CN115293812A

CN115293812A - E-commerce platform session perception recommendation prediction method based on long-term and short-term interests

Info

Publication number: CN115293812A
Application number: CN202210967561.3A
Authority: CN
Inventors: 肖云鹏; 黄于洋; 李暾; 王蓉; 贾朝龙; 陶禹冲; 朱宇
Original assignee: Chongqing University of Post and Telecommunications
Current assignee: Chongqing University of Post and Telecommunications
Priority date: 2022-08-12
Filing date: 2022-08-12
Publication date: 2022-11-04

Abstract

The invention belongs to the technical field of internet application, and particularly relates to a conversation perception recommendation prediction method of an E-commerce platform based on long-term and short-term interests, which comprises the steps of obtaining online data, wherein the data comprises basic information of a user, basic information of an article and a conversation sequence of user behaviors; extracting user behaviors and user preferences through the acquired online data, and constructing a long-term interest set of the user; acquiring the short-term interest of the user at the current stage from the long-term interest set of the user through interest matching; constructing a prediction model, taking a click sequence of a user in a session, namely article input, and short-term interest of the user in the current stage as input, and outputting and recommending the predicted articles by the prediction model; the method and the system can effectively mine the long-term and short-term interest information of the user in the user behavior sequence, more accurately express the interest preference of the user, and simultaneously improve the recommendation accuracy of the e-commerce platform.

Description

E-commerce platform session perception recommendation prediction method based on long-term and short-term interests

Technical Field

The invention belongs to the technical field of internet application, and particularly relates to a prediction method for conversation perception recommendation of an e-commerce platform based on long-term and short-term interests.

Background

In recent years, with the explosion of artificial intelligence technology, it has been the "intelligence" factor that makes recommendations more interesting and useful. Intelligence is a key core of personalization that can learn about a user's preferences, predict preferences unknown to the user, and ultimately provide recommendations beyond simple searches by matching queries and content. Recommendation system research combines a variety of Artificial Intelligence (AI) techniques including machine learning, data mining, user modeling, and case-based reasoning. The idea of having an intelligent system that can think and learn like humans has led to a more humanized technique called Computational Intelligence (CI). CI is a branch of AI that explores adaptive mechanisms to enable intelligent operation in complex and changing environments. Such intelligent "recommendations" may come from a variety of factors, including the digital habits of the user, and the history, preferences, interests, and behaviors of similar users. Recommendation systems have rapidly become one of the most important traffic centers for modern e-commerce websites and any websites with a large amount of content and users. In short, the recommender system is a complex filtering system that predicts consumer preferences in a digital environment.

At the beginning of the recommender system invention, it was easy to discover explicit similarities between people and products, but recommender systems have used a way to look at the similarities of potential attributes by using matrix factorization. Briefly, all of the attributes of an item or a customer are combined in a way that reveals relationships that have not yet been implemented, but this is very limiting and the advent of artificial intelligence allows the recommendation system to discover more potential attributes and hide relationships.

Although a great deal of research is conducted on the session-aware recommendation model by numerous scholars, which can minimize the information loss of the existing recommendation model due to ignoring short-term transactions, some challenges still remain:

1. the interests of the user dynamically evolve over time. User preferences for items under e-commerce platforms are typically dynamic rather than static, and may change significantly or slowly over time.

2. The user's selection of an item depends not only on long-term established preferences but also on short-term recent preferences. Both long-term and short-term interests of the user are important, and it is clearly a problem how to distinguish and exploit these two interests simultaneously.

3. The impact of different interests is different. The user interests are only continuously and equally utilized, so that the model is not real, and the accuracy of the recommendation result is influenced.

Disclosure of Invention

In order to solve the problems, the application provides a conversation perception recommendation prediction method of an e-commerce platform based on long-term and short-term interests, which specifically comprises the following steps:

acquiring online data comprising basic information of a user, basic information of an article and a conversation sequence of user behaviors;

extracting user behaviors and user preferences through the acquired online data, and constructing a long-term interest set of the user;

acquiring the short-term interest of the user at the current stage from the long-term interest set of the user through interest matching;

and constructing a prediction model, taking a click sequence of a user in one session, namely item input, and short-term interest of the user in the current stage as input, and outputting and recommending the predicted items by the prediction model.

