calculated attribute important component estimation parameter is stored in the attribute important component estimation parameter storage unit 230.

  The database registration unit 140 registers an item attribute important component vector and item data in the music database 150. The attribute / attribute important component conversion unit 244 is the same as the attribute / attribute important component conversion unit 223 of the parameter learning unit 120, inputs item attribute data, and outputs an item attribute important component vector p. The feature extraction unit 142 is the same as the feature extraction unit 121 of the recommendation list creation unit 120, and receives an item acoustic signal and outputs an item feature vector f. The attribute important component estimation unit 243 receives the item feature vector f output from the feature extraction unit 142 and the attribute important component estimation parameter stored in the attribute important component estimation parameter storage unit 130, and the item attribute important component vector p is output.

FIG. 7 is a diagram illustrating a concept of a method for calculating an item attribute important component vector from an item feature vector. FIG. 7A is an example in which items X and Y are plotted on a three-dimensional feature space represented by features a, b, and c. The n _f- dimensional item feature vectors representing the positions of the items X and Y are respectively represented as f _X and f _Y. In this example, n _f = 3. FIG. 7B is an example in which items X and Y are plotted on a two-dimensional attribute important component space. The n _p- dimensional item attribute vectors representing the positions of the items X and Y in the attribute important component space are denoted as p _X and p _Y , respectively. In this example, n _p = 2. FIG. 7C shows a calculation formula for converting from the feature space to the attribute important component space. The item feature vector is f, the item attribute important component vector is p, and a conversion formula p = P (f) is used to calculate f _X to p _X and f _Y to p _Y as estimated values.

  When the vector is converted using the equation (5) as the conversion equation, the matrix C and the vector d are stored in the attribute important component estimation parameter storage unit 230 (FIG. 5). Then, the attribute important component estimation unit 243 uses the item feature vector f generated by the feature extraction unit 142 and the attribute important component estimation parameters C and d stored in the attribute important component estimation parameter storage unit 230, The item attribute important component vector p is calculated.

  The database registration unit 140 accepts both an item having no item attribute data and an item having item attribute data as input. In the case of an item having no item attribute data, the feature extraction unit 142 receives an item acoustic signal and outputs an item feature vector f. The attribute important component estimation unit 243 receives the item feature vector f output from the feature extraction unit 142 and outputs an item attribute important component vector p. The output item attribute important component vector p is stored in the music database 150. In the case of an item having item attribute data, the attribute / attribute important component conversion unit 244 inputs the item attribute data and outputs the item attribute important component vector p. The attribute important component registration unit 241 registers the item attribute important vector p output from the attribute / attribute important component conversion unit 244 in the music database 150.

FIG. 8 is a flowchart showing processing of the recommendation list creation unit 110 according to the second embodiment.
First, the feature extraction unit 111 reads an item acoustic signal (content signal) (step S101). Next, the feature extraction unit 111 extracts features from the read item acoustic signal and generates an item feature vector f (step S102). Next, the attribute important component estimation unit 212 uses the item feature vector f generated by the feature extraction unit 111 and the attribute important component estimation parameter stored in the attribute important component estimation parameter storage unit 230 to calculate the item attribute important calculating a component vector _{p in} (step S103).

Next, the list creation unit 113 searches the item database 150 for item attribute important component vectors for which no read completion flag is set (step S104). If there is an item attribute important component vector read completion flag is not set in the music database 150, the list creation unit 113 reads the item attribute important component vector p _d from the music database 150, the read item attributes important component vector p _d A read completion flag is set (step S105). List creation unit 113 calculates the similarity of the item attributes important component vector _{p d} read from the music database 150, an attribute item attribute important component estimator 212 calculates important component vector _{p in} (step S106).

Similarity S ^p (p _in , p _d ) between the item attribute important component vectors p _in and p _d is similar to the similarity of the item attribute component in the first embodiment, the inner product on the attribute space, cosine, Also, a Gaussian radial basis function such as the equation (6) can be used.

Here, r _p is a parameter set in advance. The list creation unit 113 performs the processing from step S104 onward until all item attribute important component vectors are read from the music database 150. When all the item attribute important component vectors are read, the list creation unit 113 outputs a list of item data of items with high similarity on the music database 150 (step S107). When the item sound signal is divided into a plurality of short-time segments in the feature extraction unit 111, the similarity is calculated for each segment, and the item data of the item to which the segment having a high similarity belongs is output as a list.

