KR102222935B1 - Playlist Recommendation Method Using Collaborative Autoencoders - Google Patents

Abstract

The method for recommending a music playlist based on collaborative filtering using an automatic encoder according to the present invention comprises: calculating a first recommendation score, calculating a second recommendation score, and a recommended playlist based on the first recommendation score and the second recommendation score. It includes the step of generating. The step of calculating the first recommendation score includes a playlist vector including track information matching the order in which the sound sources of the playlist are played on the basis of the content recognition-based self-encoder and singer information matching the order in which the sound sources are played. And receiving a mantissa vector. The step of calculating the second recommendation score includes receiving a playlist title based on a character-level convolutional neural network. The step of generating the recommended playlist includes combining the first recommendation score and the second recommendation score according to a linear combination method based on collaborative filtering.

Description

Playlist Recommendation Method Using Collaborative Autoencoders}

The present invention relates to a method of recommending a music playlist based on collaborative filtering using an automatic encoder, and to a method of recommending a playlist based on a recommendation score generated through two sub-models.

Music recommendation methods can be largely divided into content-based recommendation and collaborative filtering-based recommendation. Content-based recommendation is a method of recommending similar types of music that a user listens to by utilizing the inherent properties of music, such as a sound source including songs and lyrics. The content-based recommendation method can recommend songs so that sound sources with a similar atmosphere are shared in one playlist, but this does not satisfy the needs of users who consume various types of music. The collaborative filtering-based recommendation is a method of introducing music preferred by other users who have similar tendencies to the user by analyzing the music consumption patterns of users.

Collaborative filtering-based recommendation has higher user preference than content-based recommendation method, but when the frequency of appearance in the playlist is small or the number of songs constituting the playlist is small, the accuracy of recommendation is lowered. In addition, collaborative filtering-based recommendation becomes impossible when the user sets only the playlist and does not save the song.

An object of the present invention is to provide a collaborative filtering-based recommendation method capable of recommending related music even for unfamous music having a small number of appearances in playlists.

The present invention is to provide a collaborative filtering-based recommendation method capable of improving recommendation accuracy even when the number of songs included in a playlist is small.

In addition, the present invention is to provide a collaborative filtering-based recommendation method in which a user can recommend music even when only the title of a playlist is set and music is not stored.

The method for recommending a music playlist based on collaborative filtering using an automatic encoder according to the present invention comprises: calculating a first recommendation score, calculating a second recommendation score, and a recommended playlist based on the first recommendation score and the second recommendation score. It includes the step of generating. The step of calculating the first recommendation score includes a playlist vector including track information matching the order in which the sound sources of the playlist are played on the basis of the content recognition-based self-encoder and singer information matching the order in which the sound sources are played. And receiving a mantissa vector. The step of calculating the second recommendation score includes receiving a playlist title based on a character-level convolutional neural network. The step of generating the recommended playlist includes combining the first recommendation score and the second recommendation score according to a linear combination method based on collaborative filtering.

The present invention uses a model that learns based on information from which schedule information is omitted in the encoding process, and recommends accuracy even when the number of music related to information that appears in playlists is small or the number of music included in the playlist is small. Can increase.

In addition, since the present invention calculates the recommended score based on the playlist title, the user can recommend music even when only the playlist title is set and no music is stored.

1 is a diagram illustrating an overview of a music playlist recommendation system based on collaborative filtering using an automatic encoder according to the present invention.
2 is a diagram for explaining an overview of the operation of a music playlist recommendation system based on collaborative filtering using an automatic encoder according to the present invention.
3 is a flowchart illustrating a method of recommending a music playlist based on collaborative filtering using an automatic encoder according to the present invention.

The technology described below may be changed in various ways and may have various embodiments, and specific embodiments will be illustrated in the drawings and described in detail. However, this is not intended to limit the technology to be described below with respect to a specific embodiment, and it should be understood to include all changes, equivalents, or substitutes included in the spirit and scope of the technology to be described below.

