KR102190304B1 - preferred content recommending apparatus based on GPU computing Using deep learning and method therefor - Google Patents

Abstract

In the present invention, the content user propensity vector vectorized content user propensity and the content propensity vector vectorized content propensity vector are input into the content recommendation model and loss function, and calculation is performed using only the GPU, so that the content evaluation score for each content user It relates to a technology that provides, and may include a vector conversion unit, a GPU operation unit, and a content evaluation score prediction unit. Compared to the recommended model that performed calculation using both a CPU and a GPU, efficient distribution of computing resources and It can achieve the purpose of improving computational speed by preventing bottlenecks, and collects information on dwell time for content used by content users, and calculates a customized score that converts dwell time into relative scores for each content user using this By providing preferences, there is an effect that more accurate individual preferences can be provided.

Description

A preferred content recommending apparatus based on GPU computing Using deep learning and method therefor.

The present invention relates to a technology for providing content that meets the preferences of content users using deep learning, and more specifically, a content user propensity vector that vectorizes the propensity of the content user and a content propensity vector that vectorizes the content propensity. An object of the present invention is to provide a preferred content recommendation device and method using deep learning based on GPU computation that provides content evaluation scores for each content user to content users by inputting into a recommendation model and loss function and performing calculations using only GPU.

As the speed of wired and wireless communication and the speed of development of operator personnel have been rapidly achieved, there are a huge amount of contents online enough to be inundated.

Therefore, there is a great need for selecting and providing content preferred by a specific user from among such enormous content, and accordingly, various content recommendation models exist.

Conventional content recommendation systems have recommended content with high similarity to the analyzed content to users by performing calculation of a deep learning model using both a CPU and a GPU.

In such a conventional content recommendation model, a bottleneck occurs when the CPU and the GPU respectively perform computations and problems due to inefficient distribution of computational resources.

In addition, by calculating the content preference for each content user based on the content user's staying time for the content, the relative time of using the content is different for each content user, so when the preference is calculated based on the absolute time, an error occurs in the preference determination. There was a problem.

The present invention uses a content recommendation model that performs computation using only a GPU and provides content evaluation scores for each content user to content users, and is more efficient in terms of computational resources than a recommendation model that previously performed computation using both a CPU and a GPU. The present invention relates to an apparatus and method for recommending preferred contents using deep learning based on GPU computation that can provide distribution and bottleneck prevention.

According to an embodiment of the present invention, a content recommendation device using deep learning based on GPU computation uses a deep learning model consisting of a neural network operated by a processor, and a content user propensity vector u vectorized content users' propensities, and a content propensity A vector converting unit that calculates a vectorized content propensity vector v; A GPU calculation execution unit that performs calculation by inputting the calculated u and v to a preset loss function in order to implement a complete deep learning structure so that calculation is possible using only a GPU; And a content evaluation score prediction unit that predicts a content evaluation score for a specific content for each content user based on the calculation result.

According to an embodiment of the present invention, when performing the calculation, the GPU calculation unit may input at least one content user propensity vector u to a weight matrix generated using multiple dense layers.

According to an embodiment of the present invention, when the content is mainly text, a dictionary may be formed and a word to vector model may be trained to vectorize texts included in the content.

According to an embodiment of the present invention, texts included in content are crawled and registered in a dictionary for each sentence, and exception words stored in advance are deleted by parsing the sentences registered in the dictionary, and only parts of speech in a predetermined format are stored in advance. A word to vector model can be generated using the extracted and stored sentences.

According to an embodiment of the present invention, the GPU computational unit uses a CNN model for document processing when the content is text-oriented, and uses a CNN model for video or music processing when the content is an image or music owner. It can be used by replacing it with the corresponding CNN model for each type of.

According to an embodiment of the present invention, a content that provides content preference for each content user by collecting residence time information for each content used by each content user and converting the collected stay time into a relative score to determine the personal preference of the content It may further include a preference providing unit and a content recommendation unit for predicting a content evaluation score for the content having a high content preference and recommending a content for which a high evaluation score is predicted.

