KR20200103148A - Method and apparatus for recommending items based on multi-type pair-wise preference - Google Patents

Abstract

Disclosed is a method for recommending items based on multi-type pair-wise preference, capable or recommending items to users with high precision. A method of recommending an item using an evaluation result, which is a result of evaluating at least one item among a plurality of items by a plurality of users, according to an embodiment of the present invention includes the steps of: generating a comprehensive result by synthesizing a prediction result, which is a result of predicting an evaluation result of a plurality of unevaluated items included in the evaluation result, and the evaluation result using a collaborative filtering technique; classifying the plurality of unevaluated items into a plurality of non-information items and a plurality of indifferent items from the synthesis result; learning a model parameter based on the synthesis result and a plurality of item pairs selected from among item pairs consisting of two of an evaluation item, an unevaluated item, a no-information item, and an indifferent item included in the synthesis result; and recommending a predetermined number of items to at least one recommendation target among the plurality of users by using a final prediction result, which is a result of predicting evaluation results for the plurality of unevaluated items based on the learned model parameters.

Description

Method and device for recommending items based on multi-type pairwise {METHOD AND APPARATUS FOR RECOMMENDING ITEMS BASED ON MULTI-TYPE PAIR-WISE PREFERENCE}

The present invention relates to a method and apparatus for dividing an item to be evaluated into an evaluated item, an unevaluated item, an uninformed item, and an indifference item, and recommending an item to a user using a multi-type item pair based thereon. will be.

Collaborative filtering (CF) is a technology that recommends the top N items to users by using past user actions such as explicit evaluation results such as ratings and implicit evaluation results such as clicks. This technology has already been applied to the system.

In recent years, as the demand to process datasets related to single class setting increases, a one-class collaborative filtering (OCCF) technique has been proposed. OCCF techniques can be divided into two types according to the method of learning user preferences (ie, evaluation results).

First, it is a method of minimizing the error between the actual evaluation result and the predicted evaluation result of an individual user on an individual item (point-wise) basis. Second, the difference between the evaluation result of the item evaluated in the item pair of the individual user and the evaluation result of the predicted item is maximized in an item pair (pair-wise) unit.

Here, the second method is called a single-type pair-wise approach, and is evaluated to exhibit superior performance in predicting an evaluation result of an item than the first method.

However, the single type pairwise-based approach assumes that the user has the same degree of negative preference for all unrated items (P1) and that the user prefers the rated items over all unrated items (P2). Is based on. However, since these assumptions do not always correspond to reality, it is true that there is a possibility of introducing errors in item recommendation.

Accordingly, there is a need for a method and apparatus for recommending items based on a multi-type pairwise to solve this problem.

R. Steffen, F. Christoph, G. Zeno, and S.-T. Lars. 2009. BPR: Bayesian Personalized Ranking from Implicit Feedback. In Proc. of the 25th Int'l Conf. on Uncertainty in Artificial Intelligence (UAI). 452-461.

The present invention divides items into evaluated items, unevaluated items, no-information items, and indifferent items, and recommends items with higher accuracy to users by using model parameters learned based on multiple types of item pairs. It is intended to provide a method and apparatus.

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, the multi-type pairwise-based item recommendation method according to an embodiment of the present invention is a method of recommending an item using an evaluation result that is a result of a plurality of users evaluating at least one item among a plurality of items. A method comprising: generating a comprehensive result by synthesizing a prediction result, which is a result of predicting an evaluation result of a plurality of unevaluated items included in the evaluation result, using a collaborative filtering technique and the evaluation result; Classifying the plurality of unevaluated items into a plurality of non-information items and a plurality of indifferent items from the synthesis result; Learning a model parameter based on the synthesis result and a plurality of selection item pairs selected from among item pairs consisting of two of an evaluation item, an unevaluated item, an uninformed item and an item of no interest included in the synthesis result; And recommending a predetermined number of items to at least one recommendation target among the plurality of users using a final prediction result that is a result of predicting evaluation results for the plurality of unevaluated items based on the learned model parameters. Includes.

