KR102195894B1 - Hybrid item recommending method and apparatus - Google Patents

Abstract

According to an embodiment of the present invention, a method for recommending a hybrid item includes generating a rating vector based on a result of a plurality of users evaluating at least one item among a plurality of items; Generating a preference vector including predicted preferences of the plurality of users for each of a plurality of item groups in which the plurality of items are classified according to a predetermined criterion based on the rating vector; Generating an output vector including predicted ratings of the plurality of users for each of the plurality of items by using the rating vector and the preference vector; And recommending a predetermined number of recommended items from among the plurality of items to at least one user among the plurality of users based on the output vector.

Description

Hybrid item recommendation method and device {HYBRID ITEM RECOMMENDING METHOD AND APPARATUS}

The present invention relates to a method and apparatus for recommending a hybrid item, and more particularly, to a method and apparatus for recommending a hybrid item using a rating given by a user to an item and a preference for each content including the item.

The recommendation system is essentially used in most online shopping malls today to promote users' purchases, and is being actively studied in academia. Recommendation algorithms can be largely divided into collaborative filtering, content-based recommendation, and hybrid recommendation algorithm.

Collaborative filtering analyzes the user's taste and product propensity based on the user's rating, and then predicts the rating to be given to the products that the user has not yet purchased or the top N items worth purchasing. Content-based recommendation is a method of recommending a product purchased by a user and other products having similar content information. The hybrid recommendation method uses both rating information of a user and content information of a product purchased by the user.

However, existing hybrid recommendation techniques simply use the known content information of the product, and a study to predict how much preference each user will show to the content of each product, and to use it as additional information for recommendation. Has not been done.

Korean Patent Publication No. 10-2014-0027866 (2015.12.04)

An object of the present invention is to provide a method and apparatus for recommending items in a hybrid method using a rating given by a user to an item and a preference for each content including the item.

In addition, the present invention is to provide a method and apparatus for recommending items in a hybrid method based on deep learning.

The problem to be solved by the present invention is not limited to the problem(s) mentioned above, and another problem(s) not mentioned will be clearly understood by those skilled in the art from the following description.

In order to achieve the above object, a method for recommending a hybrid item according to an embodiment of the present invention includes: generating a rating vector based on a result of a plurality of users evaluating at least one item among a plurality of items; Generating a preference vector including predicted preferences of the plurality of users for each of a plurality of item groups in which the plurality of items are classified according to a predetermined criterion based on the rating vector; Generating an output vector including predicted ratings of the plurality of users for each of the plurality of items by using the rating vector and the preference vector; And recommending a predetermined number of recommended items from among the plurality of items to at least one user among the plurality of users based on the output vector.

Preferably, the preference vector and the output vector may be generated using a deep learning model.

Preferably, the preference vector may be generated based on at least one of an average and frequency of ratings given to an item by the plurality of users.

Preferably, the step of generating a preference vector including predicted preferences of the plurality of users comprises generating preference data including preferences of each of the plurality of users for each of the plurality of item groups using the rating vector. Step to do; Learning a first autoencoder using the preference data; And generating the preference vector based on the first autoencoder and the preference data.

Preferably, generating an output vector including predicted ratings of the plurality of users includes: learning a second autoencoder using the rating vector and the preference vector; And generating the output vector based on the second autoencoder, the rating vector, and the preference vector.

In addition, a hybrid item recommendation apparatus according to an embodiment of the present invention for achieving the above object includes: an evaluation unit generating a rating vector based on a result of a plurality of users evaluating at least one item among a plurality of items; Generates a preference vector including predicted preferences of the plurality of users for each of a plurality of item groups in which the plurality of items are classified according to a predetermined criterion based on the rating vector, and uses the rating vector and the preference vector A generator configured to generate an output vector including predicted scores of the plurality of users for each of the plurality of items; And a recommendation unit for recommending a predetermined number of recommended items from among the plurality of items to at least one user of the plurality of users based on the output vector.

Preferably, the preference vector and the output vector may be generated using a deep learning model.

Preferably, the generator may generate a preference vector based on at least one of an average and a frequency of ratings given by the plurality of users to the item.

Preferably, the generation unit generates preference data including preferences of each of the plurality of users for each of the plurality of item groups using the rating vector, and learns a first autoencoder using the preference data And, based on the first autoencoder and the preference data, the preference vector may be generated.