According to an embodiment of the present invention, the content preference providing unit includes: a residence time information collection unit for collecting residence time information for each content used by a content user; A normal distribution function matching unit for correlating a normal distribution function to each of the collected information on the residence time for each content; A residence time distribution estimation unit for estimating a residence time distribution of the content user by linearly combining a normal distribution function for the corresponding plurality of residence time information; A customized score calculator configured to calculate a cumulative distribution function by accumulating the estimated distribution of residence time, and calculating a customized score obtained by converting the residence time into a relative score for each content user using the calculated cumulative distribution function; And a preference determining unit for determining a personal preference of the corresponding content by generating a preference graph for a plurality of contents for each content user based on the calculated customized score.

According to an embodiment of the present invention, the customized score calculator may calculate a customized score by transforming the calculated cumulative distribution function into a quartile.

According to an embodiment of the present invention, the preference determination unit may determine that the content user prefers as the calculated customized score increases.

According to an exemplary embodiment of the present invention, as the number of contents used by the content user increases, the estimated residence time distribution may follow a distribution corresponding to an actual result.

According to an embodiment of the present invention, a content recommendation method using deep learning based on GPU computation uses a deep learning model consisting of a neural network operated by a processor, and a content user propensity vector u vectorized content users' propensities, and a content propensity Calculating a vectorized content propensity vector v; Performing an operation by inputting the calculated u and v to a preset loss function in order to implement a complete deep learning structure so that computation is possible using only a GPU; And predicting a content evaluation score for a specific content for each content user based on the calculation result.

According to an embodiment of the present invention, in the calculating of the content propensity vector v, in performing the operation, at least one content user propensity vector u may be input to a weight matrix generated using multiple dense layers. have.

According to an embodiment of the present invention, when the content is mainly text, a dictionary may be formed and a word to vector model may be trained to vectorize texts included in the content.

According to an embodiment of the present invention, texts included in content are crawled and registered in a dictionary for each sentence, and exception words stored in advance are deleted by parsing the sentences registered in the dictionary, and only parts of speech in a predetermined format are stored in advance. A word to vector model can be generated using the extracted and stored sentences.

According to an embodiment of the present invention, the calculating of the content propensity vector v includes using a document processing CNN model when the content is mainly text, and a video or music processing CNN when the content is an image or music owner. By using the model, it can be used by replacing it with the corresponding CNN model for each type of content.

According to an embodiment of the present invention, providing content preference for each content user by collecting residence time information for each content used by each content user and converting the collected stay time into a relative score to determine personal preference of the corresponding content. And predicting a content evaluation score for the content having the high content preference, and recommending content for which a high evaluation score is predicted.

According to an embodiment of the present invention, the providing of the content preference may include: collecting information on residence time for each content used by a content user; Correlating a normal distribution function to each of the collected information on the residence time for each content; Estimating a residence time distribution of the content user by linearly combining a normal distribution function for the corresponding plurality of residence time information; Calculating a cumulative distribution function by accumulating the estimated distribution of dwell time, and calculating a customized score obtained by converting the dwell time into a relative score for each content user by using the calculated cumulative distribution function; And determining a personal preference of the corresponding content by generating a preference graph for a plurality of contents for each content user based on the calculated customized score.

According to an embodiment of the present invention, in the calculating of the customized score, the customized score may be calculated by transforming the calculated cumulative distribution function into a quartile.

According to an embodiment of the present invention, in determining the personal preference, as the calculated customized score increases, it may be determined as the content preferred by each content user.

According to an embodiment of the present invention, in the step of estimating the distribution of the dwell time, as the number of contents used by the content user increases, the estimated distribution of the dwell time may follow a distribution corresponding to an actual result.

According to the present invention, it is possible to achieve the object of improving computation speed through inefficient distribution of computational resources and prevention of bottlenecks, compared to a recommended model that has performed computation using both a CPU and a GPU. There is an effect that it is possible to provide more accurate personal preferences by providing content preferences by calculating a customized score that converts the residence time into a relative score for each content user by collecting information on the residence time for the content and using it.