Further, constructing the user long-term interest set includes: acquiring a historical commodity matrix according to a conversation sequence of historical user behaviors, performing convolution calculation on the historical commodity matrix by utilizing convolution checks with different scales, obtaining preference feature mapping of a user by each convolution check, converting all preferences into long-term interests of the user through a full connection layer after splicing, and forming a long-term interest set by the long-term interests obtained by all conversations of the user.

Further, when the convolution operation is performed on each of the historical item matrices by using τ convolution kernels, the convolution kernel matrix is expressed as Ω = { ω = } ω ₁ ,ω ₂ ,…,ω _τ And then, the long-term interest of the user obtained at the ith session is expressed as:

wherein ReLU () represents a ReLU activation function; concat () represents a splicing operation; h is a historical commodity matrix and is expressed as H = [ v = ₁ ,v ₂ ,…,v _i ,…] ^T ，v _i Representing an interactive item vector, W, representing an ith sub-session map ^l Is the weight matrix of the fully-connected layer, and b is the offset term of the fully-connected layer.

Further, for a value in the historical item matrix, the value is forced to be set to 0 with a probability expressed as:

wherein, O represents the probability of setting a certain value in the historical commodity matrix to 0; u shape _u Is a user vector representation; v _v Is an item vector representation;

representing a user vector representation resulting from user interaction learning with the item;

representing an item vector representation derived from user interaction learning with the item.

Further, in the process of user vector representation and article vector representation obtained by interactive learning of the user and the article, the loss function is gradually reduced until the loss function reaches a set threshold, and the loss function is represented as:

wherein l is a loss function; y belongs to {0,1}, and represents that the user has interacted with the current item when y =1 and does not interact with the current user when y = 0;

representing an item vector representation derived from user interaction learning with the item and exceeding an average dwell period with a period of dwell at the item;

the item vector representation resulting from user interaction learning with the item and used to stay on the item does not exceed the average stay time.

Further, the short-term interest of the user at the current stage is obtained from the long-term interest set of the user through interest matching, and is represented as:

wherein the content of the first and second substances,

for short-term interest at the nth stage in the process of the s-th session, a user may interact with a plurality of items in one session, wherein each time the user interacts with one item in the session is defined as one stage;

indicating the long-term interest gained in the ith session; m represents the number of elements in the long-term interest set; n is a radical of _s In order to be the number of stages,

a weight representing the short-term interest of the nth stage during the s-th session.

Further, a vector matched from the user's long-term interest set is calculated by an attention mechanism

For is to

Influence of (2), a vector matching from the user's long-term interest set

Expressed as:

wherein, the first and the second end of the pipe are connected with each other,

conversation sequence of embedded vector representing user long-term interest set and current user behavior

A difference projected to the embedding space vector;

indicating the long-term interest gained in the ith session;

representing quantized

<,>Means inner product calculation, w () means weight function;

show to obtain

Taking the average value in the range of the set Q; q represents all predecessor behaviors of the user within the current session.

Further, a prediction model is built based on the gated loop unit, the output of the gated loop unit output gate is used for predicting the probability of each item being clicked through a softmax function layer, and the items are recommended to the user from high to low according to the probability, and the process is represented as follows:

inputting a vector for the item;

in order to be of short-term interest to the user,

in order to be in the intermediate hidden state,

in the state of being hidden, the mobile phone is in a hidden state,

in order to update the door,

in order to reset the gate, the gate is reset,

is an output vector; w _p 、W _q 、W _h A transformation matrix, W, representing GRU units ₁ 、W ₂ 、W ₃ Representing a weight matrix of the GRU unit; σ denotes a sigmoid activation function, an by-bit multiplication operation;

represents the probability that the | V | th item in the set of items V was clicked on, | V | represents the number of items in the set of items V, and W represents the weight matrix that the hidden layer is connected to the output layer.

The method and the system can effectively mine the long-term and short-term interest information of the user in the user behavior sequence, more accurately express the interest preference of the user, and simultaneously improve the recommendation accuracy of the e-commerce platform.

