CN113704627B - Session recommendation method based on time interval graph - Google Patents

Abstract

The invention discloses a session recommendation method based on a time interval graph. The method models user interests given the current session and recommends items that the current user is most likely to be interested in in the next step. It mainly consists of four parts: the first part is to construct a session graph with time interval attributes according to the item sequence in the user's current session; the second part is to use the time interval graph neural network to update the item vector representation; the third part is based on the user's current session. The sequence of items in the session is used to obtain the vector representation of the user's interest; finally, the items are recommended according to the representation of the user's interest.

Description

A Session Recommendation Method Based on Time Interval Graph

technical field

The invention belongs to the technical field of Internet services, and in particular relates to a session recommendation method based on a time interval graph.

Background technique

Session is the user's interactive behavior in a period of time, and session-based recommendation is to recommend the item that the user clicks next based on the current session. The sequence of items within a session is ordered, and it is necessary to model the sequentiality of items. For example, after buying green plants, there may be a need to buy flower pots. Traditional conversational recommender systems use recurrent neural networks to model user interests, but recurrent neural networks ignore more complex item contextual relationships in conversations. The last clicked item and the next clicked item of an item in the session are called the context of the item. In the e-commerce platform, in the same session, users will repeatedly click and browse the same item. There will be multiple contexts for the same item. For example, if the session is modeled only by the recurrent neural network, there is no connection between multiple contexts of an item. To model the connection between the contexts, the session needs to be constructed as a session. picture. Conversation graphs can capture rich item transfer relationships in conversations.

It is not enough to only consider the sequence of items in a session, and the difference in the interval time between items in a session will also lead to different recommendation results. Such as: the same behavior occurred two hours ago and half an hour ago must have different effects on the current. Therefore, consider the time interval between item interactions when constructing the conversation graph. The time interval is first normalized by the minimum-maximum normalization method, and then the time interval is discretized to learn the influence of the time interval on the representation of user interest. This method first builds a session graph with time interval attributes based on the session, and updates the item vector representation based on the session graph; then uses a gated recurrent neural network (GRU) to model the session to obtain the user interest vector representation; finally, according to the user interest representation, Recommend the next item the user might be interested in.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to model the user's interest given the current session, and recommend the items that the current user is most likely to be interested in in the next step. In order to capture the rich item transfer relationship in the session, the current session is constructed as a session graph. And when building the conversation graph, the time interval between item interactions is taken into account to learn the impact of time interval on user interest modeling.

A session recommendation method based on time interval graph, including the following steps:

According to the sequence of items in the user's current session, a session graph with time interval attributes is constructed. Given a session s={v ₁ ,v ₂ ,...,v _|s| }, any item v _j is a node of the session graph T, (v _j-1 , v _j ) is a directed edge of the graph network T , indicating that a user clicks on item v _j-1 and then clicks on item v _j . And the edge property of the graph is the time interval between clicking on item v _j-1 and clicking on item v _j .

Update item vector representations using a time-interval graph neural network. The edge attribute time interval in the conversation graph is normalized by the minimum-maximum normalization method, and then the time interval is discretized. When the item node information is transmitted in the conversation graph, the node information and edge information are connected as a whole for transmission. The specific formula is as follows:

和

分别表示该会话中时间间隔的最大值和最小值，t_i→j表示经过最小最大归一化后的时间间隔。函数bucket_id(t_i→j,{·})表示时间间隔t_i→j处在集合参数中的下标，如t_i→j＝0.15，那么

表示0.15落在第2个区间[0.1,0.2)内。emb(t_i→j)表示t_i→j的向量表征，emb(v_i ) 表示物品v_i向量表征，

表示向量的连接。Among them, t _j represents the time when the item v _j is clicked,

and

represent the maximum and minimum value of the time interval in the session, respectively, and t _i→j represents the time interval after the minimum and maximum normalization. The function bucket _id (t _i→j ,{}) represents the subscript of the time interval t _i→j in the set parameter, such as t _i→j =0.15, then

Indicates that 0.15 falls within the second interval [0.1, 0.2). emb(t _i→j ) represents the vector representation of t _i→j , emb(vi ) _{represents the vector representation of item v i ,}

Represents a concatenation of vectors.

