CN113781139A

CN113781139A - Item recommendation method, item recommendation device, equipment and medium

Info

Publication number: CN113781139A
Application number: CN202011121208.0A
Authority: CN
Inventors: 王颖帅
Original assignee: Beijing Jingdong Century Trading Co Ltd; Beijing Wodong Tianjun Information Technology Co Ltd
Current assignee: Beijing Jingdong Century Trading Co Ltd; Beijing Wodong Tianjun Information Technology Co Ltd
Priority date: 2020-10-19
Filing date: 2020-10-19
Publication date: 2021-12-10

Abstract

The embodiment of the invention provides an article recommendation method, which comprises the following steps: acquiring a feature set of a user; processing the feature set by using a pre-trained deep learning network to obtain a consumption capability prediction score of the user; determining a consumer capability category for the user based on the consumer capability prediction score; and recommending the items for the user based on the item information pool aiming at the consumption capacity category. The embodiment of the invention also provides an article recommending device, computer equipment and a medium.

Description

Item recommendation method, item recommendation device, equipment and medium

Technical Field

The embodiment of the invention relates to the technical field of computers, in particular to an article recommendation method, an article recommendation device, equipment and a medium.

Background

With the development of big data and artificial intelligence, more and more users enjoy online shopping. In order to attract more users, the e-commerce platform increases the user viscosity through various strategies or algorithms, and improves the user experience. A sinking group refers to a group of users who have little shopping behavior, low user level, or are in a low-line city in the e-commerce platform, that is, a group of users with low consumption capability, and may be referred to as a sinking user. At present, sinking groups are income to families or cities below three lines at the middle and low ends, and occupy a large proportion, and the internet consumption of the sinking groups is relatively slow in development, so that the sinking groups have a large promotion space. In order to further mine the value of the sinking group, the method has very important significance for item recommendation of the sinking user.

In one processing mode, a data analyst of the e-commerce platform sets a strategy formula according to experience, and then calculates the subsidence score of the user according to the strategy consensus and the user characteristics. And recommending different commodities for the user according to different sinking scores. However, this approach has the following disadvantages: with the continuous development of services, the user characteristics will change continuously, and the formula needs to be updated continuously, which consumes more labor cost. Moreover, the experience of people is limited, and factors possibly considered by a design formula are not comprehensive, so that the calculated sinking score is not accurate, and further the misjudgment recommendation is not accurate.

Disclosure of Invention

In view of the above, the present invention provides an item recommendation method, an item recommendation apparatus, a device and a medium.

One aspect of the present invention provides an item recommendation method, including: acquiring a feature set of a user; processing the feature set by using a pre-trained deep learning network to obtain a consumption capability prediction score of the user; determining a consumer capability category for the user based on the consumer capability prediction score; and recommending the items for the user based on the item information pool aiming at the consumption capacity category.

According to an embodiment of the present invention, the method further includes: constructing an initial network; acquiring a sample feature set, and determining label scores of the sample feature set; and training the initial network by using the sample feature set and the corresponding label scores to obtain the pre-trained deep learning network.

According to an embodiment of the invention, the initial network comprises: an embedding layer, a feature extraction layer and an output layer. The feature extraction layer includes: the low-order feature extraction sub-network comprises a linear layer and a factorization machine; the medium-order feature extraction sub-network comprises a convolutional neural network and a recombination layer; and a higher order feature extraction sub-network, the higher order feature extraction sub-network comprising a deep neural network.

According to an embodiment of the present invention, the training of the initial network by using the sample feature set and the label score includes: converting the sample feature set into a feature vector by using an embedding layer; extracting a linear feature combination of the feature vectors by using a linear layer, extracting a second-order cross feature combination of the feature vectors by using a factor decomposition machine, and forming a first feature by using the linear feature combination and the second-order cross feature combination; extracting local features of the feature vectors by using a convolutional neural network, and combining the local features by using a recombination layer to obtain second features; extracting a third feature of the feature vector by using a deep neural network; outputting, by the output layer, a consumer prediction score for the sample feature set based on the first feature, the second feature, and the third feature; determining a loss value of the initial network based on the label score and the consumption capacity prediction score of the sample feature set; and determining that the initial network is trained as the pre-trained deep learning network when the loss value reaches convergence.

According to an embodiment of the present invention, the extracting a second-order cross feature combination of feature vectors by using a factorization machine includes: and extracting a second-order cross feature combination by a factorization machine based on a Dicefactor mechanism.

According to an embodiment of the invention, a convolutional neural network comprises: convolutional layers and pooling layers. The above extracting local features of the feature vector by using the convolutional neural network includes: and carrying out convolution operation on the feature vector by the convolution layer to obtain high-order local features, and carrying out dimensionality reduction operation on the high-order local features by the pooling layer to obtain the local features. The combining the local features by the recombination layer includes: the local features are combined in a fully connected manner by the recombination layer.

According to an embodiment of the present invention, the extracting the third feature of the feature vector by using the deep neural network includes: a third feature is extracted by the deep neural network based on a multi-headed self-attention mechanism.

According to an embodiment of the invention, the initial network further comprises an input layer. The training of the initial network further comprises: and inputting the sample feature set into an input layer, and preprocessing the sample feature set by using the input layer. The converting the sample feature set into the feature vector by using the embedding layer comprises: the preprocessed sample feature set is converted into feature vectors using an embedding layer.

According to an embodiment of the invention, the pre-processing comprises at least one of: outlier truncation processing, and normalization processing.

According to an embodiment of the present invention, the method further includes: and setting sample weight for the sample feature set. The determining the loss value of the initial network based on the label score and the consumption capability prediction score of the sample feature set comprises: determining a loss value of the sample feature set based on a loss function of the initial network and a difference between the label score and the consumption capability prediction score of the sample feature set; and determining a loss value of the initial network based on the sample weight of the sample feature set and the loss value of the sample feature set.

According to an embodiment of the invention, the above-mentioned loss function comprises a focus loss function.

According to an embodiment of the present invention, the obtaining of the sample feature set includes: a plurality of sample feature sets is obtained. The method further comprises the following steps: before the loss value of the initial network is determined, smoothing is carried out on the label scores of the sample feature sets to obtain updated label scores of the sample feature sets. The determining the loss value of the initial network based on the sample weight of the sample feature set and the loss value of the sample feature set includes: and carrying out weighted summation on the loss values of the plurality of sample characteristic sets by utilizing the sample weights of the plurality of sample characteristic sets to obtain the loss value of the initial network.