Preferably, the generator may learn a second autoencoder using the rating vector and the preference vector, and generate the output vector based on the second autoencoder, the rating vector, and the preference vector. .

According to the present invention, in a method and apparatus for recommending an item, the accuracy of recommendation can be improved by using both rating information on a user's item and preference information for each content including the item.

In addition, the present invention has an effect of improving the accuracy of recommendation by recommending items in a hybrid manner based on deep learning.

1 is a flowchart illustrating a hybrid item recommendation method according to an embodiment of the present invention.
2 is a flowchart illustrating a method of generating a preference vector according to an embodiment of the present invention.
3 is a flowchart illustrating a method of generating an output vector according to an embodiment of the present invention.
4 and 5 are configuration diagrams of a conventional autorec model and a hybrid autorec model.
6 is a block diagram of a hybrid item recommendation apparatus according to an embodiment of the present invention.

In the present invention, various modifications may be made and various embodiments may be provided, and specific embodiments will be illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to a specific embodiment, it is to be understood to include all changes, equivalents, and substitutes included in the spirit and scope of the present invention. In describing each drawing, similar reference numerals have been used for similar elements.

Terms such as first, second, A, and B may be used to describe various components, but the components should not be limited by the terms. These terms are used only for the purpose of distinguishing one component from another component. For example, without departing from the scope of the present invention, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element. The term and/or includes a combination of a plurality of related listed items or any of a plurality of related listed items.

When a component is referred to as being "connected" or "connected" to another component, it is understood that it may be directly connected or connected to the other component, but other components may exist in the middle. Should be. On the other hand, when a component is referred to as being "directly connected" or "directly connected" to another component, it should be understood that there is no other component in the middle.

The terms used in the present application are used only to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In the present application, terms such as "comprise" or "have" are intended to designate the presence of features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, but one or more other features. It is to be understood that the presence or addition of elements or numbers, steps, actions, components, parts, or combinations thereof, does not preclude in advance.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms as defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and should not be interpreted as an ideal or excessively formal meaning unless explicitly defined in this application. Does not.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a flowchart illustrating a hybrid item recommendation method according to an embodiment of the present invention.

In step S110, the hybrid item recommendation apparatus generates a rating vector based on a result of a plurality of users evaluating at least one item among a plurality of items.

In this case, the evaluation result is information including an evaluation result value obtained by evaluating at least one of the plurality of items by each of the plurality of users, and there may be no rating, that is, an unevaluated item in the rating vector.

For example, the evaluation result may be information including an evaluation result value evaluated for each of the N movies by each of the M users in an internet movie site in which M users and N movies exist. However, in reality, it is extremely unlikely that all of the M users watch and rate all N movies. Accordingly, the evaluation result may include evaluation result values for N or less movies watched for each user, but may not include evaluation result values for the remaining movies that were not watched.

Meanwhile, the rating vector may be generated based on an evaluation result value obtained by evaluating at least one item among a plurality of items by each of a plurality of users.

For example, in an internet movie site having 4 users a1 to a4, four movies are provided, and when these movies are s1 to s4, respectively, the rating of the movie for each user may be illustrated as shown in Table 1 below.

s1 s2 s3 s4 a1 3.0 5.0 2.0 1.0 a2 N/A 2.0 3.0 N/A a3 5.0 5.0 2.0 N/A a4 N/A 2.0 3.0 4.0

In this case, the rating vector may be composed of values in Table 1, and N/A is a rating for a movie that the user has not rated.

In step S120, the hybrid item recommendation apparatus generates a preference vector including predicted preferences of a plurality of users for each of a plurality of item groups in which a plurality of items are classified according to a predetermined criterion based on the rating vector.

That is, the hybrid item recommendation apparatus may generate a preference vector by calculating predicted preferences of a plurality of users for each of a plurality of item groups using the rating vector. Meanwhile, the apparatus for recommending a hybrid item may include a predicted preference even for an item group in which no data evaluated by a user exists in the preference vector. For example, the hybrid item recommendation apparatus may calculate the predicted preference by using preference information of a user other than any user.

