CN113781134A

CN113781134A - Item recommendation method and device and computer-readable storage medium

Info

Publication number: CN113781134A
Application number: CN202010739271.4A
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: 2020-07-28
Filing date: 2020-07-28
Publication date: 2021-12-10

Abstract

The disclosure provides an article recommendation method and device and a computer-readable storage medium, and relates to the technical field of computers. According to the method, the article recommendation model is obtained by using the mixed logistic regression model, and article recommendation is performed on the article recommendation model based on similar groups by means of the characteristic that the mixed logistic regression model is clustered firstly and then classified, so that the recommendation effectiveness is improved; and moreover, the convergence mode of the hybrid logistic regression model is optimized by calculating the second-order gradient of the loss function, so that the hybrid logistic regression model converges faster, and the training efficiency is improved.

Description

Item recommendation method and device and computer-readable storage medium

Technical Field

The present disclosure relates to the field of computer technologies, and in particular, to an article recommendation method and apparatus, and a computer-readable storage medium.

Background

With the development of internet technology and big data, online shopping becomes an increasing choice of users. The inventor finds that how to furthest mine the requirements of users in the face of huge quantities of commodities so as to realize personalized recommendation is a problem which is urgently needed to be solved by each large e-commerce platform.

Disclosure of Invention

The present disclosure provides a scheme capable of automatically recommending an item for a user.

In the embodiment of the disclosure, the article recommendation model is obtained by using the mixed logistic regression model, and by means of the characteristic that the mixed logistic regression model is clustered firstly and then classified, the article recommendation model carries out article recommendation based on similar groups, so that the recommendation effectiveness is improved; and moreover, the convergence mode of the hybrid logistic regression model is optimized by calculating the second-order gradient of the loss function, so that the hybrid logistic regression model converges faster, and the training efficiency is improved.

According to some embodiments of the present disclosure, there is provided an item recommendation method including:

acquiring a candidate item set corresponding to a user to be recommended;

inputting the feature data of the user to be recommended and the feature data of all articles in the candidate article set into an article recommendation model for processing to obtain the click through rate of the user to be recommended on each article in the candidate article set, wherein the article recommendation model is obtained by training through a mixed logistic regression model;

and recommending the articles for the user to be recommended according to the corresponding articles with the click through rate meeting the preset conditions.

In some embodiments, training the item recommendation model using a hybrid logistic regression model comprises:

inputting a training data set into a mixed logistic regression model for processing, and outputting the click through rate of a user corresponding to each piece of training data in the training data set to a corresponding article, wherein the training data set comprises a plurality of pieces of training data, and each piece of training data comprises a user, feature data of an article associated with the user, and a grade label of a behavior of the user to the article;

determining the loss of a preset loss function according to the click through rate of a user corresponding to the training data on a corresponding article and the grade label of the behavior of the user on the article corresponding to the training data;

and updating parameters of the mixed logistic regression model based on the second-order gradient value of the loss function of the loss calculation, continuing training until a preset termination condition is met, and taking the mixed logistic regression model obtained by training as the article recommendation model.

In some embodiments, according to the behavior characteristics of the user on the article, determining the grade label of the behavior of the user on the article; wherein, the behavior characteristics of the user to the article comprise: one or more of a user search behavior characteristic of the item, a user browse behavior characteristic of the item, a user click behavior characteristic of the item, and a user purchase behavior characteristic of the item.

In some embodiments, the obtaining of the candidate item set corresponding to the user to be recommended includes: and acquiring a candidate item set corresponding to the user to be recommended by recalling the corresponding item of the preset recall strategy.

In some embodiments, the recall policy is configured to recall an item corresponding to a preset event; or the recall strategy is configured to recall the articles with similarity higher than a preset value with the articles in a preset article set, wherein the preset article set comprises the articles with direct actions of the user to be recommended; or, the recall policy is configured to recall an item having a same preset feature value as an item in the preset item set; or, the recall strategy is configured to recall the items of which the preset characteristic values are larger than a preset value in the preset item set; or, the recall strategy is configured to recall the articles with the statistical characteristic values meeting the preset conditions; or the recall strategy is configured to recall the item of which the user to be recommended has direct action; or the recall strategy is configured to recall corresponding articles in a preset area or in a preset service scene within a preset time; alternatively, the recall policy is configured to recall the captured clipboard contents for the corresponding item; or the recall strategy is configured to recall the items corresponding to the preset item set, wherein the weight of the relationship between the items and the preset item set is greater than the preset value.

In some embodiments, recommending an item to the user to be recommended according to a corresponding item for which the click through rate satisfies a preset condition includes: sorting the items in the candidate item set according to the click through rate; recommending the corresponding articles with the sequencing results meeting the preset conditions to the user.

In some embodiments, recommending an item to the user to be recommended according to a corresponding item for which the click through rate satisfies a preset condition includes: according to a preset filtering strategy, filtering corresponding articles of which the click through rate in the candidate article set meets a preset condition; recommending the filtered articles to the user to be recommended.

In some embodiments, the filtering policy is configured to filter out items of which the inventory quantity related to the delivery address of the user to be recommended is less than a preset value; or the filtering strategy is configured to filter out the articles purchased by the user to be recommended within a preset period; or, the filtering policy is configured to filter out one of two items having the same picture; or, the filtering policy is configured to filter out corresponding items whose pictures do not meet preset requirements; or the filtering strategy is configured to filter out corresponding articles of which preset characteristic attributes do not meet preset requirements, wherein the preset characteristic attributes comprise one or more of price characteristic attributes, heading term characteristic attributes and applicable time characteristic attributes.

