KR102161294B1 - Promotion Performance Prediction and Recommendation Apparatus in Online Shopping Mall Using Artificial Intelligence - Google Patents

Abstract

The present invention relates to an apparatus and method for predicting and recommending promotion performance in an online shopping mall using artificial intelligence. According to an embodiment of the present invention, the promotion of the online shopping mall is elaborated and precise by learning of artificial intelligence, thereby increasing the effect of the promotion.

Description

Promotion Performance Prediction and Recommendation Apparatus in Online Shopping Mall Using Artificial Intelligence}

The present invention relates to an apparatus and method for predicting and recommending promotion performance in an online shopping mall using artificial intelligence.

Promotions in online shopping malls are conducted in the form of product discounts, giveaway events, discount coupons, and mileage payments. Since such promotion always causes a decrease in net profit from the viewpoint of online shopping malls, the type of promotion, promotion period, and degree of promotion must be precisely designed according to the persona. However, in the past, all promotion designs were empirically determined by the operators of online shopping malls or promotion agencies.

For online shopping in Korea, there are various types of companies such as open market/social commerce/general shopping mall/large mart shopping mall/card point mall, and most of the sellers sell multiple stores in the above online shopping malls. There is a problem that it is difficult to analyze sales performance based on the obtained data.

The components of the promotion are related to the date and time and conditions of the promotion and have a causal relationship with the promotion result. Since it is included, it is very difficult for a person to calculate the contribution of the component based on the result.

In particular, in the case of the promotion depreciation rate, which will serve as a benefit to consumers, the larger the depreciation rate, the greater the purchase benefit, so the quantity of sales will increase and sales will increase. However, when considering the cost of the product, it is important to apply a balance between the increase in sales quantity and the depreciation rate at an appropriate point, and it is very difficult for a person to calculate whether the expected sales profit will be achieved. .

Web advertisement recommendation method and system based on a click pattern using a collaborative filtering system having a neural network, Korean Patent Registration 10-0792700, Naver Co., Ltd.

Accordingly, an object of the present invention is to predict the performance of promotions using artificial intelligence and recommend the characteristics of promotions in an online shopping mall using artificial intelligence in order to solve the low precision of promotion of an online shopping mall that was previously designed empirically. It is to provide a device and method for predicting and recommending promotion performance of the company.

Hereinafter, specific means for achieving the object of the present invention will be described.

An object of the present invention is a memory module for storing a program code of an artificial neural network previously learned based on product attribute data, promotion attribute data, other performance data and promotion; And a processing module for predicting the performance of a specific promotion by processing the program code of the artificial neural network. Including, the program code, the product attribute data (100), promotion attribute data (200), an input step of receiving the performance data (300); A performance prediction step of generating promotion performance prediction data by predicting a performance for the specific promotion based on the product attribute data, the promotion attribute data, and the performance data; And an output step of outputting the predicted promotion performance prediction data. It may be achieved by providing a promotion performance prediction and recommendation device in an online shopping mall using artificial intelligence, which is configured to be performed on a computer.

In addition, a memory module that stores a program code of a promotion performance prediction artificial neural network module and a promotion recommendation module previously learned based on promotion attribute data and performance data; And a processing module that processes the promotion performance prediction artificial neural network module and the program code of the promotion recommendation module to predict the performance of a specific promotion and recommends promotion attributes; wherein the program code of the promotion recommendation module comprises: the promotion A promotion performance prediction step of receiving promotion performance prediction data, which is information on promotion performance predicted for a specific promotion attribute, by the performance prediction artificial neural network module; A state information receiving step of receiving state information (s _t ) for episode t; Attribute probability output for generating reward possibility information in the value network based on the status information and the promotion performance prediction data, and then outputting the recommended promotion attribute probability that is a plurality of actions in the policy network based on the reward possibility information step; A recommended promotion attribute data output step of selecting and outputting recommended promotion attribute data (a _t ) for episode t through optimal promotion search based on the plurality of recommended promotion attribute probabilities; A status and reward information receiving step of receiving status information (s _t+1 ) for episode t+1 and reward information (r _t+1 ) for a _t after the promotion according to the recommended promotion attribute data is applied; And updating the policy network and the value network based on the received state information (s _t+1 ) and the compensation information (r _t+1 ). It is configured to be performed on a computer including, and the status information includes performance data (Quantity, Sales Amount, Page View), Conversion Rate, Average Sales Price, or Sales profit) and environmental data (date information, day of the week, weather information, product information in progress, shopping mall information or other product information of the shopping mall), and the recommended promotion attribute data includes promotion category, promotion period, promotion shopping mall, Providing a promotion performance prediction and recommendation device in an online shopping mall using artificial intelligence, characterized in that it means the color of the exposure account or the promotion image, and the compensation information means the rise or fall of the performance data for a specific period. Can be achieved.

