CN110458627B

CN110458627B - Commodity sequence personalized recommendation method for dynamic preference of user

Info

Publication number: CN110458627B
Application number: CN201910763463.6A
Authority: CN
Inventors: 黄震华; 汤庸; 刘海; 李丁丁; 蔡立群; 廖晓鹏
Original assignee: Guangdong Zhixin Information Technology Co ltd; South China Normal University
Current assignee: Guangdong Zhixin Information Technology Co ltd; South China Normal University
Priority date: 2019-08-19
Filing date: 2019-08-19
Publication date: 2023-06-30
Anticipated expiration: 2039-08-19
Also published as: CN110458627A

Abstract

The invention discloses a personalized commodity sequence recommendation method for dynamic preference of users, which comprises the steps of extracting commodity sequences under the condition of similar scores of the same user to construct commodity scoring vectors, extracting the user sequences with similar scores of the same commodity to obtain the user scoring vectors, combining personal information of the user, basic attribute information of the commodity, comments of the user and the commodity and commodity pictures, respectively realizing feature extraction of the user and the commodity based on multi-task learning, taking the feature vectors of the user and the historical commodity sequences thereof as input, realizing generation of the commodity sequences through a training encoder-decoder, and accurately learning recommendation of the optimal commodity sequences by combining search strategies. Based on multi-mode user-commodity data, the method and the device for optimizing the commodity sequence based on the multi-mode user-commodity data are used for highly extracting and fusing the user characteristics and the commodity characteristics, realizing personalized recommendation of the commodity sequence oriented to user preference and improving user experience.

Description

Commodity sequence personalized recommendation method for dynamic preference of user

Technical Field

The invention relates to the field of intelligent business, in particular to a commodity sequence personalized recommendation method for dynamic preference of users.

Background

With the advent of the big data age, a large number of users are generating a large amount of information every moment, and these information items are very huge, and how to effectively extract effective features from huge data to make recommendation is a problem to be solved urgently. Information preferences are different between different users in the recommendation field, but there is a certain correlation between similar users and goods. How to put forward a high-efficiency reasonable personalized recommendation method under the background of big data is a very worthy research hotspot.

The existing recommendation algorithm often uses a relation matrix between a user and a commodity and combines other characteristics between the user and the commodity to recommend, and although a certain reference can be provided for a recommendation result, the recommendation systems have the following disadvantages: the method has low processing efficiency on large-data-volume users and article information, dynamic historical preferences of the users and evaluation of historical commodities by the users are easily ignored, and the conventional recommendation results cannot be fed back well; secondly, the traditional recommendation system has the advantages of simple model, single user characteristic, certain limitation in evaluation by adopting a probability statistical method, long recommendation time and larger error; the traditional recommendation system is not combined with the picture features of the commodities, has single task, cannot qualitatively analyze the historical comments of the user and the pictures of the commodities, and cannot effectively evaluate the recommendation features; finally, the sequence information between the commodity sequences with the same user scores and the corresponding commodity sequences is ignored in the traditional recommendation algorithm, and the recommendation results are not reasonably ordered and fused, so that the overall recommendation effect is poor.

Disclosure of Invention

The invention mainly aims to provide a commodity sequence personalized recommendation method oriented to dynamic preference of users, and aims to overcome the problems.

In order to achieve the above purpose, the invention provides a commodity sequence personalized recommendation method for dynamic preference of users, which comprises the following steps:

s10, obtaining personal information of a user, commodity attribute information, commodity pictures, scores of the same user corresponding to different commodities, scores of different users of the same commodity, comment texts of the same user on the commodities and comment texts of the same commodity from the scoring information of the user on the commodities;

s20, constructing a user scoring vector generator according to the scoring of different users received by the same commodity, and generating a user scoring vector; constructing a commodity grading vector generator according to the grading corresponding to different commodities by the same user, and generating commodity grading vectors;

s30, inputting comment texts of the same user on all commodities into a user scoring vector generator to generate user comment vectors, and carrying out mean value clustering on the user comment vectors to generate user comment categories; inputting comment texts of users born by the same commodity into a commodity grading vector generator to generate commodity comment vectors, and carrying out mean value clustering on the commodity comment vectors to generate commodity comment categories; extracting picture feature vectors of the commodity by adopting a depth convolution network, and carrying out mean value clustering on the picture feature vectors to generate commodity categories corresponding to the picture;

S40, inputting the personal information of the user and the scoring vector of the user into a user feature extraction network for joint training to obtain a user ID and a user comment prediction type; inputting commodity attribute information and commodity grading vectors into a commodity feature extraction network for joint training to obtain commodity IDs, commodity comment prediction categories and commodity picture prediction categories;

s50, inputting the feature vectors of the user feature vectors and the feature vectors of the historical commodity sequences into a commodity sequence recommendation network, training by a training encoder-decoder in the commodity sequence recommendation network to generate user feature vectors corresponding to each user ID, taking the commodity feature vectors corresponding to each commodity ID as personalized recommended commodity sequences, and learning by combining a search strategy to obtain the recommendation of the optimal commodity sequences.

Preferably, the S30 includes:

the sentence embedded type end 2Vec tool takes Chinese Wikipedia data as a training corpus, extracts each comment text in the training corpus into one-dimensional comment vectors by adopting the sentence embedded type end 2Vec tool, performs t-means clustering on all the comment vectors to obtain N types of user comment categories, extracts one-dimensional commodity comment vectors from the comment text under the same commodity by adopting the sentence embedded type end 2Vec tool, and performs t-means clustering on all the commodity comment vectors to obtain M commodity comment categories; and extracting feature vectors of the commodity pictures by adopting a deep convolution network VGG16, and t-means clustering the comment vectors to obtain I commodity picture categories.