According to further embodiments of the present disclosure, there is provided an item recommendation device including:

the acquisition module is configured to acquire a candidate item set corresponding to a user to be recommended;

the processing module is configured to input the feature data of the user to be recommended and the feature data of all articles in the candidate article set into an article recommendation model for processing to obtain the click through rate of the user to be recommended on each article in the candidate article set, wherein the article recommendation model is obtained by utilizing a mixed logistic regression model for training;

and the recommending module is configured to recommend the articles to the user to be recommended according to the corresponding articles of which the click through rate meets the preset conditions.

In some embodiments, further comprising: the output module is configured to input a training data set into a hybrid logistic regression model for processing, and output the click through rate of a user corresponding to each piece of training data in the training data set on a corresponding article, wherein the training data set comprises a plurality of pieces of training data, and each piece of training data comprises a user, feature data of an article associated with the user, and a grade label of the behavior of the user on the article; the loss determining module is configured to determine the loss of a preset loss function according to the click through rate of the user corresponding to the training data on the corresponding article and the grade label of the behavior of the user on the article corresponding to the training data; and the updating module is configured to update the parameters of the mixed logistic regression model based on the second-order gradient value of the loss function of the loss calculation, continue training until a preset termination condition is met, and take the mixed logistic regression model obtained through training as the item recommendation model.

According to still other embodiments of the present disclosure, there is provided an item recommendation apparatus, including: a memory; and a processor coupled to the memory, the processor configured to perform the item recommendation method of any of the embodiments based on instructions stored in the memory.

According to still further embodiments of the present disclosure, there is provided a non-transitory computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the item recommendation method of any of the embodiments.

Drawings

The drawings that will be used in the description of the embodiments or the related art will be briefly described below. The present disclosure can be understood more clearly from the following detailed description, which proceeds with reference to the accompanying drawings.

It is to be understood that the drawings in the following description are merely exemplary of the disclosure, and that other drawings may be derived from those drawings by one of ordinary skill in the art without undue inventive faculty.

FIG. 1 illustrates a flow diagram of a method of obtaining an item recommendation model according to some embodiments of the present disclosure.

FIG. 2 illustrates a schematic diagram of an apparatus for obtaining an item recommendation model according to some embodiments of the present disclosure.

FIG. 3 shows a schematic diagram of an apparatus for obtaining an item recommendation model according to further embodiments of the present disclosure.

FIG. 4 illustrates a flow diagram of an item recommendation method in accordance with some embodiments of the present disclosure.

Fig. 5 illustrates a schematic diagram of an item recommendation device, according to some embodiments of the present disclosure.

FIG. 6 shows a schematic view of an item recommendation device according to further embodiments of the present disclosure.

FIG. 7 illustrates a schematic diagram of an item recommendation system, according to some embodiments of the present disclosure.

Detailed Description

The technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the drawings in the embodiments of the present disclosure.

FIG. 1 illustrates a flow diagram of a method of obtaining an item recommendation model according to some embodiments of the present disclosure. The method may be performed, for example, by an apparatus that obtains an item recommendation model.

As shown in fig. 1, the method of this embodiment comprises steps 101-103.

In step 101, the training data set is input into the hybrid logistic regression model for processing, and the click through rate of the user corresponding to each piece of training data in the training data set to the corresponding article is output.

The training data set comprises a plurality of pieces of training data, and each piece of training data comprises characteristic data of a user and an article related to the user and a grade label of behavior of the user on the article. The feature data in the training data includes, for example: the unilateral characteristic data of the user, the unilateral characteristic data of the article and the bilateral characteristic data of the article by the user.

The unilateral characteristic data of the user comprises basic characteristics of the user (such as gender and age of the user), purchasing power characteristics (such as price preference of the user), purchasing habit characteristics (such as rating preference of the user, shop preference of the user, brand preference of the user) and the like; the unilateral characteristic data of the user may further include historical behavior characteristics of the user, for example, statistical data of behaviors of the user browsing items, paying attention to items, buying (i.e., adding to a shopping cart) items, placing orders and the like within a preset time window may be used. Wherein, the preference characteristic data of the user's rating is determined according to the rating of the corresponding item of the user's purchase behavior; the user shop preference can be determined according to the shop of the corresponding article of which the user generates the behaviors such as searching behavior, clicking behavior, browsing behavior or purchasing behavior, and the user shop preference can be determined according to the corresponding shop of which the user generates the behaviors such as searching behavior, clicking behavior and browsing behavior; the user brand preferences may be determined, for example, based on brand terms of the corresponding item for which the user generated a search action, a click action, a browse action, or a purchase action.

For example, the unilateral characteristics of users U1, U2 (e.g., including gender, age, price preferences) may be represented as table 1:

TABLE 1

User' s	Sex	Age (age)	Price preference
				U1	1	20	100-500
U2	2	31	300-6000

The unilateral characteristic data of the article comprises an image characteristic of the article, such as a product word of the article, a search word of the article, a brand word of the article, a modifier of the article, a price of the article and the like. The product words of the articles, the search words of the articles, the brand words of the articles and the modifiers of the articles can be subjected to numerical preprocessing and expressed in a numerical form, so that subsequent calculation is facilitated. The unilateral characteristic data of the article may further include the number of browsed articles within a preset time window, sales volume, order placement volume, evaluation volume, and the like.

