KR102645901B1 - Method and apparatus for personalized item recommendation applying knowledge graph embedding - Google Patents

Abstract

The present invention discloses a personalized item recommendation method and device using knowledge graph embedding. According to the present invention, it includes a processor and a memory connected to the processor, wherein the memory includes a purchase item list, a purchase history including purchase times of items included in the purchase item list, and an item purchase log including a buyer ID. receives input, generates a first buyer vector reflecting the weight calculated by applying a self-attention mechanism to the purchase history, and establishes a correlation between at least some items included in the purchase item list at different times based on the purchase history. Generate a graph representing the relationship, generate an item vector using the generated graph, generate an updated second buyer vector using the first buyer vector and the item vector, and create a knowledge graph with the second buyer vector. A personalized item recommendation device is provided that stores program instructions executable by the processor to determine a recommended item by calculating a similarity score using embedding.

Description

Method and apparatus for personalized item recommendation applying knowledge graph embedding}

The present invention relates to a personalized item recommendation method and device using knowledge graph embedding.

Recently, as transactions on online platforms such as Amazon have increased, interest in recommendation systems using sequential data (for example, a buyer's click log or a buyer's purchase history over time) is increasing.

The sequential data-based recommendation system analyzes the stored buyer history and recommends the item the buyer will purchase next. Recommendation systems based on sequential data are being studied either based on deep learning or non-deep learning.

Deep learning-based recommendation systems include GRU4Rec based on RNN (Recursive Neural Network) or CASER (Convolutional Sequence Embedding) based on CNN (Convolutional Neural Network).

Non-deep learning-based recommendation systems use Factorized Personalized Markov Chains (FPMC), which express changes over time in the interaction state between a buyer and an item or items as a conditional probability, to purchase an item based on the most recently purchased item. I recommend it to

However, since deep learning-based recommendation systems predict items to be purchased by buyers based on data that summarizes all past actions, it is difficult to expect good performance when the amount of data is sparse.

Additionally, non-deep learning-based recommendation systems have the disadvantage that the prediction process is too simple and therefore only works well on sparse data. SASRec (Self-Attentive Sequential Recommendation), a newly researched method, receives the list of previously purchased items and the purchase order of items, applies a Self-Attention mechanism to give weight to the purchased items, and provides a list of recommended items and purchase order. The degree of relationship in an item list is calculated and the item with the highest degree of relationship is recommended.

Although this method showed higher recommendation performance than existing methods, it has the problem of not being able to accurately represent the correlation between items or between buyers and items.

Republic of Korea Patent Publication 10-2020-0114969

In order to solve the problems of the prior art described above, the present invention seeks to propose a personalized item recommendation method and device applying knowledge graph embedding that can precisely represent the correlation between items or between buyers and items. do.

In order to achieve the above-mentioned object, according to an embodiment of the present invention, there is provided a personalized item recommendation device, comprising: a processor; and a memory connected to the processor, wherein the memory receives an item purchase log including a purchase history and purchaser ID including a purchase item list and purchase times of items included in the purchase item list, and Applying a self-attention mechanism to the history generates a first buyer vector reflecting the calculated weight, and generates a graph showing the correlation between at least some items included in the purchase item list at different time points based on the purchase history. An item vector is generated using the generated graph, an updated second buyer vector is generated using the first buyer vector and the item vector, and a similarity score is calculated using the second buyer vector and knowledge graph embedding. A personalized item recommendation device is provided that stores program instructions executable by the processor to calculate and determine a recommended item.

The graph is generated at different time points at preset time intervals, is connected to a first node corresponding to the item purchased at the first time point, and the first node is connected to the first node, and is connected to the first node at a predetermined time interval before and after the first time point. It may include one or more second nodes corresponding to the item purchased at the corresponding time.

The program instructions may determine characteristic values of all items included in the purchase item list by embedding nodes included in the graph generated at different viewpoints in the transition space using a graph neural network algorithm.