According to an embodiment of the invention, the feature set comprises: a set of attribute features and/or a set of behavior features. The sample feature set includes: a set of attribute features and/or a set of behavior features. The attribute feature set includes at least one of: a user purchasing power rating characteristic, a user gender characteristic, a user value rating characteristic, a user loyalty characteristic, a user activity characteristic, a login time characteristic, a user promotion sensitivity characteristic, and a receiving city rating characteristic; and/or the behavioral characteristic data includes at least one of: a quantity of items viewed within a first predetermined period of time characteristic, a quantity of brands viewed within a first predetermined period of time characteristic, a time of view characteristic within a first predetermined period of time characteristic, an average price of items viewed within a first predetermined period of time characteristic, a quantity of items purchased within a second predetermined period of time characteristic, an average order amount of money characteristic within a second predetermined period of time characteristic, and a time of purchase characteristic.

According to an embodiment of the present invention, the determining the label score of the sample feature set includes: determining a first score according to the purchasing power grade characteristics of the user in the sample characteristic set; determining a second score according to the user loyalty grade characteristics in the sample characteristic set; randomly selecting a numerical value from a preset numerical value interval to serve as a third score; and summing the first score, the second score, and the third score to obtain a label score for the sample feature set.

According to an embodiment of the present invention, the method further includes: building a plurality of item information pools respectively aiming at a plurality of consumption capacity categories; and for any item, determining a consumption capability category matched with the any item according to at least one of historical order data about the any item, historical browsing data about the any item, an item category to which the any item belongs and a price interval of the any item, and adding information of the any item to an item information pool for the matched consumption capability category.

According to an embodiment of the present invention, the recommending items to the user based on the item information pool for the consumption capability category includes: and selecting information of at least one item from the item information pool aiming at the consumption capacity category, and pushing the information of the at least one item to the client of the user for displaying.

Another aspect of the present invention provides an article recommendation apparatus, including: the device comprises an acquisition module, a deep learning module, a category determination module and a recommendation module. The acquisition module is used for acquiring the feature set of the user. And the deep learning module is used for processing the feature set by utilizing a pre-trained deep learning network to obtain the consumption capability prediction score of the user. The category determination module is used for determining the consumption capability category of the user based on the consumption capability prediction score. And the recommending module is used for recommending the articles to the user based on the article information pool aiming at the consumption capacity category.

Another aspect of the invention provides a computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method as described above when executing the program.

Another aspect of the invention provides a computer-readable storage medium storing computer-executable instructions for implementing the method as described above when executed.

Another aspect of the invention provides a computer program comprising computer executable instructions for implementing a method as described above when executed.

According to the embodiment of the invention, a sinking user scene analysis scheme improved based on deep learning is provided, when item recommendation is expected to be carried out on a specified user, a consumption capability category to which the specified user belongs is determined by utilizing a deep learning network based on various dimensional characteristics of the specified user, so that items can be selected from an item information pool aiming at the consumption capability category and recommended to the user. The process relies on the automatic learning and classification of the deep learning network on the service data and the user characteristics, so that the manual participation can be reduced, and the accuracy of the sinking score prediction can be improved, thereby improving the accuracy of the consumption capability category judgment. The method and the system can realize accurate positioning item recommendation aiming at the needs of users (particularly sinking users), improve the purchasing power of sinking users in different degrees, and better assist the e-commerce platform in excavating the value of sinking groups.

Drawings

The above and other objects, features and advantages of the present invention will become more apparent from the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:

FIG. 1 schematically illustrates an exemplary system architecture for applying the item recommendation method and apparatus according to an embodiment of the present invention;

FIG. 2 schematically shows a flow chart of an item recommendation method according to an embodiment of the invention;

FIG. 3 schematically illustrates an example flow diagram of a training process for a deep learning network in accordance with an embodiment of this disclosure;

FIG. 4 schematically illustrates an example architectural diagram of an initial network in accordance with an embodiment of this invention;

FIG. 5 schematically shows a block diagram of an item recommendation device according to an embodiment of the present invention; and

FIG. 6 schematically shows a block diagram of a computer device according to an embodiment of the invention.

Detailed Description

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. It is to be understood that such description is merely illustrative and not intended to limit the scope of the present invention. In the following detailed description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the embodiments of the invention. It may be evident, however, that one or more embodiments may be practiced without these specific details. Moreover, in the following description, descriptions of well-known structures and techniques are omitted so as to not unnecessarily obscure the concepts of the present invention.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The terms "comprises," "comprising," and the like, as used herein, specify the presence of stated features, steps, operations, and/or components, but do not preclude the presence or addition of one or more other features, steps, operations, or components.

All terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art unless otherwise defined. It is noted that the terms used herein should be interpreted as having a meaning that is consistent with the context of this specification and should not be interpreted in an idealized or overly formal sense.

Where a convention analogous to "at least one of A, B and C, etc." is used, in general such a construction is intended in the sense one having skill in the art would understand the convention (e.g., "a system having at least one of A, B and C" would include but not be limited to systems that have a alone, B alone, C alone, a and B together, a and C together, B and C together, and/or A, B, C together, etc.). Where a convention analogous to "A, B or at least one of C, etc." is used, in general such a construction is intended in the sense one having skill in the art would understand the convention (e.g., "a system having at least one of A, B or C" would include but not be limited to systems that have a alone, B alone, C alone, a and B together, a and C together, B and C together, and/or A, B, C together, etc.).

The embodiment of the invention provides an article recommendation method and device. The item recommendation method can comprise a feature acquisition process, a deep learning processing process, a consumption capability category determination process and an item recommendation process. And acquiring a feature set of the user in a feature acquisition process. And then, carrying out a deep learning processing process, and processing the feature set by using a pre-trained deep learning network to obtain the consumption capability prediction score of the user. The base consumption capability category determination process may determine a consumption capability category of the user based on the consumption capability prediction score. Then in the item recommendation process, item recommendation can be performed on the user based on the item information pool for the consumption capability category.

With the development of big data and artificial intelligence, more and more users enjoy online shopping. In order to attract more users, the e-commerce platform increases the user viscosity through various strategies or algorithms, and improves the user experience. At present, sinking groups are income to families or cities below three lines at the middle and low ends, and occupy a large proportion, and the internet consumption of the sinking groups is relatively slow in development, so that the sinking groups have a large promotion space. In order to further mine the value of the sinking group, the method has very important significance for item recommendation of the sinking user.

In one processing mode, a data analyst of the e-commerce platform sets a strategy formula according to experience, and then calculates the subsidence score of the user according to the strategy consensus and the user characteristics. And recommending different commodities for the user according to different sinking scores. However, this approach has the following disadvantages: with the continuous development of services, the user characteristics will change continuously, and the formula needs to be updated continuously, which consumes more labor cost. Moreover, the experience of people is limited, and factors possibly considered by a design formula are not comprehensive, so that the calculated sinking score is not accurate, and further the misjudgment recommendation is not accurate. The embodiment of the invention applies the front-edge deep learning algorithm, can automatically learn the change of the characteristics, automatically trains the model regularly, and can overcome the defects of the prior art.