In addition, in the search for the optimal promotion in the step of outputting the recommended promotion attribute data, a Monte Carlo tree search is used, and in the Monte Carlo tree search, each node of the tree represents a state, and each edge represents a specific action for a specific state. It represents the expected value according to the current state as a root node, and it is composed of a structure in which a leaf node is expanded each time a new action is taken to transition to a new state, and the recommended promotion attribute data output step is A selection step of selecting an action with the highest value among selectable actions from the current state until the leaf node appears; An expansion step of adding a new leaf node by performing an action according to the probability of the policy network learned by supervised learning when the simulation proceeds to the leaf node; An evaluation step of evaluating the value of the new leaf node by calculating the reward possibility information determined from the new leaf node using the value network and the reward obtained by proceeding from the leaf node to the end of the promotion episode using the policy network; And an update step of reevaluating the values of the nodes visited during the simulation by reflecting the value of the new leaf node and updating the visit frequency; further comprising, selecting the recommended promotion attribute data (a _t ) based on the visit frequency It can be characterized by that.

In addition, the policy network may be configured to perform supervised learning based on the promotion attribute data and the performance data including a random vector.

Another object of the present invention is a promotion performance prediction step of receiving, by a promotion recommendation module, promotion performance prediction data, which is information on promotion performance predicted for a specific promotion attribute, from a promotion performance prediction artificial neural network module; A state information receiving step of the promotion recommendation module receiving state information (s _t ) for episode t; After the promotion recommendation module generates reward possibility information in the value network based on the status information and the promotion performance prediction data, the recommendation promotion attribute probability that is a plurality of actions in the policy network is calculated based on the reward possibility information. An attribute probability output step of outputting; A recommended promotion attribute data output step of selecting and outputting, by the promotion recommendation module, recommended promotion attribute data (a _t ) for episode t through optimal promotion search based on the plurality of recommended promotion attribute probabilities; After the promotion according to the recommended promotion attribute data is applied, the promotion recommendation module receives status information for episode t+1 (s _t+1 ) and reward information for a _t (r _t+1 ) and rewards Receiving information; And updating the policy network and the value network based on the status information (s _t+1 ) and the reward information (r _t+1 ) received by the promotion recommendation module. It is configured to be performed on a computer including, and the status information includes performance data (Quantity, Sales Amount, Page View), Conversion Rate, Average Sales Price, or Sales profit) and environmental data (date information, day of the week, weather information, product information in progress, shopping mall information or other product information of the shopping mall), and the recommended promotion attribute data includes promotion category, promotion period, promotion shopping mall, By providing a method of predicting and recommending promotion performance in an online shopping mall using artificial intelligence, characterized in that it means the color of an exposure account or promotion image, and the compensation information means an increase or decrease in the performance data for a specific period. Can be achieved.

As described above, the present invention has the following effects.

First, according to an embodiment of the present invention, the promotion of the online shopping mall is elaborated and precise by learning of artificial intelligence, thereby increasing the effect of the promotion.

Second, according to an embodiment of the present invention, an effect of updating a policy network that outputs a probability of a recommended promotion attribute according to the value network 111 can be performed every episode. In the existing reinforcement learning, there is a problem in that the update of the reinforcement learning model proceeds after all episodes are finished, so it was difficult to apply it to the promotion recommendation model.

Third, according to an embodiment of the present invention, promotion recommendation has short-term and long-term goals, and thus, despite the difficulty in applying the existing DQN because the assumption of MDP is not established, the policy network and the value network depend on the relationship. Promotional recommendations also have the effect of enabling reinforcement learning.

