CN114861050A

CN114861050A - Feature fusion recommendation method and system based on neural network

Info

Publication number: CN114861050A
Application number: CN202210454110.XA
Authority: CN
Inventors: 边根庆; 李婷
Original assignee: Xian University of Architecture and Technology
Current assignee: Xian University of Architecture and Technology
Priority date: 2022-04-27
Filing date: 2022-04-27
Publication date: 2022-08-05

Abstract

The invention discloses a neural network-based feature fusion recommendation method and system, belongs to the field of intelligent commodity recommendation of new retail enterprises, is different from a traditional algorithm which only considers the behavior of a user on an article for recommendation, fuses the features of the user and the article, and provides a neural network-based feature fusion recommendation model. The method comprises the steps of selecting user and article features, converting the user and article features into feature vectors through coding, then performing deeper feature extraction through a neural network embedding layer and a full connection layer to generate feature representations of the user and the articles, and finally multiplying two matrixes to predict the scores of the user on the articles.

Description

Feature fusion recommendation method and system based on neural network

Technical Field

The invention belongs to the field of intelligent commodity recommendation of new retail enterprises, and particularly relates to a neural network-based feature fusion recommendation method and system.

Background

Nowadays, more and more people use network technology to acquire needed knowledge and information, which leads to the proliferation of information quantity and the problem of 'information overload'. For users, the problem of information overload on the internet often makes users unable to find the content desired by the users, and the existence of a search engine can relieve the information overload, but the users are required to have clear search requirements; when the user has no specific requirements, the recommendation system provides a new selection path for the user, so that personalized service is possible. The recommendation system is not only an effective information processing tool, but also an intelligent search platform based on user interests, and can help users to find out items of potential interest, so that the search time of the users is saved. For enterprises, the recommendation system can actively recommend potentially interesting commodity information to users, and improves the utilization rate of the information and the satisfaction degree of the users, so that the stickiness of the users is increased, and the purpose of increasing the commercial benefits of companies is achieved. In a word, the recommendation system brings value to enterprises, brings convenience to work and life of users, and influences development of the whole society to a certain extent. Therefore, in the future intelligent society, the recommendation system is an indispensable technology, and has great application value in academic research, business fields, practical applications and the like.

Traditional recommendation algorithms are mainly classified into three major categories: content based recommendation algorithms (CBs), Collaborative Filtering recommendation algorithms (CFs), and Hybrid Recommendations (Hybrid Recommendations). Collaborative Filtering recommendations include Item-based Collaborative Filtering (Item CF) and User-based Collaborative Filtering (User CF). Both Item CF and CB recommend items based on the similarity of the items, except that the calculation method of the similarity is different, the former is based on the historical preference of the user, and the latter is based on the characteristics of the items. Among them, the CB recommendation algorithm directly makes use of feature information of items to make recommendations, and there is a problem of feature extraction. On one hand, the feature description of the project is generally unstructured text data, which undoubtedly increases the difficulty of feature extraction; on the other hand, the attribute information of the project is less, and the difficulty of feature selection is increased. Although the problem of feature extraction can be solved by recommending the CF according to the behaviors of the user on the items, the algorithm requires a user-item matrix to have higher saturation degree, so that a more accurate recommendation result can be generated.

Through the analysis of the background, the traditional recommendation algorithm faces the problems of feature extraction, low recommendation accuracy and the like. Meanwhile, the requirements of people on personalized services are higher and higher, and the behavior characteristics of users are often difficult to obtain. Therefore, how to extract useful information from a large amount of network data and make a correct decision is one of the problems that the current recommendation technology needs to consider and explore intensively.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention aims to provide a feature fusion recommendation method and system based on a neural network, so as to solve the problems of difficulty in feature extraction, low recommendation precision and single utilization mode of feature information in the prior art.