  By using the second embodiment, the item attribute data is converted into the item attribute important component for the item having the item attribute data, and the item attribute important component is changed for the item having no item attribute data. In order to estimate, an item having no item attribute data and an item having item attribute data can be compared and recommended using the same scale. In the first embodiment, since all dimension values in the item attribute vector t are calculated and the similarity is calculated, when the number of dimensions in the attribute space is large, the calculation time is increased and the similarity is calculated. In the second embodiment, the item attribute important component vector p having a dimension lower than that of the item attribute vector t can be used to reduce the calculation time and improve the accuracy. it can.

  In the second embodiment, the case where the attribute / attribute important component conversion units 223 and 244 store parameters used for conversion has been described, but the present invention is not limited to this. For example, the attribute important component estimation parameter storage unit 230 stores the parameters used by the attribute / attribute important component conversion units 223 and 244 for conversion, and reads them when the attribute / attribute important component conversion units 223 and 244 perform calculation. It may be.

(Third embodiment)
Hereinafter, the third embodiment of the present invention will be described in detail. The third embodiment is an example in which the recommendation list creation unit accepts item attribute data.

  FIG. 9 is a schematic block diagram showing the configuration of the similar item search device 1 according to the third embodiment of the present invention. Hereinafter, as in the first embodiment, an example in which the content signal of an item is a music acoustic signal (hereinafter referred to as an item acoustic signal) will be described. The similar item search device 1 according to the third embodiment is different from the similar item search device 1 according to the first embodiment in input and processing to the list creation unit 113. Since the other processing units are the same, description will be made using the same reference numerals.

FIG. 10 is a flowchart showing the process of the recommendation list creation unit 110 according to the third embodiment. First, the recommendation list creation unit 110 determines whether the data type of the input data is item attribute data or an item sound signal (content signal), and sets an input completed flag for the data type to be processed (Step S110). S201). If the data type of the input data is the item attribute data list creating unit 113 reads the item attribute data, to create an item attribute vector t _in (step S202). The item attribute vector t, and is equivalent to the item attribute vector t _in the attribute estimating unit 112 according to the first embodiment is output.
If the data type of the input data is an item acoustic signal, the feature extraction unit 111 reads the item acoustic signal (content signal) (step S203). Next, the feature extraction unit 111 extracts features from the read item acoustic signal and generates an item feature vector f (step S204). Next, the attribute estimating unit 112 uses the item feature vector f by the feature extraction unit 111 to generate a parameter for the attribute estimation attribute estimation parameter storage unit 130 stores, calculates the item attribute vector t _in ( Step S205).

Next, the list creating unit 113 searches the item database 150 for an item attribute vector for which no read completion flag is set (step S206). If the item attribute vector read completion flag is not set in the music database 150, the list creation unit 113 reads the item attribute vector t _d corresponding the music database 150, a read completion flag in the read item attribute vector t _d Stand up (step S207). List creation unit 113 calculates the item attribute vector _{t d} read from the music database 150, the similarity between the item attribute vector _{t in} the attribute estimating unit 112 is calculated (step S208). The list creation unit 113 performs the processing from step S206 onward until all item attribute vectors are read from the music database 150. When all the item attribute vectors are read, the list creation unit 113 creates a list of item data of items with high similarity on the music database 150 (step S209). When the item sound signal is divided into a plurality of short-time segments in the feature extraction unit 111, similar to the first embodiment, the similarity is calculated for each segment, and the item to which the segment with the higher similarity belongs is calculated. Output item data as a list.

  Next, for the input data, it is confirmed whether there is a data type for which an input flag is not set (step S210). If there is a data type for which the input completion flag is not set, the processing from step S201 is performed on the corresponding data type. When the processing for all data types of the input data is completed, the list creation unit 113 creates and outputs an integrated list such as arranging all the lists in order of similarity (step S211).

  By using the third embodiment, an item that does not have item attribute data can be compared and recommended using the same scale as an item that has item attribute data. Further, for items having item attribute data and item acoustic signals, similar item search is performed with higher accuracy than in the first embodiment by further calculating similarity based on the input item attribute data. be able to.

(Fourth embodiment)
Hereinafter, the fourth embodiment of the present invention will be described in detail. The fourth embodiment is an example in which the recommendation list creation unit accepts item attribute data and extracts an item attribute important component vector.

  FIG. 11 is a schematic block diagram showing the configuration of the similar item search device 1 according to the fourth embodiment of the present invention. Hereinafter, as in the first embodiment, an example in which the content signal of an item is a music acoustic signal (hereinafter referred to as an item acoustic signal) will be described. Unlike the similar item search apparatus 1 according to the second embodiment, the similar item search apparatus 1 according to the fourth embodiment is different from the similar item search apparatus 1 according to the second embodiment in terms of input and processing. 414. Since the other processing units are the same, description will be made using the same reference numerals.