具体公式为：Then, the two-layer graph network is used to update the item vector representation, and finally the vector representation x _j of the item v _j is obtained as the graph network update result, that is,

The specific formula is:

表示向量的连接。k代表在图网络T中的搜索深度，k的最大值为搜索深度L，

代表节点v_j在k层的向量表征，B(j) 为会话图T中物品v_j的邻居集合。

表示物品节点x_j的邻居节点v_k传递过来的信息，

融合了物品节点v_j的所有邻居节点信息。

融合了物品v_j上一层的向量表征信息和物品v_j邻居节点信息。where W _pool and W _h are transition matrices, b is a vector, σ is a sigmoid function, and max represents an element-level max operation,

Represents a concatenation of vectors. k represents the search depth in the graph network T, the maximum value of k is the search depth L,

represents the vector representation of node v _j at layer k, and B(j) is the set of neighbors of item v _j in the conversation graph T.

Represents the information passed by the neighbor node v _k of the item node x _j ,

All neighbor node information of item node v _j is fused.

It combines the vector representation information of the upper layer of item v _j and the neighbor node information of item v _j .

According to the item sequence in the user's current session, the user's interest vector representation is obtained. A gated recurrent neural network (GRU) is used to characterize the session to obtain an interest representation. That is, the current session is used as the input of the gated recurrent neural network (GRU), and the output of the gated recurrent neural network (GRU) is obtained as the user's interest p _u :

z _i =σ(W _xz ·x _i +W _hz ·h _i-1 )

r _i =σ(W _xr ·x _i +W _hr ·h _i-1 )

是当前记忆内容。x_i是当前层的节点输入，也就是物品v_i的向量表征。W_xa、W_ha、W_xr和W_hr分别是控制更新门z_i和重置门r_i的参数。W_xh和W_hh是控制前记忆内容

的参数。⊙是元素级别的矩阵相乘，σ是sigmoid函数。门控制循环神经网络(GRU)的最后一个隐藏层的输出就是用户兴趣p_u。Among them, _ri is the reset gate (reset gate), _zi is the update gate (update gate), these two gate vectors determine which information can be used as the output of the gated recurrent unit.

is the current memory content. x _i is the node input of the current layer, that is _, the vector representation of the item vi. W _xa , W _ha , W _xr and W _hr are parameters that control the update gate _zi and the reset gate _ri , respectively. W _xh and W _hh are the pre-control memory contents

parameter. ⊙ is the element-wise matrix multiplication and σ is the sigmoid function. The output of the last hidden layer of the gated recurrent neural network (GRU) is the user interest p _u .

Recommend items based on user interest representations. Multiply the vector v _j of item v _j by the final vector representation of the user, and then apply the softmax function to calculate the score of item v _j :

代表物品v_j成为下一个交互物品的可能性。同时根据

的对数似然函数值，计算损失函数：where p _u represents the user's interest vector,

represents the possibility of item v _j being the next interactive item. At the same time according to

The log-likelihood function value of , calculates the loss function:

函数用梯度下降法来最优化。Among them, y _j represents the one-hot encoding of v _j ,

The function is optimized using gradient descent.

The beneficial technical effects of the present invention are as follows:

(1) The present invention is a recommendation method based on a conversation graph, which constructs a conversation into a conversation graph, and allows the model to learn more complex item transfer relationships in the conversation.

(2) The present invention is a recommendation method based on a time interval session graph, which can not only learn the complex item transfer relationship in the session, but also learn the influence of the item time interval on the user's interest.