Drawings

FIG. 1 is a flow chart of a long-short term interest-based session-aware recommendation prediction model of the present invention;

FIG. 2 is a schematic diagram of a probabilistic model for predicting purchased items at each stage of a session according to the present invention;

FIG. 3 is a schematic diagram of a CNN convolutional network according to the present invention;

FIG. 4 is a schematic diagram of the SiM algorithm calculating short-term interest preferences according to the present invention;

FIG. 5 is a diagram of a long-term interest evolution model according to the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The invention provides a long-term and short-term interest-based E-commerce platform session awareness recommendation prediction method, which comprises the following steps of:

In this embodiment, as shown in fig. 1, the performing of recommendation prediction mainly includes the following steps:

s1: the data is acquired online. The data may be obtained from a public data set website or by direct query of the enterprise to provide real-time sales data in a database. What needs to be obtained here is a conversation sequence of basic information of a user, basic information of an article and user behavior, the conversation sequence comprises a plurality of elements, the conversation sequence is converted into a fixed length, and data is preprocessed.

S2: and extracting the relevant attributes. And extracting relevant attributes according to the acquired user basic information, the acquired article basic information and the user historical conversation sequence under the E-commerce platform. And the user preference is described by fusing the characteristics from the two aspects of the long-term interest and the short-term interest of the user.

S3: and (5) establishing a model. And constructing a prediction model, predicting to obtain a candidate article sequence, and pushing the sequence to a user.

The embodiment provides a specific method for acquiring a data source, which mainly comprises the following steps:

s11: raw data is acquired. The raw data can be obtained through real-time inquiry of an enterprise database or through a public data set website.

S12: simple data cleaning. The raw data that is typically acquired is unstructured and cannot be used directly for data analysis. Most unstructured data can be structured by simple data cleansing. For example, duplicate data is deleted, invalid information is cleared, and the like.

S13: and (4) storing the data. The data after the simple data cleaning is stored by using the database, the data is further normalized by the table structure, and the data retrieval efficiency and the mapping of the relationship among the tables can be greatly improved by the database.

In the E-business platform, the ordering behavior of the user on the marketing campaign is influenced by various factors, such as the hobbies of the user, the participation behavior of similar users, the influence of a marketing campaign reward mechanism on the impulsive consumption of the user and the like. Based on this, the present embodiment extracts factors affecting user behavior from both internal and external factors, that is, to extract user behavior as an external factor and user preference as an internal factor, and the present embodiment provides a specific extraction process, including:

s21: and extracting the user behavior.

S211: set of user sessions E

Each record of interaction between a user and an item is associated with a session number s, using e _s To represent the set of all items interacted by a given user in the s-th session, all interaction information including and in the current session s is represented as E = { E = ₁ ,e ₂ ,…,e _s }。

S212: user-item interaction matrix T

User set and item set are denoted as U = { U =, respectively ₁ ,u ₂ 8230 = and V = { V ₁ ,v ₂ 8230j, according to the user to the articleDefines an interaction matrix T = { T ] between a user and an item _uv L U belongs to U and V belongs to V, i.e. if user U clicks on item V, then t _uv =1, otherwise t _uv ＝0。

And S22, extracting the user preference.

S221: user item Attention

The Attention of the user to the item can be determined by two conditions of the stay time t of the user on the item page and the number of click times Num (click), which is expressed as Attention = t × Num (click).

S222: user long-term interest set P

With the first s-1 sessions, we prepare a long-term interest set P for each user to characterize the user's long-term interest preferences. The long-term interest set P stores m hidden preference vectors for a user, denoted as

S223: short-term user interest

The long-term interest set of the user before the s-th session is expressed as

Short-term interest of the user is matched to the user's predecessor behavior by each stage n within the s-th session

Derived from the long-term interest set P and can be defined as

The method for mining and memorizing the long-term interest of the user by the constructed prediction model and estimating the candidate item sequence based on the interest evolution of the user comprises the following steps:

mining long-term interest of the user with fine granularity from the user behavior sequence in the conversation, and storing the user behavior sequence to a long-term interest set;

further matching Short-term interest dynamic representation of the user at the current stage from a long-term interest set of the user by designing a SiM (Short-interest Match) interest matching method;

the method comprises the steps of realizing the process of constructing user interest change generated along with the time by the GRU;

a gated cyclic neural network is introduced to model a click sequence of a user in a session, and a candidate article sequence is estimated.