Preferably, the scoring information of the users used by the user scoring vector generator is divided into 5 groups, and the construction method of the user scoring vector generator comprises the following steps:

grouping into 5 groups according to the user scoring information under the same commodity, wherein the users with the same score are in the same group to obtain a user composition sequence with the same score of the same commodityX is 50, which means that 50 users are selected for each group, and the sequence is selectedAs training sample, use user u _r As input, divide user u in the output prediction sequence _r Scoring vectors of other users, such that other user idsIf the probability product of (a) is maximum, if the total number of users is w=8000, the loss function y of the user scoring vector generator ₁ The method comprises the following steps:

obtaining the scoring direction length of each user to be 200;

the construction method of the commodity grading vector generator comprises the following steps:

dividing the same user with the same scoring commodity into a group, and obtaining the commodity sequence with the same scoring from each groupJ is 20, which means that 20 commodities are selected in each group, and the commodity sequences are selectedAs training sample, commodity i _g As input, divide i in the output prediction sequence _g If the total number of commodities R=120000, obtaining the loss function y of the commodity grading vector generator by using a maximum likelihood estimation method ₂ The method comprises the following steps:

the score vector length for each commodity was found to be 200.

Preferably, the user feature extraction network in S40 includes:

a first convolution layer adopts 2 convolution kernels with the size of 3*1, and the sliding step length is 1;

the first user pooling layer window size is 2*1;

the second user convolution layer adopts 4 convolution layers with the size of 3*1 and the convolution kernel step length of 2; the third user convolution layer adopts 5 convolution kernels with the size of 3*1, and the sliding step length is 1;

the fourth user convolution layer adopts 10 convolution kernels with the size of 5*1, and the step length of the convolution kernels is 2;

the pooling layer window size of the second user pooling layer is 2*1;

the fifth user convolution layer adopts 6 convolution kernels with the size of 3*1, and the step length of the convolution kernels is 2;

the window size of the third user pooling layer is 4*1;

the number of neurons in the user full connectivity layer is 200,

the user randomly deactivates the Dropout layer,

the user personal information and the user scoring vector are convolved through a first user convolution layer, the convolved vector is pooled through a first user pooling layer, then enters a third user convolution layer through batch standardization after being convolved through a second user convolution layer, enters a fourth user convolution layer after being convolved through batch standardization again, enters a user random inactivation Dropout layer after being output by the fourth user convolution layer, enters a second user pooling layer, then enters a user random inactivation Dropout layer for being output, enters a fifth user layer convolution layer, and then enters a user full connection layer after being pooled through the third user pooling layer, and the user characteristic vector is output.

Preferably, the user feature vector is accessed to a user index maximum function softmax layer after being output by a full connection layer, the softmax layer has 8000 nodes, 8000 users in the system are represented, the formed user set is U, the user feature vector is accessed to a comment index maximum function softmax layer after being output by the full connection layer, the softmax layer has 30 nodes, 30 user comment categories are represented, the formed user comment category set is T, the user index maximum function softmax layer and the comment index maximum function softmax layer both adopt a softmax multi-classification loss function,

let the current user be u _r Obtaining user u through a user index maximum function softmax layer _r Prediction of (2)

Probability P (u) _r )：

Wherein sx (u) _r ) Is u _r Input of corresponding node in user index maximum function softmax layerEntering a value;

the comment category of the current user is z _r Obtaining a user comment category z through a comment index maximum function softmax layer _r Is P (z) _r )：

Wherein sx (z) _r ) Is z _r Input values of corresponding nodes in the comment index maximum function softmax layer;

the total loss function of the user feature extraction network is:

wherein eta is L ₁ The value of the regulating coefficient is between 0.7 and 0.8.

Preferably, the personal information of the user includes at least the following 9 attributes: the user age, the user gender, the user occupation, the region where the user is located, the average score, the purchase amount, the average browsing time, the collection amount and the average purchase amount are respectively represented by independent heat vectors, all the attributes are spliced into a long vector, and the long vector is reduced to a vector with the length of 50 through five layers of self-encoder tools; the commodity attribute information at least comprises the following 9 attributes: daily order volume, sales volume, collection volume, online time, use season, average life, average score, price interval.

Preferably, the commodity feature extraction network in S40 includes:

the first commodity convolution layer adopts 5 convolution kernels with the size of 3*1 and the step length of 2;

a second commodity convolution layer which adopts 73 convolution kernels with the size of 3*1, and the sliding step length is 3;

a first commodity pooling layer having a window size of 2*1;

a third commodity convolution layer which adopts 10 convolution kernels with the size of 5*1 and the step length of 3;

a second commodity pooling layer having a window size of 4*1;

a fourth commodity convolution layer which adopts 5 convolution kernels with the step length of 2 and the size of 3*1;

a third commodity pooling layer with a window size of 2*1;

a commodity random inactivation layer;

a fifth commodity convolution layer, which adopts 6 convolution kernels with the size of 3*1 and the step length of 3;

a commodity full-connection layer, the number of the neurons of which is 200,

the commodity attribute information and the commodity grading vector are input into a first convolution layer of a commodity feature extraction network and then are connected into a second convolution layer, the second convolution layer is input into a first pooling layer after output and is input into a third convolution layer after input into a commodity random inactivation layer after input into the second pooling layer after output, the second pooling layer is input into a fourth convolution layer after output, the fourth convolution layer is input into the third pooling layer after output, the third pooling layer is input into the commodity random inactivation layer after output into a fifth convolution layer, the commodity full-connection layer is input after output into the fifth convolution layer, and finally the commodity feature vector is output.