1 is a diagram showing a conventional recommended model using deep learning that simultaneously performs CPU and GPU calculations.
2 is a block diagram of a preferred content recommendation apparatus using deep learning based on GPU computation according to an embodiment of the present invention.
3 is a detailed configuration diagram of a content preference providing unit disclosed in FIG. 2.
4 is a diagram illustrating a calculation process of an apparatus for recommending preferred content based on GPU calculation according to an embodiment of the present invention.
FIG. 5 is a diagram illustrating a data flow of a device for recommending preferred content based on a GPU calculation, which uses a model for recommending preferred content by analyzing content in an embodiment of the present invention.
6 is a flowchart of a method for recommending preferred content using deep learning based on GPU computation according to an embodiment of the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art may easily implement the present invention. However, the present invention may be implemented in various forms and is not limited to the embodiments described herein.

In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and similar reference numerals are assigned to similar parts throughout the specification.

Throughout the specification, when a part "includes" a certain component, it means that other components may be further included rather than excluding other components unless specifically stated to the contrary.

Hereinafter, an apparatus and method for recommending contents using deep learning based on GPU calculation according to an embodiment of the present invention will be described with reference to the drawings.

1 is a diagram showing a conventional recommended model using deep learning that simultaneously performs CPU and GPU calculations.

Referring to FIG. 1, a matrix decomposition that predicts an evaluation score for a specific content of a content user by constructing a user and a characteristic vector of the content using evaluation information on the existing content among the recommendation models based on conventional deep learning. The recommended model using (Matrix Factorization) can be divided into a matrix decomposition operation part using a CPU operator source and a CNN operation part using a GPU operator source.

The matrix decomposition operation part can be calculated as the inner product of the content user's propensity vector u and the content propensity vector v, which can be expressed as Equation 1 below.

[Equation 1]

In addition, since vectorization of all content users and content is difficult in reality, a random initialization method as shown in Equation 2 can be used to solve the problem through a generally probabilistic approach.

[Equation 2]

As shown in FIG. 1, the matrix decomposition operation part is calculated by a CPU, and the content characteristic vector propensity part may be calculated using a GPU based on a CNN operation.

In the CNN-based content propensity vector generation part disclosed in FIG. 1, X is the raw information of the content, and this can be processed into the content propensity vector V through the CNN, and W is the weights existing in each layer of the CNN. Indicates.

The loss function to be optimized in the recommendation model shown in FIG. 1 can be expressed as Equation 3 below.

[Equation 3]

2 is a block diagram of a preferred content recommendation apparatus 1000 using deep learning based on GPU computation according to an embodiment of the present invention.

Referring to FIG. 2, the preferred content recommendation apparatus 1000 using deep learning based on GPU calculation according to the first embodiment of the present invention includes a vector conversion unit 100, a GPU calculation execution unit 200, and a content evaluation score prediction unit. It may include (300).

According to an embodiment of the present invention, the preferred content recommendation apparatus 1000 using deep learning based on GPU computation may use a deep learning model consisting of a neural network operated by a processor.

The vector conversion unit 100 may calculate a content user propensity vector u obtained by vectorizing the propensity of the content user and a content propensity vector v vectorized by the content propensity.

The content user propensity vector u and the content propensity vector v obtained by vectorizing the content propensity can be calculated as described with reference to FIG. 1.

The GPU operation execution unit 200 may perform an operation by inputting the calculated u and v to a preset loss function in order to implement a complete deep learning structure so that the operation can be performed using only the GPU.

According to an embodiment of the present invention, the preset loss function may be expressed as Equation 4 below.

[Equation 4]

는 회귀 손실(regression loss)을 의미하며, Absolute error, Squared error, Huber loss 등이 사용될 수 있다.here

Denotes a regression loss, and Absolute error, Squared error, Huber loss, etc. can be used.

The loss function as shown in Equation 4 does not require a separate CPU calculation and can be calculated using only GPU calculation, so it is possible to prevent problems such as delays in total calculation time caused by bottlenecks and efficient distribution of calculation resources between the CPU and GPU. have.