In the process of obtaining the long-term interest set of the user, a long-term interest set module P is defined for each user in s sessions, and from a feature level, by performing feature extraction on historical articles interacted by the user and storing m hidden interest vectors, preference features of each user are described, as shown in fig. 3, the method specifically includes the following steps:

inputting an embedding layer according to a historical conversation sequence of a user to obtain a historical commodity matrix H = [ v ] ₁ ,v ₂ ,…,v _i ,…] ^T In which

Multiple convolution kernels

Performing convolution calculation on the matrix H, wherein each convolution kernel can obtain a preference feature mapping p of one user _i Namely:

wherein, w _t Denotes the element of the convolution kernel ω with index t, H _t,i Representing the element in the ith row and ith column of the matrix,

an offset term that is a function of the above; n is _I Is the dimension of the embedding vector; v. of _i Interactive item vectors representing the ith sub-session map, the vector packageL features are included, each feature being an element in the vector.

Generating different user preference feature mappings by using a plurality of convolution kernels, and setting omega and tau as convolution kernel matrixes and numbers, obtaining preference feature mappings p of tau users according to the numbers of the convolution kernels, splicing the preference feature mappings p and the preference feature mappings p, and converting the preference feature mappings p into long-term interests of the users through a full connection layer, wherein the calculation is as follows:

where and concat represent convolution calculations and vector splicing operations respectively,

is a weight matrix for the fully-connected layer,

is the bias term for the fully connected layer,

indicating long-term interest to the user. The convolutional neural network can capture the global relationship among the same dimensional features in the user behavior, and can effectively extract the long-term interest of the user.

Since the long-term interest of the user is continuously increased along with the behavior data of the user, the user preference characteristics are written after each session is ended, and the accuracy of the recommendation system is further enriched. While at the same time to prevent overfitting, the present embodiment introduces Dropout to reduce the sensitivity of the model to noise. For example, the user may sometimes inadvertently click on some items that the user is not interested in, and the model may be overfit due to this noisy portion of the data. Specifically, the input article feature is forced to be 0 according to a certain probability, and the probability is expressed as:

wherein the content of the first and second substances,

is a representation of the user's vector learned by the model,

is a vector representation of the object learned by the model, U _u And V _v User and item vector representations, respectively, externally input as supervisory signals;

a vector for the predicted user;

a vector for predicting the item;

probability of clicking for a user;

is the probability of the item being clicked; u represents the value range of the user; v represents the value range of the article; t denotes the operation of matrix row-column transposition.

In order to simplify the training process, the present embodiment uses the interaction duration between the user and the article as a supervision signal, and if the residence time of the user on the article exceeds the average time, the user and the article form a positive sample; those that do not exceed the average duration are constructed as negative samples. Through the design of the loss function, the long-term interest of the user, the model of which learns the positive sample, can be more accurate, and the loss function is as follows:

representing an item vector representation resulting from user interaction learning with the item and exceeding an average dwell period with a duration of dwell at the item;

representing an item vector representation resulting from user interaction learning with the item and not exceeding an average dwell period with a duration of dwell at the item; v. of ⁺ An item indicating that the user stays on the item for a period of time exceeding an average period of time; v. of ^- An item that indicates that the user has stayed on the item for a period that does not exceed the average period.

The introduction of the user long-term interest set has two purposes: one of the short-term interest sets is at each stage of a conversation, the long-term interest set can match the short-term interest of a user through the SiM, the matched short-term interest vector can describe the current behavior of the user in a finer granularity by combining the current hidden state of the user, and the other is that after one conversation is finished, the long-term interest set can store the information of the user in the conversation and realize the modeling of the long-term interest evolution of the user through the sequence learning capability of the recurrent neural network.

In the interaction process of the user and the platform, the user conveys a very focused and clear intention to the model through the current clicked item, so that the recommended related result cannot be generalized; however, at the same time, the gathering of the smell will reduce the effectiveness of the distribution, so that the user will feel tired during the browsing process, and therefore the "clear intention, moderate divergence" policy should be followed.

Given a long-term interest set P of a user in the s-th session, the task of this embodiment is to extract the short-term interest of the user at this stage from the long-term interest set through the preamble behavior in the session, and abstract this operation as a matching (Match) operation:

what is shown is the short-term interest of the user at the nth stage in the s-th session. One common practice is by averaging all the interest vectors of the user:

wherein, N _s Is the number of stages and is also the number of interactions with the item in the current session.