Fourth, according to an embodiment of the present invention, since the policy network and the value network are connected to the promotion performance prediction artificial neural network device and learn based on the promotion performance prediction data, the learning speed is improved.

The following drawings attached to the present specification illustrate preferred embodiments of the present invention, and serve to further understand the technical idea of the present invention together with the detailed description of the present invention, so the present invention is limited to the matters described in such drawings. And should not be interpreted.
1 shows a promotion performance prediction and recommendation device 1 in an online shopping mall using artificial intelligence according to an embodiment of the present invention.
2 is a schematic diagram showing a learning process of the artificial neural network device for predicting promotion performance according to an embodiment of the present invention;
3 is a schematic diagram showing an inference process of the artificial neural network device for predicting promotion performance according to an embodiment of the present invention;
4 is a schematic diagram showing a promotion recommendation device 11 according to an embodiment of the present invention,
5 is a flowchart showing an operation example of the promotion recommendation device 11 according to an embodiment of the present invention;
6 is a schematic diagram showing a promotion image correction device 13 according to an embodiment of the present invention,
7 is a schematic diagram showing an example of a ConvNet encoder according to an embodiment of the present invention,
8 is a flowchart illustrating a promotion recommendation method according to an embodiment of the present invention.

Hereinafter, exemplary embodiments in which the present invention can be easily implemented by those of ordinary skill in the art will be described in detail with reference to the accompanying drawings. However, when it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted.

In addition, the same reference numerals are used for parts having similar functions and functions throughout the drawings. Throughout the specification, when a specific part is said to be connected to another part, this includes not only the case that it is directly connected, but also the case that it is indirectly connected with another element interposed therebetween. In addition, the inclusion of a specific component does not exclude other components unless specifically stated to the contrary, but means that other components may be further included.

A device and method for predicting and recommending promotion performance in online shopping malls using artificial intelligence

1 is a diagram showing a promotion performance prediction and recommendation device 1 in an online shopping mall using artificial intelligence according to an embodiment of the present invention. As shown in FIG. 1, the apparatus for predicting and recommending promotion performance in an online shopping mall using artificial intelligence according to an embodiment of the present invention is a pre-learned promotion based on product attribute data, promotion attribute data, performance data, and promotion. It may include a performance prediction artificial neural network device 10 and a promotion recommendation device 11.

Related to the promotion performance prediction artificial neural network device 10, FIG. 2 is a schematic diagram showing a learning process of the promotion performance prediction artificial neural network device according to an embodiment of the present invention, and FIG. 3 is a promotion performance prediction according to an embodiment of the present invention. It is a schematic diagram showing the inference process of the artificial neural network device. As shown in FIGS. 2 and 3, a method for predicting promotion performance in an online shopping mall using artificial intelligence is an artificial neural network for predicting promotion performance by receiving product attribute data 100, promotion attribute data 200, and performance data 300. The device 10 is trained, and product attribute data 100 and promotion attribute data 200 are input to the previously learned promotion performance prediction artificial neural network device 10 to output the predicted promotion performance prediction data 400. have. In particular, when the artificial neural network device 10 for predicting promotion performance is learning, the weight of the hidden layer may be updated based on an error between the promotion performance data and the output data. In FIG. 2, each attribute constituting the product attribute data 100 and the promotion attribute data 200 is input to each node of the input layer such as x1, x2, and x3, and based on weights such as w1, h1, h2, and h3 The promotion performance prediction data 400 predicted based on a cost function such as softmax after passing through the hidden layer such as is output as an output layer y1. Based on an error (error, -Sigma(y _i log p _i )) between the predicted performance data and the actual performance data 300, the weight of the artificial neural network device 10 for predicting promotion performance may be updated as back propagation.

Alternatively, a separate promotion outcome prediction artificial neural network device 10 is configured for each product, and promotion property data 200 and performance data 300 are received to learn the promotion outcome prediction artificial neural network device 10, Promotional outcome prediction data 400 predicted by inputting promotion attribute data 200 to the previously learned promotion outcome prediction artificial neural network device may be output. In addition to the artificial neural network that can be used at this time, machine learning such as Multi-layer perceptron, Naive-Bayesian Classification, and Random Forest Classification can be used.