For example, the unilateral characteristics of the item A, B (e.g., including price, sales, quantity evaluated) may be represented by table 2:

TABLE 2

Article with a cover	Price	Amount of sales	Number of evaluations
				A	99	230	80
B	160	100	76

The bilateral characteristic data of the user to the article comprises the characteristic data of a certain user to a certain article. For example, the characteristic may be the number of times that a user performs a search action, a click action, a browse action, and the like on an item within a preset time (for example, within a half year, a month, or an hour), or may be the preference degree of a user on an item within a preset time. Wherein, the preference degree of a certain user for a certain article in a preset time may be determined according to a score value calculated by a search behavior, a click behavior, a browse behavior, an attention behavior, a purchase adding behavior, an order placing behavior, and other behaviors of the certain user for the certain article in the preset time, for example, in one month, the user U1 may generate 1 search behavior, 4 click behaviors, 3 browse behaviors, 1 attention behavior, 2 purchase adding behaviors, and 2 order placing behaviors for the article a, assuming that the search behavior, the click behavior, the browse behavior, the attention behavior, the purchase adding behavior, and the order placing behavior are respectively assigned with a weight value of 0.1, 0.05, 0.15, 0.2, and 0.35, then calculating the score value of the user U1 for the article a in one month may be represented as 1 × 0.1+4 × 0.05+3 × 0.15+1 × 0.15+ 2.35 ═ 0.35, that is, the user U1 has a preference degree for item a of 2. Referring to the same method, for example, the preference degree of the user U2 for the article B is calculated to be 3.

For example, the bilateral characteristics (degrees of preference) of user U1 for item A, and user U2 for item B may be represented as Table 3:

TABLE 3

User' s	Article with a cover	Degree of preference
			U1	A	2
U2	B	3

The method of determining the rating label includes, for example: and determining the grade label of the behavior of the user on the article according to the behavior characteristics of the user on the article. The behavior characteristics of the user on the article may include, for example: one or more of a user search behavior characteristic of the item, a user browse behavior characteristic of the item, a user click behavior characteristic of the item, and a user purchase behavior characteristic of the item. For example, the level labels are set to five levels of 0, 1, 2,3, 5, where 0 represents exposure, 1 represents a low quality click, 2 represents a high quality click, 3 represents a low quality order, and 5 represents a high quality order. If the item A is only exposed to the user U1, and the user U1 does not perform any action on the item A, the grade label of the user U1 on the item A is 0; if the user U1 has only click behavior on the item A, for example, the number of clicks is less than 5, the rating label of the user U1 on the item A is 1, for example, the number of clicks is not less than 5, the rating label of the user U1 on the item A is 2; if the user U1 has purchase behavior for the item A, for example, the number of purchases is less than 2 and/or there is a return behavior, the user U1 has a rating label of 3 for the item A, for example, when the number of clicks on the purchase is not less than 2, the user U1 has a rating label of 5 for the item A. The values and behavior characteristics, etc. used in determining the rating labels herein are not limited to the illustrated examples.

In some embodiments, each training datum is characterized by associating a user's unilateral characteristic data (e.g., gender, age, price preference), an item's unilateral characteristic data (e.g., price, sales, quantity of evaluations), and a user's bilateral characteristic data (e.g., degree of preference) for the item, with a user's rating label for the item, resulting training data, which may be represented, for example, in table 4:

TABLE 4

After the training data set is obtained, the click through rate of the corresponding user to the corresponding article in the training data is obtained by using the mixed logistic regression model. The mixed logistic regression model in the present disclosure refers to a logistic regression model mixed with a clustering algorithm, that is, a clustering algorithm is used to cluster a training data set, and then logistic regression is performed on each cluster included in a clustering result.

In some embodiments, using the hybrid logistic regression model, the function for calculating the click through rate of the corresponding user to the corresponding item of the training data may be defined as:

wherein, P () represents the click through rate of the corresponding item by the user corresponding to the training data. x represents the characteristics of the user and its corresponding item in the training data, and y represents the value associated with the rank label in the training data. g () represents a normalization function. And m is a preset hyper-parameter and represents the fragment number of the clustering function.

The representative clustering function may be, for example, a cascade clustering function.

Representing a logistic regression function. Mu.s_jAnd w_jAre parameters of the mixed logistic regression model that need to be trained.

And

respectively represent mu_jAnd w_jThe transposing of (1).

M in the present disclosure may be set to an integer greater than or equal to 1, for example, when m > 1 (e.g., m is 2,3, …,10, …), it means that all training data are clustered into m clusters by a clustering algorithm; when m is 1, the number of slices is 1, that is, all training data are regarded as 1 slice as a result of clustering, that is, clustering is not performed.

In other embodiments, the function for calculating the click through rate of the corresponding item by the user corresponding to the training data may be further defined as:

wherein, P () represents the click through rate of the corresponding item by the user corresponding to the training data. x represents the characteristics of the user and its corresponding item in the training data, and y represents the value associated with the rank label in the training data. And m is a preset hyper-parameter and represents the fragment number of the clustering function. Mu.s_jAnd w_jAre parameters of the mixed logistic regression model that need to be trained.

And

respectively represent mu_jAnd w_jThe transposing of (1).

E is used as the base to show

Is a function of the exponent. Wherein the content of the first and second substances,

expression solution

The vector inner product with x.

E is used as the base to show

expression solution

The vector inner product with x.

In step 102, the loss of the preset loss function is determined according to the click through rate of the user corresponding to the training data on the corresponding article and the grade label of the user behavior on the article corresponding to the training data.

And under the condition that the click through rate of the user corresponding to the training data on the corresponding article is obtained, determining the loss of the preset loss function according to the click through rate of the user corresponding to the training data on the corresponding article and the grade label of the behavior of the user corresponding to the training data on the article.

In some embodiments, the preset loss function may be defined as, for example:

wherein, log (p (y)_t＝0|x_t))＝1-log(p(y_t＝1|x_t))。y_t0 means that the value associated with the rank label is 0, y_t1 indicates that the value associated with the rank label is 1. n represents the total number of pieces of training data in the training data set. For example, when a rank label is greater than 0 (e.g., rank labels are 1, 2,3, 5), the value y associated with that rank label_tIs set to 1; when the rank label is equal to 0, the value y associated with the rank label_tIs set to 0.