According to an embodiment of the present invention, in performing an operation, a value of at least one content user propensity vector u may be input to a weight matrix generated using multiple dense layers.

This will be described in more detail with reference to FIG. 4.

According to an embodiment of the present invention, when the content is mainly text, a dictionary may be formed and a word to vector model may be trained for vectorization of texts included in the content.

According to an embodiment of the present invention, texts included in content are crawled and registered in a dictionary for each sentence, and exception words stored in advance are deleted by parsing the sentences registered in the dictionary, and only parts of speech in a predetermined format stored in advance are extracted. Thus, a word to vector model can be generated using the stored sentences.

According to an embodiment of the present invention, in order to generate a word to vector model, text included in content may be parsed as a Mecab, and sentences may be stored according to a preset rule.

For example, you can only store parts of speech such as body language, verbal verbs, modifiers, and independent words, and can be used for'Reporter','Photo','Cut','Report','Dcom','Offer','News','Capture', etc. Sentences with the same specific word and length less than 3 words can be excluded.

Using these stored sentences, a word-to-vector model can be created.

It may be the same as the above-described embodiment, but is not limited thereto. According to another embodiment of the present invention, a CNN model for document processing is used when the content is text-oriented, and when the content is a video or music owner, video or music processing Using the CNN model for each content type, it can be replaced with the corresponding CNN model.

The content evaluation score prediction unit 300 may predict a content evaluation score for a specific content for each content user based on the calculation result.

Here, the content evaluation score for a specific content may mean a score predicted through deep learning as to what evaluation score (R) a content user will give for a specific content that the content user has not yet used.

According to the second embodiment of the present invention, the preferred content recommendation device 1000 using deep learning based on GPU calculation includes not only the vector conversion unit 100, the GPU calculation execution unit 200, and the content evaluation score prediction unit 300 A content preference providing unit 400 and a content recommendation unit 500 may be further included.

The content preference providing unit 400 can provide content preference for each content user by collecting residence time information for each content used by each content user and converting the collected residence time into a relative score to determine the personal preference of the corresponding content. have.

The content preference providing unit 400 will be described in more detail with reference to FIG. 3.

The content recommendation unit 500 may predict a content evaluation score for a content having the highest content preference and recommend a content for which the highest evaluation score is predicted.

3 is a detailed configuration diagram of the content preference providing unit 400 disclosed in FIG. 2.

Referring to FIG. 3, a residence time information collection unit 410, a normal distribution function matching unit 420, a residence time distribution estimation unit 430, a customized score calculation unit 440, and a preference determination unit 450 may be included. have.

The residence time information collection unit 410 may collect residence time information for each content used by a content user.

According to an embodiment of the present invention, a dwell time at which one content user stays on a corresponding web page to read the content may be measured and defined as dwell time information.

According to the above embodiment, the residence time may be measured for each content included in a plurality of contents subscribed by the content user, and the residence time information may be generated based on the measured stay time.

At this time, the time information included in the generated stay time information is for the absolute time, and the speed and usage habits of each content user are not considered at all, so when the preference is calculated based on this, the accuracy decreases due to individual differences. exist.

Therefore, there is a need to calculate the relative time for each individual based on the residence time measured as in the present invention.

The normal distribution function matching unit 420 may correspond to a normal distribution function to each of the collected information on the residence time for each content.

According to an embodiment of the present invention, the content user A is the content 1, 2, 3, ... Assuming that is used, an event with a rare probability of using a specific content among numerous contents has occurred, so each event can have an important meaning.

According to an embodiment of the present invention, if the content user A shows other specific content similar in taste to content 1, assuming that it will have a similar stay time as when using content 1, it is estimated that similar stay time will follow a normal distribution. In addition, a normal distribution function can be applied to each of the collected information on the time of stay for each content.

The residence time distribution estimating unit 430 may linearly combine a normal distribution function for the corresponding plurality of residence time information to estimate the residence time distribution of the content user.