The product attribute data 100 may include a product category, a product selling price, a product cost, and the like.

The promotion attribute data 200 may include a promotion category (discount promotion, gift promotion, etc.), a promotion period, a promotion shopping mall, an exposure account, a promotion image, and the like. Among the promotion attribute data according to an embodiment of the present invention, the promotion image data may mean a result value classified into a specific atmosphere category and a specific color sense category through a previously learned Convolutional Neural Network (CNN).

The performance data 300 may include a quantity, a sales amount, a page view, a conversion rate, an average sales price, and a sales profit. According to an embodiment of the present invention, in order to prevent a problem due to multicollinearity of the performance data 300, PCA is applied to each data constituting the performance data 300 in the form of a diagnol matrix. It can offset collinearity.

In relation to the promotion attribute recommendation, the promotion recommendation device 11 may be configured to output recommended promotion attribute data so that the highest promotion performance prediction data 400 is output using the promotion performance prediction artificial neural network device.

Promotion recommendation by the promotion recommendation device 11 according to an embodiment of the present invention may be performed using reinforcement learning. 4 is a schematic diagram showing a promotion recommendation device 11 according to an embodiment of the present invention. As shown in Fig. 4, the promotion recommendation device 11 according to an embodiment of the present invention includes a value network 111, which is an artificial neural network that learns a value function that outputs a value in a specific state, and the probability of a recommendation promotion attribute. It may include a policy network 110 for learning a policy function that outputs, and the policy network 110 and the value network 111 according to an embodiment of the present invention are connected to the artificial neural network device 10 for predicting promotion performance. It can be configured to be. In addition, the policy network 110 and the value network 111 may be connected to the optimal promotion search module 12 to output recommended promotion attribute data 500.

In terms of reinforcement learning, the objective of the promotion recommendation device 11 according to an embodiment of the present invention is to improve the performance of the promotion, and the state is current performance data (Sales Amount, Sales Amount). ), Inflow (Page View), Conversion Rate, Average Sales Price, Sales Profit, etc.) and environmental data (date, day of the week, weather, products currently being promoted, shopping malls, other product information of shopping malls) ), and Action can mean promotion attribute data (promotion category (discount promotion, free gift promotion, etc.), promotion period, promotion shopping mall, exposure account, promotion image color, etc.), and reward (Reward ) Can mean a rise or fall in performance data over a specific period.

The policy network 110 is an artificial neural network that determines the probability of specific attributes of a recommended promotion in each state of the promotion recommendation device 11, and learns a policy function to output a recommended promotion attribute probability. The cost function of the policy network may be a function obtained by multiplying the policy function and the cost function of the value network to calculate the cross entropy and then taking the policy gradient. The policy network can be back propagated based on the product of the cross entropy and the time difference error, which is a cost function of the value network.

In Equation 1, π is a policy function, θ is a policy network parameter, π _θ (a _i │s _i ) is the probability of performing a specific action (promotion of a specific attribute) in the current episode, V is a value function, and w is a value network. Parameter, s _i is the status information of the current episode i, S _i+1 is the status information of the next episode i+1, r _i+1 is the reward expected to be acquired in the next episode, and V _w (s _i ) is The possibility of compensation in the current episode, V _w (s _i+1 ) may indicate the possibility of compensation in the next episode, and γ may indicate a depreciation rate. In this case, r _i+1 may be configured to be received by the artificial neural network device 10 for predicting promotion performance. As a result, the policy network 110 according to an embodiment of the present invention initially outputs a promotion attribute that simulates a user's promotion history through a policy gradient.

The policy network 110 according to an embodiment of the present invention is supervised learning based on the existing promotion attribute data and the corresponding performance data before reinforcement learning proceeds, and the weight of the policy network is updated. You can learn the basics. That is, the weight of the policy network may be supervised and set based on existing promotion attribute data and performance data. According to this, there is an effect that the policy network can be learned very quickly by the record of the existing promotion.