In the embodiment, the loss of the model is determined by calculating the error between the real grade label of the training data and the click through rate predicted by the logistic regression model, so that the accuracy of the model can be accurately evaluated, and a foundation is laid for subsequent model optimization.

In step 103, parameters of the hybrid logistic regression model are updated based on the second order gradient value of the loss function of the loss calculation, training is continued until a preset termination condition is met, and the hybrid logistic regression model obtained through training is used as an article recommendation model.

The preset termination condition is, for example, that the loss is smaller than a preset loss value, or the number of updates (i.e., the number of iterations) of the model parameter reaches a preset value.

In some embodiments, the method of updating the parameters of the hybrid logistic regression model based on the second order gradient values of the loss function of the loss calculation may include, for example:

μ_i+1＝μ_i-(G_i·g_i)·η_i,i＝0,1,…

W_i+1＝W_i-(G_i·g_i)·η_i,i＝0,1,4

wherein, mu_i+1And W_i+1Respectively representing the parameters, mu, of the corresponding mixed logistic regression model at the i +1 th iteration_iAnd W_iAnd respectively representing two parameters of the corresponding mixed logistic regression model at the ith iteration. When i is 0, μ₀And W₀Respectively, represent the initialized values of two parameters. G_iA matrix representing the determination of the second order gradient values based on the loss function can be obtained, for example, by the DFP (Davidon-Fletcher-Powell Formula) algorithm, or the BFGS (Broyden-Fletcher-Goldfarb-Shanno Formula) algorithm. In some embodiments, for example, a near sea plug matrix (e.g., denoted as H) may be calculated_i) Wherein the sea plug matrix can be expressed as

g_iRepresenting a determination of a first step value based on a loss functionThe matrix can be expressed, for example, as

η_iThe learning rate is indicated, and may be set to a constant (e.g., to 1) in advance, for example.

Compared with the method for optimizing the hybrid logistic regression model by utilizing the first-order gradient value of the loss function, the embodiment determines the optimization direction of the hybrid logistic regression model by calculating the second-order gradient value of the loss function, so that the convergence speed of the hybrid logistic regression model can be increased, and the training efficiency of the model can be improved.

Mu determined from losses based on current mixed logistic regression model_i，W_i，G_i，g_i，η_iUpdating the parameters of the mixed logistic regression model to obtain the value of mu_i+1And W_i+1. Next, the updated parameter (i.e., μ) is utilized_i+1And W_i+1) And (3) calculating new loss by the mixed logistic regression model, repeatedly executing the step of updating the parameters of the mixed logistic regression model until a preset termination condition is met, and taking the mixed logistic regression model obtained by training as an article recommendation model.

In the embodiment, the article recommendation model is obtained by using the mixed logistic regression model, and article recommendation is performed on the article recommendation model based on similar groups by using the characteristic that the mixed logistic regression model is clustered firstly and then classified, so that the recommendation effectiveness is improved; and moreover, the convergence mode of the hybrid logistic regression model is optimized by calculating the second-order gradient of the loss function, so that the hybrid logistic regression model converges faster, and the training efficiency is improved.

FIG. 2 illustrates a schematic diagram of an apparatus for obtaining an item recommendation model according to some embodiments of the present disclosure. The apparatus for obtaining the item recommendation model may be deployed in an item recommendation apparatus as shown in fig. 5 or fig. 6, for example.

As shown in fig. 2, the apparatus 200 for obtaining an item recommendation model of this embodiment includes: an output module 201, a loss determination module 202, and an update module 203.

The output module 201 is configured to input a training data set into the hybrid logistic regression model for processing, and output a click-through rate of a user corresponding to each piece of training data in the training data set to a corresponding article, where the training data set includes a plurality of pieces of training data, and each piece of training data includes a user, feature data of an article associated with the user, and a rating label of a behavior of the user on the article.

In some embodiments, according to the behavior characteristics of the user on the article, determining a grade label of the behavior of the user on the article; wherein, the behavior characteristics of the user to the article comprise: one or more of a user search behavior characteristic of the item, a user browse behavior characteristic of the item, a user click behavior characteristic of the item, and a user purchase behavior characteristic of the item.

The loss determining module 202 is configured to determine the loss of the preset loss function according to the click through rate of the user corresponding to the training data on the corresponding item and the level label of the behavior of the user corresponding to the training data on the item.

And the updating module 203 is configured to update parameters of the hybrid logistic regression model based on the second-order gradient value of the loss function of the loss calculation, continue training until a preset termination condition is met, and use the trained hybrid logistic regression model as the article recommendation model.

The article recommendation model is obtained by using the mixed logistic regression model, and article recommendation is performed on the article recommendation model based on similar groups by means of the characteristic that the mixed logistic regression model is clustered firstly and then classified, so that the recommendation effectiveness is improved; and moreover, the convergence mode of the hybrid logistic regression model is optimized by calculating the second-order gradient of the loss function, so that the hybrid logistic regression model converges faster, and the training efficiency is improved.

FIG. 3 shows a schematic diagram of an apparatus for obtaining an item recommendation model according to further embodiments of the present disclosure. The apparatus for obtaining the item recommendation model may be deployed in an item recommendation apparatus as shown in fig. 5 or fig. 6, for example.

As shown in fig. 3, the apparatus 300 for obtaining an item recommendation model of this embodiment includes: a memory 301 and a processor 302 coupled to the memory 301, the processor 302 being configured to execute a method of obtaining an item recommendation model in any of the embodiments of the present disclosure based on instructions stored in the memory 301.

The memory 301 may include, for example, a system memory, a fixed nonvolatile storage medium, and the like. The system memory stores, for example, an operating system, an application program, a Boot Loader (Boot Loader), and other programs.