According to an embodiment of the present invention, a content user A associates a normal distribution function with one event (content, residence time) using the content, and linearly combines the corresponding normal distribution functions to determine the distribution of the content user A's residence time. Can be estimated.

, 콘텐츠 이용자 A의 사건 개수

, 콘텐츠 이용자 A의 체류시간 표준편차

일 때 평균 및 표준 편차는 아래와 같다.According to the above embodiment, each normal distribution function is the residence time for article N

, Number of incidents of content user A

, Standard deviation of content user A's residence time

When is, the mean and standard deviation are as follows.

According to an embodiment of the present invention, by using Equation 5 below based on the above-described variables, the distribution of the content user's residence time (pdf) can be estimated.

[Equation 5]

According to an embodiment of the present invention, it may be assumed that the average stay time for the distribution of stay time is the same as the average stay time of the recorded content user, and the distribution of stay time for the stay time distribution (pdf) is as Equation 6 below.

[Equation 6]

According to an embodiment of the present invention, as the number of contents subscribed by a content user increases, the estimated distribution of residence time may follow a distribution corresponding to an actual result.

In other words, as the number of contents read by a content user increases, the residence time distribution (pdf) can increase the accuracy of the actual result.

The customized score calculation unit 440 may calculate a cumulative distribution function by accumulating the estimated distribution of staying time, and calculate a custom score obtained by converting the staying time into a relative score for each content user using the calculated cumulative distribution function.

Here, the customized score is not a score having a certain perfect score, but may mean a value that can determine whether or not a preference is based on a relative size, but is not limited thereto, and any number capable of representing a relative size may be used without limitation.

According to an embodiment of the present invention, the cumulative distribution function cdf may be calculated using Equation 7 below.

[Equation 7]

According to an embodiment of the present invention, a customized score may be calculated by transforming a calculated cumulative distribution function into a quartile.

According to an embodiment of the present invention, after assuming the distribution of the ideal score, a custom score may be calculated by transforming a quartile. According to an embodiment, a custom score may be calculated by performing a quartile transformation as shown in Equation 8 below. Can be calculated.

[Equation 8]

The preference determination unit 450 may determine a personal preference of the corresponding content by generating a preference graph for a plurality of contents for each content user based on the calculated customized score.

According to an embodiment of the present invention, the preference graph for a plurality of contents for each content user may be drawn so that the customized score for each content can be viewed at a glance by using the X-axis as the residence time and the Y-axis as a customized score. You can judge individual preferences.

According to an embodiment of the present invention, as the calculated customized score is higher, it may be determined as the content preferred by each content user.

4 is a diagram illustrating a calculation process of an apparatus for recommending preferred content based on GPU calculation according to an embodiment of the present invention.

According to an embodiment of the present invention, the GPU calculation-based preferred content recommendation apparatus 1000 may perform a calculation process as shown in FIG. 4.

According to an embodiment of the present invention, in the case of text-oriented content, as shown in FIG. 4, in order to minimize the bottleneck of the operation by performing an operation process as shown in FIG. 4, time required for operation may be minimized using only the GPU.

Referring to FIG. 4, when data of specific content is input, evaluation scores for each content user can be predicted and provided, and at least one content user propensity vector u is input in a weight matrix created using multiple dense layers. Can be.

Here, in the case of the second computational process, if the type of content is not text-oriented, it can be replaced with a separate deep learning model.

FIG. 5 is a diagram illustrating a data flow of a device for recommending preferred content based on a GPU calculation in which a model for recommending preferred content by analyzing text-oriented content in one embodiment of the present invention is an example.

5, a word-to-model is generated by inputting a news article, which is a text-oriented content, and is calculated through a CNN-based text processing model that reflects a loss function, and a recommendation model is calculated through multi-dense layer calculations for specific content. Content evaluation score can be predicted.

6 is a flowchart of a method for recommending preferred content using deep learning based on GPU computation according to an embodiment of the present invention.

The content user propensity vector and the content propensity vector are calculated (610).

According to an embodiment of the present invention, a content user propensity vector u obtained by vectorizing the propensity of a content user and a content propensity vector v vectorized by a content propensity may be calculated.