In addition, according to an embodiment of the present invention, during supervised learning of the policy network 110, it may be configured to perform supervised learning based on existing promotion attribute data and result data including a random vector. The random vector may use, for example, a Gaussian distribution. Accordingly, there is an effect that the policy network can output a challenging promotion policy with a random probability. When supervised learning of the policy network 110 is configured to be supervised learning based on the existing promotion attribute data and corresponding performance data, a result of the promotion recommendation is optimized within the existing policy. However, according to an embodiment of the present invention, if a random vector is included in the supervised learning of a policy network, the more effective reinforcement learning proceeds, the more effective the policy network can learn promotion than the existing policy.

The value network 111 is an artificial neural network that derives the possibility of achieving a reward in each state that the promotion recommendation device 11 may have, and learns a value function. The value network 111 presents the direction in which the promotion recommendation device 11, which is an agent, is to be updated. To this end, the input variable of the value network 111 is set as status information, which is information on the status of the promotion recommendation device 11, and the output variable of the value network 111 is the promotion recommendation device 11 It may be set as information on the possibility of compensation, which is a possibility. Compensation possibility information according to an embodiment of the present invention may be calculated by a Q-function as shown in the following equation.

In Equation 2 above, Q _π may mean information about the total reward possibility expected in the future in the case of a state s and an action a in a specific policy π, R represents a reward for a specific period, and gamma represents a depreciation rate. S _t may represent a state of time t, A _t may represent an action of time t, and E may represent an expected value. Compensation possibility information (Q value) according to an embodiment of the present invention defines the update direction and size of the policy network 110.

In this case, the cost function of the value network may be a mean square error (MSE) function for the value function, and may be configured as in Equation 3 below. The value network 111 may be back propagated based on a time difference error, which is a cost function of the value network.

In Equation 2, V is a value function, w is a value network parameter, s _i is the state information of the current episode i, S _i+1 is the state information of the next episode _i+1 , and r _i+1 is the next episode. Reward expected to be obtained, V _w (s _i ) may indicate a possibility of compensation in the current episode, V _w (s _i+1 ) may indicate a possibility of compensation in the next episode, and γ indicates a depreciation rate. In this case, r _i+1 may be configured to be received by the artificial neural network device 10 for predicting promotion performance.

Accordingly, the value network may update the cost function of Equation 1 in the direction of gradient descent when the state of the promotion recommendation device changes.

According to an embodiment of the present invention, while learning the value network separately from the policy network, the Q value of the value network is supervised instead of starting at random, so that rapid learning is possible. According to this, the effect of being able to greatly reduce the burden of exploration in an action of selecting a combination of promotion attributes with high complexity occurs.

According to the promotion recommendation device 11 according to an embodiment of the present invention, when the policy network 110 that has completed supervised learning recommends the promotion attribute of the current episode i, the value network 111 proceeds with the recommended promotion attribute. Is learned to predict the reward of. The policy network 110 and the value network 111 of the promotion recommendation device 11 that has finished learning are combined with a simulation using the optimal promotion search module 112 and are finally used to select promotion attributes.

In addition, according to the value network 111 according to an embodiment of the present invention, an effect of updating a policy network that outputs a probability of a recommended promotion attribute may be performed every episode. In the existing reinforcement learning, there is a problem in that the update of the reinforcement learning model proceeds after all episodes are finished, so it was difficult to apply it to the promotion recommendation model.

The optimal promotion search module 112 is a component that searches for an optimal promotion attribute by performing a plurality of simulations for various states and various actions based on a plurality of agents calculated in the policy network and the value network. The optimal promotion search module according to an embodiment of the present invention may utilize, for example, a Monte Carlo tree search, each node of the tree is a state, and each edge is a specific action for the state. It represents the expected value accordingly, and it is a structure in which the leaf node expands whenever a new action is taken and transition to a new state with the current state as the root node. In the optimal promotion search module according to an embodiment of the present invention, when a Monte Carlo tree search is used, the optimal promotion search may be processed in four stages of selection, expansion, evaluation, and backup.

The Selection step of the optimal promotion search module 112 is a step in which the action with the highest value among selectable actions is selected and proceeded from the current state until the leaf node appears. At this time, the value of the value function stored in the edge and the visit frequency value to balance inquiry-use are used. The equation for action selection in the selection step is as follows.