Preferably, the learning of the model parameters comprises: generating an objective function based on a difference in evaluation results between item types included in each of the plurality of selected item pairs; Acquiring a plurality of evaluation result information including evaluation results for each item type included in the plurality of selected item pairs from the synthesis result for one selected user selected from among the plurality of users; Updating the model parameter value to maximize the value of the objective function using the objective function and the plurality of evaluation result information; And repeating the steps of obtaining the plurality of evaluation result information and updating the model parameter values until a predetermined completion condition is satisfied.

Preferably, the updating of the model parameter values comprises: calculating a gradient value of the objective function using the plurality of evaluation result information; And updating the model parameter value by applying the gradient value based on a stochastic gradient descent (SGD) technique.

Preferably, the plurality of evaluation result information may include one evaluation result for each item type included in the plurality of selection item pairs.

Preferably, the plurality of evaluation result information includes a group evaluation result that is a plurality of evaluation results for each item type included in the plurality of selected item pairs, and the objective function may be based on a difference between the group evaluation results. .

Preferably, the objective function is that the evaluation result of the evaluation item is higher than the evaluation result of the unevaluated item, the no-information item and the indifferent item, and the evaluation result of the no-information item is higher than the evaluation result of the indifferent item. It can be based on assumptions.

Preferably, the plurality of selection item pairs include an item pair consisting of the evaluation item and the unevaluated item and an item pair consisting of the no-information item and the indifferent item, or the evaluation item and the no-information item It may include an item pair composed of, an item pair composed of the evaluation item and the indifferent item, and an item pair composed of the no-information item and the indifferent item.

In addition, the multi-type pairwise-based item recommendation apparatus according to an embodiment of the present invention for achieving the above object recommends an item using an evaluation result that is a result of a plurality of users evaluating at least one item among a plurality of items. What is claimed is: 1. An apparatus comprising: a generator configured to generate a comprehensive result by synthesizing a prediction result, which is a result of predicting an evaluation result of a plurality of unevaluated items included in the evaluation result, and the evaluation result using a collaborative filtering technique; A classification unit for classifying the plurality of unevaluated items into a plurality of non-information items and a plurality of non-interested items in the synthesis result; A learning unit that learns model parameters based on the synthesis result and a plurality of selection item pairs selected from among item pairs consisting of two of an evaluation item, an unevaluated item, a no-information item, and an indifferent item included in the synthesis result; And a recommendation for recommending a predetermined number of items to at least one recommendation target among the plurality of users using a final prediction result that is a result of predicting evaluation results for the plurality of unevaluated items based on the learned model parameters. Includes wealth.

Preferably, the learning unit generates an objective function based on an evaluation result difference between item types included in each of the plurality of selected item pairs, and for one selected user selected among the plurality of users, from the synthesis result Obtaining a plurality of evaluation result information including evaluation results for each item type included in a plurality of selected item pairs, and using the objective function and the plurality of evaluation result information, the model parameter value so that the value of the objective function is maximized. The process of acquiring the plurality of evaluation result information and the process of updating the model parameter value may be repeated until a predetermined completion condition is satisfied.

Preferably, when the learning unit updates the model parameter value, the gradient value of the objective function is calculated using the plurality of evaluation result information, and the gradient value is calculated based on a stochastic gradient descent (SGD) technique. By applying, the model parameter value can be updated.

Preferably, the plurality of evaluation result information may include one evaluation result for each item type included in the plurality of selection item pairs.

Preferably, the plurality of evaluation result information includes a group evaluation result that is a plurality of evaluation results for each item type included in the plurality of selected item pairs, and the objective function may be based on a difference between the group evaluation results. .

Preferably, the objective function is that the evaluation result of the evaluation item is higher than the evaluation result of the unevaluated item, the no-information item and the indifferent item, and the evaluation result of the no-information item is higher than the evaluation result of the indifferent item. It can be based on assumptions.

Preferably, the plurality of selection item pairs include an item pair consisting of the evaluation item and the unevaluated item and an item pair consisting of the no-information item and the indifferent item, or the evaluation item and the no-information item It may include an item pair composed of, an item pair composed of the evaluation item and the indifferent item, and an item pair composed of the no-information item and the indifferent item.

The present invention divides the item to be evaluated into an evaluated item, an unevaluated item, an uninformed item, and an indifference item, and uses model parameters learned based on multiple types of item pairs. It has the effect of recommending items.