Preferably, the commodity feature vector is output from the commodity full-connection layer and then is connected to a commodity first-index maximum-function softmax layer, the softmax layer has 120000 nodes, 120000 commodities in the system are represented, the formed user set is V, the commodity feature vector is output from the commodity full-connection layer and then is connected to a commodity second-index maximum-function softmax layer, the softmax layer has 15 nodes, 15 commodity types are represented, and the formed commodity type set is A; the commodity feature vector is output from the commodity full-connection layer and then is connected to a commodity third index maximum function softmax layer, the softmax layer is provided with 50 nodes, 50 commodity picture categories are represented, and a formed commodity picture category set is I; the commodity feature vector is output from the commodity full-connection layer and then is connected into a commodity fourth index maximum function softmax layer, the softmax layer is provided with 30 nodes, 30 commodity comment categories are represented, the formed commodity comment category set is M, the commodity first index maximum function softmax layer, the commodity second index maximum function softmax layer, the commodity third index maximum function softmax layer and the commodity fourth index maximum function softmax layer all adopt softmax multi-classification loss functions,

Let the current commodity be i _x Obtaining commodity i through a commodity first index maximum function softmax layer _x Is P (i) _x )：

Wherein sx (i) _x ) Is i _x Input values for corresponding nodes in the first softmax layer;

let the current commodity category be a _x Obtaining commodity category a through commodity second index maximum function softmax layer _x Is P (a) _x )：

Wherein sx (a) _x ) Is a as _x Input values for corresponding nodes in the second softmax layer;

let the current commodity picture category be v _x Obtaining commodity picture category v through commodity second index maximum function softmax layer _x Is P (v) _x )：

Wherein sx (v) _x ) V is _x Input values of corresponding nodes in the third softmax layer;

let the current commodity comment category be m _x Obtaining commodity comment category m through commodity second index maximum function softmax layer _x Is P (m) _x )：

Wherein sx (m) _x ) Is m _x The input values of the corresponding nodes in the fourth softmax layer,

the total loss function of the commodity feature extraction network is:

wherein lambda is ₁ ～λ ₄ Is L ₂ And lambda is the adjustment coefficient of (1) ₁ +λ ₂ +λ ₃ +λ ₄ ＝1，λ ₁ The value is between 0.1 and 0.3, lambda ₂ The value is between 0.2 and 0.5, lambda ₃ The value is between 0.2 and 0.4, lambda ₄ The value is between 0.2 and 0.3.

Preferably, the step S50 specifically includes:

acquiring g commodities currently purchased by a user, and firstly using the previous g in the g commodities ₁ The commodity sequence composed of the individual commodities is used as the input of the encoder and passes g ₁ The gating cyclic GRU unit obtains the background vector of the encoder, then the background vector is spliced with the user characteristic vector and is used as the input of a decoder, and the decoder is composed of g-g ₁ The gate control circulation GRU units are sequentially linked to form output nodes, and the output nodes respectively correspond to the rear g-g in g commodities ₁ The number of the neural networks of each gating cycle GRU unit in the commodity sequence, the encoder and the decoder is set to be 1000, and the length of the background vector generated by the encoder is 200;

for the decoder, assume that the 20 product sequences output are , let s _k (1 is less than or equal to k is less than or equal to 20) is commodity i _k Corresponding decoder hidden variables s _k ＝GRU(i _k-1 ,c,s _k-1 ) Then commodity i _k Output probability p (i) _k ) The representation is:

p(i _k |i ₁ ,i ₂ ,...,i _k-1 ,c)＝p(i _k-1 ,s _k ,c)，

the joint probability of the maximized output sequence of the kth node of the decoder is obtained as follows:

the loss function of the decoder output sequence is: y is ₃ ＝-log P(i ₁ ,i ₂ ,...,i _k )，

Adopting a Beam Search strategy Beam Search, setting the Beam Size beam_Size to be 2, taking an evaluation index as a return, calculating a return value generated by each commodity sequence on each decoding node, selecting the commodity sequence with the largest return value as a current recommended sequence, and calculating the return value through a normalized damage accumulated gain index NDCG, a Recall index Recall and an average precision average value index MAP to obtain the commodity sequence:

reward(i ₁ ,i ₂ ,i _k )＝0.4×NDCG(i ₁ ,i ₂ ,i _k )+0.4×Recall(i ₁ ,i ₂ ,i _k )+0.2×MAP(i ₁ ,i ₂ ,i _k )，

The total loss function of the encoder is thus expressed as:

L ₃ ＝-log P(i ₁ ,i ₂ ,...,i _k )-reward(i ₁ ,i ₂ ,...,i _k )，

back propagation is carried out according to the error of the recommended sequence, and the weight parameter of the encoder-decoder is adjusted;

the feature vector of each user and the feature vector of each commodity in the commodity sequence recommendation training are stored in a database, the user feature vector of each user is stored in a guide table of the database by taking the user id as a first column and the corresponding user feature vector as a second column, the commodity feature vector of each commodity is stored in an entry table of the database by taking the commodity id as a first column and the corresponding commodity feature vector as a second column.

Preferably, after S50, the method further includes:

s60, acquiring g currently purchased by the user according to the purchase log file ₁ The commodity sequence of each commodity is obtained by inquiring the databaseFeature vectors of the commodity are input to an encoder in time sequence to generate background vectors; inquiring a database to obtain a characteristic vector of a current user, splicing a background vector and the characteristic vector of the current user, inputting the spliced background vector and the characteristic vector of the current user into a decoder, and generating g-g ₁ And recommending the commodity sequence formed by the commodities to the current user as an optimal commodity sequence.