Taking the average results in the short-term interest being static across the session. In order to more truly simulate the current preference of a user in a real scene, we need to obtain a finer-grained preference of short-term interest. Therefore, a module for matching the short-term interest of the user based on the long-term interest of the user is designed in the model, and the preface behavior sequence of the user is obtained

And carrying out efficient vector similarity search (ScaNN) with the long-term interest set P of the user, namely the operation shown in the figure 4. The method specifically comprises the following steps:

firstly, the method is to

Mapping to an embedding space, and then finding an embedding vector closest to a query in all P embedding vectors; the vector search process is as follows:

suppose there are two embedded vectors x ₁ And x ₂ Quantize each to one of two centers c ₁ Or c ₂ ；

Each x is _i Is quantified as

To make the inner product

As close as possible to the original inner product<q,x _i >This process can be visualized as

The projection amplitude on q is as similar as possible to x _i The amplitude of the projection on q;

in the conventional quantization method, for each x _i Selecting the nearest center will result in incorrect relative rankings of the two points:

is greater than

But in practice it is<q,x ₁ >Is less than<q,x ₂ >. Thus using anisotropic vector quantization, x ₁ Is assigned to c ₁ Will x ₂ Is given to c ₂ The inner product error is reduced, and the precision is improved, which is shown as follows:

wherein h is _|| (w,||x _i ||)、h _⊥ (w,||x _i | |) represents x _i The scale parameters of (1);

denotes x _i And

the parallel residuals of (1);

represents x _i And

the quadrature residual of (a); the | | table calculates the vector modulo length.

In this embodiment, the following components

As an embedding vector, will

As a difference, find the user interest vector that best matches the long-term interest set, i.e., find the difference

The vector of the smallest size is the vector of the smallest size,

represents:

finally, using f _att The normalized impact value is calculated as an attention function, and as a softmax function:

attention function f _att Reading a vector obtained by matching from a user long-term interest set

And

as an input, and output the influence value between them, a larger value of α indicates that the interest will have a more dominant influence on the user at the current stage. The user's short-term interest can be expressed as a weighted average according to the value of α:

the interest of one user usually changes among different sessions, and in order to simulate the interest evolution phenomenon, an interest evolution module is provided, which has the advantages that more historical related information can be provided for the final interest representation through the interest evolution module, and meanwhile, the user click item sequence prediction can be better carried out through the interest evolution trend.

The following two characteristics can exist in the evolution process due to the long-term interest of users: one of the characteristics is that as the diversity of interests is increased, the interests of the user can drift, and the interest drift can generate certain influence on the behavior of the user; the other is that although the user interests may influence each other, each interest has its own evolution process, so the embodiment only focuses on the development process related to the target item. As shown in fig. 5, in the interest evolution module, an Attention mechanism is adopted to overcome the phenomenon of interest drift, which is shown as:

wherein e is _a Is a vector representation of the candidate item(s),

is a weight matrix, calculated a _t Weight values for candidate items for different interests; nA is the embedding dimension, i.e. the size of the item vector; nH is the degree of the hidden layer.

After the addition of the Attention mechanism, the weight of the update gate is controlled by the Attention Score. The original updating direction is kept, and the updating strength of the hidden layer state can be controlled according to the degree of correlation with the candidate object. In this way, accurately controlling when to update, the degree and direction of the update can solve the problem of information loss caused by directly multiplying the Attention Score on each interest vector as the input of the next layer of ordinary GRU, which is expressed as:

u′ _t ＝u _t *a _t

in AUGRU, the calculated update factor u _t Needs to be multiplied by an a _t Weight factor as final update factor u _t '. Wherein

An implicit state of AUGRU, which is a bit-wise multiplication operation.

In this model, GRU is selected as the RNN implementation, modeling the user's click sequence in one session. Due to its unique gate unit design, the GRU is an RNN architecture that can efficiently extract information from time series with long distance dependencies. The method specifically comprises the following steps:

defined in the nth phase of the input sequence (N =1,2, \8230;, N) _s ) The GRU unit is as follows: article input

User short-term interest input

Intermediate hidden state

Hidden state

Updating door

Reset door

Output vector

Wherein the gate vector

And

has a value range of [1,0 ]]The transfer equation for the GRU unit is as follows:

wherein W _p ，W _q ，W _h A conversion matrix, W, representing GRU units ₁ ，W ₂ And W ₃ Is a matrix used to calculate the output values; σ is a sigmoid activation function, and |, is a bit-wise multiplication operation.