According to the above embodiment, the content user disposition vector u and the content disposition vector v obtained by vectorizing the content disposition may be calculated as described with reference to FIG. 1.

An operation is performed by inputting a vector value to a preset loss function (620).

According to an embodiment of the present invention, in order to implement a complete deep learning structure so that computation can be performed using only a GPU, the calculation may be performed by inputting the calculated u and v to a preset loss function.

According to an embodiment of the present invention, the preset loss function may be expressed as Equation 4 above.

According to an embodiment of the present invention, in performing an operation, a value of at least one content user propensity vector u may be input to a weight matrix generated using multiple dense layers.

According to an embodiment of the present invention, when the content is mainly text, a dictionary may be formed and a word to vector model may be trained for vectorization of texts included in the content.

According to an embodiment of the present invention, texts included in content are crawled and registered in a dictionary for each sentence, and exception words stored in advance are deleted by parsing the sentences registered in the dictionary, and only parts of speech in a predetermined format stored in advance are extracted. Thus, a word to vector model can be generated using the stored sentences.

According to an embodiment of the present invention, in order to generate a word to vector model, text included in content may be parsed as a Mecab, and sentences may be stored according to a preset rule.

The content evaluation score for each content user is predicted (630).

According to an embodiment of the present invention, a content evaluation score for a specific content may be predicted for each content user based on an operation result.

Here, the content evaluation score for a specific content may mean a score predicted through deep learning as to what evaluation score (R) a content user will give for a specific content that the content user has not yet used.

Content preference for each content user is determined (640).

According to an embodiment of the present invention, it is possible to provide content preference for each content user by collecting information on the residence time for each content used by each content user and converting the collected residence time into a relative score to determine the personal preference of the content have.

According to an embodiment of the present invention, it is possible to collect information about the residence time for each content used by a content user.

According to an embodiment of the present invention, a dwell time at which one content user stays on a corresponding web page to read the content may be measured and defined as dwell time information.

According to the above embodiment, the residence time may be measured for each content included in a plurality of contents subscribed by the content user, and the residence time information may be generated based on the measured stay time.

According to an embodiment of the present invention, a normal distribution function may be applied to each of the collected information about the dwell time for each content.

According to an embodiment of the present invention, the content user A is the content 1, 2, 3, ... Assuming that is used, an event with a rare probability of using a specific content among numerous contents has occurred, so each event can have an important meaning.

According to an embodiment of the present invention, if the content user A shows other specific content similar in taste to content 1, assuming that it will have a similar stay time as when using content 1, it is estimated that similar stay time will follow a normal distribution. In addition, a normal distribution function can be applied to each of the collected information on the time of stay for each content.

According to an embodiment of the present invention, by linearly combining a normal distribution function for a plurality of corresponding dwell time information, it is possible to estimate the distribution of the dwell time of a content user.

According to an embodiment of the present invention, a content user A associates a normal distribution function with one event (content, residence time) using the content, and linearly combines the corresponding normal distribution functions to determine the distribution of the content user A's residence time. Can be estimated.

, 콘텐츠 이용자 A의 사건 개수

, 콘텐츠 이용자 A의 체류시간 표준편차

일 때 평균 및 표준 편차는 아래와 같다.According to the above embodiment, each normal distribution function is a residence time for content N

, Number of incidents of content user A

, Standard deviation of content user A's residence time

When is, the mean and standard deviation are as follows.

According to an embodiment of the present invention, a content user's residence time distribution (pdf) can be estimated by using Equation 5 above based on the above-described variables.

According to an embodiment of the present invention, it may be assumed that the average stay time for the distribution of stay time is the same as the average stay time of the recorded content user, and the distribution of stay time for the stay time distribution (pdf) is as Equation 6 above. .

According to an embodiment of the present invention, as the number of contents subscribed by a content user increases, the estimated distribution of residence time may follow a distribution corresponding to an actual result.

According to an embodiment of the present invention, a cumulative distribution function may be calculated by accumulating the estimated distribution of dwell time, and a customized score obtained by converting the dwell time into a relative score for each content user may be calculated using the calculated cumulative distribution function.