Compared with the prior art, the invention has the beneficial effects that:

1. The multi-mode characteristics of the user and the commodity are introduced, the comment, the score and the personal information of the user are effectively used, and meanwhile, the score, the comment, the picture and the attribute information of the commodity are fully utilized, so that the recommendation result is more accurate and reliable.

2. Based on deep learning and multitask learning technology, the invention respectively fuses the user and commodity features through the deep convolutional neural network, and adopts a multitask learning mode to jointly train a plurality of classification tasks, thereby being capable of deeply fusing the user features and commodity features instead of simply splicing.

3. The invention adopts an efficient sequence-to-sequence generation model, is based on the user history commodity sequence, and is integrated with the user characteristic vector, so that personalized recommendation can be realized.

4. The invention has the advantages of clear structure, reasonable logic, lower coupling degree between modules, easy realization and deployment, and capability of being rapidly expanded into a distributed and parallelized development environment, being beneficial to expansion, test maintenance and improving user experience.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to the structures shown in these drawings without inventive effort for a person skilled in the art.

Figure 1 is a logic structure diagram of the commodity sequence personalized recommendation offline training of the invention,

the achievement of the objects, functional features and advantages of the present invention will be further described with reference to the accompanying drawings, in conjunction with the embodiments.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

It should be noted that, if directional indications (such as up, down, left, right, front, and rear … …) are included in the embodiments of the present invention, the directional indications are merely used to explain the relative positional relationship, movement conditions, etc. between the components in a specific posture (as shown in the drawings), and if the specific posture is changed, the directional indications are correspondingly changed.

In addition, if there is a description of "first", "second", etc. in the embodiments of the present invention, the description of "first", "second", etc. is for descriptive purposes only and is not to be construed as indicating or implying a relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include at least one such feature. In addition, the technical solutions of the embodiments may be combined with each other, but it is necessary to base that the technical solutions can be realized by those skilled in the art, and when the technical solutions are contradictory or cannot be realized, the combination of the technical solutions should be considered to be absent and not within the scope of protection claimed in the present invention.

As shown in fig. 1, the method for personalized recommendation of commodity sequences for dynamic preference of users provided by the invention comprises the following steps:

The multi-mode characteristics of the user and the commodity are introduced, the comment, the score and the personal information of the user are effectively used, and meanwhile, the score, the comment, the picture and the attribute information of the commodity are fully utilized, so that the recommendation result is more accurate and reliable. Based on deep learning and multitask learning technology, the invention respectively fuses the user and commodity features through the deep convolutional neural network, and adopts a multitask learning mode to jointly train a plurality of classification tasks, thereby being capable of deeply fusing the user features and commodity features instead of simply splicing. The invention adopts an efficient sequence-to-sequence generation model, is based on the user history commodity sequence, and is integrated with the user characteristic vector, so that personalized recommendation can be realized. The invention has the advantages of clear structure, reasonable logic, lower coupling degree between modules, easy realization and deployment, and capability of being rapidly expanded into a distributed and parallelized development environment, and is beneficial to expansion and test maintenance.

Preferably, the S30 includes:

grouping into 5 groups according to the user scoring information under the same commodity, wherein the users with the same score are in the same group to obtain a user composition sequence with the same score of the same commodityX is 50, which means that 50 users are selected for each group, and the sequence is selectedAs training sample, use user u _r As input, divide user u in the output prediction sequence _r Other users 'scoring vectors are excluded so that the probability product of other user ids is maximized, and if the total number of users w=8000, the user scoring vector generator's loss function y ₁ The method comprises the following steps:

obtaining the scoring direction length of each user to be 200;

the score vector length for each commodity was found to be 200.

Preferably, the user feature extraction network in S40 includes:

the first user pooling layer window size is 2*1;

a second user convolution layer adopts 4 convolution kernels with the size of 3*1 and the convolution kernel step length of 2

A layer; the third user convolution layer adopts 5 convolution kernels with the size of 3*1, and the sliding step length is 1;

the pooling layer window size of the second user pooling layer is 2*1;

the window size of the third user pooling layer is 4*1;

the number of neurons in the user full connectivity layer is 200,

the user randomly deactivates the Dropout layer,

Probability P (u) _r )：

Wherein sx (u) _r ) Is u _r Input values of corresponding nodes in a user index maximum function softmax layer;

the comment category of the current user is z _r By comment on the maximum function of the indexThe softmax layer gets user comment category z _r Is P (z) _r )：

the total loss function of the user feature extraction network is:

Preferably, the commodity feature extraction network in S40 includes:

a first commodity pooling layer having a window size of 2*1;

a second commodity pooling layer having a window size of 4*1;

a third commodity pooling layer with a window size of 2*1;

a commodity random inactivation layer;

a commodity full-connection layer, the number of the neurons of which is 200,

Wherein sx (m) _x ) Is m _x Input of corresponding node in fourth softmax layerThe value of the sum of the values,

the total loss function of the commodity feature extraction network is:

Preferably, the step S50 specifically includes:

p(i _k |i ₁ ,i ₂ ,...,i _k-1 ,c)＝p(i _k-1 ,s _k ,c)，

The total loss function of the encoder is thus expressed as:

L ₃ ＝-log P(i ₁ ,i ₂ ,...,i _k )-reward(i ₁ ,i ₂ ,...,i _k )，

Preferably, after S50, the method further includes:

s60, acquiring g currently purchased by the user according to the purchase log file ₁ The commodity sequence of each commodity is inquired in a database to obtain the characteristic vector of each commodity in the commodity sequence, and the characteristic vectors are input to an encoder in time sequence to generate a background vector; querying a database to obtain a current userIs input into a decoder after the background vector and the characteristic vector of the current user are spliced, and g-g is generated ₁ And recommending the commodity sequence formed by the commodities to the current user as an optimal commodity sequence.