In Equation 4 above, a _t is an action (promotion performed) at time t, Q(s _t ,a) is the value of the value function stored in the tree, and u(s _t ,a) is the corresponding state-action pair. This value is inversely proportional to the number of visits, and is used to balance exploration and use.

The expansion step of the optimal promotion search module 112 is a step of adding a new node as a leaf node by performing an action according to the probability of the policy network learned by supervised learning when the simulation proceeds to the leaf node.

In the evaluation step of the optimal promotion search module 112, the leaf node's value (reward possibility) determined using the value network from the newly added leaf node and the reward obtained by proceeding from the leaf node to the end of the promotion episode using the policy network This is the step of evaluating the value. The following equation is an example of evaluating the value of a new leaf node.

In Equation 5 above, V(s _L ) is the value of the leaf node, λ is the mixing parameter, v _θ (s _L ) is the value obtained through the value network, and z _L is the reward obtained by continuing the simulation.

The backup step of the optimal promotion search module 112 is a step of re-evaluating the values of the nodes visited during the simulation by reflecting the values of the newly added leaf nodes and updating the visit frequency. The following equation is an example of node value re-evaluation and visit frequency update.

In Equation 6 above, s _L ⁱ represents the leaf node in the i-th simulation, 1(s,a,i) represents whether the connection (s,a) was visited in the i-th simulation, and when the tree search is completed, the algorithm It can be configured to select the most visited connection (s,a) from the root node. According to the optimal promotion search module 112 according to an embodiment of the present invention, it is possible to select an optimal promotion attribute by performing a simulation multiple times based on the value network for a plurality of promotion attributes selected by the policy network. Occurs.

According to an embodiment of the present invention, the promotion recommendation device 11 may be configured such that a plurality of agents are configured. When a plurality of agents are configured, promotions recommended by the promotion recommendation device for each of a specific state and specific promotion attribute compete with each other, and the most optimal product and promotion thereof can be recommended within a certain budget.

5 is a flowchart showing an example of the operation of the promotion recommendation device 11 according to an embodiment of the present invention. As shown in FIG. 5, when the state s(t) is input by the Internet shopping mall data hub device 12 that tracks the state data in a plurality of Internet shopping malls, the plurality of policy networks 110 are Various promotion attributes are input to the optimal promotion search module 112 by two agents and are promoted by the recommended promotion attribute probability a(t), which is an action output by the optimal promotion search module 112 With this progression, episode t ends and episode t+1 begins. In episode t+1, s(t+1), which is a state change due to a(t), is input by the internet shopping mall data hub device 12, and r(t+1), which is a compensation according to a(t), is It is immediately input to update the value network 111 and the policy network 110.

In addition, according to an embodiment of the present invention, in the case of promotion image color information among data included in the recommended promotion attribute data 500, according to an embodiment of the present invention, the actual promotion image is used in an online shopping mall using artificial intelligence. The promotion performance prediction and recommendation device 1 may receive and correct the color of the promotion image based on color information of the promotion image and output the modified promotion image. 6 is a schematic diagram showing a promotion image correction device 13 according to an embodiment of the present invention. As shown in Fig. 6, the promotion image modification device 13 may be configured to include a generator and a ConvNet encoder.

The generator may be composed of an image generator composed of encoders and decoders such as VAE and GAN, and may receive the actual promotional image 210 and output the generated promotional image 212 with a color changed. The generated promotion image 212 is again received by the ConvNet encoder, and the ConvNet encoder outputs the encoded promotional image 211 to calculate the color information of the promotion image and cross entropy of the recommended promotion attribute data 500 to output an error. have. This error may be updated so that the promotion image 212 generated by being fed back to the generator is further generated to match the color information of the promotional image.