FIG. 12 is a flowchart showing the process of the recommendation list creation unit 110. First, the recommendation list creation unit 110 determines whether the data type of the input data is item attribute data or an item sound signal (content signal), and sets an input completed flag for the data type to be processed (Step S110). S301). If the data type of the input data is the item attribute data, the attribute / attributes important component conversion unit 114 reads the item attribute data, to create an item attribute vector t _in (step S302). Attribute / attribute important component conversion section 414, from the item attribute vector _{t in,} it calculates the item attributes important component vector _{p in} (step S303). List creation unit 113, attribute / attributes important item attribute component conversion unit 414 has output key component vector p _in, equivalent to item attribute important component vector p _in the attribute important component estimator 212 according to the second embodiment outputs Suppose that
If the data type of the input data is an item acoustic signal, the feature extraction unit 111 reads the item acoustic signal (content signal) (step S304). Next, the feature extraction unit 111 extracts features from the read item acoustic signal, and generates an item feature vector f (step S305). Next, the attribute important component estimation unit 212 uses the item feature vector f generated by the feature extraction unit 111 and the attribute important component estimation parameter stored in the attribute important component estimation parameter storage unit 230 to calculate the item attribute important calculating a component vector _{p in} (step S306).

Next, the list creation unit 113 searches the song database 150 for an item attribute important component vector for which no read completion flag is set (step S307). If the item attributes important vector read completion flag is not set in the music database 150, the list creation unit 113 reads the item attribute important component vector p _d corresponding the music database 150, the read item attributes important component vector p _d A read completion flag is set at (Step S308). List creation unit 113 calculates the similarity of the item attributes important component vector _{p d} read from the music database 150, an attribute item attribute important component estimator 212 calculates important component vector _{p in} (step S309). The list creation unit 113 performs the processing from step S307 onward until all item attribute important component vectors are read from the music database 150.
When all the item attribute important component vectors have been read, the list creation unit 113 creates a list of item data of items with high similarity on the music database 150 (step S310). When the item sound signal is divided into a plurality of short-time segments in the feature extraction unit 111, similar to the second embodiment, the similarity is calculated for each segment, and the item to which the segment having a high similarity belongs is calculated. Output item data as a list.

  Next, it is confirmed whether or not there is a data type for which the input completion flag is not set for the input data (step S311). If there is a data type for which the input completion flag is not set, the processing after step S1 is performed for the corresponding data type. When the processing for all the data types of the input data is completed, the list creation unit 413 creates and outputs an integrated list such as arranging all the lists in order of similarity (step S312).

  By using the item attribute important component vector p by using the fourth embodiment, calculation time can be shortened and accuracy can be improved. In addition, for an item having item attribute data and an item acoustic signal, the similarity calculation based on the item attribute important component vector obtained from the input item attribute data is further performed, thereby making it more than the second embodiment. A similar item search with high accuracy can be performed.

  As described above, the embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to the above, and various design changes and the like can be made without departing from the scope of the present invention. It is possible to

The similar item search apparatus described above has a computer system therein. Each process described above is stored in a computer-readable recording medium in the form of a program, and the above process is performed by the computer reading and executing the program. Here, the computer-readable recording medium means a magnetic disk, a magneto-optical disk, a CD-ROM, a DVD-ROM, a semiconductor memory, or the like.
Alternatively, the computer program may be distributed to the computer via a communication line, and the computer that has received the distribution may execute the program.

  The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

It is a schematic block diagram which shows the structure of the similar item search device by the 1st Embodiment of this invention. It is a figure showing the concept of the method of calculating an item feature vector from an item acoustic signal. It is a figure showing the concept of the method of calculating an item attribute vector from an item feature vector. It is a flowchart showing the process of the recommendation list preparation part by the 1st Embodiment of this invention. It is a schematic block diagram which shows the structure of the similar item search device by the 2nd Embodiment of this invention. It is a figure showing the concept of the method of calculating an item attribute important component vector from an item attribute vector. It is a figure showing the concept of the method of calculating an item attribute important component vector from an item feature vector. It is a flowchart showing the process of the recommendation list preparation part by the 2nd Embodiment of this invention. It is a schematic block diagram which shows the structure of the similar item search device by the 3rd Embodiment of this invention. It is a flowchart showing the process of the recommendation list preparation part by the 3rd Embodiment of this invention. It is a schematic block diagram which shows the structure of the similar item search device by the 4th Embodiment of this invention. It is a flowchart showing the process of the recommendation list preparation part by the 4th Embodiment of this invention. It is a figure which shows the outline | summary of the similar item search method based on item attribute data. It is a figure which shows the outline | summary of the similar item search method based on a content signal.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Similar item search device 110 ... Recommendation list creation part 111 ... Feature extraction part 112 ... Attribute estimation part 113 ... List creation part 120 ... Parameter learning part 121 ... Feature extraction part 122 ... Parameter learning part 130 ... Parameter accumulation part for attribute estimation 140 ... Database registration unit 141 ... Attribute registration unit 142 ... Feature extraction unit 143 ... Attribute estimation unit 150 ... Music database 212 ... Attribute important component estimation unit 222 ... Attribute important component estimation parameter learning unit 223 ... Attribute / attribute important component conversion unit 230 ... Attribute important component estimation parameter storage unit 241 ... Attribute important component registration unit 243 ... Attribute important component estimation unit 244 ... Attribute / attribute important component conversion unit 414 ... Attribute / attribute important component conversion unit