In this case, the predetermined criterion may be a criterion for classifying a category in which an item is classified. For example, assuming that the item is a movie, the criterion may be a genre, and the category may be a horror, a thriller, a drama, a documentary, an action, and the like.

Meanwhile, one item may be included in at least one item group. That is, one item may be included in only one item group, but it may be included in two or more item groups.

For example, if each of the three users rated different movies t1 to t3 classified into genres such as horror, thriller, drama, documentary, action, t1 is horror, thriller, t2 is thriller, drama, t3 is drama , If applicable to the documentary, t1 may be included in the item group corresponding to fear, t1 and t2 in the item group corresponding to the thriller, t2 and t3 in the item group corresponding to the drama, and t3 may be included in the item group corresponding to the documentary.

In another embodiment, the preference vector may be generated based on at least one of an average and frequency of ratings given to an item by a plurality of users.

For example, in the hybrid item recommendation device, a user assigned 5 points to a “Harry Potter” movie, 4 points to a “Star Wars” movie, and 1 point to a “Murder Memories” movie, and the genre of “Harry Potter” is “Fantasy,” Assuming that the genre of “Drama” and “Star Wars” is “Fantasy, Action”, and the genre of “Murder Memories” is “Horror, Drama”, the user's preference for each genre is “Fantasy” when calculated as an average. It can be predicted as 4.5 points for “Action”, 3 points for “Drama” and 1 point for “Fear”. The above is an example of an average, and a frequency may be used instead of the average, or a numerical value obtained by linearly combining the average and the frequency may be used, but the present invention is not limited thereto.

In step S130, the hybrid item recommendation apparatus generates an output vector that is a predicted rating of a plurality of users of each of the plurality of items by using the rating vector and the preference vector.

That is, the hybrid item recommendation apparatus may generate an output vector by calculating predicted scores of each of a plurality of users for each of a plurality of items. In this case, the hybrid item recommendation apparatus may use not only the rating vectors of the plurality of users, but also the preference vectors for the item groups of the plurality of users. Meanwhile, the hybrid item recommendation apparatus may include a predicted rating even for items not evaluated by the user in the output vector.

Meanwhile, a detailed method of generating the output vector will be described later in detail in the description of FIG. 3.

Finally, in step S140, the hybrid item recommending apparatus recommends a predetermined number of recommended items from among the plurality of items to at least one of the plurality of users based on the output vector.

For example, the hybrid item recommendation apparatus may sort items not evaluated by a certain user according to the order of predicted ratings using the output vector. In addition, the hybrid item recommending apparatus may recommend a predetermined number of recommended items to a certain user according to the sorted order.

In this case, the hybrid item recommendation device may recommend to at least one user among a plurality of users, and the predetermined number may be a preset number.

2 is a flowchart illustrating a method of generating a preference vector according to an embodiment of the present invention.

In step S210, the hybrid item recommendation apparatus generates preference data, which is preference information of a plurality of users for a plurality of item groups, using the rating vector.

For example, the hybrid item recommendation device may generate preference data using a rating vector. In this case, there may be an item group in which the preference does not exist in the preference data.

More specifically, it may be assumed that a user evaluates four movies s1 to s4 in an Internet movie site having four users a1 to a4. At this time, if the genre of s1 to s4 is at least one of fantasy, drama, action, horror and documentary, s1 to s4 may be included in at least one of a total of five item groups classified into fantasy, drama, action, horror, and documentary. . In this case, the hybrid item recommendation apparatus may generate preference data for each item group of each of the plurality of users by calculating the average rating of movies included in the item group.

The result can be exemplified as in Table 2 below, and N/A represents the preference of the item group in which the user's preference does not exist.

fantasy drama action fear Documentary a1 4 3 3 1.5 1.5 a2 2 N/A 2 3 3 a3 5 5 5 2 2 a4 2 2 4 3 3.5

In step S220, the hybrid item recommendation apparatus learns the first autoencoder using the preference data.

That is, the apparatus for recommending a hybrid item may train the first autoencoder to produce an output similar to the preference data as possible by receiving the preference data. In this case, the hybrid item recommendation device may repeatedly learn the first autoencoder.

Finally, in step S230, the hybrid item recommendation apparatus generates preference vectors, which are preferences of a plurality of users, based on the first autoencoder and preference data.