In addition, according to the present invention, when a model parameter is trained using multiple types of item pairs, evaluation results for the same type of item are grouped and used, so that an item can be recommended to a user with higher accuracy.

FIG. 1 is a diagram for explaining (a) a single type pairwise preference and (b) a multiple type pairwise preference.
2 is a flowchart illustrating a method of recommending a multi-type pairwise-based item according to an embodiment of the present invention.
3 is a flowchart illustrating a method of learning model parameters according to an embodiment of the present invention.
4 is a block diagram illustrating a multi-type pairwise-based item recommendation apparatus according to an embodiment of the present invention.
5 is a diagram illustrating a pseudocode of a model parameter learning algorithm according to an embodiment of the present invention.
6 and 7 are block diagrams for explaining the effect of a multi-type pairwise-based item recommendation method 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 elements, but the elements 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 only used 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.

In the present invention, when there is an evaluation result (e.g., a matrix) that is a result of evaluating at least one of a plurality of items subject to evaluation by a plurality of users, the item evaluated by a user in the evaluation result is a rated item It is defined as, and items not rated by the user are defined as unrated items. Here, among the unevaluated items, an item that is expected to have a low evaluation result by the user is defined as an uninteresting item, and among the unevaluated items, the remaining items except for the indifference item are items that are difficult to predict the preference. item).

Meanwhile, the present invention can be mainly applied to a single class environment based on implicit feedback (eg, click). However, the present invention is not limited thereto, and can be applied to a multi-class environment based on explicit feedback (eg, rating) through appropriate data transformation and processing.

FIG. 1 is a diagram for explaining (a) a single type pairwise preference and (b) a multiple type pairwise preference.

First, referring to FIG. 1(a), a single type pairwise preference according to a conventional technique (ie, Bayesian Personalized Ranking, BPR) is shown. At this time, in the conventional technology, a single type of pair is composed of an evaluation item and an unevaluated item, and under the premise that the evaluation result for the evaluation item is higher than the evaluation result for the unevaluated item (PP), the model parameter is to maximize the difference. Can be learned.

Further, referring to FIG. 1(b), a multi-type pairwise preference according to the present invention is shown. At this time, in the present invention, not only a pair of evaluation items and unevaluated items, but also a pair of evaluation items and non-information items, a pair of evaluation items and indifference items, and a pair of non-information items and indifference items. have. In addition, in the present invention, the premise that the evaluation result for the evaluation item is higher than the evaluation result for unevaluated items, no information items, and indifferent items (PP1, PP2, PP3) and the evaluation results for no information items are evaluated for indifferent items. Under the premise of higher than the result (PP4), model parameters can be trained to maximize the difference. At this time, i ₁ , i ₂ , i ₃ are evaluation items, k ₁ and k ₂ are non-information items, and l ₁ , l ₂ , l ₃ and l ₄ are indifferent items.

2 is a flowchart illustrating a method of recommending a multi-type pairwise-based item according to an embodiment of the present invention.

In step S210, the item recommendation device generates a comprehensive result by synthesizing a prediction result, which is a result of predicting the evaluation result of a plurality of unevaluated items included in the evaluation result, using a collaborative filtering technique and the evaluation result.

In this case, the item recommendation device may roughly predict the user's evaluation result for the unevaluated item using a collaborative filtering technique. Preferably, the item recommendation device may use a weighted regularized matrix factorization (WRMF) technique. Meanwhile, WRMF is a well-known OCCF technique, and can represent unevaluated items included in the prediction result as 1 (positive) or 0 (negative).

In addition, the item recommendation device may generate a comprehensive result by synthesizing the evaluation result and the prediction result. For example, the overall result is a sum of the evaluation result of the evaluation item included in the evaluation result and the evaluation result of the unevaluated item included in the prediction result, and may be in the form of a matrix.

In step S220, the item recommendation device classifies the plurality of unevaluated items into non-information items and indifferent items from the aggregate result.

For example, the item recommendation device may classify unevaluated items of a predetermined ratio having a low predicted evaluation result through a collaborative filtering technique as indifferent items. In addition, the item recommendation apparatus may classify the rest of the plurality of unevaluated items excluding a plurality of indifferent items into a plurality of non-information items.