The apparatus 300 for obtaining an item recommendation model may further include an input/output interface 303, a network interface 304, a storage interface 305, and the like. These

interfaces

303, 304, 305 and the memory 301 and the processor 302 may be connected by a bus 306, for example. The input/output interface 303 provides a connection interface for input/output devices such as a display, a mouse, a keyboard, and a touch screen. The network interface 304 provides a connection interface for various networking devices. The storage interface 305 provides a connection interface for external storage devices such as an SD card and a usb disk.

In some embodiments, the processor 302 is configured to input a training data set into the hybrid logistic regression model for processing, and output a click through rate of the corresponding item by the user corresponding to each piece of training data in the training data set, wherein the training data set includes a plurality of pieces of training data, each piece of training data includes feature data of one user and one item associated with the user, and a rating label of a behavior of the user on the item. And then, determining the loss of the preset loss function according to the click through rate of the user corresponding to the training data on the corresponding article and the grade label of the user behavior on the article corresponding to the training data. And finally, updating parameters of the hybrid logistic regression model based on the second-order gradient value of the loss function of the loss calculation, continuing training until a preset termination condition is met, and taking the hybrid logistic regression model obtained through training as an article recommendation model.

For example, the level label of the behavior of the user on the article can be determined according to the behavior characteristics of the user on the article; wherein, the behavior characteristics of the user to the article comprise: one or more of a user search behavior characteristic of the item, a user browse behavior characteristic of the item, a user click behavior characteristic of the item, and a user purchase behavior characteristic of the item.

In the device for obtaining the item recommendation model in the embodiment, the mixed logistic regression model is used for obtaining the item recommendation model, and the item recommendation model is recommended based on similar groups by means of the characteristic that the mixed logistic regression model is clustered firstly and then classified, so that the recommendation effectiveness is improved; and moreover, the convergence mode of the hybrid logistic regression model is optimized by calculating the second-order gradient of the loss function, so that the hybrid logistic regression model converges faster, and the training efficiency is improved.

FIG. 4 illustrates a flow diagram of an item recommendation method in accordance with some embodiments of the present disclosure. The method may be performed by a recommendation model device, for example.

As shown in fig. 4, the method of this embodiment includes

steps

401 and 403.

In step 401, a candidate item set corresponding to a user to be recommended is obtained.

The method for acquiring the candidate item set corresponding to the user to be recommended comprises the following steps: and acquiring a candidate item set corresponding to the user to be recommended by recalling the corresponding item of the preset recall strategy.

For example, the recall policy is configured to recall an item corresponding to a preset event. For example, during the new coronavirus epidemic period, a mask, a protective clothing, goggles, a disinfectant, a hand sanitizer and the like become necessary articles, and then the articles (such as the mask, the protective clothing, the goggles, the disinfectant and the hand sanitizer) related to the new coronavirus epidemic situation are taken as the articles corresponding to the preset event and recalled, so that a candidate article set is obtained. The article recalling strategy for recalling the article corresponding to the preset event can recommend the article related to the preset event according to the event (such as epidemic situation), so that the recommendation is closely related to the event which occurs in real time, and a foundation is laid for subsequently improving the recommendation effectiveness.

For another example, the recall policy is configured to recall an item having a similarity greater than a preset value with an item in a preset item set, where the preset item set includes an item having a direct behavior with the user to be recommended, and the direct behavior may include one or more of a search behavior, a browse behavior, a click behavior, a purchase behavior, and an attention behavior, for example. In addition, the preset item set may be history data extracted offline, or may be real-time data extracted online. The similarity includes an article similarity, an article brand similarity, a product word similarity corresponding to the article, and the like, for example, an article with a similarity greater than a preset value to an article in a preset article set is calculated, or an article with a similarity greater than a preset value to an article of a corresponding brand to an article in the preset article set is calculated, or a corresponding article with a similarity greater than a preset value to a product word of an article in the preset article set is calculated, and the corresponding article is recalled to obtain a candidate article set. In some embodiments, items with a similarity greater than a preset value to the items in the preset item set may be selected, for example, using a K-nearest neighbor algorithm or a deep neural network algorithm. And the recall strategy of recalling the articles with the similarity between the articles and the articles in the preset article set larger than the preset value lays a foundation for subsequently improving the accuracy of the recommendation result.

For another example, the recall policy is configured to recall items that have the same preset characteristic values as items in the preset set of items. The preset item set comprises items having direct behaviors with the user to be recommended, and the direct behaviors may comprise one or more of searching behaviors, browsing behaviors, clicking behaviors, purchasing behaviors and attention behaviors. In addition, the preset item set may be history data extracted offline, or may be real-time data extracted online. For example, the same item as the preset feature value (e.g., the brand name and the category of the item) of the item in the preset item set is recalled. The requirements and interest preferences of the user can be better mined according to the real-time behaviors or the historical behaviors of the user, and the possibility of subsequent accurate recommendation is improved.

In some embodiments, for example, an article that has the same preset characteristic value as an article in a preset article set may be selected by using a knowledge graph algorithm, where the knowledge graph algorithm constructs a relationship between the article and the article, a node represents a name of a tertiary class, a secondary class, or a tertiary class to which the article belongs (for example, the primary class is "home appliance", the secondary class is "refrigerator", and the tertiary class is "shelf refrigerator"), and an edge represents a relationship between these nodes, for example, nodes included in the constructed knowledge graph include "mobile phone", "Huacheng", "millet", and "electronic product", then a relationship between "mobile phone" and "Huacheng" is created (i.e., an edge is connected), and the relationship is named as "brand"; and establishing a relation (namely connecting one edge) between the mobile phone and the millet, wherein the relation is named as 'brand'. When a recall strategy of related recalls is utilized, if a user to be recommended only has direct behavior with the Huaqi mobile phone, at the moment, the related mobile phone with the millet brand can be recalled to the user to be recommended through the established knowledge graph. The articles with the same preset characteristic values as the articles in the preset article set are recalled, so that the demands and interest preferences of the users can be better mined, and the possibility of subsequent accurate recommendation is improved.