In order to achieve the purpose, the invention adopts the following technical scheme to realize the purpose:

the invention discloses a feature fusion recommendation method based on a neural network, which comprises the following steps:

acquiring commodity characteristics and user characteristic information;

vectorizing the commodity characteristics and the user characteristic information to form a commodity characteristic matrix and a user characteristic matrix;

inputting the commodity characteristic matrix and the user characteristic matrix into a neural network characteristic fusion recommendation model to generate a characteristic fusion recommendation model of a three-layer neural network;

initializing the parameters of the feature fusion recommendation model of the three-layer neural network, and generating a recommendation list with prediction scores from high to low.

Preferably, the commodity characteristics include a code, a class, a name, a secondary classification, an applicable age, an applicable part, an applicable skin, and a user unit price of the commodity; the user characteristics comprise sex and skin, wherein the sex is male and female 2 types, and the skin is dry, oily, sensitive muscle and mixed 4 types.

Preferably, the commodity feature vectorization includes:

1) selecting commodity characteristics to form a commodity characteristic label library;

2) vectorizing the text information of the commodity feature tag library through unique hot coding;

3) and carrying out normalization processing and embedding processing on the unit price of the user to obtain deep characteristic matrix representation of the commodity.

Preferably, the selection of commodity feature data is performed by a variance analysis method, a mutual information method, a classification and regression tree embedding method, namely CART, and a recursive feature elimination cross validation method, namely RFECV packaging method.

Preferably, the selection of the commodity features is performed by an analysis of variance method, a mutual information method, a classification and regression tree embedding based method or a recursive feature elimination cross validation based packaging method.

Preferably, in the step one, the vectorization of the user features includes:

user feature vectorization, comprising:

a) selecting user characteristics to form a user characteristic label library;

b) vectorizing the text information of the user feature tag library through one-hot coding;

c) and carrying out normalization processing and embedding processing on the feature vectors to obtain the deep feature matrix representation of the user.

Preferably, the feature fusion recommendation model of the three layers of neural networks comprises a first layer of neural network which is an embedded layer for converting feature vectors into feature representations with low latitude density; the second layer of neural network is a full connection layer which splices all the feature representations together to obtain the feature vectors of the users and the members; the third layer of neural network is a full-connection layer which takes the commodity and the user characteristics obtained by the first two layers as input, obtains an output value by multiplying the two input values in a matrix mode and returns the output value to the real score.

Preferably, the constructing of the feature fusion recommendation model of the three-layer neural network comprises the following steps:

the method comprises the following steps: vectorizing commodity characteristics and vectorizing user characteristics to form a commodity characteristic matrix and a user characteristic matrix;

step two: the commodity characteristic matrix and the user characteristic matrix are used as input of a first layer of neural network of the characteristic fusion recommendation model, and low latitude and density embedded layer characteristic representation is formed through iterative training of the first layer of neural network;

step three: inputting the embedded layer feature direction into a second full-connection layer, and splicing to form a full-connection layer feature vector;

step four: and inputting the feature vector of the full-connection layer into a third full-connection layer, and forming a feature fusion recommendation model passing through a three-layer neural network through iteration and regression.

Preferably, the evaluation of the feature fusion recommendation model of the three-layer neural network comprises:

s1: initializing model parameters, and generating a recommendation list with prediction scores from high to low;

s2: and evaluating the precision of the feature fusion recommendation model through the difference value of the prediction score and the real score.

Preferably, the accuracy of the feature fusion recommendation model is evaluated by MSE, RMSE and MAE,

wherein u represents a user, i represents an item, r _ui Is the actual rating of item i by user u,

is to recommendThe prediction score given by the algorithm.

A neural network-based feature fusion recommendation system, comprising:

the characteristic information acquisition module is used for acquiring commodity characteristics and user characteristic information;

the vectorization module is used for vectorizing the commodity characteristics and the user characteristic information to form a commodity characteristic matrix and a user characteristic matrix;

the model generation module is used for inputting the commodity feature matrix and the user feature matrix into the neural network feature fusion recommendation model to generate a feature fusion recommendation model of the three-layer neural network;

and the parameter initialization module is used for initializing the parameters of the feature fusion recommendation model of the three-layer neural network and generating a recommendation list with prediction scores from high to low.