In step S230, the item recommendation device is based on the overall result and a plurality of selected item pairs selected from item pairs consisting of two of an evaluation item, an unevaluated item, a no-information item, and an indifference item included in the total result, Train model parameters.

That is, the item recommendation device may learn model parameters used in the item recommendation method by using the aggregate result and a plurality of selected item pairs.

At this time, the item pair is composed of two items selected from evaluation items, unevaluated items, no information items, and indifferent items. For example, the item pair may be {evaluated item, unevaluated item}, {evaluated item, no information item}, {evaluated item, indifferent item}, {no information item, indifferent item}, etc.

Meanwhile, details of the item recommendation device learning model parameters will be described in detail later in the description of FIG. 3.

In another embodiment, the plurality of selection item pairs may include an item pair consisting of an evaluation item and an unevaluated item, and an item pair consisting of an uninformation item and an indifference item.

In addition, the plurality of selection item pairs may include an item pair consisting of an evaluation item and a non-information item, an item pair consisting of an evaluation item and an indifference item, and an item pair consisting of non-information items and indifferent items.

That is, the item recommendation device may select a plurality of selection item pairs in order to more efficiently and accurately recommend an item to a user.

To this end, the plurality of selection item pairs may include {evaluated items, unevaluated items}, and {no information items, indifferent items}. In addition, the plurality of selection item pairs may include {evaluation item, no information item}, {evaluation item, indifferent item}, and {no information item, indifferent item}.

Finally, in step S240, the item recommendation device uses the final prediction result, which is a result of predicting evaluation results for a plurality of unevaluated items based on the learned model parameters, to at least one recommendation target among the plurality of users. Recommend a number of items.

That is, the item recommendation device may generate a final prediction result by predicting evaluation results for a plurality of unevaluated items by using the learned model parameters. The final prediction result is a result predicted more accurately than the above-described prediction result, and the item recommendation device may recommend a high-order predetermined number of items to a recommendation target by using the final prediction result.

As described above, the multi-type pairwise-based item recommendation method according to an embodiment of the present invention divides a plurality of items into evaluated items, unevaluated items, no-information items, and indifferent items, and is based on multi-type item pairs. Thus, by using the learned model parameters, it is possible to recommend items to users with higher accuracy.

3 is a flowchart illustrating a method of learning model parameters according to an embodiment of the present invention.

In step S310, the item recommendation device generates an objective function based on a difference in evaluation results between item types included in each of the plurality of selected item pairs.

For example, it is assumed that a plurality of selection item pairs include {evaluated items, unevaluated items}, and {no information items, indifferent items}. In this case, the item recommendation device may generate an objective function representing the sum of the difference between the evaluation results of the evaluation item and the non-evaluation item and the difference between the evaluation result of the non-information item and the indifferent item for a plurality of users.

In another embodiment, the objective function may be based on the assumption that the evaluation result of the evaluation item is higher than the evaluation result of the unevaluated item, the no-information item and the indifference item, and that the evaluation result of the no-information item is higher than the evaluation result of the indifference item .

At this time, referring to FIG. 1(b), the assumption that the evaluation result of the evaluation item is higher than the evaluation result of the unevaluated item corresponds to the PP1, and the assumption that the evaluation result of the evaluation item is higher than the evaluation result of the non-information item corresponds to PP2. Corresponds, and the assumption that the evaluation result of the evaluation item is higher than the evaluation result of the indifferent item corresponds to PP3, and the assumption that the evaluation result of the no-information item is higher than the evaluation result of the indifferent item corresponds to PP4.

That is, the item recommendation device may train the model parameter to maximize the difference in the objective function based on the above-described assumption.

In step S320, the item recommendation apparatus acquires a plurality of evaluation result information including evaluation results for each item type included in the plurality of selected item pairs from the aggregate result for one selected user selected from the plurality of users.

In this case, it may be preferable that the item recommendation device selects one selected user from among the plurality of users at random. In addition, the item recommendation device may randomly extract one or more evaluation results for each item type included in the plurality of selected item pairs from the evaluation results of the selected user included in the comprehensive result, and obtain a plurality of evaluation result information. .