For another example, the recall policy is configured to recall items in the preset set of items for which the preset characteristic value is greater than the preset value. The preset item set comprises items having direct behaviors with the user to be recommended, and the direct behaviors may include one or more of a search behavior, a browse behavior, a click behavior, a purchase behavior, and an attention behavior. In addition, the preset item set may be history data extracted offline, or may be real-time data extracted online. The preset feature value may be, for example, a purchase number, i.e., an item recall in which the purchase number reaches a preset number. The preset feature value may be, for example, a weighted average score value (i.e., a degree of preference of the user for the item) calculated according to the behavior of the user for the item (e.g., search behavior, browse behavior, click behavior, purchase behavior, attention behavior), and the corresponding item with the weighted average score value being greater than the preset value is recalled. According to the high-frequency real-time behavior or historical behavior of the user, the requirement of the user can be better mined, and the possibility of subsequent accurate recommendation is improved.

For another example, the recall policy is configured to recall items for which the statistical characteristic value satisfies a preset condition. For example, recall all corresponding items for which the user never acted (i.e., the statistical item characteristic attribute was empty), i.e., recall items that were newly shelved (also referred to as item cold start); for another example, the articles with the purchase times larger than the preset value in the preset area or within the preset time are recalled (i.e. hot door article recall); for another example, for a new user, the corresponding item (also called user cold start) of the direct action of the user having the same or similar feature data (i.e. the same or similar age, or the same or similar payroll income) as the new user is recalled according to the unilateral feature data (e.g. age, payroll income, etc.) of the new user, wherein the statistical feature attribute is the same or similar age, or the same or similar payroll income. Aiming at the recall strategy under special scenes (cold starting of articles, cold starting of users and the like), the follow-up recommendation steps are free of invalidation under the special scenes, and the recommendation effectiveness is improved.

As another example, the recall policy may be configured to recall an item for which the user to be recommended has direct action. The direct behavior may include one or more of a search behavior, a browse behavior, a click behavior, a purchase behavior, and an attention behavior, for example. In addition, the behaviors may be historical behaviors or real-time behaviors extracted online. According to the real-time behavior or the historical behavior of the user, the demand and interest preference of the user can be better mined, and the possibility of subsequent accurate recommendation is improved.

For another example, the recall policy is configured to recall the corresponding item in a preset service scenario in a preset area or within a preset time. The preset service scenario may include, for example, various promotional activities (e.g., second killing activities), and the like, that is, corresponding items participating in a certain service scenario are recalled. The items can be actively selected for the user, so that new requirements or interest preferences of the user can be mined, and the possibility of subsequent accurate recommendation is improved.

As another example, the recall policy is configured to recall the captured clipboard contents for the corresponding item. Namely, according to the clipboard content of the user to be recommended, the corresponding article in the clipboard content is determined to be recalled. Corresponding articles can be actively selected for the user according to the clipboard content of the user, and before the clipboard content is pasted into the recommendation platform by the user, even under the condition that the user forgets to paste the clipboard content, the corresponding articles can still be actively selected for the user, so that the user experience is improved.

For another example, the recall policy is configured to recall an item whose weight of relationship with an item in the preset set of items is greater than that corresponding to the preset value. In some embodiments, by constructing the graph neural network, the graph neural network is used to recall the items corresponding to the relationship weights of the items in the preset item set, which are greater than the preset value. Firstly, according to the feature data of the articles in the preset article set, namely the feature data of the articles of which all users have direct behaviors on the articles in the preset article set, a graph neural network related to the articles is constructed, wherein nodes in the graph neural network represent the articles, and edges represent the relationship between the articles. For example, assuming that the preset item set includes A, B, C items, and the click action counted 100 times is that item a is clicked and then item B is clicked, the weights of the directional edges of items a to B (i.e., the relationship values of items a to B) are set to 100, for example, and referring to the same method, the weights of the directional edges of items B to a are 200, the weights of the directional edges of items B to C are 20, and the weights of the directional edges of items C to B are 10, for example. Assuming that the preset value (i.e., the threshold) is set to be 50, the edges with the weight less than 50 are pruned by using a pruning method, and then, in the finally pruned graph neural network, the articles corresponding to the nodes with the edges connected are the article a and the article B, and then the article a and the article B are recalled. Corresponding articles can be recalled by constructing a graph neural network, articles with implicit relations can be found, corresponding associated articles can be recommended to users better, and user experience is improved.

In some embodiments, for example, multiple recall policies may also be adopted, for example, multiple recall policies may result in a union of multiple candidate item sets as the final candidate item set. Through the combination of various recalling strategies, a proper amount of candidate articles can be obtained, the problems that the number of recalled articles is too small and the number of articles recommended to a user is too small can be avoided, and the conversion rate of the recommendation result is improved.

In the above embodiment, the candidate item set corresponding to the user to be recommended is obtained through a plurality of recall strategies in different manners, and the click through rate is subsequently calculated for all items in the candidate item set, so that the recommendation is completed. The method and the device can avoid the process of calculating and even sequencing a large number of articles under the condition of not adopting a recall strategy, reduce the calculated amount, realize the personalized recommendation aiming at different users to be recommended and improve the effectiveness of the recommendation result.

In step 402, the feature data of the user to be recommended and the feature data of all the articles in the candidate article set are input into the article recommendation model for processing, so as to obtain the click through rate of the user to be recommended on each article in the candidate article set.

The item recommendation model is obtained by a method for obtaining the item recommendation model in any embodiment of the disclosure.