Fig. 1 schematically illustrates an exemplary system architecture 100 to which the item recommendation method and apparatus may be applied, according to an embodiment of the present invention. It should be noted that fig. 1 is only an example of a system architecture to which the embodiment of the present invention may be applied, so as to help those skilled in the art understand the technical content of the embodiment of the present invention, and it does not mean that the embodiment of the present invention may not be applied to other devices, systems, environments or scenarios.

As shown in fig. 1, a system architecture 100 according to an embodiment of the present invention may include

terminal devices

101, 102, 103, a network 104, and a server 105. The network 104 serves as a medium for providing communication links between the

terminal devices

101, 102, 103 and the server 105. Network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables, to name a few.

The

terminal apparatuses

101, 102, 103 communicate with the server 105 through the network 104 to receive or transmit messages and the like. The

terminal devices

101, 102, 103 may have installed thereon client applications having various functions, such as shopping-like applications, web browser applications, search-like applications, instant messaging tools, mailbox clients, social platform software, etc. (by way of example only).

The

terminal devices

101, 102, 103 may be various electronic devices including, but not limited to, smart phones, tablet computers, laptop portable computers, desktop computers, and the like.

The server 105 may be a server providing various services, such as a background management server providing support for various client applications in the

terminal devices

101, 102, 103. The background management server may receive the request message sent by the

terminal device

101, 102, 103, perform a response such as analysis processing on the received request message, and feed back a response result (for example, a web page, information, or data generated according to the request message or the like) to the

terminal device

101, 102, 103, where the

terminal device

101, 102, 103 outputs the response result to the user.

It should be noted that the item recommendation method according to the embodiment of the present invention may be implemented in the

terminal devices

101, 102, and 103, and accordingly, the item recommendation device according to the embodiment of the present invention may be disposed in the

terminal devices

101, 102, and 103. Alternatively, the item recommendation method according to the embodiment of the present invention may be implemented in the server 105, and accordingly, the item recommendation apparatus according to the embodiment of the present invention may be provided in the server 105. Alternatively, the item recommendation method according to the embodiment of the present invention may be implemented in other computer devices capable of communicating with the

terminal devices

101, 102, 103 and/or the server 105, and accordingly, the item recommendation apparatus according to the embodiment of the present invention may be provided in other computer devices capable of communicating with the

terminal devices

101, 102, 103 and/or the server 105.

It should be understood that the number and types of terminal devices, networks, and servers in fig. 1 are merely illustrative. There may be any number and any type of terminal devices, networks, and servers, depending on the actual needs.

According to the embodiment of the invention, an article recommendation method is provided, and is a sinking user scene analysis scheme based on deep learning improvement, so that the value of sinking groups can be better assisted for an e-commerce platform. The method is illustrated by the figure below. It should be noted that the sequence numbers of the respective operations in the following methods are merely used as representations of the operations for description, and should not be construed as representing the execution order of the respective operations. The method need not be performed in the exact order shown, unless explicitly stated.

Fig. 2 schematically shows a flow chart of an item recommendation method according to an embodiment of the invention.

As shown in fig. 2, the method may include operations S201 to S204.

In operation S201, a feature set of a user is acquired.

Illustratively, a user interacts with a server of an e-commerce platform through a client, the server recording a large amount of information indexed by the user identification of the user. When article recommendation needs to be performed on a user, the user identification of the user can be used as an index, and the feature set of the user can be obtained from a large amount of information recorded by the server. One or more features may be included in the feature set. The feature set may include, for example, at least one of an attribute feature set, a behavior feature set, and the like. The attribute feature set represents the static features of the user, and the behavior feature set represents the dynamic features of the user.

In operation S202, the feature set is processed by using a pre-trained deep learning network to obtain a consumption capability prediction score of the user.

For example, the consumption ability prediction score can be used to represent the level of consumption ability and further the level of sinking, and in this case, the consumption ability prediction score can also be referred to as the sinking score. Wherein, the mapping relation between the consumption capability prediction score and the sinking degree can be set according to the requirement. For example, the higher the consumer ability prediction score, the higher the characterized sinking level, and the lower the corresponding consumer's consumption level in the e-commerce platform. In addition, the deep learning network can be trained regularly according to the business data, so that the processing performance of the deep learning network can adapt to the development change of the actual business.

In operation S203, a consuming capacity category of the user is determined based on the consuming capacity prediction score.

For example, a plurality of consumption capability categories, which may also be referred to as sinking categories in a scenario for sinking user item recommendation, may be classified in advance according to different sinking degrees. For example, the sinking scores may be divided into four sinking categories of "non-sinking user", "advanced sinking user", "standard sinking user" and "deep sinking user" in order from low to high. The sink categories may be divided, for example, as follows.

The sinking score of the sinking category of deep sinking users is in the interval of [0, 19], wherein 78% of users have no portrait information, and among the users with portrait information, the predetermined user grade score of 98% of users is less than one thousandth, and the users in the first-line city only account for 8%. Inactive users account for 37%, normal active users for 22%, active users for 28%, and very active users for 12% in 90 days.

The sink score for the sink category "standard sink users" is in the [20, 39] interval, where 60% of users have no portrait information and 12% of users with portrait information are from the first-line city. Inactive users account for 16%, normal active users for 21%, active users for 40%, and very active users for 23% in 90 days.

The sink score for the sink category "advanced sink users" is in the [40, 69] interval, with 35% of users without portrait information and 17% of users with portrait information coming from the first-line city. Inactive users account for 10%, normal active users for 10%, active users for 30%, and very active users for 50% within 90 days.

The sinking score of the sinking category "non-sinking users" is within the [70, 100] interval, where 20% of users have no portrait information, and among users with portrait information, 43% of users have a predetermined user rating score greater than ten-thousandths, and 24% come from a first-line city. Inactive users account for 3%, normal active users for 4%, active users for 16%, and very active users for 75% within 90 days.

The above sinking category division manner is only an example, and in an actual situation, the sinking category division strategy may be performed according to a data statistical rule embodied by an actual service. Therefore, after the subsidence score of a user is obtained, the subsidence category to which the user belongs can be determined according to the partitioning strategy.

Next, in operation S204, item recommendation is made to the user based on the item information pool for the consumption capability category.