The trunk line may be given a weight according to the number of items purchased at a time corresponding to a predetermined time interval before and after the first time point.

The second buyer vector may be generated by calculating a weight through entropy calculation to the value obtained by adding the first buyer vector and the item vector, and assigning the calculated weight to the first buyer vector.

The first buyer vector may be generated by considering the purchase order of items included in the purchase item list.

The similarity score is calculated according to the following equation, and the item whose similarity score is ranked higher than a preset rank among the candidate items can be determined as the recommended item.

[Equation]

here, is the first buyer vector, is the item vector, is the candidate item vector, is the second buyer vector

The second buyer vector can be generated according to the equation below.

[Equation]

here, is the sigmoid function ego, and is a learnable parameter, b is a bias vector

According to another aspect of the present invention, there is a method of recommending personalized items in a device including a processor and memory, including a purchase item list, purchase history including the purchase time of the items included in the purchase item list, and a buyer ID. Step of receiving an item purchase log; generating a first buyer vector reflecting weights calculated by applying a self-attention mechanism to the purchase history; generating a graph representing a correlation between at least some items included in the purchased item list at different times based on the purchase history; generating an item vector using the generated graph; generating an updated second buyer vector using the first buyer vector and the item vector; and determining a recommended item by calculating a similarity score using the second buyer vector and knowledge graph embedding.

According to another aspect of the present invention, a computer readable program for performing the above method is provided.

According to the present invention, when creating a graph, the edges between item nodes are not represented based on the latest purchase criteria, but the point in time that has passed a certain period of time from the time of purchase is flexibly applied, and the correlation between items is calculated at the time of embedding to create new weights. Because it is given, the item vector has the advantage of being able to well represent the dependency relationship with other items, inheritance relationship, interaction relationship, and the situation when the customer purchases the item.

In addition, according to the present invention, since the self-attention mechanism is applied to the purchase history and expressed as a buyer vector, the buyer information is expressed in the purchased item and the temporal order of the purchased item, which has the advantage of better representing the buyer's characteristics. there is.

Furthermore, according to the present invention, by applying a knowledge graph embedding technique and applying a similarity score function to find items with high similarity to the sum of the buyer vector and the item vector, it is possible to recommend items with high accuracy.

Figure 1 is a diagram illustrating the process of recommending an item by applying TransRec according to the prior art.
Figure 2 shows the definition of TransE, a representative algorithm of knowledge graph embedding.
Figure 3 is a diagram showing the configuration of a personalized item recommendation device according to a preferred embodiment of the present invention.
Figure 4 is a flowchart showing an item recommendation based on purchase history according to this embodiment.
Figure 5 is a diagram showing the weight calculation process for generating the first buyer vector according to this embodiment.
FIG. 6 is a diagram illustrating an example of creating a graph by connecting item nodes based on the purchase history of the item to generate an item vector according to this embodiment.
Figure 7 graphically illustrates the process of determining a recommended item using the second buyer vector and the item vector according to this embodiment.

Since the present invention can make various changes and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail.

However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all changes, equivalents, and substitutes included in the spirit and technical scope of the present invention.

This embodiment proposes an algorithm that embeds buyer information and item information in a transition space, represents them as buyer vectors and item vectors, and recommends the item with the highest similarity to the item the buyer purchased.

According to this embodiment, the problems of the existing SASRec can be solved by recommending items based on correlations between items or between buyers and items by embedding data in the conversion space.

It may be similar to TransRec (Translation-based Recommendation) in that it recommends items with high similarity by embedding buyer information and previously purchased item information into buyer vectors and item vectors, but TransRec uses additional item vectors and buyer vectors. Because it does not contain information and is based on recently purchased items, it has the disadvantage that vectors do not have a significant impact on time information or relationships between purchased items in item recommendation.

In this embodiment, a list of items purchased by a buyer at different times is displayed as a graph, and then a GNN (Graph Neural Network) algorithm is used to represent the relationship between previously purchased items and subsequently purchased items using the structure of the graph. The difference is that the buyer vector is created and embedded by embedding it as an item vector, and weighting each item through a self-attention mechanism in the purchase history containing information on purchasing items sequentially.