In step 403, recommending the item for the user to be recommended according to the corresponding item whose click through rate meets the preset condition.

In some embodiments, the items in the candidate item set are ordered according to click through rate; recommending the corresponding articles of which the sorting results meet the preset conditions to the user, for example, recommending the articles with the top 10 of the click through rate sorting to the user.

In some embodiments, recommending an item by a user to be recommended according to a corresponding item for which the click through rate meets a preset condition includes: according to a preset filtering strategy, filtering corresponding articles of which the click through rate in the candidate article set meets a preset condition; and recommending the filtered articles to the user to be recommended.

For example, the filtering policy is configured to filter out items whose inventory number associated with the shipping address of the user to be recommended is less than a preset value (e.g., set to 1). By filtering out the articles with insufficient inventory, the problem that the user cannot place an order or purchase recommended articles is avoided, the user experience is improved, and the effectiveness of the recommendation result is improved.

For another example, the filtering policy is configured to filter out items that have been purchased by the user to be recommended within a preset period. The conversion of the recommended results can be improved.

As another example, the filtering policy is configured to filter out one of two items having the same picture. Repeated recommendation can be avoided, user experience is improved, and the conversion rate of recommendation results is improved.

As another example, the filtering policy is configured to filter out corresponding items for which the picture does not meet preset requirements. For example, filtering out articles with poor picture quality, or articles containing sensitive pictures. And the effectiveness of the recommendation result is improved.

For another example, the filtering policy is configured to filter out corresponding items for which the preset characteristic attribute does not meet the preset requirement, wherein the preset characteristic attribute includes one or more of a price characteristic attribute, a heading term characteristic attribute, and an applicable time characteristic attribute. For example, articles whose price does not satisfy the price preference of the user to be recommended are filtered out, or articles containing sensitive subject words are filtered out, or articles which do not conform to the season, such as a warmer, warm clothes, and the like, are filtered out in summer. The effectiveness of the recommendation results can be improved.

In the embodiment, the article recommendation model is obtained by using the mixed logistic regression model, and article recommendation is performed on the article recommendation model based on similar groups by using the characteristic that the mixed logistic regression model is clustered and then classified, so that the recommendation effectiveness is improved. And a recall strategy and/or a filtering strategy are/is adopted, so that the recommendation effectiveness is further improved, the user experience is improved, and the conversion rate of the recommendation result is improved.

As shown in fig. 5, the item recommendation device 500 of this embodiment includes: an acquisition module 501, a processing module 502, and a recommendation module 503.

The obtaining module 501 is configured to obtain a candidate item set corresponding to a user to be recommended.

In some embodiments, the obtaining module 501 is configured to obtain a candidate item set corresponding to the user to be recommended by recalling an item corresponding to a preset recall policy.

The recall strategy is configured to recall the item corresponding to the preset event; or the recall strategy is configured to recall the articles with similarity greater than a preset value with the articles in a preset article set, wherein the preset article set comprises the articles with direct actions with the user to be recommended; or, the recall strategy is configured to recall the items with the same preset characteristic values as the items in the preset item set; or the recall strategy is configured to recall the articles with the preset characteristic values larger than the preset value in the preset article set; or the recall strategy is configured to recall the articles with the statistical characteristic values meeting the preset conditions; or the recall strategy is configured to recall the item with the direct action of the user to be recommended; or the recall strategy is configured to recall corresponding articles in a preset area or in a preset service scene within preset time; alternatively, the recall policy is configured to recall the captured clipboard contents for the corresponding item; or the recall strategy is configured to recall the items corresponding to the preset value, wherein the relation weight of the items in the preset item set is greater than the preset value.

And acquiring a candidate item set corresponding to the user to be recommended through a plurality of recall strategies in different modes, and calculating click through rates aiming at all items of the candidate item set subsequently to finish the recommendation. The method and the device can avoid the process of calculating and even sequencing a large number of articles under the condition of not adopting a recall strategy, reduce the calculated amount, realize the personalized recommendation aiming at different users to be recommended and improve the effectiveness of the recommendation result.

The processing module 502 is configured to input the feature data of the user to be recommended and the feature data of all the articles in the candidate article set into the article recommendation model for processing, so as to obtain the click through rate of each article in the candidate article set by the user to be recommended. The item recommendation model is obtained by a method for obtaining the item recommendation model in any embodiment of the disclosure.

The recommending module 503 is configured to recommend an item to the user to be recommended according to the corresponding item whose click through rate meets a preset condition.

In some embodiments, the recommendation module 503 is configured to rank the items in the candidate item set according to the click through rate; recommending the corresponding articles with the sequencing results meeting the preset conditions to the user.

In some embodiments, the recommending module 503 is configured to perform filtering processing on the corresponding items in the candidate item set, where the click through rate of the corresponding items meets a preset condition, according to a preset filtering policy; and recommending the filtered articles to the user to be recommended. Wherein the filtering strategy is configured to filter out the articles of which the stock quantity related to the receiving address of the user to be recommended is less than a preset value; or the filtering strategy is configured to filter out the articles purchased by the user to be recommended in a preset period; alternatively, the filtering policy is configured to filter out one of two items having the same picture; or the filtering strategy is configured to filter out corresponding articles of which the pictures do not meet preset requirements; or the filtering strategy is configured to filter out corresponding articles of which the preset characteristic attributes do not meet the preset requirements, wherein the preset characteristic attributes comprise one or more of price characteristic attributes, heading term characteristic attributes and applicable time characteristic attributes.

As shown in fig. 6, the item recommendation apparatus 600 of this embodiment includes: a memory 601 and a processor 602 coupled to the memory 601, the processor 602 configured to execute the item recommendation method in any of the embodiments of the present disclosure based on instructions stored in the memory 601.