Illustratively, a plurality of item information pools, i.e., a plurality of item information storage areas, may be pre-constructed in the server of the e-commerce platform. Different item information pools correspond to different consumer capability categories. For example, the item information pool of the "deep sink user" holds item information preferred by users belonging to the consumption capability category of the "deep sink user". The same applies to the item information pools of other consumer capability categories. Therefore, one or more item information can be selected from the item information pool of the consumption capability category to which the user belongs and pushed to the client of the user, so that the user can know the recommended item information through the client. Because the recommended article information is selected from the article information pool matched with the sinking attribute of the user, the probability that the recommended article information can obtain the attention and the favorite of the user is high, and therefore the internet consumption enthusiasm of the users with different consumption capability categories can be called.

It can be understood by those skilled in the art that the item recommendation method according to the embodiment of the present invention substantially provides a sinking user scenario analysis scheme improved based on deep learning, and when item recommendation is desired for a specific user, a consumption capability category to which the specific user belongs is determined based on various dimensional features of the specific user by using a deep learning network, so that an item can be selected from an item information pool for the consumption capability category to be recommended to the user. The process relies on the automatic learning and classification of the deep learning network on the service data and the user characteristics, so that the manual participation can be reduced, and the accuracy of the sinking score prediction can be improved, thereby improving the accuracy of the consumption capability category judgment. The method and the system can realize accurate positioning item recommendation aiming at the needs of users (particularly sinking users), improve the purchasing power of sinking users in different degrees, and better assist the e-commerce platform in excavating the value of sinking groups.

According to an embodiment of the present invention, the feature set of the user may include: a set of attribute features and/or a set of behavior features.

Illustratively, the set of attribute features may include, for example, at least one of: a user purchasing power rating feature, a user gender feature, a user value rating feature, a user loyalty feature, a user activity feature, a login time feature, a user promotion sensitivity feature, and a city of receipt rating feature. For example, the behavior feature data may include, for example, at least one of: a quantity of items viewed within a first predetermined period of time characteristic, a quantity of brands viewed within a first predetermined period of time characteristic, a time of view characteristic within a first predetermined period of time characteristic, an average price of items viewed within a first predetermined period of time characteristic, a quantity of items purchased within a second predetermined period of time characteristic, an average order amount of money characteristic within a second predetermined period of time characteristic, and a time of purchase characteristic. The login time characteristic may be, for example, a characteristic determined according to the latest login time of the user, and the receiving city grade characteristic may be a characteristic determined according to a grade to which a receiving city most frequently used by the user belongs. The first predetermined period of time and the second predetermined period of time may be set according to actual needs, for example, the first predetermined period of time is the last month, the second predetermined period of time is the last year, and the like. The characteristic of the number of items viewed in the first predetermined period of time may be, for example, a characteristic determined according to the number of SKUs (Stock Keeping Unit) numbers of items viewed by the user in the last month. The quantity characteristic of items purchased during the second predetermined period of time may be, for example, a characteristic determined from the quantity of SKU numbers of items purchased by the user in the last year. The quantity characteristic of the purchased item class within the second predetermined period of time may be, for example, a characteristic determined according to the quantity of items of the third class purchased by the user in the last year. The purchasing time characteristic may be, for example, a time interval between a time when the last purchasing behavior of the user occurs and a current time, such as a number of days from the last purchasing behavior to today. The characteristics can describe the user more comprehensively, abundantly and stereoscopically from a plurality of different dimensions so as to assist the subsequent deep learning process to obtain a more accurate prediction result.

Further, before the item recommendation method according to the embodiment of the present invention is performed, an item information pool for different consumption capability categories needs to be established for performing the final item recommendation. Exemplarily, the method may further include: a plurality of item information pools respectively for a plurality of consumption capability categories are constructed. For any item, determining a consumption capability category matched with the any item according to at least one of historical order data about the any item, historical browsing data about the any item, an item category to which the any item belongs and a price interval of the any item, and adding information of the any item to an item information pool for the matched consumption capability category. This process may be referred to as a sinking selection process, such as finding a good with a high click-through rate and conversion rate for each consumer capability category.

On this basis, according to an embodiment of the present invention, the recommending items to the user based on the item information pool for the consumption capability category includes: and selecting information of at least one item from the item information pool aiming at the consumption capacity category, and pushing the information of the at least one item to the client of the user for displaying.

Before establishing an item information pool for a consumption capability category, item features need to be analyzed from multiple dimensions so that matching relationships between items and consumption capability categories can be determined. For example, for any item, at least one of the following characteristics of the item is obtained: data such as UV (Unique viewer, number of independent visitors), PV (Page View, volume of access) where the item was viewed within a third predetermined period of time (e.g., 1 day) in the past; the amount of articles to be put into close attention, the amount of comments, the amount of good comments, the amount of comments with drawings, etc.; total sales of the goods; introducing an order of an article and completing the order; the order conversion rate after the article is purchased; the attention amount of the articles accounts for the third-class category; the average proportion of the unit price of the article in the order amount, and the like.

A consumer capability category matching the item may then be determined based on historical order data for the item and/or based on historical browsing data for the item. The selection from the order dimension is taken as an example for explanation. The following information may be obtained from historical order data for the item: (1) the order data of the item in the fourth predetermined period (for example, the previous 1 day) is taken from the basic order table, and the type of the equipment used by the user who purchases the item can be determined according to the value of the field "order _ type", for example, whether the user equipment is a personal computer or a mobile terminal can be distinguished, and the brand of the user equipment can also be distinguished. (2) Since the consumption capability categories of some users are predicted in the historical business operation process and recorded in the sinking user table, the consumption capability categories to which the users who purchase the item belong can be known from the sinking user table, and include 'sinking users' and 'non-sinking users', for example. (3) And counting the price interval of the average price of the items from the dimension of the SKU number. For example, a plurality of price intervals, such as 0 to 10, 10 to 100, etc., may be divided. (4) And counting the order quantity of different user groups about the item. (5) And counting the order quantity of different user equipment about the item. Similarly, the information related to the browsing behavior of the user as described above may also be obtained according to the historical browsing data about the item, and will not be described herein again.

Based on the characteristics of the articles and the information, the matching relationship between the articles and the consumption capability categories can be analyzed.

According to the embodiment of the present invention, before the feature set is processed by using the deep learning network trained in advance in operation S202, a training process of the deep learning network, that is, a model training process, needs to be performed. The method is illustrated by the figure below. It should be noted that the sequence numbers of the respective operations in the following methods are merely used as representations of the operations for description, and should not be construed as representing the execution order of the respective operations. The method need not be performed in the exact order shown, unless explicitly stated.

FIG. 3 schematically shows an example flow diagram of a training process for a deep learning network according to an embodiment of the invention. It should be understood that the training process and the item recommendation process of the above embodiments may be implemented on the same device, or may be implemented on different devices, and are not limited herein.