According to an embodiment of the present invention, a cumulative distribution function (cdf) may be calculated using Equation 7 above.

According to an embodiment of the present invention, a customized score may be calculated by transforming a calculated cumulative distribution function into a quartile.

According to an embodiment of the present invention, after assuming the distribution of the ideal score, a custom score may be calculated by transforming a quartile. According to an embodiment, a custom score may be calculated by performing a quartile transformation as in Equation 8 above. Can be calculated.

According to an embodiment of the present invention, a preference graph for a plurality of contents for each content user may be generated based on the calculated customized score to determine the personal preference of the corresponding content.

According to an embodiment of the present invention, the preference graph for a plurality of contents for each content user may be drawn so that the customized score for each content can be viewed at a glance by using the X-axis as the residence time and the Y-axis as a customized score. You can judge individual preferences.

According to an embodiment of the present invention, as the calculated customized score is higher, it may be determined as the content preferred by each content user.

The content evaluation score for the content having the highest content preference is predicted (650).

According to an embodiment of the present invention, a content evaluation score for content having a high content preference may be predicted by repeating the above process.

Contents for which the highest evaluation score is predicted are recommended (660).

According to an embodiment of the present invention, content for which a high evaluation score is predicted may be recommended to content users.

The embodiments of the present invention are not implemented only through the apparatus and/or method described above, and the embodiments of the present invention have been described in detail above, but the scope of the present invention is not limited thereto, and the following claims Various modifications and improvements of those skilled in the art using the basic concept of the present invention defined in are also within the scope of the present invention.

100: vector conversion unit 200: GPU operation execution unit
300: content evaluation score prediction unit 400: content preference providing unit
500: content recommendation unit 1000: content recommendation device

Claims (20)