Actual operation example of the present invention:

The invention is composed of a first commodity sequence personalized recommendation offline training module and a second commodity sequence personalized recommendation online application module.

In the first stage (multi-mode big data preprocessing) of the first module, after the original data of the recommendation system are cleaned and tidied, effective 8000 pieces of user data and 120000 pieces of commodity data are selected, commodity categories are 15, effective comments are about 250000 pieces, effective scores are 3000000 pieces, wherein the scores are divided into 1-5 grades, the larger the numerical value is, the higher the score is, 1-3 commodity images are selected for feature extraction, and about 310000 effective images are selected for each commodity image. In the data preprocessing process, users with the same scores under the same commodity are used as user sequencesHere, x is 50, and commodity scores are according to commodity composition sequences of the same scores of the same userHere, taking 20, this procedure resulted in about 200000 pieces of merchandise sequence data, 80% of which were selected, about 160000 pieces for training. The remaining 40000 strips are used as a test set for evaluation and verification of indexes such as accuracy, recall rate and the like.

The user personal information used by the invention mainly comprises user age, user gender, user occupation, region, average score, purchase amount, average browsing time, collection amount and average purchase amount, and totally selects 9 attributes; the commodity attribute information mainly comprises daily order quantity, sales quantity, collection quantity, online time, use season, average service life, average score, price and price interval (three grades are respectively represented by 3, 2 and 1 numbers according to the average price), sales quantity and 9-dimensional attributes. For processing user comments, the invention adopts the ent2Vec tool to extract comment texts into one-dimensional comment vectors with the length of 200, and the ent2Vec tool trains corpus into Chinese wikipedia data. After all comment texts are processed, clustering is carried out on all comment vectors by using a t-means clustering method, and the t-means clustering method is simple and high in efficiency, and scores of end users are totally divided into 30 categories, namely t=70. And processing comments under the commodity by adopting the same method, and clustering after processing to obtain 30 classes. For commodity picture processing, the invention uses VGG16 tool to extract the characteristics of the pictures of the same commodity, if there are multiple pictures, the final characteristic vector is normalized, then t-means clustering is carried out on commodity pictures, and 50 kinds of clustering results are obtained after clustering and are used for distinguishing commodity categories corresponding to the pictures.

The invention uses the grading information of the commodity by the user, the grading information of the user is divided into 5 grades which are 1-5 grades, and the grading vector generator of the user and the grading vector generator of the commodity are constructed based on the grading data of the commodity by the user.

The specific construction process of the user scoring vector generator is as follows: firstly, for the same commodity, extracting the grading information of all users on the commodity, then grouping the grading information according to 1-5 grades, wherein the users with the same grades are in the same group, thus obtaining the grading sequence based on the users asHere x takes 50, meaning that 50 users are selected from the group. U for the current user in the sequence _r Divide u in predicted sequence _r Other user ids maximize the probability product of the other user ids. Since the total number of users w=8000, the loss function y of the scoring user vector generator can be obtained ₁ The method comprises the following steps:

wherein log () is a logarithmic function with a base of 2, exp () is an exponential function.

After the user scoring vector generator is constructed, the scoring vector of each user can be obtained, and the length is 200.

The specific construction process of the commodity grading vector generator is as follows: firstly, grouping commodities with the same scores of the same user, wherein each group can obtain commodity sequences with the same scores Here j takes 20, representing 20 items from the group. Assume that the current commodity in the sequence is i _g The final goal is to predict the division i in the current sequence _g The probability of occurrence of other commodities than the commodity total number r=120000, so that the loss function y of the commodity grading vector generator can be obtained by using the maximum likelihood estimation method ₂ The method comprises the following steps:

/>

after the commodity grading vector generator is constructed, the grading vector of each commodity can be obtained, and the length is 200.

In the second stage of module one (user and item feature extraction), the user feature extraction network takes the user personal information and the user scoring vector obtained in the first stage as input, and jointly trains two tasks: task one is used for predicting user id, task two is used for predicting user comment category.

The 9 attributes of the personal information of the user are respectively represented by one-hot vectors, the obtained 9 one-hot vectors are spliced into a long vector, and the long vector is reduced to a vector with the length of 50 through a five-layer autoencoder tool. Thus, the input to the user feature extraction network is a vector of length equal to 250.

In a user feature extraction network, firstly, convolving a user feature extraction network by adopting 2 convolution layers with the convolution kernel size of 3*1 and the sliding step length of 1; then, a pooling layer is adopted to operate the vector after convolution, and the window size of the pooling layer is 2*1; the convolution 2 layer adopts 4 convolution kernels with the size of 3*1 and the step length of 2 to carry out convolution operation; after batch standardization, the third layer convolution is carried out, the convolution kernel size is 3*1, the step length is 1, and the number of convolution kernels is 5; after batch normalization, the method reaches a fourth layer of convolution, the step length of a fourth layer of convolution kernel is set to be 2, the size of the convolution kernel is 5*1, and the number of the convolution kernels is 10; then the second layer of pooling layer is reached after Dropout, and the window size of the pooling layer is 2*1; then, dropout operation is adopted and then input into layer 5 convolution, the convolution kernel size is 3 x 1, and the step length of 6 convolution kernels is 2; then passing through a third pooling layer, wherein the window size is 4*1; finally, the data are input into a full connection layer with the number of the neurons being 200, and the values of the neurons of 200 form the characteristic vector of the user.