As shown in fig. 3, the training process of the deep learning network may include operations S301 to S303.

In operation S301, an initial network is constructed.

An exemplary structure of the initial network is explained below with reference to fig. 4, and fig. 4 schematically shows an exemplary structural diagram of the initial network according to an embodiment of the present invention.

In the example shown in fig. 4, the initial network 400 may include: an Embedding Layer (Embedding Layer)410, a feature extraction Layer, and an Output Layer (Output Layer) 430. The feature extraction layer may include a low-order feature extraction sub-network 421, a medium-order feature extraction sub-network 422, and a high-order feature extraction sub-network 423, among others.

Illustratively, when a certain set of features is input to the embedding layer 410, the embedding layer 410 is used to convert the input features into feature vectors.

The low-order feature extraction sub-network 421 may include a Linear layer 4211 and a Factoring Machine (FM) 4212. The low-order feature extraction sub-network 421 is configured to extract a linear feature combination of feature vectors by using the linear layer 4211, extract a second-order cross feature combination of the feature vectors by using the factoring machine 4212, and form a first feature by the linear feature combination and the second-order cross feature combination.

The medium-order feature extraction subnetwork 422 can include a Convolutional Neural Network (CNN) 4221 and a Recombination Layer (Recombination Layer) 4222. The medium-order feature extraction sub-network 422 is configured to extract local features of the feature vectors using the convolutional neural network 4221, and combine the local features using the reconstruction layer 4222 to obtain second features.

The higher-order feature extraction sub-Network 423 may include a Deep Neural Network (DNN) for calculating a third feature of the feature vector using a matrix of the Deep Neural Network.

Output layer 430 is to compute a prediction score based on the first feature, the second feature, and the third feature. Specifically, the results of the neural network forward propagation are passed to the output layer 430 to calculate a consumer prediction score, i.e., a subsidence score, for the feature set.

It is understood that the first feature, the second feature and the third feature respectively represent information with different granularities so as to more fully describe the characteristics of the input. In order to extract information with different granularities, the embodiment of the invention designs a neural network of a multilayer perceptron.

According to the embodiment of the present invention, the factorization machine 4212 is provided with a random factor (Dicefactor) mechanism, which is a random inactivation (dropout) strategy, and can randomly discard part of interaction paths in a second-order interaction path formed by pairwise interaction of elements in the feature vector, so as to prevent mutual influence between the elements, thereby alleviating the problem of gradient coupling and preventing overfitting.

According to an embodiment of the present invention, the convolutional neural network 4221 may include a convolutional Layer (Convolution Layer) and a Pooling Layer (Pooling Layer). The above process of extracting local features of the feature vector by using the convolutional neural network may include: and carrying out convolution operation on the feature vector by the convolution layer to obtain high-order local features, and carrying out dimensionality reduction operation on the high-order local features by the pooling layer to obtain the local features. The process of combining local features using a recombination layer may include: the local features are combined in a fully connected manner by the recombination layer.

According to an embodiment of the present invention, the deep neural network is provided with a Multi-head Self-Attention (Multi-head Self-Attention) mechanism. On this basis, the process of extracting the third feature of the feature vector by using the deep neural network may include: a third feature is extracted by the deep neural network based on a multi-headed self-attention mechanism. The multi-head self-attention mechanism can better capture high-order features with mutual information.

According to an embodiment of the invention, the initial network 300 may further comprise an input layer, which may be arranged before the embedding layer. The input feature set may first enter an input layer for preprocessing the input feature set. The embedding layer is used to convert the preprocessed sample feature set into feature vectors.

Illustratively, the above-described preprocessing process may include, for example, an outlier truncation process, a normalization process, and the like. The normalization process is to make the data distribution of each feature in the feature set in the same interval, and the maximum and minimum normalization and the normal distribution normalization are commonly used. For example, the embodiment of the present invention may adopt maximum and minimum normalization, and the calculation formula is: (max-min)/(current-min).

After the initial network, such as that shown in fig. 4, is constructed, reference is continued below with respect to fig. 3.

In operation S302, a sample feature set is obtained, and a label score of the sample feature set is determined.

According to an embodiment of the present invention, the above process of determining the label score of the sample feature set may be implemented as follows: determining a first score according to the purchasing power grade characteristics of the user in the sample characteristic set; determining a second score according to the user loyalty grade characteristics in the sample characteristic set; randomly selecting a value from the predetermined value interval as a third score. The first score, the second score, and the third score are then summed to obtain a label score for the sample feature set.

For example, when acquiring sample data of an initial network, user delineation may be performed first. For example, an active user in the e-commerce platform is selected as a Sample (Sample) user, defined as a user who has made an order within the last 1 year or has browsed within the last 1 month. The time and the setting of the behavior type in the definition can be changed according to actual needs, and are not limited herein. Construction of sample features and labels (Label) was then performed.

In the process of constructing the sample feature and the label, the feature construction and the label construction can be performed in sequence. And acquiring a sample feature set of the selected sample user during feature construction. According to an embodiment of the invention, the sample feature set may comprise an attribute feature set and/or a behavior feature set. The attribute feature set includes at least one of: a user purchasing power rating characteristic, a user gender characteristic, a user value rating characteristic, a user loyalty characteristic, a user activity characteristic, a login time characteristic, a user promotion sensitivity characteristic, and a receiving city rating characteristic; and/or the behavioral characteristic data includes at least one of: a quantity of items viewed within a first predetermined period of time characteristic, a quantity of brands viewed within a first predetermined period of time characteristic, a time of view characteristic within a first predetermined period of time characteristic, an average price of items viewed within a first predetermined period of time characteristic, a quantity of items purchased within a second predetermined period of time characteristic, an average order amount of money characteristic within a second predetermined period of time characteristic, and a time of purchase characteristic. The sample feature set and the feature set of the user are obtained in the same manner, and are not described herein again.

Tag construction is one of the difficulties of the embodiment of the invention, namely how to give a Label which is in accordance with the expectation under the condition that the sinking score is not existed at present, and the Label is provided for model training. First, it can be known by analyzing the model requirements that the final goal is to give a prediction score of the degree of subsidence according to the user characteristics. Among the existing features, the features related to the feature of "sinking" include: "purchasing power level", "promotion sensitivity", "user loyalty", etc., it is desirable for embodiments of the present disclosure to determine a subsidence score (which may also be referred to as a label score) as a sample label based on these characteristics associated with the "subsidence" characteristic for training of the initial network. Illustratively, the purchasing power level features are obtained by extracting user history data, such as the average price of the user's past one month order, the average price of the clicked item, the average price of the interested item, the clicked item class, the clicked store, the clicked brand, etc., and then calculating the purchasing power level of the user using a linear weighting model. The promotion sensitivity degree is characterized in that a score is calculated by integrating a plurality of factors such as user preference for promotion commodities, user preference for discounted commodities, user acceptance of commodity price reduction range and the like. The user loyalty feature is primarily to mine some information from the user profile, such as the user's retention time, the user's predetermined user rating score, whether the user is entitled to a certain title, etc.