For example, it is assumed that a plurality of selection item pairs include {evaluated items, unevaluated items}, and {no information items, indifferent items}. In this case, the item recommendation device may obtain a plurality of evaluation result information by extracting evaluation results for evaluation items, non-evaluated items, non-information items, and indifferent items from among the evaluation results of the selected user. In addition, the evaluation result for the unevaluated item may be an evaluation result for an item randomly selected among the no-information item and the indifference item.

In another embodiment, the plurality of evaluation result information may include one evaluation result for each item type included in the plurality of selection item pairs.

That is, the item recommendation device may calculate an objective function value by using the evaluation result of one item for each item type for the selected user, and include each evaluation result in order to learn model parameters.

This is a change to the existing Bayesian Personalized Ranking (BPR) based on the result of multi-type pairwise evaluation for individual items, and is referred to as M-BPR (Multi-type pair-wise preference BPR) in the present invention.

In another embodiment, the plurality of evaluation result information includes a group evaluation result that is a plurality of evaluation results for each item type included in the plurality of selection item pairs, and the objective function may be based on a difference between the group evaluation results.

That is, the item recommendation device calculates the objective function value by using the evaluation results of a plurality of items (i.e., group evaluation results) for each item type for the selected user, and calculates each group evaluation result to learn model parameters. Can be included. In this case, the objective function may be generated not based on differences in evaluation results of individual items, but differences in group evaluation results. Further, the group evaluation result may be calculated using an average value of the evaluation results of the plurality of items.

This is a modification of the M-BPR of the present invention described above to be based on an item in a group unit, and is referred to as M-BPR ^G in the present invention.

In step S330, the item recommendation device updates the model parameter value so that the value of the objective function is maximized by using the objective function and a plurality of evaluation result information.

For example, the item recommendation device may use a plurality of evaluation result information to determine a model parameter value that maximizes a difference in evaluation results between item types (ie, an objective function). In this case, the item recommendation device may additionally use a method such as gradient descent.

In another embodiment, after the item recommendation device calculates a gradient value of an objective function using a plurality of evaluation result information, the model parameter value may be updated by applying the gradient value based on the SGD technique.

First, the item recommendation device calculates a gradient value of the objective function by using a plurality of evaluation result information.

For example, the item recommendation device may calculate a gradient value by inputting a plurality of evaluation result information to the gradient of the objective function.

In addition, the item recommendation device updates the model parameter value by applying a gradient value based on a stochastic gradient descent (SGD) technique.

That is, the item recommendation device may update the existing model parameter value with the newly calculated model parameter value by applying the calculated gradient value based on the SGD technique.

Finally, in step S340, steps S320 and S330 are repeated until the item recommendation device satisfies a predetermined completion condition.

In this case, the completion condition may be a condition for repeating a predetermined number of times. Alternatively, the completion condition may be a condition in which the updated model parameter value is compared with the existing model parameter value and converged within a predetermined threshold range.

That is, the item recommendation device may continuously update the model parameter value by repeating steps S320 and S330 until such a completion condition is satisfied.

On the other hand, in the M-BPR of the present invention, in the case where a plurality of selection item pairs include {evaluated items, unevaluated items} and {no information items, indifferent items}, the objective function is as shown in Equation 1. At this time, referring to FIG. 1(b), Equation 1 is an objective function when PP1 and PP4 are used.

[Equation 1]

Here, f(θ) is the objective function, θ is the model parameter, R(θ) is the normalization term of the model parameter, U is the user set, I _u ^rt is the evaluation item set of user u, and I\I _u ^rt is the set of ^unevaluated items of user u, I _u ^unk is the set of ^uninformed items of user u, I _u ^uin is the set of indifference items of user u, w is the weight, σ is the sigmoid function, and r _uij is r -r _{ui uj, uk} r _ukl is r -r _ul, the evaluation results for _ui r, r _{uj, uk} r, r _ul item eunneun user u i, j, k, l .

In addition, in the M-BPR of the present invention, in the case where a plurality of selection item pairs include {evaluation item, no information item}, {evaluation item, indifferent item}, {no information item, indifferent item}, the objective function is Same as 2. At this time, referring to FIG. 1(b), Equation 2 is an objective function in the case of using PP2, PP3, and PP4.