Description of drawings

1 is a schematic flowchart of a session recommendation method based on a time interval diagram of the present invention;

FIG. 2 is a model frame diagram of a session recommendation method based on a time interval graph according to the present invention.

Detailed ways

In order to further understand the present invention, a session recommendation method based on a time interval graph provided by the present invention will be specifically described below with reference to specific embodiments, but the present invention is not limited to this, and those skilled in the art will make a The non-essential improvements and adjustments still belong to the protection scope of the present invention.

First, the relevant definitions of the variables and formulas used need to be given.

Definition 1. V: an item set, and V={v ₁ , v ₂ ,...,v _|V| }, |V| represents the number of items in the item set.

Definition 2.s: the current session of the current user, the session is the set of all interactive items in the current time period s={v ₁ , v ₂ ,...,v _|s| }, where |s| represents the number of items in the session.

Definition 3.S: session set in the system, S={s ₁ , s ₂ ,...,s _|S| }, |S| represents the number of sessions in the session set.

Definition 4.T: A session graph with time interval attribute constructed based on the sequence of interacting items in the current session.

Definition 5.B(j): The set of neighbors of item v _j in the conversation graph T with time interval attribute.

物品v_j的初始化的向量表征。Definition 6.

Initialized vector representation of item v _j .

物品v_j的经过会话图更新后的向量表征。Definition 7.

The vector representation of item v _j after the session graph update.

用户兴趣向量表征。Definition 8.

User interest vector representation.

Combined with the above variable definitions, the final problem is defined as: given the current session, model user interests, and recommend the items that the current user is most likely to be interested in in the next step. The items are a subset of the set V. In order to capture the rich item transfer relationship in the session, the current session is constructed as a session graph. When constructing the conversation graph, the time interval between item interactions is taken into account to learn the impact of time interval on the representation of user interest.

To this end, the present invention proposes a session recommendation method based on a time interval graph. As shown in FIG. 2 , the forward propagation part of the method mainly consists of four parts. The first part is to construct a session graph with time interval attributes according to the item sequence in the user's current session; the second part is to use the time interval graph neural network to update the item vector representation; the third part is to obtain the user's Interest vector representation; finally, according to the user's interest representation, recommend items.

As shown in FIG. 1 , according to an embodiment of the present invention in e-commerce, the method includes the following steps:

S100, construct a session graph with a time interval attribute according to the item sequence in the user's current session. Given a session s = {v ₁ ,v ₂ ,...,v _|s| }, any item v _j is a node of the session graph T, and (v _j-1 ,v _j ) is a directed edge of the graph network T , indicating that a user clicks on item v _j-1 and then clicks on item v _j . And the edge property of the graph is the time interval between clicking on item v _j-1 and clicking on item v _j .

S200, using the time interval graph neural network to update the item vector representation. The edge attribute time interval in the conversation graph is normalized by the minimum-maximum normalization method, and then the time interval is discretized. When the item node information is transmitted in the conversation graph, the node information and edge information are connected as a whole for transmission. The specific formula is as follows:

和

分别表示该会话中时间间隔的最大值和最小值，t_i→j表示经过最小最大归一化后的时间间隔。函数bucket_id(t_i→j,{·})表示时间间隔t_i→j处在集合参数中的下标，如t_i→j＝0.15，那么

表示0.15落在第2个区间[0.1,0.2)内。emb(t_i→j)表示t_i→j的向量表征，emb(v_i ) 表示物品v_i向量表征，

表示向量的连接。Among them, t _j represents the time when the item v _j is clicked,

and

represent the maximum and minimum value of the time interval in the session, respectively, and t _i→j represents the time interval after the minimum and maximum normalization. The function bucket _id (t _i→j ,{}) represents the subscript of the time interval t _i→j in the set parameter, such as t _i→j =0.15, then

Indicates that 0.15 falls within the second interval [0.1, 0.2). emb(t _i→j ) represents the vector representation of t _i→j , emb(vi ) _{represents the vector representation of item v i ,}

Represents a concatenation of vectors.