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

the invention discloses a characteristic fusion recommendation method based on a neural network, which utilizes auxiliary information to extract distributed characteristic representation of a user and an article, and combines the advantages of the neural network in the aspects of characteristic extraction and large-scale data analysis to solve the problems that the user learns an interest model of the user according to the historical scoring of the user on the article, and the user can be predicted to score the article when seeing the article which is not scored by the user in the future. The prediction accuracy of the score prediction is typically calculated by mean square error, root mean square error and mean absolute error, which are suitable for evaluating a data set that possesses a score. For the problem of difficult feature extraction, the invention uses an analysis of variance method, a mutual information method, a classification-based and regression tree embedding method and the like to extract features, and inputs the feature data into a neural network embedding layer after the feature data is subjected to unique hot coding, and the embedded data has lower dimensionality and can map discrete sequences into continuous vectors. The advantage is that the relation between nearest neighbors and mining variables is found in the embedding space. The accuracy of the recommendation result is effectively improved through MSE, RMSE and MAE.

The feature fusion recommendation model is a three-layer neural network feature fusion recommendation model, and a first embedding layer can convert feature vectors into low-latitude dense feature representation; the second layer is a full connection layer, and all the characteristic vectors are spliced together by the layer to obtain the characteristic vectors of the user and the member. The third layer is also a full-connection layer, commodity and user characteristics obtained from the first two layers are used as input, an output value is obtained by multiplying the two input values in a matrix mode, and the output value is regressed to a real score to optimize loss. The model can gradually reduce the error between the predicted value and the true value after one iteration, and a more accurate prediction result is obtained.

The invention discloses a method for constructing a feature fusion recommendation model based on a neural network, which is characterized in that commodity and user features extracted through feature engineering are input into a neural network embedding layer after being subjected to unique hot coding, the embedded data has low dimensionality, and a discrete sequence can be mapped into a continuous vector. If the characteristics of the commodity are vectorized only by means of the one-hot code, the characteristic vectors of the commodity are extremely sparse. However, the neural network cannot process sparse vectors, and therefore the sparse vectors must be further processed to obtain a lower-dimensional and dense feature vector representation. Embedding, namely Embedding, is a method for converting discrete variables into dense vectors, and the commodity and user feature vectors are further trained through a neural network Embedding layer, so that the dimensionality of the feature vectors can be reduced, and the method has the advantage that the relationship between nearest neighbors and mined variables is found in an Embedding space. The fully-connected layer then serves to map the learned feature representation to the label space of the sample. In other words, features are integrated together, i.e., highly refined, and are conveniently handed over to a final classifier or regression function.

Further, because the machine cannot directly recognize the received identifier, such as a word, phrase, character, etc., it is necessary to digitize the identifier. The invention adopts single hot coding to carry out numerical processing of commodity characteristics. By utilizing the one-hot coding, on one hand, the low-dimensional space of the discrete features can be expanded to a limited n-dimensional space, so that the inner product, the distance and the like of the vector can be calculated, and the output of the vector can be used for machine learning; on the other hand, the features subjected to the unique hot coding can be regarded as continuous features in each dimension, and the features in each dimension can be normalized. Secondly, the method carries out normalization processing on the numerical information such as the price in the characteristic information, and avoids the adverse effect on subsequent calculation caused by the existence of particularly large or small values in the sample.

Further, for the problem of difficult feature extraction, the invention uses an analysis of variance method, a mutual information method, a classification and regression tree embedding based method and the like to extract features, thereby relieving the influence on algorithm precision due to improper feature selection.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a diagram of user profile information extraction;

FIG. 3 is a diagram of commodity feature information extraction;

FIG. 4 is a diagram of a feature fusion recommendation model;

FIG. 5 is a training diagram of a RMSE-based feature fusion recommendation model;

FIG. 6 is a feature fusion recommendation model training diagram based on MSE;

FIG. 7 is a feature fusion recommendation model training diagram based on MAE;

FIG. 8 is a graph comparing experimental results;

FIG. 9 is a diagram of a system;

fig. 10 is a system architecture diagram.