For example, the user purchasing power level feature and the user loyalty feature may be mapped to corresponding Label scores. For example, the mapping between the purchasing power level characteristic and the first score includes: the "local tyrant" grade → 90 points, "advanced white collar" grade → 70 points, "common white collar" grade → 50 points, "blue collar" grade → 25 points, "little income" grade → 10 points, etc. The first score may be written as label _ purch. User loyalty characteristics can also be introduced, and people with different purchasing powers can be subdivided through purchasing frequency and money amount. The mapping between the user loyalty characteristic and the second score comprises: "high loyalty" → 10 point, "moderate loyalty" → 5 point, "general" → 0 point, "occasional" → -5 point, "thrower type" → -10 point, and the like. The second score may be written as label _ loyal. To increase the robustness of the model, an integer may be randomly chosen as a third score, also called a random label score, denoted label _ rand, in a predetermined value interval, e.g., -5, 5. Combining the above factors, Label score Label is Label _ purch + Label _ royal + Label _ layout. In other embodiments, the above first score, second score, and third score may also be weighted and summed to obtain the label score.

In operation S303, the initial network is trained using the sample feature set and the corresponding label scores to obtain the above-mentioned pre-trained deep learning network.

The training process described above is illustratively described below with continued reference to fig. 4. As shown in fig. 4, the sample Feature set may be divided into a Category Feature (Category Feature) group 441 and a numerical Feature (numerical Feature) group 442 according to different properties of the features. For example, gender characteristics of sample users may be classified into a category characteristic group, and the number of purchased items of the sample users in the last year may be classified into a numerical characteristic group, thereby avoiding interaction between characteristics of different nature. The sample feature set may be input to the input layer first to perform preprocessing such as outlier truncation and normalization, and the preprocessed sample feature set may be input to the embedding layer 410. The sample feature set is converted into feature vectors by using the embedding layer 410, linear feature combinations of the feature vectors are extracted by using the linear layer 4211, second-order cross feature combinations of the feature vectors are extracted by using the factor decomposition machine 4212, and first features are formed by the linear feature combinations and the second-order cross feature combinations. Local features of the feature vectors are extracted using the convolutional neural network 4221, and the local features are combined using the reconstruction layer 4222 to obtain second features. And extracting a third feature of the feature vector by using the deep neural network. The sink scores for the sample feature set are output by the output layer 430 based on the first feature, the second feature, and the third feature. Then, determining a loss value of the initial network 400 based on the label score and the subsidence score of the sample feature set; when the loss value converges, it is determined that the initial network is trained as the previously trained deep learning network. The internal mechanism of each part in the initial network is described above, and repeated parts are not described again.

According to the embodiment of the invention, in order to avoid the problem that the prediction capability of the deep learning network is reduced due to Sample imbalance, the tendency of the training process can be adjusted by adjusting the Sample Weight (Sample Weight). For example, after selecting a plurality of sample users, a plurality of sample feature sets may be obtained, and a sample weight may be set for each sample feature set. For example, the sample weight is added according to the label to which the sample belongs, for example, the sample represents an exposure sample with label of 0, the sample represents a click sample with weight of 10, the sample represents a label of 1, the sample represents an order sample with weight of 20, the sample represents an order sample with label of 3, and the like. The sample feature set can then be input into an initial network for training, for any sample feature set, a dip score of the sample feature set is output by the initial network, and a loss value of the sample feature set is determined based on a loss function of the initial network and a difference between the label score and the dip score of the sample feature set. Similarly, a loss value for each of a plurality of sample feature sets may be obtained. And carrying out weighted summation on the loss values of the plurality of sample characteristic sets by utilizing the sample weights of the plurality of sample characteristic sets, so as to determine the loss value of the initial network. And determining whether the loss value of the initial network reaches convergence, if not, optimizing and updating the weight of the initial network, and training the updated initial network by using the sample feature set. And repeating the steps until the loss value is converged, determining that the training is finished, and obtaining the deep learning network after the initial network is subjected to iterative optimization.

In order to further improve the prediction performance of the deep learning network obtained by training, a Label smoothing (Label smoothing) process can be added to avoid overfitting. For example, before determining the loss value of the initial network, smoothing the label scores of the plurality of sample feature sets to obtain updated label scores of the plurality of sample feature sets. And (3) performing the initial network input and iterative updating on the basis of the updated label scores of the sample feature sets until a deep learning network is finally obtained. For example, the tags identified in the sample are hard targets, and tag smoothing uses a weighted average of the hard targets and the data distribution on the tags as soft targets. The concrete formula is as follows: y x (1-alpha) + alpha/K.

According to the embodiment of the invention, the Loss function of the initial network can comprise a Focal Loss (Focal local) function, and the problem of poor training effect caused by sample imbalance can be solved.

Furthermore, according to the embodiment of the invention, sinking data analysis can be performed through the following four dimensions, so as to better guide model training of the sinking user prediction, extract more valuable features into the model and construct more reliable labels. Examples include: and data fluctuation analysis shows that about 5% of users have fluctuation every day, and few users with the convergence prediction score exceeding 10 points exist in the fluctuation users. Permeability analysis, permeability in this example is defined as: permeability for a city (electricity merchant platform visited UV over 30 days)/(city resident population). Data analysis shows that the permeability of resident population of monthly active users of the e-commerce platform in the whole country is 10%, and the active users are very concentrated in first-line, new first-line and second-line cities. Illustrating that cities have very much room for growth below three lines. Urban distribution analysis shows that in a two-line city, deep sinking accounts for 10% of users, middle sinking accounts for 17% of users, and non-sinking accounts for 38%; in a three-four-line city, the deep sinking users account for 17 percent, the middle sinking accounts for 36 percent, and the middle sinking does not account for non-sinking; in the city under four lines, deep sinking accounts for 40% of users, and middle sinking accounts for 25%. And (4) analyzing the purchase categories, wherein the willingness of the sinking user to purchase a certain first-level category is equal to the amount of orders of the sinking user in the first-level category/the total amount of orders of the sinking user. According to the purchase intention, deeply sinking the categories which the user likes to purchase: food beverage (20.8%), household cleaning/paper (15.2%), dress underwear (5.6%), personal care (5.5%), and daily household use (5.5%). Mid-sink users like the categories purchased: food beverage (18%), dress underwear (12%), makeup skin care (5.4%), personal care (5.3%), books (5.05%).