[Equation 2]

Here, α, β, and γ are confidence coefficients for controlling the relative learning weights for each of PP2, PP4, and PP4.

In addition, each of Equations 3 and 4 below is a result of optimization by applying the SGD algorithm to Equations 1 and 2.

[Equation 3]

[Equation 4]

Here, γ is the learning rate, λθ is a parameter related to the normalized term, and the model parameter θ includes Uu· which is a latent factor of user u, Vi· which is a latent factor of item i, and bias bi of item i.

In addition, the relationship as shown in Equation 5 below is additionally established.

[Equation 5]

In addition, in the case where a plurality of selection items in the M-BPR ^G of the present invention includes {evaluated items, unevaluated items}, and {no information items, indifferent items}, the objective function is shown in Equation 6. At this time, referring to FIG. 1(b), Equation 1 is an objective function when PP1 and PP4 are used.

[Equation 6]

In addition, in the case of the M-BPR ^G of the present invention, when a plurality of selection item pairs include {evaluation item, no-information item}, {evaluation item, indifferent item}, and {no information item, indifferent item}, the objective function is mathematical Same as Equation 7. At this time, referring to FIG. 1(b), Equation 2 is an objective function in the case of using PP2, PP3, and PP4.

[Equation 7]

In addition, each of Equations 8 and 9 below is a result of optimization by applying the SGD algorithm to Equations 6 and 7.

[Equation 8]

[Equation 9]

Here, the relationship as in Equation 10 below is established.

[Equation 10]

Meanwhile, referring to FIG. 5, a pseudocode of a model parameter learning algorithm according to an embodiment of the present invention is shown. That is, the entire process of M-BPR and M-BPR ^G of the present invention is shown.

First, the item recommendation device predicts an approximate evaluation result for a plurality of unevaluated items included in the training data set R = {(u, i)}. For each user u, the item set I is classified into I _u ^rt , I _u ^unk , and I _u ^uni . Next, the model parameter θ is updated while repeating T times. At this time, if group-based (G=TRUE), three types of group evaluation results for users u and u are randomly obtained for each iteration, and if not group-based (G=FALSE), users u and u for each iteration The evaluation results of 3 types of individual items are obtained at random.

In addition, when MODE = A, as a case where the plurality of selection item pairs include {evaluated items, unevaluated items}, {no information items, indifferent items}, evaluation results for unevaluated items can be further obtained. In addition, if MODE=B, it is a case in which the plurality of selection item pairs include {evaluation item, no information item}, {evaluation item, indifferent item}, and {no information item, indifferent item}.

Finally, the item recommendation device updates the model parameter θ for the user u. For this, the item recommendation device may use Equations 3, 4, 8, and 9.

Further, referring to Fig. 6, the accuracy of the conventional BPR, the two types of M-BPR of the present invention, and the two types of M-BPR ^G of the present invention are shown.

At this time, the two types of M-BPR are according to the type of the selection item pair, and (A) is the case where the plurality of selection item pairs include {evaluated items, unevaluated items}, and {no information items, indifferent items}. , (B) is a case where a plurality of selected item pairs include {evaluation item, no information item}, {evaluation item, indifferent item}, and {no information item, indifferent item}.

Specifically, M-BPR and M-BPR ^G exhibit superior performance than conventional BPR regardless of the number of recommended items (N). In addition, it can be seen that between M-BPR and M-BPR ^G, M-BPR ^G exhibits better performance than M-BPR.

Also, referring to FIG. 7, recommendation accuracy of a total of eight methods is shown.

At this time, it can be seen that the M-BPR of the present invention showed superior performance with a significant difference overall for all datasets than in the case of all other methods. More specifically, the GAIN of M-BPR is 12.84% (MovieLens100K), 27.41% (Watcha), 61.88% (Ciao), and 17.82% (Last.fm).

4 is a block diagram illustrating a multi-type pairwise-based item recommendation apparatus according to an embodiment of the present invention.

4, a multi-type pairwise-based item recommendation device 400 according to an embodiment of the present invention includes a generation unit 410, a classification unit 420, a learning unit 430, and a recommendation unit 440. Include.