In addition, this embodiment can express the information of the items more precisely by graphically expressing the interaction between items purchased at different times, and utilizes the buyer information expressed through the self-attention mechanism of the buyer's purchase history. It has the advantage of showing higher recommendation performance.

Before explaining this embodiment, existing TransRec will be described.

Figure 1 is a diagram illustrating the process of recommending an item by applying TransRec according to the prior art.

Referring to Figure 1, TransRec recommends the most similar item by embedding information on items purchased so far and buyer information in the conversion space.

TransRec uses a knowledge graph embedding method. Knowledge graph embedding means that if there is a knowledge expression such as (Mona Lisa (h), was created (r), Leonardo da Vinci (t)), then "Mona Lisa", "was created", and "Leonardo da Vinci (t)". The concept is that when the "Vinci" concept is embedded in the transition space, the sum of the Mona Lisa vector and the created vector will be similar to the Leonardo da Vinci vector.

Figure 2 shows the definition of TransE, a representative algorithm of knowledge graph embedding.

Referring to Figure 2, given h and r, this is an algorithm that finds the optimal knowledge expression by finding t that minimizes h+r-t. TransRec applies the knowledge graph embedding method to find the item vector with the highest similarity to the sum of the item vector embedding previously purchased item information and the buyer vector embedding buyer information, and determines the item as a recommended item.

In this embodiment, the same method is used to recommend an item with a value most similar to the sum of vectors embedding buyer information and item information.

However, the buyer vector is weighted by applying a deep learning algorithm to the items purchased so far and the purchase order, and the item vector represents the items purchased by the buyer in a graph, so that each item is a list of previously purchased items and the purchase order later. It is expressed as a vector, including the characteristics of the item list, the correlation with items purchased before and after a specific point in time, and the characteristics of the item.

Figure 3 is a diagram illustrating the configuration of a personalized item recommendation device according to a preferred embodiment of the present invention, and Figure 4 is a flowchart showing an item recommendation based on purchase history according to this embodiment.

As shown in FIG. 3, the device according to this embodiment may include a processor 300 and a memory 302.

Here, the processor 300 may include a central processing unit (CPU) capable of executing a computer program or another virtual machine.

Memory 302 may include a non-volatile storage device, such as a non-removable hard drive or a removable storage device. Removable storage devices may include compact flash units, USB memory sticks, etc. Memory 302 may also include volatile memory, such as various types of random access memory.

Program instructions executable by the processor 300 are stored in the memory 302, and the process shown in FIG. 4 is performed through the program instructions.

Referring to FIG. 4, the device according to this embodiment receives a purchase item list, a purchase history including purchase times of items included in the purchase item list, and an item purchase log including a buyer ID (step 400), A weight is calculated by applying a self-attention mechanism to the purchase history (step 402).

Step 402 is a process of calculating a weight for each item by considering the purchase order of each item included in the purchase item list.

Next, a first buyer vector reflecting the weight calculated in step 402 is generated (step 404).

Figure 5 is a diagram showing the weight calculation process for generating the first buyer vector according to this embodiment.

The number k of purchase items to be referenced to generate the first buyer vector can be arbitrarily set by the recommendation algorithm user. For example, if you set k=50, the recommendation algorithm will use the buyer's purchase history from the 50th time ago.

Referring to FIG. 5, since the temporal order information of purchased items plays an important role in the recommendation algorithm using sequential data according to this embodiment, the temporal location information of each item is calculated using a location embedding algorithm to determine the weight for each item. Calculate

To calculate the weight, a self-attention mechanism is applied to the purchase history including the purchase item list and the purchase time of the items included in the purchase item list.

In addition, the device according to this embodiment generates a graph representing the correlation between at least some items included in the purchase item list at different time points based on the above-described purchase history (step 406), and adds a GNN (GNN) to the generated graph. Graph Neural Network) is applied to generate an item vector (step 408).