As shown in fig. 2, the GRU unit adopted in the present embodiment is different from the conventional GRU unit in three places, which specifically includes the following:

initial state value of GRU

The value is assigned by the interest evolution module, and the interest evolution module contains the interest information of the user in the previous session, namely a long-term interest set, so that the cold start problem of the traditional GRU can be reduced;

each GRU unit has an additional input supervisory signal indicating a user's short-term interest input

Different from the traditional GRU which directly uses the hidden state as the output, additionally uses

And

to construct an output vector;

experimental results show that compared with the traditional model, the model provided by the application can effectively improve the performance of the model through the improvement.

At the nth stage

Then, a softmax function layer is used to calculate the predicted probability values of all candidate items:

wherein | V | is the number of items in the item set V;

the method is characterized in that the probability that the items i are clicked in the nth stage in the s-th conversation is predicted, and the items are recommended to the user after being sorted in a descending order according to the probability value.

Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims

1. A long-term and short-term interest-based E-commerce platform session awareness recommendation prediction method is characterized by comprising the following steps:

2. The E-commerce platform session awareness recommendation prediction method based on long-term and short-term interests as claimed in claim 1, wherein constructing the user long-term interest set comprises: acquiring a historical commodity matrix according to a conversation sequence of historical user behaviors, performing convolution calculation on the historical commodity matrix by utilizing convolution checks with different scales, obtaining preference feature mapping of a user by each convolution check, converting all preferences into long-term interests of the user through a full connection layer after splicing, and forming a long-term interest set by the long-term interests obtained by all conversations of the user.

3. The method as claimed in claim 2, wherein if τ convolution kernels are used to perform convolution operation on historical item matrices respectively, the convolution kernel matrices are represented as Ω = { ω = ₁ ,ω ₂ ,…,ω _τ And then, the long-term interest of the user obtained in the ith session is represented as:

wherein ReLU () represents a ReLU activation function; concat () represents a splicing operation; h is a historical commodity matrix and is expressed as H = [ v = ₁ ,v ₂ ,…,v _i ,…] ^T ，v _i Representing an interactive item vector, W, representing an ith sub-session map ^l Is made ofThe weight matrix of the connection layer, b is the bias term of the full connection layer.

4. The E-commerce platform session awareness recommendation prediction method based on long and short term interests as claimed in claim 2, wherein for the values in the historical item matrix, the values are forced to be set to 0 with a probability expressed as:

5. The E-commerce platform conversation awareness recommendation prediction method based on long-term and short-term interests as claimed in claim 3, wherein the user vector representation and the item vector representation obtained by interactive learning of the user and the item are processes of making a loss function gradually smaller until a set threshold is reached, and the loss function is represented as:

6. The E-commerce platform session awareness recommendation prediction method based on long-short term interest as claimed in claim 1, wherein the short term interest of the user at the current stage is obtained from the user long term interest set through interest matching, and is represented as follows:

short-term interest at the nth stage during the s-th session;

indicating the long-term interest gained in the ith session; m represents the number of elements in the long-term interest set; n is a radical of hydrogen _s In order to be the number of stages,

7. The E-commerce platform session awareness recommendation prediction method based on long-term and short-term interests as claimed in claim 5, wherein a vector matched from the user's long-term interest set is calculated through an attention mechanism

For is to

Influence of (2), a vector obtained from matching the user's long-term interest set

Expressed as:

wherein the content of the first and second substances,

A difference projected to the embedded space vector;

indicating the long-term interest gained in the ith session;

representing quantized

<,>Means inner product calculation, w () means weight function;

expression solution

Taking the average value in the range of the set Q; q represents the current meetingAll predecessor behaviors of the user within the conversation.

8. The E-commerce platform conversation awareness recommendation prediction method based on long-term and short-term interests as claimed in claim 1, wherein a prediction model is built based on a gating cycle unit, the output of an output gate of the gating cycle unit is used for predicting the probability of each item being clicked through a softmax function layer, and the items are recommended to a user according to the probability from high to low, and the process is represented as:

wherein the content of the first and second substances,

inputting a vector for the article;

in order to be of short-term interest to the user,

in order to be in a middle hidden state,

in the state of being hidden, the mobile phone is in a hidden state,

in order to update the door,

in order to reset the gate, the gate is reset,

is an output vector; w is a group of _p 、W _q 、W _h A transformation matrix, W, representing GRU units ₁ 、W ₂ 、W ₃ Representing a weight matrix of the GRU unit; σ denotes a sigmoid activation function, an-denotes a bit-wise multiplication operation;