That is, the hybrid item recommendation apparatus may input preference data to the first autoencoder in which the learning is completed, and obtain a preference vector as an output. In this case, a prediction preference may be input to the preference vector for an item group in which the user's preference did not exist (that is, indicated by N/A in Table 2).

3 is a flowchart illustrating a method of generating an output vector according to an embodiment of the present invention.

In step S310, the hybrid item recommendation apparatus learns the second autoencoder using the rating vector and the preference vector.

That is, the apparatus for recommending a hybrid item may train the second autoencoder to produce an output that is as similar as possible to the rating vector and the preference vector by inputting the rating vector and the preference vector. In this case, the hybrid item recommendation apparatus may repeatedly learn the second autoencoder.

Finally, in step S320, the hybrid item recommendation apparatus generates output vectors that are predicted scores of a plurality of users based on the second auto-encoder, the rating vector, and the preference vector.

That is, the hybrid item recommendation apparatus may input the rating vector and the preference vector to the second autoencoder in which the learning is completed, and obtain an output vector as an output. In this case, a prediction score may be input to the output vector even for items in which the user's preference did not exist (ie, indicated by N/A in Table 1). As a result, the hybrid item recommendation device can recommend a recommended item to a user by using a predicted rating for an item not evaluated by the user.

Experimental Example: Item Recommendation Accuracy Comparison Data

MovieLens 100K data was used to evaluate the accuracy.

The data consisted of 943 users, 1682 movies, and 100,000 ratings. Of the 100,000 ratings, a randomly selected 80% rating was used as data for deep learning model training, and the remaining 20% data was used as data for evaluating recommendation accuracy.

The existing U-AutoRec, I-AutoRec, Hybrid AutoEncoder models and the proposed model were trained using training data, respectively, and the top-5 recommendation accuracy of each model was calculated and compared using the evaluation data. The accuracy of the recommended method was also compared.

As a measure for calculating accuracy, precision, recall, nDCG and MRR, which are widely used in the past, were used.

As shown in Table 3 below, it can be seen that the hybrid item recommendation method according to an embodiment of the present invention always performs more accurate recommendation than other existing models.

This is because the hybrid item recommendation method according to an embodiment of the present invention predicts the preferences that users may have for each content (i.e., corresponding to an item group) and utilizes this for deep learning model learning, unlike existing item recommendation methods. to be.

Precision Recall nDCG MRR Popularity 0.081 0.102 0.117 0.204 U-AutoRec 0.141 0.170 0.202 0.308 I-AutoRec 0.127 0.141 0.172 0.271 HybridAE 0.138 0.156 0.181 0.274 The present invention 0.153 0.177 0.209 0.311

Deep Learning Model Learning Method: Auto Encoder Operation Mechanism

In an embodiment of the present invention, training a deep learning model of the first and second autoencoders may be based on a conventional AutoRec model.

4 is a configuration diagram of a conventional autorec model, and FIG. 5 is a configuration diagram of a conventional hybrid autorec model.

In a general operation method of Autorec, Autorec encodes the user's rating vector into a hidden layer, and then decodes the encoded information again. In this case, the encoding model parameter and the decoding model parameter may be learned so that the decoded result becomes the same as the initially inputted rating vector. When the learning of the encoding model parameter and the decoding model parameter is completed through several iterations, Autorec encodes the user's rating vector and decodes it again, so that the empty result value of each user's rating vector (i.e., unrated The result of the item) can be predicted and calculated.

In addition, the conventional hybrid autorec model simply uses the known content information of the item, predicts how much preference each user will show for the content of each item (ie, corresponding to the item group), and recommends it. It was not intended to be used as additional information.

For example, in the conventional hybrid autorec model, if the genre of a movie is "Action" or "Fantasy", the content information is expressed in the form of a vector whose dimension is 1 and the rest is 0, and simply I was just using it. However, in the hybrid item recommendation method according to an embodiment of the present invention, there is a difference in that a user's content preference is specifically predicted and used for recommendation, and the difference is shown in Table 3 above.

6 is a block diagram of a hybrid item recommendation apparatus according to an embodiment of the present invention.

Referring to FIG. 6, a hybrid item recommendation apparatus 600 according to an embodiment of the present invention includes an evaluation unit 610, a generation unit 620, and a recommendation unit 630.