FIG. 6 is a diagram illustrating an example of creating a graph by connecting item nodes based on the purchase history of the item to generate an item vector according to this embodiment.

In FIG. 6, the graph is a directed graph, is generated at different time points at preset time intervals, is connected to a first node corresponding to the item purchased at the first time point, and the first node is connected to the first point in time. It includes one or more second nodes corresponding to items purchased at a time corresponding to a predetermined time interval before and after .

Based on the nodes indicated by gray rectangles in Figure 6, items purchased later are connected with orange arrows, and items purchased previously are connected with blue arrows.

The edges between item nodes are separated not based on the recently purchased item, but based on the time when a preset period has elapsed from the time of purchase.

The preset period can be arbitrarily set by the recommendation algorithm user.

At this time, if the existing period is less than the period set by the recommendation algorithm user, the item node can be connected to other nodes for the existing period.

The right side of Figure 6 shows the node corresponding to the third purchased item connected to the nodes of other items.

The edges connecting each node in the graph have weights. If n is 3, the edge connecting the nodes is defined as 1/3.

However, when drawing a graph based on the third purchased item in Figure 6, since there are only two items at the previous time, the edge with the item at the previous time is defined as 1/2.

According to this embodiment, based on the graph generated in step 406, nodes in the graph representing purchase items over time are embedded using a graph neural network algorithm mainly used in graph analysis. Item nodes have unique characteristic values, and more useful information can be obtained from item nodes with similar characteristic values than from items that are completely unrelated to the item.

Therefore, the connection relationship between an item node and an item node connected by a trunk line is calculated, and more weight is given if the connected item is the same type of item as the item node or is an item that is usually purchased in the same situation.

Additionally, because the recommendation system is influenced by the most recently purchased item, the item node is re-embedded by adding the characteristic value of the recently purchased item node to each item node, allowing the item node to retain time characteristics.

In order to consider the relationship between purchased items, the item vector is expressed using a graph neural network algorithm that shows high performance in graph structure analysis, so the item vector not only has correlations with previously and subsequently purchased items, but also itself. Information can be displayed effectively, allowing item information to be displayed flexibly.

Referring again to FIG. 4, the device according to this embodiment generates an updated second buyer vector using the first buyer vector and the item vector (step 410), and generates a similarity score using the second buyer vector and knowledge graph embedding. Calculate to determine a recommended item (step 412).

As described above, the device according to this embodiment defines buyer information and item information and then embeds each piece of information in the conversion space to determine the candidate item with the most similar values to the buyer information and item information as the recommended item.

으로 표현할 수 있다. When the buyer's information is p, the item information is r, and the information of the item to be recommended is q,

It can be expressed as

In other words, among the list Q of items to be recommended, q with the highest similarity between p+r and q is recommended.

Since the recommendation system is not accuracy as a measure of evaluation, but rather relevance to the product desired by the buyer, it requires flexibility in embedding buyer and item information.

As shown in the equation below, the similarity score is calculated according to the following equation, and the item whose similarity score is ranked higher than the preset rank among the candidate items is determined as the recommended item.

here, is the first buyer vector, is the item vector, is the candidate item vector, is the second buyer vector.

Additionally, the second buyer vector is generated according to the equation below.

here, is the sigmoid function ego, and is a learnable parameter, and b is a bias vector.

Figure 7 graphically illustrates the process of determining a recommended item using the second buyer vector and the item vector according to this embodiment.

In Figure 7, p is the first buyer vector, r is the item vector, p' is the second buyer vector, and q is the recommended item determined through the second buyer vector and the item vector.

As shown in FIG. 7, this embodiment uses the updated second buyer vector rather than the first buyer vector.

In other words, by defining a second buyer vector to which a weight reflecting the correlation with the item is applied to the first buyer vector, when calculating a recommended item through the correlation between buyer information and item information, the item is evaluated in a broader sense than the existing recommendation system. The recommendation was made to increase customer satisfaction with the recommendation.