Detailed Description

In order to make the technical solutions of the present invention better understood, 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.

It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the invention described herein are capable of operation in sequences other than those illustrated or described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

The invention is described in further detail below with reference to the accompanying drawings:

acquiring commodity characteristics and user characteristic information;

inputting the commodity feature matrix and the user feature matrix into a neural network feature fusion recommendation model to generate a feature fusion recommendation model of a three-layer neural network;

Referring to fig. 1, a flow chart of the present invention includes:

s1, vectorization of commodity features

The commodity feature vectorization includes feature extraction and feature vectorization, and fig. 2 is a flow of commodity feature vectorization, which is described in detail as follows:

1) the invention adopts an analysis of variance method, a mutual information method, a classification and regression tree embedding method, namely CART, and a recursive feature elimination cross validation, namely RFECV packaging method to select commodity feature data, and finally selects eight features of commodity codes, categories, names, secondary classification, applicable ages, applicable positions, applicable skin types and user unit prices to form a commodity feature label library.

2) Because the machine can not directly receive the text identifier, the selected features are digitalized before the feature fusion recommendation model is constructed, and the text information of the commodity is vectorized through the unique hot code. The commodity features needing digital processing include secondary classification, category, suitable age, suitable part and suitable skin. Wherein the secondary classification comprises 15 types, and the categories comprise 37 types; applicable parts are divided into 11 types, which are respectively: head, face, eyes, lips, hands, feet, torso, hair, body hair, whole body skin, nails, and specific parts; applicable age is divided into 5 intervals, which are respectively: under 18 years old, 18 to 25 years old, 25 to 30 years old, 3 to 40 years old, and over 40 years old; suitable skin types are divided into 5 types, namely dry, neutral, oily, mixed and sensitive.

3) After all the feature information is coded, the user unit price needs to be normalized, so that the adverse effect on subsequent calculation caused by the fact that a particularly large or small value exists in a sample is avoided. The normalized data is limited to a certain range, such as [0, 1 ]. Common normalization methods include linear normalization and standard deviation normalization. Linear normalization, also known as min-max normalization, is a linear transformation of the raw data such that the resulting values map between [0, 1 ]. The transfer function is shown in (1).

Where min (x) is the minimum value of the original data set, and max (x) is the maximum value of the original data set, such normalization is commonly applied to scenes with more concentrated values, and if the difference between min and max is larger, the normalization result is easily unstable, so in practical applications, the min and max are generally replaced by empirical constants. Normalization of the standard deviation, also known as Z-score normalization, involves the correlation of the raw measurements with the standard deviation and the calculation of the sum of the squares of the differences to obtain the absolute value. The processed data are in accordance with the standard normal distribution, namely the mean value is 0 and the standard deviation is 1. The transfer function is as follows:

where μ is the mean of the raw data and σ is the standard deviation of the raw data. After multiple times of verification, the commodity price is processed in a linear normalization mode.

S2, vectorizing user characteristics

The user feature vectorization includes feature extraction and feature vectorization, and fig. 3 is a flow of user feature vectorization, which is described in detail as follows:

1) and selecting four characteristics of the mobile phone number, the gender, the age and the skin of the user to form a user characteristic label library.

2) And vectorizing the text information of the user through one-hot coding. The numerical value is required to be processed to obtain sex and skin, wherein the sex is male and female, and the skin is dry, neutral, oily, mixed and sensitive.

3) After all the feature information is converted into numerical values, the user age needs to be normalized. The invention selects a linear normalization mode to process the age of the user.