具体公式为：Then, the two-layer graph network is used to update the item vector representation, and finally the vector representation x _j of the item v _j is obtained as the graph network update result, that is,

The specific formula is:

表示向量的连接。k代表在图网络T中的搜索深度，k的最大值为搜索深度L，

代表节点v_j在k层的向量表征，B(j) 为会话图T中物品v_j的邻居集合。

表示物品节点v_j的邻居节点v_k传递过来的信息，

融合了物品节点v_j的所有邻居节点信息。

融合了物品v_j上一层的向量表征信息和物品v_j邻居节点信息。本方法中采用两层的图网络，所以搜索深度L＝2。where W _pool and W _h are transition matrices, b is a vector, σ is a sigmoid function, and max represents an element-level max operation,

Represents a concatenation of vectors. k represents the search depth in the graph network T, the maximum value of k is the search depth L,

represents the vector representation of node v _j at layer k, and B(j) is the set of neighbors of item v _j in the conversation graph T.

Represents the information passed by the neighbor node v _k of the item node v _j ,

All neighbor node information of item node v _j is fused.

It combines the vector representation information of the upper layer of item v _j and the neighbor node information of item v _j . In this method, a two-layer graph network is used, so the search depth L=2.

S300, according to the sequence of items in the current session of the user, obtain the vector representation of the user's interest. A gated recurrent neural network (GRU) is used to characterize the session to obtain an interest representation. That is, the current session is used as the input of the gated recurrent neural network (GRU), and the output of the gated recurrent neural network (GRU) is obtained as the user's interest p _u :

z _i =σ(W _xz ·x _i +W _hz ·h _i-1 )

r _i =σ(W _xr ·x _i +W _hr ·h _i-1 )

是当前记忆内容。x_i是当前层的节点输入，也就是物品v_i的向量表征。W_xa、W_na、W_xr和W_hr分别是控制更新门z_i和重置门r_i的参数。W_xh和W_hh是控制前记忆内容

的参数。⊙是元素级别的矩阵相乘，σ是sigmoid函数。门控制循环神经网络(GRU)的最后一个隐藏层的输出就是用户兴趣p_u。Among them, _ri is the reset gate (reset gate), _zi is the update gate (update gate), these two gate vectors determine which information can be used as the output of the gated recurrent unit.

is the current memory content. x _i is the node input of the current layer, that is _, the vector representation of the item vi. W _xa , W _na , W _xr and W _hr are parameters that control the update gate _zi and the reset gate _ri , respectively. W _xh and W _hh are the pre-control memory contents

parameter. ⊙ is the element-wise matrix multiplication and σ is the sigmoid function. The output of the last hidden layer of the gated recurrent neural network (GRU) is the user interest p _u .

S400, recommend items according to the user's interest representation. Multiply the vector v _j of item v _j by the final vector representation of the user, and then apply the softmax function to calculate the score of item v _j :

代表物品v_j成为下一个交互物品的可能性。同时根据

的对数似然函数值，计算损失函数：where p _u represents the user's interest vector,

represents the possibility of item v _j being the next interactive item. At the same time according to

The log-likelihood function value of , calculates the loss function:

函数用梯度下降法来最优化。Among them, y _j represents the one-hot encoding of v _j ,

The function is optimized using gradient descent.

The above description of the embodiments is for the convenience of those skilled in the art to understand and apply the present invention. It will be apparent to those skilled in the art that various modifications to the above-described embodiments can be readily made, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above-mentioned embodiments, and improvements and modifications made to the present invention by those skilled in the art according to the disclosure of the present invention should all fall within the protection scope of the present invention.