Using a deep learning model consisting of a neural network operated by a processor,
A vector conversion unit that calculates a content user propensity vector u obtained by vectorizing a propensity of a content user and a content propensity vector v vectorized by a content propensity;
A GPU calculation execution unit that performs calculation by inputting the calculated u and v to a preset loss function in order to implement a complete deep learning structure so that calculation is possible using only a GPU;
A content evaluation score predictor for predicting a content evaluation score for a specific content for each content user based on the calculation result; And
Including; a content preference providing unit that collects information on the residence time for each content used by each content user and converts the collected residence time into a relative score to determine the personal preference of the corresponding content and provides the content preference for each content user; and
The content preference providing unit,
A residence time information collection unit that collects residence time information for each content used by a content user;
A normal distribution function matching unit for correlating a normal distribution function to each of the collected information on the residence time for each content;
A residence time distribution estimation unit for estimating a residence time distribution of the content user by linearly combining a normal distribution function for the corresponding plurality of residence time information;
A customized score calculator configured to calculate a cumulative distribution function by accumulating the estimated distribution of residence time, and calculating a customized score obtained by converting the residence time into a relative score for each content user using the calculated cumulative distribution function; And
Contents using deep learning based on GPU computation, comprising: a preference determining unit that generates a preference graph for a plurality of contents for each content user based on the calculated customized score and determines personal preference of the corresponding content. Recommended device. The method of claim 1, wherein the GPU operation performing unit,
A content recommendation apparatus using deep learning based on GPU computation, wherein a value of at least one content user propensity vector u is input to a weight matrix generated using multiple dense layers in performing the calculation. The method of claim 1,
When the content is text-oriented, a content recommendation device using deep learning based on GPU calculation, characterized in that a dictionary is formed and a word to vector model is learned for vectorization of texts included in the content. . The method of claim 3,
Texts included in the content are crawled and registered in a dictionary for each sentence, and exception words stored in advance are deleted by parsing the sentences registered in the dictionary, and only pre-stored parts of speech in a predetermined format are extracted, and the stored sentences are used as words. A content recommendation device using deep learning based on GPU computation, characterized in that it generates a word to vector model. The method of claim 1, wherein the GPU operation performing unit,
When the content is text-oriented, a document processing CNN model is used, and when the content is a video or music owner, a video or music processing CNN model is used, and a corresponding CNN model is used for each content type. Content recommendation device using deep learning based on GPU computation. The method of claim 1,
A content recommendation device using deep learning based on GPU computation, further comprising: a content recommendation unit that predicts a content evaluation score for a content having the highest content preference and recommends a content whose highest evaluation score is predicted. delete The method of claim 1, wherein the customized score calculation unit,
A content recommendation device using deep learning based on GPU computation, characterized in that the calculated cumulative distribution function is transformed into a quartile to calculate a customized score. The method of claim 1, wherein the preference determination unit,
A content recommendation device using deep learning based on GPU calculation, characterized in that, as the calculated customized score is higher, it is determined as the content preferred by each content user. The method of claim 1, wherein the residence time distribution estimation unit,
Content recommendation apparatus using deep learning based on GPU computation, characterized in that as the number of contents used by the contents user increases, the estimated distribution of residence time follows a distribution corresponding to an actual result. Using a deep learning model consisting of a neural network operated by a processor,
Calculating a content user propensity vector u that vectorizes the propensity of content users and a content propensity vector v that vectorizes the content propensity;
Performing an operation by inputting the calculated u and v to a preset loss function in order to implement a complete deep learning structure so that computation is possible using only a GPU;
Predicting a content evaluation score for a specific content for each content user based on the calculation result; And
Including; collecting residence time information for each content used by each content user, converting the collected stay time into a relative score, determining personal preference of the corresponding content, and providing content preference for each content user; including,
Providing the content preference,
Collecting information on a residence time for each content used by a content user;
Correlating a normal distribution function to each of the collected information on the residence time for each content;
Estimating a residence time distribution of the content user by linearly combining a normal distribution function for the corresponding plurality of residence time information;
Calculating a cumulative distribution function by accumulating the estimated distribution of dwell time, and calculating a customized score obtained by converting the dwell time into a relative score for each content user by using the calculated cumulative distribution function; And
Generating a preference graph for a plurality of contents for each content user based on the calculated customized score, determining a personal preference of the corresponding content; and a content recommendation method using deep learning based on GPU computation, comprising: . The method of claim 11, wherein calculating the content propensity vector v,
A content recommendation method using deep learning based on GPU calculation, characterized in that at least one content user propensity vector u is input to a weight matrix generated using multiple dense layers in performing the calculation. The method of claim 11,
When the content is text-oriented, a content recommendation method using deep learning based on GPU computation, characterized in that a dictionary is formed and a word to vector model is learned for vectorization of texts included in the content. . The method of claim 13,
Texts included in the content are crawled and registered in a dictionary for each sentence, and exception words stored in advance are deleted by parsing the sentences registered in the dictionary, and only pre-stored parts of speech in a predetermined format are extracted, and the stored sentences are used as words. A content recommendation method using deep learning based on GPU computation, characterized by generating a word to vector model. The method of claim 11, wherein calculating the content propensity vector v,
When the content is text-oriented, a document processing CNN model is used, and when the content is a video or music owner, a video or music processing CNN model is used, and a corresponding CNN model is used for each content type. Content recommendation method using deep learning based on GPU computation. The method of claim 11,
A content recommendation method using deep learning based on GPU computation, further comprising: predicting a content evaluation score for a content having the highest content preference and recommending a content for which the highest evaluation score is predicted. delete The method of claim 11, wherein calculating the customized score,
A content recommendation method using deep learning based on GPU calculation, characterized in that the calculated cumulative distribution function is transformed into a quartile to calculate a customized score. The method of claim 11, wherein determining the personal preference,
A content recommendation method using deep learning based on GPU computation, characterized in that, as the calculated customized score is higher, it is determined as the content preferred by each content user. The method of claim 11, wherein estimating the residence time distribution,
A content recommendation method using deep learning based on GPU computation, characterized in that as the number of contents used by the content user increases, the estimated distribution of residence time follows a distribution corresponding to an actual result.

Images