7 is a schematic diagram showing an example of a ConvNet encoder according to an embodiment of the present invention. As shown in Fig. 7, the ConvNet encoder can be constructed with [INPUT-CONV-RELU-POOL-FC]. In the case of the generated promotion image 212 as input data, when the INPUT input image has 32 horizontally, 32 vertically, and RGB channels, the size of the input is [32x32x3]. The CONV layer (Conv. Filter, 101) is connected to a partial area of the input image, and the dot product of the connected area and its weight is calculated. The resulting volume will have the same size as [32x32x12]. The RELU layer is an activation function applied to each element, such as max(0,x). The RELU layer does not change the volume size ([32x32x12]). As a result, Activation map 1 (102) is generated. The POOL layer (pooling, 103) performs downsampling on the "horizontal and vertical" dimensions to output a reduced volume (Activation map 2, 104) such as [16x16x12]. The fully-connected (FC) layer 105 calculates class scores and outputs an output layer 106 having a size of [1x1x10]. "10" corresponds to a class score for 10 categories (promotional image color information according to an embodiment of the present invention). The FC layer is connected to all elements of the previous volume.

In this way, ConvNet transforms the original image consisting of pixel values into class scores (promotional image color information according to an embodiment of the present invention) through each layer. Some layers have parameters, while others do not. In particular, CONV/FC layers are activation functions that include not only the input volume but also weight and bias. On the other hand, RELU/POOL layers are fixed functions. The parameters of the CONV/FC layer are learned with gradient descent so that the class score for each image is the same as the label of the image.

The parameters of the CONV layer consist of a series of learnable filters. Each filter is small in the horizontal/vertical dimension but covers the entire depth in the depth dimension. In the forward pass, each filter is slid (convolved precisely) to the horizontal/vertical dimensions of the input volume, and a two-dimensional activation map is created. When sliding the filter over the input, a dot product is performed between the filter and the input volume. Through this process, ConvNet learns a filter that activates for a specific pattern at a specific location of the input data. The stacking of these activation maps in the depth dimension becomes the output volume. Therefore, each element of the output volume only handles a small area of the input, and neurons in the same activation map share the same parameters as the result of applying the same filter.

8 is a flowchart illustrating a promotion recommendation method according to an embodiment of the present invention. As shown in FIG. 8, the promotion recommendation method according to an embodiment of the present invention includes a promotion performance prediction step (S10), a status information reception step (S11), a recommendation promotion attribute data output step (S12), a status and a reward. It may include an information reception step (S13), a policy network and a value network update step (S14).

In the promotion performance prediction step (S10), the promotion performance prediction artificial neural network device receives the existing promotion attribute data and performance data, predicts the promotion performance for a specific promotion attribute, and transmits it to the promotion recommendation device 11. The predicted promotion performance data can be used as initial values for the policy network and the value chain.

The status information reception step S11 is a step in which the promotion recommendation device 11 receives status information s _t for the episode t from the internet shopping mall data hub device.

The recommended promotion attribute data output step (S12) is a step in which the promotion recommendation device 11 outputs recommended promotion attribute data (a _t ) including the recommended promotion attribute probability for episode t through the optimal promotion search module 112. .

After the promotion is applied to the internet shopping mall according to the recommended promotion attribute data, the status and reward information reception step (S13) is performed by the promotion recommendation device 11 from the internet shopping mall data hub device to the status information on episode t+1 (s _t+1). ) And compensation information for a _t (r _t+1 ).

The policy network and value network update step (S14) is a step of updating the policy network and the value network based on the received state information (s _t+1 ) and compensation information for a _t (r _t+1 ).

In terms of reinforcement learning, the objective of the promotion recommendation method according to an embodiment of the present invention is to improve the performance of the promotion, and the state is the current performance data (Quantity, Sales Amount, Inflow (Page View), Conversion Rate, Average Sales Price, Sales Profit, etc.) and environmental data (date, day of the week, weather, products currently being promoted, shopping malls, other product information of shopping malls) Action can mean promotion attribute data (promotion category (discount promotion, free gift promotion, etc.), promotion period, promotional shopping mall, exposure account, promotion image color, etc.), and reward is a specific It can mean a rise or fall in performance data over a period.

As described above, those skilled in the art to which the present invention pertains will appreciate that the present invention can be implemented in other specific forms without changing the technical spirit or essential features thereof. Therefore, the above-described embodiments are illustrative in all respects and should be understood as non-limiting. The scope of the present invention is indicated by the claims to be described later rather than the detailed description, and all changes or modified forms derived from the meaning and scope of the claims and equivalent concepts should be interpreted as being included in the scope of the present invention.