It can be understood that in order to better mine the market value of the sinking user of the e-commerce platform, the item recommendation method according to the embodiment of the invention applies a front-edge deep learning algorithm to predict the sinking score of the user and further guide the fine operation of the sinking service. The construction of the characteristic data and the label goes deep into business logic, so that the deep learning network obtained by training has the prediction performance according with the actual business situation. In addition, the embodiment of the invention is combined with a specific scene, improves the loss function in the deep learning algorithm and further improves the prediction precision of the deep learning network.

Fig. 5 schematically shows a block diagram of an item recommendation device according to an embodiment of the present invention.

As shown in fig. 5, the item recommendation device 500 may include: an acquisition module 510, a deep learning module 520, a category determination module 530, and a recommendation module 540.

The obtaining module 510 is configured to obtain a feature set of a user.

The deep learning module 520 is configured to process the feature set using a pre-trained deep learning network to obtain a consumption capability prediction score of the user.

The category determination module 530 is configured to determine a consumer capability category for the user based on the consumer capability prediction score.

The recommending module 540 is configured to recommend the item to the user based on the item information pool for the consumption capability category.

It should be noted that the implementation, solved technical problems, implemented functions, and achieved technical effects of each module/unit/subunit and the like in the apparatus part embodiment are respectively the same as or similar to the implementation, solved technical problems, implemented functions, and achieved technical effects of each corresponding step in the method part embodiment, and are not described herein again.

Any number of modules, sub-modules, units, sub-units, or at least part of the functionality of any number thereof according to embodiments of the invention may be implemented in one module. Any one or more of the modules, sub-modules, units, and sub-units according to the embodiments of the present invention may be implemented by being divided into a plurality of modules. Any one or more of the modules, sub-modules, units, sub-units according to embodiments of the present invention may be implemented at least in part as a hardware circuit, such as a Field Programmable Gate Array (FPGA), a Programmable Logic Array (PLA), a system on a chip, a system on a substrate, a system on a package, an Application Specific Integrated Circuit (ASIC), or may be implemented in any other reasonable manner of hardware or firmware by integrating or packaging a circuit, or in any one of or a suitable combination of software, hardware, and firmware implementations. Alternatively, one or more of the modules, sub-modules, units, sub-units according to embodiments of the present invention may be at least partially implemented as computer program modules, which, when executed, may perform the corresponding functions.

For example, any of the obtaining module 510, the deep learning module 520, the category determining module 530, and the recommending module 540 may be combined and implemented in one module, or any one of them may be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules may be combined with at least part of the functionality of the other modules and implemented in one module. According to an embodiment of the present invention, at least one of the obtaining module 510, the deep learning module 520, the category determining module 530 and the recommending module 540 may be implemented at least partially as a hardware circuit, such as a Field Programmable Gate Array (FPGA), a Programmable Logic Array (PLA), a system on a chip, a system on a substrate, a system on a package, an Application Specific Integrated Circuit (ASIC), or may be implemented in hardware or firmware by any other reasonable manner of integrating or packaging a circuit, or may be implemented in any one of three implementations of software, hardware and firmware, or in any suitable combination of any of them. Alternatively, at least one of the obtaining module 510, the deep learning module 520, the category determination module 530 and the recommendation module 540 may be at least partially implemented as a computer program module, which when executed may perform a corresponding function.

FIG. 6 schematically illustrates a block diagram of a computer device suitable for implementing the model training method and/or mapping method described above, according to an embodiment of the invention. The computer device shown in fig. 6 is only an example, and should not bring any limitation to the function and the scope of use of the embodiments of the present invention.

As shown in fig. 6, a computer apparatus 600 according to an embodiment of the present invention includes a processor 601 which can perform various appropriate actions and processes according to a program stored in a Read Only Memory (ROM)602 or a program loaded from a storage section 608 into a Random Access Memory (RAM) 603. Processor 601 may include, for example, a general purpose microprocessor (e.g., a CPU), an instruction set processor and/or associated chipset, and/or a special purpose microprocessor (e.g., an Application Specific Integrated Circuit (ASIC)), among others. The processor 601 may also include onboard memory for caching purposes. Processor 601 may include a single processing unit or multiple processing units for performing the different actions of the method flows according to embodiments of the present invention.

In the RAM 603, various programs and data necessary for the operation of the apparatus 600 are stored. The processor 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. The processor 601 performs various operations of the method flow according to the embodiments of the present invention by executing programs in the ROM 602 and/or RAM 603. It is to be noted that the programs may also be stored in one or more memories other than the ROM 602 and RAM 603. The processor 601 may also perform various operations of method flows according to embodiments of the present invention by executing programs stored in the one or more memories.

Device 600 may also include an input/output (I/O) interface 605, also coupled to bus 604, according to an embodiment of the invention. The device 600 may also include one or more of the following components connected to the I/O interface 605: an input portion 606 including a keyboard, a mouse, and the like; an output portion 607 including a display such as a Cathode Ray Tube (CRT), a Liquid Crystal Display (LCD), and the like, and a speaker; a storage section 608 including a hard disk and the like; and a communication section 609 including a network interface card such as a LAN card, a modem, or the like. The communication section 609 performs communication processing via a network such as the internet. The driver 610 is also connected to the I/O interface 605 as needed. A removable medium 611 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like is mounted on the drive 610 as necessary, so that a computer program read out therefrom is mounted in the storage section 608 as necessary.

According to an embodiment of the invention, the method flow according to an embodiment of the invention may be implemented as a computer software program. For example, embodiments of the invention include a computer program product comprising a computer program embodied on a computer-readable storage medium, the computer program comprising program code for performing the method illustrated by the flow chart. In such an embodiment, the computer program may be downloaded and installed from a network through the communication section 609, and/or installed from the removable medium 611. The computer program, when executed by the processor 601, performs the above-described functions defined in the system of the embodiment of the present invention. The above described systems, devices, apparatuses, modules, units, etc. may be implemented by computer program modules according to embodiments of the present invention.

An embodiment of the present invention further provides a computer-readable storage medium, which may be included in the apparatus/device/system described in the foregoing embodiment; or may exist separately and not be assembled into the device/apparatus/system. The computer-readable storage medium carries one or more programs which, when executed, implement the model training method and/or the mapping method according to an embodiment of the present invention.

According to embodiments of the present invention, the computer readable storage medium may be a non-volatile computer readable storage medium, which may include, for example but is not limited to: a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In embodiments of the invention, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. For example, according to an embodiment of the present invention, a computer-readable storage medium may include the ROM 602 and/or the RAM 603 described above and/or one or more memories other than the ROM 602 and the RAM 603.