Meanwhile, the multi-type pairwise-based item recommendation device 400 according to an embodiment of the present invention may be mounted on a computing device such as a server, a desktop PC, a notebook, a smart phone, and a tablet PC.

The generation unit 410 generates a comprehensive result by synthesizing the prediction result and the evaluation result, which is a result of predicting the evaluation result of a plurality of unevaluated items included in the evaluation result, using a collaborative filtering technique.

The classification unit 420 classifies a plurality of unevaluated items into a plurality of non-information items and a plurality of non-interested items in the aggregate result.

The learning unit 430 is based on the overall result and a plurality of selected item pairs selected from among item pairs consisting of two of evaluation items, unevaluated items, no-information items, and non-interested items included in the overall result, and the model Learn parameters.

Finally, the recommendation unit 440 uses the final prediction result, which is a result of predicting evaluation results for a plurality of unevaluated items, based on the learned model parameters, to at least one recommendation target among the plurality of users. Recommend the item.

In another embodiment, the learning unit 430 generates an objective function based on an evaluation result difference between item types included in each of the plurality of selected item pairs, and for one selected user selected from among the plurality of users, from the synthesis result. Acquires a plurality of evaluation result information including evaluation results for each item type included in the plurality of selected item pairs, and updates the model parameter value so that the value of the objective function is maximized by using the objective function and the plurality of evaluation result information. Then, the process of acquiring the plurality of evaluation result information and the process of updating the model parameter values may be repeated until a predetermined completion condition is satisfied.

In another embodiment, when the learning unit 430 updates a model parameter value, a gradient value of an objective function is calculated using a plurality of evaluation result information, and the gradient value is based on a stochastic gradient descent (SGD) technique. By applying, it is possible to update the model parameter value.

In another embodiment, the plurality of evaluation result information may include one evaluation result for each item type included in the plurality of selection item pairs.

In another embodiment, the plurality of evaluation result information includes a group evaluation result that is a plurality of evaluation results for each item type included in the plurality of selection item pairs, and the objective function may be based on a difference between the group evaluation results.

In another embodiment, the objective function may be based on the assumption that the evaluation result of the evaluation item is higher than the evaluation result of unevaluated items, no information items, and indifferent items, and that the evaluation result of no information items is higher than the evaluation result of indifferent items. have.

In another embodiment, the plurality of selection item pairs include an item pair consisting of an evaluation item and an unevaluated item and an item pair consisting of an uninformed item and an indifferent item, or an item pair consisting of an evaluation item and an uninformed item. , An item pair consisting of an evaluation item and an indifference item, and an item pair consisting of an uninformed item and an indifference item.

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.

Claims (14)