S3, embedding layer processing

1) If the codes of the commodities are only subjected to numerical processing by means of the one-hot codes, the characteristic vectors are abnormally sparse. However, the neural network cannot process sparse vectors, and therefore the sparse vectors must be further processed to obtain a lower-dimensional and dense feature vector representation. Embedding (Embedding) is a method for converting discrete variables into dense vectors, and commodity feature vectors and user feature vectors are further trained through a neural network Embedding layer, so that the dimensionality of the feature variables can be reduced.

2) And after the embedded layer processing, obtaining deep feature matrix representation of the commodity and the user.

The input of the model is an Embedding matrix of the user and the commodity, which represents the unique characteristic information of the user and the commodity, and fig. 4 is a neural network structure diagram of the characteristic aggregation recommendation model.

1) Firstly, a user characteristic Embedding matrix and a commodity characteristic Embedding matrix after characteristic extraction are used as input of a characteristic aggregation recommendation model, wherein the user characteristic Embedding matrix comprises a mobile phone number, a gender, an age and a skin of a user, the output vector size of the mobile phone number is (N,32), and the output vector sizes of the gender, the age and the skin are (N, 16); the commodity characteristic Embedding matrix comprises commodity class, name, secondary classification, applicable age, applicable position, applicable skin and user unit price, the commodity ID output vector size is (N,32), and the name, secondary classification, class, applicable age, applicable position, applicable skin and user unit price output vector size are (N, 16).

2) The trained feature vectors via the embedding layer are used as input for the second layer fully-connected layer. For the user feature vector, the size of a mobile phone number input sample is 32, the size of an output sample is 32, the sizes of other feature input samples are 16, and the size of the output sample is 32; for commodity feature input, the size of an ID input sample is 32, the size of an output sample is 32, the sizes of other feature input samples are 16, and the size of the output sample is 32;

3) and in the third fully-connected layer, the dimension of an input feature space is 128, and the number of the final fully-connected neurons is 200.

S4, initializing model parameters and training

Initializing parameters, setting the number of single iteration epochs of data to be 10, learning rate learning _ rate to be 0.0001, embedding layer dimension embed _ dim to be 32, second layer full-connection layer dimension embed _ dim to be 32, third layer full-connection layer dimension to be 128, the number of final full-connection neurons to be 200, sample size batch _ size of each training to be 256, adopting an Adalgorithm as an optimizer am of the model, and training by taking an MAE index, an MSE index and an RMSE index as loss functions of the model.

S5, evaluating model

Besides calculating the accuracy of the recommendation result, the quality of the recommendation model can be evaluated by predicting the product rating of the user and the difference between the predicted rating and the real rating. If the historical scores of the user on the products can be acquired, the scoring result of the user on unknown products can be predicted, and whether the products are recommended to the user or not is determined according to the scores. The accuracy of the prediction model is generally evaluated by MSE, RMSE and MAE, which are suitable for evaluating datasets that possess a score. The calculation formulas of MSE, RMSE and MAE are respectively (3), (4) and (5),

where u represents a user, i represents an item, r _ui Is the actual rating of item i by user u, and

is the prediction score given by the recommendation algorithm.

Firstly, the data set of the 'certain retail enterprise' used by the invention is processed, and the number of times that the user purchases the commodity is taken as the grade of the commodity of the user, thereby meeting the application requirements of the indexes. Secondly, the feature fusion recommendation model based on the neural network, which is provided by the invention, predicts the score of the unknown commodity according to the score of the user on the known commodity, so that MSE, RMSE and MAE are used as the evaluation indexes of the experiment.

8. System applications

Through the research, a feature fusion recommendation model based on the neural network, which is superior to the traditional algorithm, is obtained. And then constructing an intelligent commodity recommendation system of the new retail enterprise, and applying the model to the intelligent commodity recommendation system. The system is mainly divided into a background management module, a data processing module, a commodity display module, a statistic recommendation module and a personalized recommendation module. The personalized recommendation module is a core module and has the main tasks of helping a user to find information which is valuable to the user, improving the satisfaction degree of the user and increasing the profit of an enterprise, so that the win-win situation of a consumer and a producer is realized.