Terms such as 1st, 2nd, A, B, etc. may be used to describe various components, but the components are not limited by the above terms, and only for the purpose of distinguishing one component from other components. Is only used. For example, a first component may be referred to as a second component, and similarly, a second component may be referred to as a first component without departing from the scope of the rights of the technology described below. The term and/or includes a combination of a plurality of related listed items or any of a plurality of related listed items.

In terms of the terms used in the present specification, expressions in the singular should be understood as including plural expressions unless clearly interpreted differently in context, and terms such as "includes" are specified features, numbers, steps, actions, and components. It is to be understood that the presence or addition of one or more other features or numbers, step-acting components, parts or combinations thereof is not meant to imply the presence of, parts, or combinations thereof.

Prior to the detailed description of the drawings, it is intended to clarify that the division of the constituent parts in the present specification is merely divided by the main function that each constituent part is responsible for. That is, two or more constituent parts to be described below may be combined into one constituent part, or one constituent part may be divided into two or more for each more subdivided function. In addition, each of the constituent units to be described below may additionally perform some or all of the functions of other constituent units in addition to its own main function, and some of the main functions of each constituent unit are different. It goes without saying that it can also be performed exclusively by.

In addition, in performing the method or operation method, each of the processes constituting the method may occur differently from the specified order unless a specific order is clearly stated in the context. That is, each of the processes may occur in the same order as the specified order, may be performed substantially simultaneously, or may be performed in the reverse order.

1 is a diagram illustrating a playlist recommendation system based on deep collaborative filtering according to the present invention.

Referring to FIG. 1, a playlist recommendation system based on deep collaborative filtering according to the present invention includes a first sub-model 100 and a second sub-model 200 for generating a recommended playlist from playlist information D_PL. do.

The playlist information D_PL includes track information TL, singer information a and content c. Each track information TL refers to information on a song title. Singer information (a) refers to information about singers of the music, for example, singer names. Content (c) refers to "Album", that is, refers to title information of the album in which the corresponding song is recorded.

The first sub-model 100 is a model implemented with an autoencoder, and in particular, is implemented with a content-aware autoencoder. The first sub-model 100 includes an encoder 110 and a decoder 120. The encoder 110 compresses the input value, and the decoder 120 restores the compressed input value. The first sub-model 100 receives a playlist vector ap and a mantissa vector p as input values, and calculates a first recommendation score based on a restored value obtained in a process of restoring the input value.

The playlist vector ap refers to a song title in which sound sources included in the playlist information D_PL are matched to the order in which they are played. The singer vector p refers to a singer name in which sound sources included in the playlist information D_PL are matched to the order in which they are played.

The first sub-model 100 adopts an unsupervised learning method of learning a latent vector compressed in a low dimension in the process of restoring itself.

The second submodel 200 is a model implemented with character level convoulutional neural networks. The second sub-model 200 calculates a second recommendation score from the title vector Tp. The second sub-model 200 calculates a second recommendation score between '0' and '1' by performing a convolution operation on the title vector Tp. The second sub-model 200 is implemented as an independent model that is distinct from the first sub-model 100.

2 is a diagram showing an overview of the operation of the playlist recommendation system according to the present invention. 3 is a flowchart illustrating a method for recommending a playlist based on deep collaborative filtering according to the present invention.

Referring to FIGS. 1 to 3, a method for recommending a playlist according to the present invention will be described as follows.

In the first step (S1), the first sub-model 100 is based on the content recognition-based self-encoder, the first recommendation from the playlist vector (ap) and mantissa vector (p) generated from the playlist information (D_PL) Generate a score. To this end, the first submodel 100 receives the combined vector Vc in which the playlist vector ap and the mantissa vector p are combined, and encodes and decodes the combination vector Vc. You can generate a recommendation score.