Claims (2)

1. A conversation recommendation method based on a time interval graph is characterized in that:

constructing a conversation graph with a time interval attribute according to an article sequence in the current conversation of a user; given a session s ═ v₁,v₂,…,v_|s|}, any item v_jIs a node of the session graph T, (v)_j-1,v_j) For directed edges of the conversation graph T, indicating that a user clicked on an item v_j-1After clicking on item v_j(ii) a And the edge attribute of the graph is click item v_j-1And click on an item v_jThe time interval in between;

updating the item vector representation using a time interval graph neural network; normalizing the edge attribute time interval in the conversation graph by adopting a minimum maximization normalization mode, and then discretizing the time interval; when transmitting article node information in a session graph, the node information and the side information are connected into a whole for transmission, and the specific formula is as follows:

wherein, t_jRepresenting click items v_jTime of occurrence, t_iRepresenting click items v_iThe time of occurrence of the reaction is,

and

respectively representing the maximum and minimum of the time interval in the session, t_i→jRepresenting the time interval after the minimum and maximum normalization; function bucket_id(t_i→jDenotes the time interval t {. The) }_i→jSubscript in set parameter and bucket_id(t_i→jAssigning a value of the function {. The) } to

To represent

Vector characterization of (c), emb (v)_i) Representing an article v_iVector characterization, # denotes the concatenation of vectors,

representing an article v_iTo an article v_jThe information of (a);

then, the object vector representation is updated by adopting a two-layer graph network, and finally, an object v is obtained_jThe vector of (2) characterizes x_jUpdating the results for the graph network, i.e.

The concrete formula is as follows:

wherein, W_poolAnd W_hThe method comprises the following steps that a is a transfer matrix, b is a vector, sigma is a sigmoid function, max represents max operation at an element level, and ^ represents connection of the vector; k represents the search depth in the session map T, the maximum value of k is the search depth L,

representative node v_jVector characterization at layer k, B (j) is item v in conversation graph T_jA neighbor set of (2);

representing an item node v_jV of a neighbor node_kThe information that is transmitted to the user,

fuse article nodes v_jIs transmitted to the mobile station, and all the neighbor node information,

is a vector;

fuse the article v_jVector characterization information and article v of the previous layer_jNeighbor node information;

obtaining a user interest vector representation according to an article sequence in a current session of a user;a gate control cyclic neural network is adopted to characterize the session to obtain an interest characterization; namely, the current conversation is taken as the input of the gate control cyclic neural network, and the output of the gate control cyclic neural network is taken as the interest p of the user_u；

Recommending the item according to the user interest representation; article v_jVector emb (v) of_j) Multiplied by the user interest p_uThen, the softmax function is applied to calculate the object v_jThe fraction of (c):

wherein p is_uRepresenting the user's interest vector, the superscript T being a transposed symbol, i.e. p_u ^TRepresenting a vector p_uThe transpose of (a) is performed,

representative article v_jThe possibility of becoming the next interactive item; at the same time according to

The log-likelihood function value of (a), calculating a loss function:

wherein, y_jRepresents v_jThe one-hot code of (a) is,

the function is optimized using a gradient descent method.

2. The conversation recommendation method based on the time interval graph as claimed in claim 1, wherein: the gate control cyclic neural network comprises the following components:

z_i＝σ(W_xz·x_i+W_hz·h_i-1)

r_i＝σ(W_xr·x_i+W_hr·h_i-1)

wherein r is_iIs a reset gate, z_iTo update the gate (update gate), these two gating vectors determine which information can be used as the output of the gated loop unit;

is the current memory content; x is the number of_iIs the node input of the current layer, i.e. the item v_iThe vector characterization of (2); w_xz、W_hz、W_xrAnd W_hrRespectively, control the update gate z_iAnd a reset gate r_iThe parameters of (1); w_xhAnd W_hhIs a parameter, W_xhAnd W_hhThe value of (A) influences the current memory content

As a matrix multiplication at the element level, σ is a sigmoid function; the output of the last hidden layer of the gated recurrent neural network (GRU) is the user interest p_u。

Images