The method is different from the traditional algorithm which only considers the behavior of the user on the article to recommend, integrates the characteristics of the user and the article, and provides a neural network-based characteristic integration recommendation model. The method comprises the steps of selecting user and article features, converting the user and article features into feature vectors through coding, then performing deeper feature extraction through a neural network embedding layer and a full connection layer to generate feature representations of the user and the articles, and finally multiplying two matrixes to predict the scores of the user on the articles.

Verification test

In order to illustrate the performance of the invention, the method of the invention and the traditional collaborative filtering-based method are utilized to carry out a comparison experiment, the experiment is mainly to carry out algorithm test on a selected data set, the data set is from a certain retail enterprise, and the data is composed of sales records, user data and commodity data of the enterprise in one year. The sales record is a record of the commodity purchased by the user in the past year, and the fields include a user number, a commodity number, a quantity, an amount, a transaction date and the like. The user data set records basic attributes of the user, including gender, year of birth, date of birth, user rating, skin type, and the like. The commodity data includes commodity name, classification, series, class, applicable age, retail price, and the like.

The key point of the feature fusion model is feature extraction of commodities and members, in order to ensure the accuracy of the experiment, the commodity data and the member data need to be cleaned, and data with null selected features in the member data and the commodity data are filtered. For member data, filtering data with null member gender, age and skin; filling the skin with neutral skin only with empty skin; for gender empty, fill is female; for the age being empty, filled to the average of all ages, there are 185696 pieces of final membership data. For the commodity data, more than four data which are empty in the commodity characteristics such as name, secondary classification, category, applicable age, applicable part, applicable skin and member unit price are filtered, and 882 pieces of final commodity data are obtained.

This experiment sets up data single iteration number epochs to be 10, learning rate learning _ rate is 0.0001, embedding layer dimension embed _ dim is 32, second floor full-link layer dimension embed _ dim is 32, the third layer full-link layer, the dimension of input feature space is 128, the last full-link neuron number is 200, it is 256 to train sample size batch _ size each time, adopt Adam algorithm as the optimizer of model, adopt random the chaos to training set data, train as the loss function of model with MAE index, MSE index and RMSE index, the experimental analysis is specifically as follows:

and (3) convergence analysis: in the experiment, different loss functions are selected to carry out multiple iterative optimization, and the obtained experimental results are shown in fig. 5, 6 and 7. Fig. 5 is a training diagram of a neural network with Loss ═ RMSE, fig. 6 is a training diagram of a neural network with Loss ═ MSE, and fig. 7 is a training diagram of a neural network with Loss ═ MAE. The loss function of the model already goes to zero when the iteration is about ten times, so that the conclusion that the recommended model is high in convergence speed and small in error can be obtained. In order to prevent the model from being over-fitted, the model is verified by applying a test set, fig. 8 is a model fitting comparison graph, a red line is a score true value, a black line is a predicted value, and the peak value basic fitting can be seen.

And (3) analyzing the accuracy: besides measuring the recommendation performance of an algorithm by MSE, RMSE and MAE, the method also needs to compare the neural network feature fusion-based recommendation constructed by the method with other recommendation algorithms to verify the accuracy of the method. Table 1 is an experimental comparison of the recommendation model proposed by the present invention with the recommendation algorithm of project-Based collaborative filtering recommendation Slope One, proposed Factorization Machines (FM), and Deep Neural Network-Based Factorization Machines (Deep FM) under a "certain retail enterprise" dataset:

TABLE 1 Experimental results chart

As can be seen from table 1, compared with the conventional project-based collaborative filtering algorithm, the recommendation algorithm based on the factorization machine model, and the deep neural network-based factorization machine recommendation model, the neural network-based feature fusion recommendation model has significant advantages in MSE, RMSE, and MAE.