When the total number of songs included in the playlist information D_PL is "M", the size of the playlist vector ap has a value of "M". Each element of the playlist vectors (ap) corresponds to one of the track information (TL). If the playlist does not include a track, it has a value of '0'. If included in the playlist, the element has a value of "1/n". For example, if there are 6 tracks in the playlist information D_PL, each of the tracks has a value of "1/6". This has the same effect as taking the average of vectors, assuming that the encoder of the magnetic encoder is composed of a set of latent vectors for each track.

The encoder 110 receives a playlist vector and compresses the playlist vector into a low-dimensional latent vector.

The decoder 120 generates a reconstructed vector by reconstructing the latent vector compressed by the encoder 110. Similar to the input vector, the reconstructed vector has a size of M, and each element is regarded as a recommended score for the corresponding song. Each element of the reconstructed vector has a value of "0" or more and "1" or less, and it can be interpreted as meaning that a higher recommendation score is recorded as it approaches 1.

In order to generate the recommended score based on the mantissa vector (p), the encoder 110 receives the mantissa vector (p). Similar to the playlist vector ap, the singer vector p is created to have a value of "1" if a singer appears in the playlist and "0" if the singer does not appear. For example, when "singer (a)-sound source" is matched in a playlist, it is recognized that "singer (a)" appears when the sound source is played, and the singer vector (p) is created to have a value of "1". do. The mantissa vector (p) is combined with the playlist vector (ap) to create an associative vector (Vc). That is, the first sub-model 100 corresponding to the self-encoder receives the combination vector Vc, which is not one of the playlist vector ap or the mantissa vector p, and generates a first recommendation score.

In the above description, the playlist vector (ap) and the mantissa vector (p) have been separately explained, but the first sub-model 100 of the present invention combines the playlist vector (ap) and the mantissa vector (p) to provide a combination vector. (Vc) is created. The binding vector (Vc) is [p; It can be expressed in the form of a _{p ].}

)의 차이를 Cross Entropy Loss를 통해서 줄이는 것을 설명하고 있다.Then, the encoder 110 receives a combination vector ([p; a _p ]). [Equation 1] below is a combination vector ([p; a _p ]) and a reconstruction vector corresponding to the output of the decoder 120 (

) To reduce the difference through Cross Entropy Loss.

[Equation 1]

In [Equation 1], p denotes a playlist vector and a _p denotes a mantissa vector.

Θ can be defined as = {W ∈ R ^d×(n+k) , W′ ∈ R(n+k)×d, b ∈ Rd, b′ ∈ Rn+k }. In this case, W denotes a weight in the artificial neural network, and b denotes a bias. W'denotes a weight of the decoder 120, and b'denotes a deflection of the decoder 120.

The first sub-model 100 learns an artificial neural network in a manner that reduces reconstruction vectors and cross entropy. The first sub-model 100 performs learning based on removing a part of information of the combination vector Vc provided as an input in the learning process. That is, when a playlist vector (ap) or an incomplete combination vector (Vc) in which some information is omitted from among the singer vector (p) is received, the learning proceeds with the aim of restoring the complete playlist order and singer information.

A process in which the first sub-model 100 performs learning based on the combination vector Vc from which a part has been removed includes two steps as follows.

The first step is the hide-and-seek step. The hide-and-seek step is a step of randomly selecting and removing one of the playlist vector (ap) and the singer vector (p).

For example, as shown in [Equation 2] below, the first sub-model 100 may delete information on the mantissa vector p and perform learning.

[Equation 2]

In [Equation 2], f denotes the encoder 110, and g denotes the decoder 120. In [ ^~ p;0], ^~ p denotes an incomplete vector in which some of the playlist vectors are denoted. Since the part corresponding to the mantissa vector has been removed in [Equation 2], the mantissa vector component in the combined vector of the encoder 110 is expressed as “0”. In conclusion, the first sub-model 100 performs learning in a direction that reduces the difference between the ^{combined vector and the reconstructed vector of [~ p; 0].}

Alternatively, the first sub-model 100 may delete information on the playlist vector ap and perform learning as shown in [Equation 3] below.