In order to realize task one, the invention accesses the user characteristic vector to a softmax (exponential maximum function) layer after the full connection layer, wherein the softmax layer has 8000 nodes, which represent 8000 users in the system, and the user set formed by the nodes is U. Meanwhile, in order to realize the task II, after the full connection layer, the user feature vector is accessed into another softmax layer, wherein the softmax layer has 30 nodes, 30 user comment categories are represented, and the formed user comment category set is T.

Both tasks employ a softmax multi-class loss function:

1) Let the current user be u _r User u can be calculated by using softmax _r Is P (u) _r )：

Wherein sx (u) _r ) Is u _r Input values for corresponding nodes in the softmax layer.

2) Let the comment category of the current user be z _r The user comment category z can be calculated by adopting softmax _r Is P (z) _r )：

Wherein sx (z) _r ) Is z _r Input values for corresponding nodes in the softmax layer.

The total loss function of the user feature extraction network is thus:

wherein eta is L ₁ The adjustment coefficient of (2) is between 0.7 and 0.8, here 0.75.

On the other hand, in the second stage of the first module (user and article feature extraction), the article feature extraction network takes the article attribute information and the article scoring vector obtained in the first stage as inputs, and performs joint training on four tasks: task one is used for predicting commodity id, task two is used for predicting commodity category, task three is used for predicting commodity picture category, task four is used for predicting commodity comment category.

The 9 attributes of the commodity attribute information are respectively represented by one-hot vectors, the 9 obtained one-hot vectors are spliced into a long vector, and the long vector is subjected to dimension reduction through a five-layer autoencoder tool to form a vector with the length of 50. Thus, the input to the commodity feature extraction network is a vector of length equal to 250.

In the article characteristic extraction network, the number of convolution kernels of a first layer of convolution layers is 5, the size of the convolution kernels is 3*1, and the step length is 2; then, carrying out second convolution, wherein the size of a convolution kernel of the second layer is 3*1, the number of convolution kernels is 73, and the sliding step length is 3; after batch standardization, the material enters a first pooling layer, and the window size of the pooling layer is 2*1; the method comprises the steps of entering a third layer of convolution after batch standardization, wherein the convolution kernel size is 5*1, the step length is 3, and the number of the convolution kernels is 10; after passing through the Dropout layer, the material enters a second pooling layer, and the window size is 4*1; then, the method enters a 4 th layer convolution layer, the fourth layer convolution parameter is set to be a convolution kernel with the convolution kernel size of 3*1, wherein 5 step sizes are 2; then passing through a pooling layer with a window size of 2*1; then entering a fifth layer of convolution layer through Dropout, wherein the convolution kernel is 3*1 in size, the step length is 3, and the number of the convolution kernels is 6; finally, the data are input into a full-connection layer with the number of the neurons being 200, and the values of the neurons of the 200 form the characteristic vector of the commodity.

In order to realize task one, the invention inserts commodity feature vectors into a first softmax (exponential maximum function) layer after the full connection layer, wherein the softmax layer has 120000 nodes which represent 120000 commodities in the system, and the user set formed by the softmax layer is V. In order to achieve task two, after the fully connected layer, the commodity feature vector is accessed to a second softmax layer, the softmax layer has 15 nodes, 15 commodity categories are represented, and the commodity category set formed is A. In order to realize task three, after the full connection layer, the commodity feature vector is accessed to a third softmax layer, wherein the softmax layer has 50 nodes, which represent 50 commodity picture categories, and the constituted commodity picture category set is I. In order to realize task four, after the full connection layer, the commodity feature vector is accessed to a fourth softmax layer, wherein the softmax layer has 30 nodes, which represent 30 commodity comment categories, and the formed commodity comment category set is M.

All four tasks use softmax multi-class loss functions:

1) Let the current commodity be i _x The commodity i can be calculated by adopting softmax _x Is P (i) _x )：

Wherein sx (i) _x ) Is i _x The input values of the corresponding nodes in the first softmax layer.

2) Let the current commodity category be a _x Commodity category a can be calculated by using softmax _x Is P (a) _x )：

Wherein sx (a) _x ) Is a as _x The input values of the corresponding nodes in the second softmax layer.

3) Let the current commodity picture category be v _x The commodity picture category v can be calculated by adopting softmax _x Is P (v) _x )：

Wherein sx (v) _x ) V is _x The input values of the corresponding nodes in the third softmax layer.

4) Let the current commodity comment category be m _x Commodity comment category m can be calculated by adopting softmax _x Is P (m) _x )：

Wherein sx (m) _x ) Is m _x The input values of the corresponding nodes in the fourth softmax layer.

The overall loss function of the commodity feature extraction network is thus:

wherein lambda is ₁ ～λ ₄ Is L ₂ And lambda is the adjustment coefficient of (1) ₁ +λ ₂ +λ ₃ +λ ₄ ＝1，λ ₁ Takes a value of 0.1-0.3, here 0.15 lambda ₂ Takes a value of 0.2-0.5, here 0.3 lambda ₃ Takes a value of 0.2-0.4, here 0.3 lambda ₄ The value is between 0.2 and 0.3, here 0.25.

In the third stage of module one (merchandise sequence recommendation training), the user history merchandise sequence for training includes g=40 merchandise, where there is g input into the encoder ₁ =20 products with g-g for decoder output prediction ₁ =20 goods. The number of neural networks per GRU unit in the encoder and decoder is set to 1000 and the length of the background vector generated by the encoder is 200.

For the decoder, assume that its output 20 commodity sequences are: < i ₁ ,i ₂ ,...,i ₂₀ The output of each commodity depends on the output of the previous commodity and also on the background vector c. Let s _k (1 is less than or equal to k is less than or equal to 20) is commodity i _k Corresponding decoder hidden variable, then s _k ＝GRU(i _k-1 ,c,s _k-1 ) Thereby commodity i _k Output probability p (i) _k ) Can be expressed as:

p(i _k |i ₁ ,i ₂ ,...,i _k-1 ,c)＝p(i _k-1 ,s _k ,c)。

thus, the joint probability of the maximized output sequence of the kth node of the decoder can be obtained as:

thus, the loss function of the decoder output sequence is: y is ₃ ＝-log P(i ₁ ,i ₂ ,...,i _k )。

In order to improve the accuracy of recommendation, the invention adopts a Beam Search strategy in the training process of a decoder, wherein the beam_size is set to be 2, and the specific process is as follows: firstly, when a first commodity is generated, selecting 2 commodities a and b with the highest probability from candidate sequences; then when generating the second commodity, combining the current two commodity sequences < a >, with the 2 commodities with the largest next prediction probability, such as c, d, to obtain four commodity sequences < ac >, < ad >, < bc >, < bd >, and then selecting the sequence with the largest 2 reward (return) values; the above process is repeated until 20 recommended products are finally produced. The Reward value is calculated from the indexes of the three main streams, namely, an NDCG (Normalized Discounted Cumulative Gain: normalized loss cumulative gain) index, a Recall index and a MAP (MeanAverage Precision: average precision average) index.

reward(i ₁ ,i ₂ ,i _k )＝0.4×NDCG(i ₁ ,i ₂ ,i _k )+0.4×Recall(i ₁ ,i ₂ ,i _k )+0.2×MAP(i ₁ ,i ₂ ,i _k )。

The total loss function of the encoder can thus be expressed as:

L ₃ ＝-log P(i ₁ ,i ₂ ,...,i _k )-reward(i ₁ ,i ₂ ,...,i _k )。

after the commodity sequence recommendation training phase is completed, the feature vector of each User is stored in a database table user_F, wherein the database table comprises two columns, the first column is a User id, and the second column is a User feature vector. Similarly, the feature vector of each commodity is stored in a database table item_f, which is also in two columns, the first column being the commodity id and the second column being the commodity feature vector.

In the second module (online application of personalized recommendation of commodity sequences), for the current user, the method firstly acquires 20 commodities recently purchased by the user from a purchase log file of a system, forms a commodity sequence, then identifies commodity ids of the 20 commodities, and sequentially acquires feature vectors of the 20 commodities from an item_F database table based on the commodity ids. The feature vectors of the 20 commodities are sequentially input into a trained encoder to generate background vectors, and meanwhile, the feature vectors of the current User are obtained from a user_F database table based on the current User id. On the basis, the background vector and the user characteristic vector are spliced, then the spliced background vector and the user characteristic vector are input into a trained decoder, a commodity sequence formed by 20 commodities is generated, and the sequence is recommended to a current user as an optimal commodity sequence.

The foregoing description is only of the preferred embodiments of the present invention and is not intended to limit the scope of the invention, and all equivalent structural changes made by the description of the present invention and the accompanying drawings or direct/indirect application in other related technical fields are included in the scope of the invention.

Claims

1. The personalized commodity sequence recommending method for the dynamic preference of the user is characterized by comprising the following steps of:

s50, inputting the characteristic vectors of the user characteristic vectors and the characteristic vectors of the historical commodity sequences into a commodity sequence recommendation network, training by a training encoder-decoder in the commodity sequence recommendation network to generate the user characteristic vectors corresponding to each user ID, taking the commodity characteristic vectors corresponding to each commodity ID as personalized recommended commodity sequences, and learning by combining a search strategy to obtain the recommendation of the optimal commodity sequences;

the scoring information of the users used by the user scoring vector generator is divided into 5 groups, and the construction method of the user scoring vector generator comprises the following steps:

grouping into 5 groups according to the user scoring information under the same commodity, wherein the users with the same score are in the same group to obtain a user composition sequence with the same score of the same commodityX is 50, which means that 50 users are selected for each group, and the sequence is selectedAs training sample, use user u _r As input, divide user u in the output prediction sequence _r Other users 'scoring vectors are excluded so that the probability product of other user IDs is maximized, and if the total number of users w=8000, the user scoring vector generator's penalty function y ₁ The method comprises the following steps:

obtaining the length of a scoring vector of each user to be 200;

wherein u is _n，j For the nth user in the j-th group of user sequences, u _j A j-th user in the recommendation system;

n is the position number of the current user of the user sequence, j is the position number of the current user in the recommendation system;

obtaining the length of a scoring vector of each commodity to be 200;

wherein i is _n',j For the nth commodity in the commodity sequence, i _j 'is the j' th merchant in the recommendation systemA product;

n 'is the position number of the current commodity of the commodity sequence, and j' is the position number of the current commodity in the recommendation system.

2. The personalized recommendation method for commodity sequences facing user dynamic preferences according to claim 1, wherein said S30 comprises:

3. The personalized recommendation method for a commodity sequence for a dynamic user preference according to claim 2, wherein said user feature extraction network in S40 comprises:

the first user pooling layer window size is 2*1;

The fourth user convolution layer adopts 10 convolution kernels with the size of 5*1, and the step length of the convolution kernels is 2; the pooling layer window size of the second user pooling layer is 2*1;

the window size of the third user pooling layer is 4*1;

the number of neurons in the user full connectivity layer is 200,

the user randomly deactivates the Dropout layer,

4. The personalized commodity sequence recommending method according to claim 3, wherein the user feature vector is connected to a user index maximum function softmax layer after being output by a full connection layer, the softmax layer has 8000 nodes for representing 8000 users in the system, the user set is U, the user feature vector is connected to a comment index maximum function softmax layer after being output by the full connection layer, the softmax layer has 30 nodes for representing 30 user comment categories, the user comment category set is T, the user index maximum function softmax layer and the comment index maximum function softmax layer both adopt softmax multi-classification loss functions,

Let the current user be u _r Obtaining a user through a user index maximum function softmax layer

u _r Is P (u) _r )：

the total loss function of the user feature extraction network is:

wherein eta is L ₁ The value of the regulating coefficient is between 0.7 and 0.8;

u _r ' z for each user in the user set U _r ' for each comment category in the set of user comment categories T.

5. The personalized recommendation method for commodity sequences facing user dynamic preferences according to claim 1, wherein said user personal information comprises at least the following 9 attributes: the method comprises the steps of expressing each attribute by a single hot vector, then splicing all the attributes into a long vector, and reducing the dimension of the long vector into a vector with the length of 50 through a five-layer self-encoder tool; the commodity attribute information at least comprises the following 9 attributes: daily order volume, sales volume, collection volume, online time, use season, average life, average score, price interval.

6. The personalized recommendation method for commodity sequences facing to user dynamic preferences according to claim 2, wherein said commodity feature extraction network in S40 comprises:

a first commodity pooling layer having a window size of 2*1;

a second commodity pooling layer having a window size of 4*1;

a third commodity pooling layer with a window size of 2*1;

a commodity random inactivation layer;

a commodity full-connection layer, the number of the neurons of which is 200,

7. The personalized recommendation method for commodity sequences facing to user dynamic preference according to claim 6, wherein the commodity feature vector is output from the commodity full-connection layer and then is connected to a commodity first index maximum function softmax layer, the softmax layer has 120000 nodes, 120000 commodities in the system are represented, the formed user set is V, the commodity feature vector is output from the commodity full-connection layer and then is connected to a commodity second index maximum function softmax layer, the softmax layer has 15 nodes, 15 commodity categories are represented, and the formed commodity category set is A; the commodity feature vector is output from the commodity full-connection layer and then is connected to a commodity third index maximum function softmax layer, the softmax layer is provided with 50 nodes, 50 commodity picture categories are represented, and a formed commodity picture category set is I; the commodity feature vector is output from the commodity full-connection layer and then is connected into a commodity fourth index maximum function softmax layer, the softmax layer is provided with 30 nodes which represent 30 commodity comment categories, the formed commodity comment category set is M,

the commodity first index maximum function softmax layer, the commodity second index maximum function softmax layer, the commodity third index maximum function softmax layer and the commodity fourth index maximum function softmax layer all adopt softmax multi-class loss functions,

the total loss function of the commodity feature extraction network is:

wherein lambda is ₁ ～λ ₄ Is L ₂ And lambda is the adjustment coefficient of (1) ₁ +λ ₂ +λ ₃ +λ ₄ ＝1，λ ₁ The value is between 0.1 and 0.3, lambda ₂ The value is between 0.2 and 0.5, lambda ₃ The value is between 0.2 and 0.4, lambda ₄ The value is between 0.2 and 0.3;

v is a set formed by all commodities in the recommendation system, A is a set formed by all commodity categories in the recommendation system, I is a set formed by all commodity picture categories in the recommendation system, and M is a set formed by all commodity comment categories in the recommendation system;

i _x ' for each commodity in the commodity set V, a _x ' v for each commodity category in commodity category set A _x ' for each commodity picture category in commodity picture category set I, m _x ' is each commodity comment category in the commodity comment category set M.

8. The personalized recommendation method for commodity sequences facing to user dynamic preference according to claim 1, wherein said S50 specifically comprises:

acquiring g commodities currently purchased by a user, and firstly using the previous g in the g commodities ₁ The commodity sequence composed of the individual commodities is used as the input of the encoder and passes g ₁ The gating cyclic GRU unit obtains the background vector c of the encoder, then the background vector c is spliced with the user characteristic vector and is used as the input of a decoder, and the decoder is composed of g-g ₁ The gate control circulation GRU units are sequentially linked to form output nodes, and the output nodes respectively correspond to the rear g-g in g commodities ₁ The number of the neural networks of each gating cycle GRU unit in the encoder and the decoder is set to be 1000, the length of the background vector c generated by the encoder is 200, and c is the background vector of the encoder;

for the decoder, assume that the 20 product sequences output are , let s _k (1 is less than or equal to k is less than or equal to 20) is commodity i _k Corresponding decoder hidden variables s _k ＝GRU(i _k-1 ,c,s _k-1 ) Then the output probability p (i) _k ) The representation is:

p(i _k |i ₁ ,i ₂ ,...,i _k-1 ,c)＝p(i _k-1 ,s _k ,c)，

the loss function of the decoder output sequence is: y is ₃ ＝-logΡ(i ₁ ,i ₂ ,...,i _k )，

the total loss function of the encoder is thus expressed as:

L ₃ ＝-logΡ(i ₁ ,i ₂ ,...,i _k )-reward(i ₁ ,i ₂ ,...,i _k )，

9. The personalized recommendation method for a sequence of items for dynamic preference of a user according to claim 8, further comprising, after S50:

s60, acquiring g currently purchased by the user according to the purchase log file ₁ The commodity sequence of each commodity is inquired in a database to obtain the characteristic vector of each commodity in the commodity sequence, and the characteristic vectors are input to an encoder in time sequence to generate a background vector; inquiring a database to obtain a characteristic vector of a current user, splicing a background vector and the characteristic vector of the current user, inputting the spliced background vector and the characteristic vector of the current user into a decoder, and generating g-g ₁ And recommending the commodity sequence formed by the commodities to the current user as an optimal commodity sequence.