The recommendation model based on neural network feature fusion disclosed by the invention fuses commodity features and user features, and model training is carried out based on a neural network. The method solves the problems that the traditional algorithm directly predicts the scores by a user-item matrix, does not fully excavate some invisible characteristics of users or commodities, and has low recommendation accuracy. Secondly, neural networks have great advantages over traditional methods in terms of feature extraction and feature representation. And finally, performing simulation experiment on the neural network feature fusion recommendation model by adopting annual sales data of a certain retail enterprise in the experiment. The method is analyzed through experiments, the accuracy of the method is effectively proved, and the method is compared with the traditional method based on the collaborative filtering in the aspect of accuracy. The result shows that the model has more remarkable advantages on MSE, RMSE and MAE, and the recommendation accuracy is higher.

Development and application

Based on the neural network feature fusion-based recommendation model provided by the invention, a retail enterprise commodity recommendation system is designed and realized. The system analyzes the user behavior information generated in the E-commerce platform of a certain retail enterprise, finds out the association between users and users, between commodities and between users and commodities, and recommends the commodities which the users may be interested in to the users. The system cleans the received user data, commodity data and transaction data of a certain retail enterprise, and imports the user data, the commodity data and the transaction data into the system to obtain data required by an experiment; and performing feature processing on the data, completing recommendation by using the feature fusion recommendation model based on the neural network, and showing the recommendation to a user at the front end.

1. Example analysis

As shown in the usage diagram of the system in fig. 9, the user roles of the system are divided into three roles, namely, a system administrator, an e-commerce platform administrator and a common user.

For the system administrator, user information management, commodity information management, transaction data management, rating data management, and recommendation result management are used.

For e-commerce platform administrators, the recommended model maintenance, data cleaning and importing are used.

For common users, personal information management, commodity browsing and statistical recommendation list checking are used, including sales lists, commodities with high browsing amount and the like. Most importantly, personalized recommendation results are checked, and the use case aims at historical purchasing behaviors of users and realizes commodities which are more consistent with characteristics and interests of the users for each user.

2. Architectural design

The system is based on a B/S framework, and Django is selected as a background development framework. Django is an open-source Web development framework that can help developers develop Web sites faster and more easily. Django adopts a three-layer architecture design, namely a template layer, a view layer and a model layer. The template layer is in direct contact with a user for displaying data and accepting user input data. The view layer is positioned in the middle layer and plays a role in data processing and transmission. The model layer has the main function of creating and maintaining a data model. The system architecture is as in fig. 10.

3 System Module design

The main functional modules of the system are a system management module, a data processing module, a statistical recommendation module and a personalized recommendation module.

1) Background management module

The background management module is mainly used for managing member data, commodity data and transaction data. The member data comprises a login account password and personal basic information when the member registers; the commodity data comprises basic information of the commodity, and an administrator can update the commodity data or check detailed information of the commodity according to screening conditions; the transaction data comprises purchase records of commodities by members, and the administrator can view the transaction records according to purchase dates, member codes, commodity codes and the like.

2) Data processing module

The data processing module comprises the cleaning and importing of member data, commodity data and transaction data.

3) Commodity display module

The member can browse all commodities, can screen the commodities according to applicable skin types and applicable ages, and sorts the commodities according to the commodity clicks. The commodity detail page can check the basic information of the commodity and the commodity recommendation of the same series with the commodity.

4) Statistical recommendation module

And the statistical recommendation is mainly based on historical scoring records, and the average scoring statistics of historical popular commodities and commodities are calculated.

5) Personalized recommendation module

The personalized recommendation module is a core part of the system and is responsible for the part of the platform for recommending the commodity for the member, and the part comprises data information acquisition, member and commodity feature extraction, model construction, recommendation candidate set screening and final recommendation results. And the data processed by the data processing module is used as the input of the recommendation module, and then the recommendation model provided by the text is called for training to obtain and store the characteristic vectors of the members and the products. Meanwhile, dictionaries of the members and the products are stored and used for acquiring information of specific members or products, and a relevant algorithm is designed to obtain a recommended product set. The system designs three recommendation schemes, wherein the recommendation of the same type of products, the recommendation of the favorite products of members and the recommendation of the favorite products of the similar type of members are included.