The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams or flowchart illustration, and combinations of blocks in the block diagrams or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

It will be appreciated by a person skilled in the art that various embodiments of the invention and/or features recited in the claims may be combined and/or coupled in a number of ways, even if such combinations or couplings are not explicitly recited in embodiments of the invention. In particular, various combinations and/or combinations of features recited in the various embodiments and/or claims of the embodiments may be made without departing from the spirit and teachings of the embodiments. All such combinations and/or associations are within the scope of embodiments of the present invention.

Claims

1. An item recommendation method comprising:

acquiring a feature set of a user;

processing the feature set by utilizing a pre-trained deep learning network to obtain a consumption capability prediction score of the user;

determining a consumer capability category for the user based on the consumer capability prediction score; and

and recommending the items for the user based on the item information pool aiming at the consumption capability category.

2. The method of claim 1, further comprising:

constructing an initial network;

acquiring a sample feature set, and determining label scores of the sample feature set; and

and training the initial network by using the sample feature set and the label scores to obtain the pre-trained deep learning network.

3. The method of claim 2, wherein the initial network comprises: embedding layer, characteristic extraction layer and output layer, the characteristic extraction layer includes:

a low-order feature extraction sub-network comprising a linear layer and a factorizer;

a medium-order feature extraction sub-network comprising a convolutional neural network and a reconstruction layer; and

a higher order feature extraction sub-network comprising a deep neural network.

4. The method of claim 3, wherein the training the initial network using the sample feature set and the label scores comprises:

converting the sample feature set into a feature vector using the embedding layer;

extracting a linear feature combination of the feature vectors by using the linear layer, extracting a second-order cross feature combination of the feature vectors by using the factorization machine, and forming a first feature by using the linear feature combination and the second-order cross feature combination;

extracting local features of the feature vectors by using the convolutional neural network, and combining the local features by using the recombination layer to obtain second features;

extracting a third feature of the feature vector by using the deep neural network;

outputting, by the output layer, a prediction score for the spending capacity of the sample feature set based on the first, second, and third features;

determining a loss value for the initial network based on the label score and the consumption capacity prediction score of the sample feature set; and

determining that the initial network is trained as the pre-trained deep learning network if the loss value reaches convergence.

5. The method of claim 4, wherein said extracting a second order cross feature combination of the feature vectors using the factorizer comprises:

and extracting the second-order cross feature combination by the factorization machine based on a stochastic factor Dicefactor mechanism.

6. The method of claim 4, wherein the convolutional neural network comprises: a convolutional layer and a pooling layer;

the extracting local features of the feature vector using the convolutional neural network comprises: performing convolution operation on the feature vector by the convolution layer to obtain a high-order local feature, and performing dimensionality reduction operation on the high-order local feature by the pooling layer to obtain the local feature;

the combining the local features with the reconstruction layer comprises: combining the local features in a fully connected manner by the recombination layer.

7. The method of claim 4, wherein the extracting third features of the feature vector using the deep neural network comprises:

extracting, by the deep neural network, the third feature based on a multi-headed self-attention mechanism.

8. The method of claim 4, wherein the initial network further comprises an input layer;

the training the initial network further comprises: inputting the sample feature set into the input layer, and preprocessing the sample feature set by using the input layer;

the converting, with the embedding layer, the sample feature set into feature vectors comprises: converting the preprocessed sample feature set into feature vectors with the embedding layer.

9. The method of claim 4, further comprising: setting sample weights for the sample feature set;

the determining a loss value for the initial network based on the label scores and the customer capacity prediction scores for the sample feature set comprises:

determining a loss value for the sample feature set based on a loss function of the initial network and a difference between the label score and the consumer performance prediction score for the sample feature set; and

determining a loss value for the initial network based on the sample weights for the sample feature set and the loss values for the sample feature set.

10. The method of claim 9, wherein,

the obtaining a sample feature set comprises: obtaining a plurality of sample feature sets;

the method further comprises the following steps: before the loss value of the initial network is determined, smoothing label scores of the sample feature sets to obtain updated label scores of the sample feature sets;

the determining a loss value for the initial network based on the sample weights for the sample feature set and the loss values for the sample feature set comprises: and carrying out weighted summation on the loss values of the plurality of sample characteristic sets by utilizing the sample weights of the plurality of sample characteristic sets to obtain the loss value of the initial network.

11. The method of claim 2, wherein the feature set comprises: a set of attribute features and/or a set of behavior features, the set of sample features comprising: attribute feature sets and/or behavior feature sets;

the set of attribute features includes at least one of: a user purchasing power rating characteristic, a user gender characteristic, a user value rating characteristic, a user loyalty characteristic, a user activity characteristic, a login time characteristic, a user promotion sensitivity characteristic, and a receiving city rating characteristic; and/or

The behavioral characteristic data includes at least one of: a quantity of items viewed within a first predetermined period of time characteristic, a quantity of brands viewed within a first predetermined period of time characteristic, a time of view characteristic within a first predetermined period of time characteristic, an average price of items viewed within a first predetermined period of time characteristic, a quantity of items purchased within a second predetermined period of time characteristic, an average order amount of money characteristic within a second predetermined period of time characteristic, and a time of purchase characteristic.

12. The method of claim 2, wherein the determining the label scores for the sample feature set comprises:

determining a first score according to the user purchasing power grade characteristics in the sample characteristic set;

determining a second score according to the user loyalty grade characteristics in the sample characteristic set;

randomly selecting a numerical value from a preset numerical value interval to serve as a third score; and

summing the first score, the second score, and the third score to obtain a label score for the sample feature set.

13. The method of claim 1, further comprising:

building a plurality of item information pools respectively aiming at a plurality of consumption capacity categories; and

for any item, determining a consumption capability category matched with the any item according to at least one of historical order data about the any item, historical browsing data about the any item, an item category to which the any item belongs and a price interval of the any item, and adding information of the any item to an item information pool for the matched consumption capability category.

14. The method of claim 1, wherein the recommending items to the user based on the pool of item information for the consumption capability category comprises:

and selecting information of at least one item from the item information pool aiming at the consumption capacity category, and pushing the information of the at least one item to the client of the user for displaying.

15. An item recommendation device comprising:

the acquisition module is used for acquiring a feature set of a user;

the deep learning module is used for processing the feature set by utilizing a pre-trained deep learning network to obtain a consumption capability prediction score of the user;

a category determination module to determine a consumption capability category of the user based on the consumption capability prediction score; and

and the recommending module is used for recommending the articles for the user based on the article information pool aiming at the consumption capability category.

16. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor when executing the program implementing:

the method of any one of claims 1 to 14.

17. A computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform:

the method of any one of claims 1 to 14.