[Equation 3]

That is, [Equation 3] expresses that the component corresponding to the playlist in the combined vector of the encoder 110 is '0'.

3 shows that the mantissa vector (p) has been removed as in [Equation 2].

The first sub-model 100 accurately restores both the playlist vector ap and the singer vector p as a restoration value when any one of the playlist vector ap and the singer vector p is omitted. Aim for that. Through this, not only patterns between playlists, but also patterns between playlists and singers can be learned together to recommend playlists closer to the user's taste.

The second step is an additional elimination step. The additional removal step is a process of removing the information of the vector left unremoved in the hide-and-seek step once more.

When selecting the removal value in the additional removal step, regardless of the order stored in the playlist, it is selected whether to hide the last information corresponding to the latest information among the information stored in the playlist, namely, whether to hide the information stored later. If users prefer to play randomly within a playlist, they use a method of randomly hiding them regardless of the order in which they are stored in the playlist. Alternatively, in order to implement a system that recommends songs to be played on top of the playlist when the user sequentially plays the playlist, that is, a system that recommends the playlist order, a method of hiding the back part corresponding to information stored later in the playlist is used.

The method of masking the removal value and the degree of masking the removal value in the additional removal step may be determined according to the system to be implemented. For example, in order to implement a system that works smoothly as the number of songs decreases, as much information as possible is removed. Accordingly, accurate prediction is possible even when the input value of the first sub-model 100 is small. On the other hand, relatively little information is removed in order to implement a system that works smoothly while songs are secured to some extent. Alternatively, in order to implement a system having performance regardless of the amount of input values, a ratio that is obscured each time learning is performed is selected as a random value.

In the second step (S2), the second sub-model 200 calculates a second recommendation score from the title vector Tp based on the character-level convolutional neural network. The title vector Tp contains information on the spelling order of the title of the playlist.

The title vector (Tp) is generated based on the number and order of the spellings of the playlist title (Title). The title vector (Tp) can be created as a vector consisting of a 9-by-k matrix when nine letters of "C,h,r,i,s,t,m,a,s" are sequentially input. have. In this case, "k" refers to the dimension of a filter for a convolution operation. For example, the title vector Tp of FIG. 2 shows an example expressed in a 9-by-4 matrix.

The second sub-model 200 performs a convolution operation using a k-dimensional filter to determine a pattern of spelling combinations, and obtains a feature vector. In the convolution operation, since "christmas" and "christa" are finally expressed as feature vectors of similar values, they are robust against typos entered by the user.

In addition, the second sub-model 200 calculates the largest feature value through a “max pooling” operation in the feature vector.

As described above, the second sub-model 200 calculates the second recommendation score from the title of the playlist based on the character-level convolutional neural network rather than the "word level CNN". The "word level CNN" is for analyzing the pattern of word arrangement in a sentence, and expresses the input sentence as a vector. In the "word level CNN", since the system operates based on words, if the word cannot be analyzed due to a typo, the model may not be smoothly driven.

In contrast, in the present invention, since the second recommendation score is generated based on the character-level convolutional neural network, smoother driving is possible even when the correct word is not input due to a typo or the like.

In the third step (S3), according to a linear combination method based on collaborative filtering, a combined recommendation score is calculated by combining the first recommendation score and the second recommendation score, and a song matching the playlist is selected according to the combined recommendation score. I recommend you. The first recommended score is restored through the first sub-model 100, and the second recommended score is restored through the second sub-model 200. The first and second recommendation scores may be regarded as reconstructed vectors of the decoder.

The combined model 300 sets a first weight (Witem) for the first recommendation score and a second weight (Wtitle) for the second recommendation score, and generates a recommended playlist based on this.

The sizes of the first and second weights Witem and Wtitle may be set to be the same. This method of setting the same size of the first and second weights (Witem, Wtitle) has the advantage of simple system implementation, but the accuracy of the first sub-model 100 or the second sub-model 200 In other cases, the accuracy of the recommended playlist is lowered.

In order to increase the accuracy of the recommended playlist, a method of dynamically allocating the first sub-model 100 and the second sub-model 200 may be used.

As for the first and second weights (Witem, Wtitle), as in [Equation 4] below, the importance of the playlist title is fixed to 1, and the weight of the first recommendation score is the amount and singer of the track information (TL). It can be set to be proportional to the amount of information (a).

[Equation 4]

In [Equation 4], N([p; ap]) means the number of data of the concatenation vector Vc, and “I(Tp)” means the number of playlist titles. For example, when a total of 20 track information (TL) and singer information (a) are given in the playlist information (D_PL), the first weight (Witem) is 20/21 and the second weight (Wtitle) is 1/21. Can be set.

It will be appreciated by those skilled in the art through the above description that various changes and modifications can be made without departing from the spirit of the present specification. Therefore, the technical scope of the present specification should not be limited to the content described in the detailed description of the specification, but should be determined by the scope of the claims.

100: content recognition-based self-encoder
110: encoder 120: decoder
200: Character level convolutional neural network model

Claims (10)

In the method of operating a playlist recommendation system based on deep collaboration filtering including a first sub-model, a second sub-model, and a combined model,
On the basis of the content recognition-based self-encoder using the first sub-model, a playlist vector including track information matching the order in which sound sources of the playlist are played, and singer information matching the order in which the sound sources are played are included. Calculating a first recommendation score based on the mantissa vector;
Calculating a second recommendation score based on a playlist title based on a character-level convolutional neural network using the second sub-model; And
Calculating a combined recommendation score by combining the first recommendation score and the second recommendation score according to a linear combination method based on collaborative filtering using the combination model, and generating a recommended playlist according to the combined recommendation score. Collaborative filtering-based music playlist recommendation method using an automatic encoder comprising a. The method of claim 1,
The playlist vector is
A method of recommending a music playlist based on collaborative filtering using an automatic encoder including an element having a size of M when the number of sound sources is "M" and classifying whether or not the sound sources have been reproduced as binary data. The method of claim 1,
The step of calculating the first recommendation score
Compressing the playlist vector and the singer vector to generate a low-dimensional latent vector. The method of claim 3,
Generating the latent vector is
A method for recommending a music playlist based on collaborative filtering using an automatic encoder, comprising the step of deleting at least one information from the playlist vector or the singer vector. The method of claim 4,
Generating the latent vector is
A method for recommending a music playlist based on collaborative filtering using an automatic encoder, further comprising the step of additionally removing at least part of information from information on the playlist vector or the singer vectors that are not deleted. The method of claim 5,
The step of calculating the first recommendation score
A method for recommending a music playlist based on collaborative filtering using an automatic encoder, further comprising the step of generating a reconstructed vector by restoring the low-dimensional latent vector. The method of claim 6,
A method for recommending a music playlist based on collaborative filtering using an automatic encoder that increases the likelihood of adopting the first recommendation score as the restoration vector is generated with a size of 0 or more and 1 or less. The method of claim 1,
The step of calculating the second recommendation score is
A method for recommending music playlists based on collaborative filtering using an automatic encoder, comprising the step of convoluting a title vector including information obtained by sequentially sorting the spellings of the playlist titles. The method of claim 8,
A method for recommending a music playlist based on collaborative filtering using an automatic encoder in which rows and columns of the title vector are each composed of a number of letters of the playlist title and a dimension of a filter used in the convolution operation. The method of claim 9,
Generating the recommended playlist comprises:
Including the step of assigning a first weight to the first recommendation score and a second weight to the second recommendation score,
A method for recommending a music playlist based on collaborative filtering using an automatic encoder in which the first weight and the second weight are assigned using a dynamic allocation method.

Images