Recommending the same type of products, namely recommending the products similar to the purchased products to the member, wherein the design idea is to calculate the cosine similarity between the products purchased by the current member and other products, then recommending the ten products with the maximum similarity to the member, and adding random selection so as to perform corresponding adjustment according to the actual condition of the member.

Recommending the products which are the favorite products of the members, namely recommending the products with the highest forecast scores, wherein the design idea is to use the product of the member feature matrix and the product feature matrix to express the scores of the forecast members on unknown commodities, and to take the ten commodities with the highest forecast scores to recommend to the members.

Recommending similar type member favorite products, namely recommending products purchased by members similar to the members, firstly extracting member feature vectors of all the members, performing dot product operation on all the product features to obtain scores of the members on the products, performing sequencing operation on the scores, returning the previous product with the highest score of each member, and storing the previous product in a lexicographic order. And then according to the member feature set, calculating cosine similarity with all other members, returning the first ten members similar to the target member after sorting, and acquiring the stored products with the highest scores of the ten members to generate a recommended product set.

The invention discloses a feature fusion recommendation system based on a neural network, which comprises:

As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting the same, and although the present invention is described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications and equivalents may be made to the embodiments of the invention without departing from the spirit and scope of the invention, which is to be covered by the claims.

Claims

1. A feature fusion recommendation method based on a neural network is characterized by comprising the following steps:

acquiring commodity characteristics and user characteristic information;

initializing parameters of a feature fusion recommendation model of the three-layer neural network, and generating a recommendation list with a prediction score from high to low.

2. The neural network-based feature fusion recommendation method according to claim 1, wherein the commodity features comprise commodity codes, commodity types, names, secondary classifications, applicable ages, applicable positions, applicable skin types and user unit prices; the user characteristics comprise sex and skin, wherein the sex is male and female 2 types, and the skin is dry, oily, sensitive muscle and mixed 4 types.

3. The neural network-based feature fusion recommendation method according to claim 1, wherein the commodity feature vectorization comprises:

4. The feature fusion recommendation method based on the neural network as claimed in claim 3, wherein the selection of the commodity features is performed by using an analysis of variance method, a mutual information method, a classification and regression tree embedding method or a recursive feature elimination cross validation-based packaging method.

5. The neural network-based feature fusion recommendation method according to claim 1, wherein the user feature vectorization comprises:

a) selecting user characteristics to form a user characteristic label library;

6. The feature fusion recommendation method based on the neural network as claimed in claim 1, wherein the feature fusion recommendation model of the three layers of neural networks is characterized in that the first layer of neural network is an embedded layer for converting feature vectors into low latitude dense feature representations; the second layer of neural network is a full connection layer which splices all the feature representations together to obtain the feature vectors of the users and the members; the third layer of neural network is a full connection layer which takes the commodity and the user characteristics obtained by the first two layers as input, obtains an output value by multiplying the two input values in a matrix mode and regresses the output value to a real score.

7. The feature fusion recommendation method based on the neural network as claimed in claim 1, wherein the building of the feature fusion recommendation model of the three-layer neural network comprises:

8. The neural network-based feature fusion recommendation method according to claim 1, wherein the evaluation of the feature fusion recommendation model of the three-layer neural network comprises:

9. The neural network-based feature fusion recommendation method according to claim 8, wherein the accuracy of the feature fusion recommendation model is evaluated through MSE, RMSE and MAE,

where u represents a user, i represents an item, r _ui Is the actual rating of item i by user u,

is the prediction score given by the recommendation algorithm.

10. A neural network-based feature fusion recommendation system, comprising: