WO2020221022A1

WO2020221022A1 - Service object recommendation method

Info

Publication number: WO2020221022A1
Application number: PCT/CN2020/085254
Authority: WO
Inventors: 彭艺; 李楠; 刘家豪; 王超; 谢淼; 王寅
Original assignee: 阿里巴巴集团控股有限公司
Priority date: 2019-04-28
Filing date: 2020-04-17
Publication date: 2020-11-05
Also published as: CN111861605A

Abstract

Disclosed is a service object recommendation method, comprising: by means of a first parameter and a second parameter comprised in a service object value evaluation model, determining, according to first feature data of a candidate service object, a score of the candidate service object; determining, according to the score, a service object set recommended to a user; and pushing the service object set to a client. By means of the use of such a processing means, the service object value evaluation model is divided into a parameterization item and a non-parameterization item, and the values of service objects with unknown feature distributions are evaluated in combination with a parameter model and a non-parameter model; the non-parameterization item causes the models to fit the unknown service object feature distributions, and a gap between the parameter model and a real environment can thus be continuously reduced; therefore, the value accuracy of the service objects with the unknown feature distributions can be effectively improved, such that the single recommendation time step loss can be converged, and the accuracy of service object recommendation can then be improved step by step.

Description

Business object recommendation method

This application claims the priority of the Chinese patent application with the application number 201910350833.3 and the invention title of "Business Object Recommendation Method" filed on April 28, 2019, the entire content of which is incorporated into this application by reference.

Technical field

This application relates to the field of data processing technology, and specifically relates to a method for recommending business objects.

Background technique

The recommendation system is to use e-commerce websites to provide customers with product information and suggestions, help users decide what products should be purchased, and simulate sales staff to help customers complete the purchase process. Product cold start refers to the recommendation of products that are lacking in user behavior. Because of the lack of a data basis for recommendation in the case of product cold start, cold start has become a classic problem in the recommendation system.

At present, a typical cold-start method of a recommendation system is based on the upper limit of the confidence limit of the multi-armed gambling machine. The processing process includes the following steps: 1) Data collection is performed to construct a product data set, and the product data in the product data set Perform preprocessing to obtain the explicit features of the products in a standardized format; construct the invisible features of the commodities based on the explicit features of the commodities, based on the latent Dirichlet algorithm, set the output invisible feature dimensions, and relabel the commodities; 2) Construct candidates based on the commodity data set Commodity set: cluster the commodity data set according to the invisible characteristics of the commodity, and cluster the commodities. The commodities in the same cluster have similar properties, and the commodities in different clusters are quite different. From each cluster A product is randomly selected to construct a candidate product set; 3) The selection of the best product from the candidate product set is regarded as a multi-armed gambling machine problem, and the product with the highest estimated score is calculated based on the upper bound algorithm of the confidence interval as the recommended product; 4) After recommending the product with the highest score in the candidate product set to the user, update the user characteristics and weight parameters according to the feedback.

However, in the process of implementing the present invention, the inventor found that the technical solution has at least the following problems: because the above solution requires that the product has sufficient user behavior characteristic data, that is, the user behavior characteristic data must be large enough to correctly evaluate the product. Therefore, it is only suitable for the application scenarios of personalized product recommendation for new users based on parameterized modeling of product value. However, in practical applications, the distribution of more product characteristics is unknown, that is, some products do not have sufficient user behavior characteristic data, and it is impossible to correctly evaluate the value of the product based on a parameterized model constructed based on user behavior data. For example, in the second-hand commodity recommendation scenario, because the new products in the second-hand transaction products account for a relatively large proportion and most of them are single products (orphan products), the corresponding transaction cycle is short, which leads to short exposure time, and because of the exposure flow on the products. The distribution is relatively uniform, so the user behavior characteristic data that can be collected for second-hand goods will be relatively insufficient, that is, the value of second-hand goods cannot be determined based on the user behavior characteristics of a limited dimension, and cold-start recommendations for products with unknown commodity characteristic distribution In application scenarios, the above solution cannot correctly evaluate the value of the product, which causes the recommendation result to fail to gradually converge, and thus it is impossible to screen out the products that users are interested in.