Preferably, the second buyer vector is calculated by adding the first buyer vector and the item vector and reflecting the weight calculated through entropy calculation in the first buyer vector.

According to this embodiment, the item vector and the buyer vector are analyzed and a weight is assigned to the buyer vector so that the buyer vector can comprehensively represent the buyer's purchase history and information on the purchased item.

The above-described embodiments of the present invention have been disclosed for illustrative purposes, and those skilled in the art will be able to make various modifications, changes, and additions within the spirit and scope of the present invention, and such modifications, changes, and additions will be possible. should be regarded as falling within the scope of the patent claims below.

Claims (10)

As a personalized item recommendation device,
processor; and
Including a memory connected to the processor,
The memory is,
Receive input of a purchase item list, a purchase history including the purchase time of the items included in the purchase item list, and an item purchase log including the buyer ID,
Generate a first buyer vector reflecting the weight calculated by applying a self-attention mechanism to the purchase history,
Generating a graph showing the correlation between at least some items included in the purchased item list at different times based on the purchase history,
Create an item vector using the generated graph,
Generating an updated second buyer vector using the first buyer vector and the item vector,
To determine a recommended item by calculating a similarity score using the second buyer vector and knowledge graph embedding,
Store program instructions executable by the processor,
The graph is generated at different time points at preset time intervals, is connected to a first node corresponding to the item purchased at the first time point, and the first node is connected to the first node, and is connected to the first node at a predetermined time interval before and after the first time point. Includes one or more second nodes corresponding to the item purchased at the time corresponding to,
The trunk line is given a weight according to the number of items purchased at a time corresponding to a predetermined time interval before and after the first time point,
The second buyer vector is a personalized item recommendation device that is generated by calculating a weight through entropy calculation to the value obtained by adding the first buyer vector and the item vector, and assigning the calculated weight to the first buyer vector. delete According to paragraph 1,
The program commands are:
A personalized item recommendation device that determines the characteristic values of all items included in the purchase item list by embedding nodes included in the graph generated at different viewpoints in the transition space using a graph neural network algorithm. delete delete According to paragraph 1,
A personalized item recommendation device in which the first buyer vector is generated by considering the purchase order of items included in the purchase item list. According to paragraph 1,
The similarity score is calculated according to the following equation, and a personalized item recommendation device determines the item with the highest similarity score higher than a preset rank among the candidate items as the recommended item.
[Equation]

here, is the first buyer vector, is the item vector, is the candidate item vector, is the second buyer vector In clause 7,
A personalized item recommendation device in which the second buyer vector is generated according to the equation below.
[Equation]

here, is the sigmoid function ego, and is a learnable parameter, b is a bias vector 1. A method for recommending personalized items on a device comprising a processor and memory, comprising:
Receiving a purchase item list, a purchase history including each purchase of an item included in the purchase item list, and an item purchase log including a buyer ID;
generating a first buyer vector reflecting weights calculated by applying a self-attention mechanism to the purchase history;
generating a graph representing a correlation between at least some items included in the purchased item list at different times based on the purchase history;
Generating an item vector using the generated graph;
generating an updated second buyer vector using the first buyer vector and the item vector; and
Including determining a recommended item by calculating a similarity score using the second buyer vector and knowledge graph embedding,
The graph is generated at different time points at preset time intervals, is connected to a first node corresponding to the item purchased at the first time point, and the first node is connected to the first node, and is connected to the first node at a predetermined time interval before and after the first time point. Includes one or more second nodes corresponding to the item purchased at the time corresponding to,
The trunk line is given a weight according to the number of items purchased at a time corresponding to a predetermined time interval before and after the first time point,
A personalized item recommendation method in which the second buyer vector is generated by calculating a weight through entropy calculation to the sum of the first buyer vector and the item vector, and assigning the calculated weight to the first buyer vector. A computer program stored in a computer-readable recording medium that performs the method according to claim 9.