In a method of recommending an item using an evaluation result that is a result of a plurality of users evaluating at least one item among a plurality of items,
Generating a comprehensive result by synthesizing a prediction result, which is a result of predicting an evaluation result of a plurality of unevaluated items included in the evaluation result, using a collaborative filtering technique and the evaluation result;
Classifying the plurality of unevaluated items into a plurality of non-information items and a plurality of indifferent items from the synthesis result;
Learning a model parameter based on the synthesis result and a plurality of selection item pairs selected from among item pairs consisting of two of an evaluation item, an unevaluated item, an uninformed item and an item of no interest included in the synthesis result; And
Recommending a predetermined number of items to at least one recommendation target among the plurality of users using a final prediction result that is a result of predicting evaluation results for the plurality of unevaluated items based on the learned model parameters
Multi-type pairwise-based item recommendation method comprising a. The method of claim 1,
Learning the model parameters
Generating an objective function based on a difference in evaluation results between item types included in each of the plurality of selected item pairs;
Acquiring a plurality of evaluation result information including evaluation results for each item type included in the plurality of selected item pairs from the synthesis result for one selected user selected from among the plurality of users;
Updating the model parameter value to maximize the value of the objective function using the objective function and the plurality of evaluation result information; And
Repeating the steps of acquiring the plurality of evaluation result information and updating the model parameter values until a predetermined completion condition is satisfied.
Multi-type pairwise-based item recommendation method comprising a. The method of claim 2,
Updating the model parameter value comprises:
Calculating a gradient value of the objective function using the plurality of evaluation result information; And
Updating the model parameter value by applying the gradient value based on a stochastic gradient descent (SGD) technique
Multi-type pairwise-based item recommendation method comprising a. The method of claim 2,
The plurality of evaluation result information
A multi-type pairwise-based item recommendation method comprising one evaluation result for each item type included in the plurality of selected item pairs. The method of claim 2,
The plurality of evaluation result information
And a group evaluation result that is a plurality of evaluation results for each item type included in the plurality of selected item pairs,
The objective function is
Multi-type pairwise-based item recommendation method, characterized in that based on the difference between the group evaluation results. The method of claim 2,
The objective function is
Characterized in that based on the assumption that the evaluation result of the evaluation item is higher than the evaluation result of the unevaluated item, the no-information item and the indifferent item, and that the evaluation result of the no-information item is higher than the evaluation result of the indifferent item Multi-type pairwise-based item recommendation method. The method of claim 1,
The plurality of selection item pairs are
An item pair consisting of the evaluation item and the unevaluated item and an item pair consisting of the non-information item and the indifferent item, or
An item pair consisting of the evaluation item and the non-information item, an item pair consisting of the evaluation item and the indifferent item, and an item pair consisting of the non-information item and the indifference item. Wise-based item recommendation method. In the apparatus for recommending an item using an evaluation result that is a result of a plurality of users evaluating at least one item among a plurality of items,
A generator for generating a comprehensive result by synthesizing a prediction result, which is a result of predicting an evaluation result of a plurality of unevaluated items included in the evaluation result, and the evaluation result using a collaborative filtering technique;
A classification unit for classifying the plurality of unevaluated items into a plurality of non-information items and a plurality of non-interested items in the synthesis result;
A learning unit that learns model parameters based on the synthesis result and a plurality of selection item pairs selected from among item pairs consisting of two of an evaluation item, an unevaluated item, a no-information item, and an indifferent item included in the synthesis result; And
A recommendation unit that recommends a predetermined number of items to at least one recommendation target among the plurality of users using a final prediction result that is a result of predicting evaluation results for the plurality of unevaluated items based on the learned model parameters
Multi-type pairwise-based item recommendation device comprising a. The method of claim 8,
The learning unit
Generate an objective function based on a difference in evaluation results between item types included in each of the plurality of selected item pairs,
For one selected user selected from among the plurality of users, a plurality of evaluation result information including evaluation results for each item type included in the plurality of selected item pairs are obtained from the comprehensive result,
Using the objective function and the plurality of evaluation result information, updating the model parameter value so that the value of the objective function is maximized,
A multi-type pairwise-based item recommendation device, characterized in that the process of acquiring the plurality of evaluation result information and the process of updating the model parameter value are repeated until a predetermined completion condition is satisfied. The method of claim 9,
When the learning unit updates the model parameter value,
Using the plurality of evaluation result information, calculate the gradient value of the objective function,
A multi-type pairwise-based item recommendation device, characterized in that the model parameter value is updated by applying the gradient value based on a stochastic gradient descent (SGD) technique. The method of claim 9,
The plurality of evaluation result information
A multi-type pairwise-based item recommendation device, comprising: one evaluation result for each item type included in the plurality of selected item pairs. The method of claim 9,
The plurality of evaluation result information
And a group evaluation result that is a plurality of evaluation results for each item type included in the plurality of selected item pairs,
The objective function is
Multi-type pairwise-based item recommendation device, characterized in that based on the difference between the group evaluation results. The method of claim 9,
The objective function is
Characterized in that based on the assumption that the evaluation result of the evaluation item is higher than the evaluation result of the unevaluated item, the no-information item and the indifferent item, and that the evaluation result of the no-information item is higher than the evaluation result of the indifferent item Multi-type pairwise-based item recommendation device. The method of claim 8,
The plurality of selection item pairs are
An item pair consisting of the evaluation item and the unevaluated item and an item pair consisting of the non-information item and the indifferent item, or
An item pair consisting of the evaluation item and the non-information item, an item pair consisting of the evaluation item and the indifferent item, and an item pair consisting of the non-information item and the indifference item. Wise-based item recommendation device.

Images