Claims (9)

Attribute important component estimation database;
The item attribute vector representing the attributes of the item indicated by the learning content signal, attribute projected to the axis representing the critical components, attribute / attributes to generate a lower item attributes important component vector of dimensionality than the item attribute vector Important component conversion means,
A first feature extraction means for generating a training feature vectors characterizing elements of the learning content signal,
Using the item attribute important component vector p and the learning feature vector f , obtain attribute important component estimation parameters C and d used to estimate the attribute important component that satisfy the relational expression p = Cf + d Parameter learning means for storing the attribute important component estimation parameter in the attribute important component estimation database;
An item database in which items and item attribute important component vectors of the items are registered;
Second feature extraction means for extracting an item feature vector whose element is the feature of the content signal from the content signal that has received the input;
By substituting the item feature vector f and the attribute important component estimation parameters C and d stored in the attribute important component estimation database into the relational expression , the item indicated by the content signal that has received the input Attribute important component estimation means for generating an item attribute important component vector;
The similarity between the item attribute important component vector of the item indicated by the content signal that has received the input and at least one item attribute important component vector stored in the item database is calculated, and based on the similarity value A list creation means for creating and outputting a list;
A similar item retrieval apparatus characterized by comprising: A database registration unit for registering an item and an item attribute important component vector of the item in the item database;
When the item is associated with an item attribute vector, the database registration unit registers an item attribute important component vector obtained by projecting the item attribute vector onto an axis representing the attribute important component in the item database. And
The database registration unit, when the item is not associated with an item attribute vector, includes an item feature vector f of the item and attribute important component estimation parameters C and d stored in the attribute important component estimation database. The item attribute important component vector calculated by substituting into the relational expression is registered in the item database.
The similar item search device according to claim 1, wherein:   The attribute / attribute important component conversion means uses, as an axis representing an attribute important component, an axis having a certain direction and a large variance in an attribute space that is a space having coordinates of each attribute of an item.
  The similar item search device according to claim 1, wherein the similar item search device is provided.   The list creating means inputs not only the content signal but also attribute data indicating the attribute, and using the attribute important component vector calculated from the attribute data, at least one item attribute important stored in the item database The similarity item search device according to claim 2, wherein the similarity is calculated with the component vector. The similar item search device according to any one of claims 1 to 4, wherein in the second feature extraction means, the input content signal is an acoustic signal.   6. The similar item search apparatus according to claim 5, wherein each element of the item feature vector to be output in the feature extraction means is an acoustic power spectrum at each frequency that is equally spaced on a logarithmic frequency axis. Attribute / attribute important component conversion means, the item attributes vector representing the attributes of the item indicated by the learning content signal, attribute projected to the axis representing the critical component, the item attribute lower item attribute dimensionless number than the vector Generate important component vectors ,
The first feature extracting means generates a learning feature vector having the feature of the learning content signal as an element ,
Parameter learning unit, said the item attributes important component vector is p, the feature vectors for learning when is f, for attributes important component estimation used for estimation of the attributes important component which satisfies the relational expression shown in p = Cf + d Obtaining parameters C and d , storing the attribute important component estimation parameters in the attribute important component estimation database;
The second feature extraction means extracts an item feature vector having the feature of the content signal as an element from the content signal that has received the input,
The attribute important component estimation means sets the item feature vector to f, and substitutes the item feature vector and the attribute important component estimation parameters C and d stored in the attribute important component estimation database into the relational expression. And generating an item attribute important component vector of the item indicated by the content signal that has received the input,
The list creating means includes at least one item attribute stored in an item database in which item attribute important component vectors of items indicated by the content signals that have received the input, item data, and item attribute important component vectors of the item data are registered . A similar item search method, wherein a similarity with an item attribute important component vector is calculated, and a list is created and output based on the similarity value. A program for causing a computer to function as the similar item search device according to claim 1 . A computer-readable recording medium on which the program according to claim 8 is recorded.

Images