The memory 601 may include, for example, a system memory, a fixed nonvolatile storage medium, and the like. The system memory stores, for example, an operating system, an application program, a Boot Loader (Boot Loader), and other programs.

The item recommendation device 600 may further include an input-output interface 603, a network interface 604, a storage interface 605, and the like. These

interfaces

603, 604, 605 and the memory 601 and the processor 602 may be connected via a bus 606, for example. The input/output interface 603 provides a connection interface for input/output devices such as a display, a mouse, a keyboard, and a touch screen. The network interface 604 provides a connection interface for various networking devices. The storage interface 605 provides a connection interface for external storage devices such as an SD card and a usb disk.

As shown in fig. 7, the item recommendation system 700 of this embodiment includes: an apparatus 701 for obtaining an item recommendation model, and an item recommendation apparatus 702.

The apparatus 701 for obtaining an item recommendation model may be, for example, an apparatus for obtaining an item recommendation model in any of the embodiments of the present disclosure (e.g., the apparatus 200 for obtaining an item recommendation model shown in fig. 2 or the apparatus 300 for obtaining an item recommendation model shown in fig. 3), and performs the method for obtaining an item recommendation model in any of the embodiments of the present disclosure. The item recommendation device 702 may be, for example, an item recommendation device in any embodiment of the present disclosure (e.g., the item recommendation device 500 shown in fig. 5 or the item recommendation device 600 shown in fig. 6), and performs an item recommendation method in any embodiment of the present disclosure.

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

The present disclosure is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the disclosure. 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.

The above description is only exemplary of the present disclosure and is not intended to limit the present disclosure, which is to be construed in any way as imposing limitations thereon, such as the appended claims, and all changes and equivalents that fall within the true spirit and scope of the present disclosure.

Claims

1. An item recommendation method comprising:

acquiring a candidate item set corresponding to a user to be recommended;

2. The item recommendation method of claim 1, wherein training the item recommendation model using a hybrid logistic regression model comprises:

3. The item recommendation method according to claim 2,

determining a grade label of the behavior of the user on the article according to the behavior characteristics of the user on the article;

wherein, the behavior characteristics of the user to the article comprise: one or more of a user search behavior characteristic of the item, a user browse behavior characteristic of the item, a user click behavior characteristic of the item, and a user purchase behavior characteristic of the item.

4. The item recommendation method according to claim 1, wherein the obtaining of the candidate item set corresponding to the user to be recommended comprises:

and acquiring a candidate item set corresponding to the user to be recommended by recalling the corresponding item of the preset recall strategy.

5. The item recommendation method according to claim 4,

the recall strategy is configured to recall the item corresponding to the preset event;

or the recall strategy is configured to recall the articles with similarity higher than a preset value with the articles in a preset article set, wherein the preset article set comprises the articles with direct actions of the user to be recommended;

or, the recall policy is configured to recall an item having a same preset feature value as an item in the preset item set;

or, the recall strategy is configured to recall the items of which the preset characteristic values are larger than a preset value in the preset item set;

or, the recall strategy is configured to recall the articles with the statistical characteristic values meeting the preset conditions;

or the recall strategy is configured to recall the item of which the user to be recommended has direct action;

or the recall strategy is configured to recall corresponding articles in a preset area or in a preset service scene within a preset time;

alternatively, the recall policy is configured to recall the captured clipboard contents for the corresponding item;

or the recall strategy is configured to recall the items corresponding to the preset item set, wherein the weight of the relationship between the items and the preset item set is greater than the preset value.

6. The item recommendation method according to claim 1, wherein recommending items for the user to be recommended according to the corresponding items of which click-through rates meet preset conditions comprises:

sorting the items in the candidate item set according to the click through rate;

recommending the corresponding articles with the sequencing results meeting the preset conditions to the user.

7. The item recommendation method according to claim 1, wherein recommending items for the user to be recommended according to the corresponding items of which click-through rates meet preset conditions comprises:

according to a preset filtering strategy, filtering corresponding articles of which the click through rate in the candidate article set meets a preset condition;

recommending the filtered articles to the user to be recommended.

8. The item recommendation method according to claim 7,

the filtering strategy is configured to filter out the articles of which the stock quantity related to the receiving address of the user to be recommended is less than a preset value;

or the filtering strategy is configured to filter out the articles purchased by the user to be recommended within a preset period;

or, the filtering policy is configured to filter out one of two items having the same picture;

or, the filtering policy is configured to filter out corresponding items whose pictures do not meet preset requirements;

or the filtering strategy is configured to filter out corresponding articles of which preset characteristic attributes do not meet preset requirements, wherein the preset characteristic attributes comprise one or more of price characteristic attributes, heading term characteristic attributes and applicable time characteristic attributes.

9. An item recommendation device comprising:

10. The item recommendation device of claim 9, further comprising:

the output module is configured to input a training data set into a hybrid logistic regression model for processing, and output the click through rate of a user corresponding to each piece of training data in the training data set on a corresponding article, wherein the training data set comprises a plurality of pieces of training data, and each piece of training data comprises a user, feature data of an article associated with the user, and a grade label of the behavior of the user on the article;

the loss determining module is configured to determine the loss of a preset loss function according to the click through rate of the user corresponding to the training data on the corresponding article and the grade label of the behavior of the user on the article corresponding to the training data;

and the updating module is configured to update the parameters of the mixed logistic regression model based on the second-order gradient value of the loss function of the loss calculation, continue training until a preset termination condition is met, and take the mixed logistic regression model obtained through training as the item recommendation model.

11. An item recommendation device, comprising:

a memory; and

a processor coupled to the memory, the processor configured to perform the item recommendation method of any of claims 1-8 based on instructions stored in the memory.

12. A non-transitory computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the item recommendation method of any one of claims 1-8.