Meanwhile, the hybrid item recommendation device 600 may be mounted on various computing devices such as a PC, a server, a smart phone, and a tablet, and used to recommend an item.

The evaluation unit 610 generates a rating vector based on a result of a plurality of users evaluating at least one item among a plurality of items.

The generation unit 620 generates a preference vector, which is a preference of a plurality of users for a plurality of item groups in which a plurality of items are classified according to a predetermined criterion based on the rating vector, and uses the rating vector and the preference vector to generate a plurality of preference vectors. An output vector, which is the predicted rating of a plurality of users of each item, is generated.

In another embodiment, the generation unit 620 generates a preference vector based on at least one of an average and frequency of ratings given by a plurality of users to an item.

In another embodiment, the generation unit 620 generates preference data, which is preference information of a plurality of users for a plurality of item groups, using a rating vector, and learns a first autoencoder using the preference data, A preference vector, which is a preference of a plurality of users, is generated based on the first autoencoder and preference data.

In another embodiment, the generator 620 learns a second autoencoder using a rating vector and a preference vector, and an output vector that is a preference of a plurality of users based on the second autoencoder, the rating vector and the preference vector. Create

The recommending unit 630 recommends a predetermined number of recommended items from among the plurality of items to at least one user among the plurality of users based on the output vector.

In another embodiment, the preference vector and the output vector may be generated using a deep learning model.

The above-described embodiments of the present invention can be written in a program that can be executed on a computer, and can be implemented in a general-purpose digital computer that operates the program using a computer-readable recording medium.

The computer-readable recording medium includes a magnetic storage medium (for example, a ROM, a floppy disk, a hard disk, etc.), and an optical reading medium (for example, a CD-ROM, a DVD, etc.).

So far, the present invention has been looked at around its preferred embodiments. Those of ordinary skill in the art to which the present invention pertains will be able to understand that the present invention can be implemented in a modified form without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments should be considered from an illustrative point of view rather than a limiting point of view. The scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the scope equivalent thereto should be construed as being included in the present invention.

600: hybrid item recommendation device
610: evaluation unit
620: generation unit
630: recommendation

Claims (10)

Generating a rating vector based on a result of a plurality of users evaluating at least one of the plurality of items;
Using a deep learning model, the first auto is generated by generating preference data including predicted preferences of the plurality of users for each of a plurality of item groups in which the plurality of items are classified according to a predetermined criterion based on the rating vector. Generating a preference vector by learning an encoder;
Generating an output vector including predicted ratings of the plurality of users for each of the plurality of items by learning a second autoencoder using a deep learning model, the rating vector and the preference vector; And
Recommending a predetermined number of recommended items from among the plurality of items to at least one user of the plurality of users based on the output vector
Hybrid item recommendation method comprising a. delete The method of claim 1,
The preference vector is,
The hybrid item recommendation method, characterized in that it is generated based on at least one of an average and a frequency of ratings given to items by the plurality of users. The method of claim 1,
Generating the preference vector,
And generating the preference vector based on the first autoencoder and the preference data. The method of claim 1,
Generating an output vector including the predicted ratings of the plurality of users,
And generating the output vector based on the second autoencoder, the rating vector, and the preference vector. An evaluation unit generating a rating vector based on a result of a plurality of users evaluating at least one item among the plurality of items;
Using a deep learning model, the first auto is generated by generating preference data including predicted preferences of the plurality of users for each of a plurality of item groups in which the plurality of items are classified according to a predetermined criterion based on the rating vector. Learning an encoder to generate a preference vector, using a deep learning model, and learning a second autoencoder using the rating vector and the preference vector to include predicted ratings of the plurality of users for each of the plurality of items A generator for generating an output vector to be determined; And
A recommendation unit for recommending a predetermined number of recommended items from among the plurality of items to at least one user of the plurality of users based on the output vector
Hybrid item recommendation device comprising a. delete The method of claim 6,
The generation unit,
And generating a preference vector based on at least one of an average and a frequency of ratings given to an item by the plurality of users. The method of claim 6,
The generation unit,
And generating the preference vector based on the first autoencoder and the preference data. The method of claim 6,
The generation unit,
And generating the output vector based on the second autoencoder, the rating vector, and the preference vector.

Images