The features and advantages described herein are not all inclusive, and in particular many additional features and advantages will become apparent to those skilled in the art in view of the drawings, specification, and claims. Moreover, it should be noted that the language used herein has been selected primarily for readability and for teaching purposes, and may not be chosen to describe or limit the subject matter of the invention.

The above description of embodiments of the present invention has been presented for purposes of illustration. It is not intended to be exhaustive or to limit the invention to the precise form disclosed. Those skilled in the art will understand that many modifications and variations are possible in light of the above disclosure.

Therefore, the scope of the invention is not limited by the detailed description, but by any claims in the application on which it is based. Accordingly, the disclosure of the embodiments of the present invention is illustrative and does not limit the scope of the present invention described in the following claims.

1: Device for predicting and recommending promotion performance in online shopping malls using artificial intelligence
10: Promotional performance prediction artificial neural network device
11: Promotional recommendation device
12: Internet shopping mall data hub device
13: Promotional image editing device
100: product attribute data
110: policy network
111: value chain
112: Optimal Promotion Search Module
200: promotion attribute data
300: performance data
400: Promotional performance prediction data
500: Recommended promotion attribute data

Claims (1)

A memory module for storing a program code of a promotion performance prediction artificial neural network module and a promotion recommendation module previously learned based on promotion attribute data and performance data for at least one online shopping mall; And
A processing module for predicting a performance of a specific promotion and recommending a promotion attribute by processing the promotion performance prediction artificial neural network module and the program code of the promotion recommendation module;
Including,
The program code of the promotion recommendation module,
A promotion performance prediction step of receiving promotion performance prediction data, which is information on promotion performance predicted for a specific promotion attribute, by the promotion performance prediction artificial neural network module;
A state information receiving step of receiving first state information (s _t ) for episode t;
Attribute for generating reward possibility information in a value network based on the first state information and the promotion performance prediction data, and then outputting a recommendation promotion attribute probability that is a plurality of actions in a policy network based on the reward possibility information Probability output step;
A recommended promotion attribute data output step of selecting and outputting recommended promotion attribute data (a _t ) for the episode t through optimal promotion search based on the plurality of recommended promotion attribute probabilities;
After the promotion according to the recommended promotion attribute data is applied, second state information (s _t+1 ) for episode t+1 and reward information (r _t+1 ) for the recommended promotion attribute data (a _t) are received. Receiving status and compensation information; And
A policy network and value network updating step of updating the policy network and the value network based on the received second state information (s _t+1 ) and the compensation information (r _t+1 );
It is configured to be performed on a computer including,
The first state information and the second state information are performance data (Quantity, Sales Amount, Page View), Conversion Rate, Average Sales Price or Sales Profit) And environmental data (date information, day of the week information, weather information, product information in progress, shopping mall information or other product information in the shopping mall),
The recommended promotion attribute data refers to a promotion category, promotion period, promotion shopping mall, exposure account or promotion image color,
The compensation information means an increase or decrease in the performance data for a specific period,
The policy network is an artificial neural network that outputs the recommended promotion attribute probability in each state, and the cost function of the policy network is configured as in Equation 1 below,
The value network is an artificial neural network that generates the reward possibility information, which is a possibility of achieving a reward in each state, and the cost function of the value network is configured as in Equation 2 below,
Device for predicting and recommending promotion performance in online shopping malls using policy and value networks:
[Equation 1]

[Equation 2]

In Equation 1, π is a policy function, θ is a policy network parameter, π _θ (a _i │s _i ) is the probability of performing a specific action (promotion of a specific attribute) in the current episode, V is a value function, and w is a value Network parameter, s _i is the state information of the current episode i, S _i+1 is the state information of the next episode i+1, r _i+1 is the reward expected to be acquired in the next episode, V _w (s _i ) Is the probability of compensation in the current episode, V _w (s _i+1 ) is the probability of compensation in the next episode, γ is the depreciation rate,
In Equation 2, V is a value function, w is a value network parameter, s _i is the state information of the current episode i, S _i+1 is the state information of the next episode _i+1 , and r _i+1 is the next episode. Reward expected to be obtained at, V _w (s _i ) is the probability of compensation in the current episode, V _w (s _i+1 ) is the probability of compensation in the next episode, and γ is the depreciation rate.

Images