Summary of the invention

This application provides a method for recommending business objects to solve the problem in the prior art that the products of interest to users cannot be filtered out in the cold-start scenario of products.

This application provides a method for recommending business objects, including:

According to the first parameter and the second parameter included in the business object value evaluation model, the score of the candidate business object is determined according to the first characteristic data of the candidate business object; the first characteristic data includes user behavior characteristic data; the first The parameter includes a weight parameter related to the first characteristic data, and the second parameter includes an unknown second characteristic data distribution parameter;

Determining a set of business objects recommended to the user according to the score;

Push the set of business objects to the client.

Optionally, the business object includes:

A business object whose business object value is determined jointly by the first characteristic data and the second characteristic data, and/or a business object whose business object value is determined by the first characteristic data.

Optional, also includes:

Acquiring first user feedback information for the set of business objects;

Updating the first parameter and the second parameter according to the first user feedback information.

Optionally, the first user feedback information includes operation behavior information and browsing behavior information of the user on the business object.

Optionally, the updating the first parameter and the second parameter according to the first user feedback information includes:

Updating the user behavior characteristic data according to the operation behavior information;

Generating training samples according to the updated user behavior characteristic data and the browsing behavior information;

The first parameter and the second parameter are updated according to the generated training samples and historical samples.

Optional, also includes:

Judging whether the model converges according to the first parameter and the second parameter before the update, and the first parameter and the second parameter after the update;

If the above judgment result is yes, stop updating the model.

Optional, also includes:

If the above judgment result is no, continue to update the model.

Optional, also includes:

Initialize the first parameter and the second parameter.

Optionally, the initializing the first parameter and the second parameter includes:

Show users at least one candidate business object;

Acquiring second user feedback information for the at least one candidate business object;

Generating training samples of the model according to the second user feedback information;

According to the training sample, the first parameter and the second parameter to be initialized are determined.

Optionally, the first parameter includes: a parameter of a linear machine learning model or a parameter of a nonlinear machine learning model;

The second parameter includes: statistical items related to the Gaussian process, statistical items related to the Dirichlet process, and statistical items related to the infinite-dimensional distribution.

Optionally, the business objects include: commodity objects, video objects, and news objects.

The present application also provides a computer-readable storage medium having instructions stored in the computer-readable storage medium, which when run on a computer, cause the computer to execute the above-mentioned various methods.

The present application also provides a computer program product including instructions, which when run on a computer, causes the computer to execute the above-mentioned various methods.

Compared with the prior art, this application has the following advantages:

In the business object recommendation method provided by the embodiment of the application, the score of the candidate business object is determined according to the first characteristic data of the candidate business object through the first parameter and the second parameter included in the business object value evaluation model; the score is determined according to the score The set of business objects recommended to the user; the set of business objects are sent back to the client; this processing method allows the business object value evaluation model to be divided into parameterized items and non-parameterized items, and comprehensive parameter models and non-parametric models Evaluate the value of business objects with unknown feature distributions. Because non-parametric terms enable the model to fit the feature distribution of unknown business objects, it can continuously narrow the gap between the parametric model and the real environment; therefore, it can effectively improve the performance of business objects with unknown feature distributions. Value accuracy, so that the single recommendation time step loss can converge, and then the accuracy of business object recommendation can be gradually improved.

Description of the drawings

FIG. 1 is a flowchart of an embodiment of a method for recommending a business object provided by this application;

FIG. 2 is a specific flowchart of an embodiment of the business object recommendation method provided by the present application;

FIG. 3 is a specific flowchart of an embodiment of the business object recommendation method provided by the present application;

FIG. 4 is a specific flowchart of an embodiment of the business object recommendation method provided by the present application;

FIG. 5 is a specific flowchart of an embodiment of a method for recommending business objects provided by this application;

Fig. 6 is a specific flowchart of an embodiment of a method for recommending a business object provided by the present application.

Detailed ways

In the following description, many specific details are explained in order to fully understand this application. However, this application can be implemented in many other ways different from those described here, and those skilled in the art can make similar promotion without violating the connotation of this application. Therefore, this application is not limited by the specific implementation disclosed below.

The business object recommendation technical solution provided by the embodiments of this application has the technical idea of dividing the business object value evaluation model into parameterized items and non-parameterized items, and comprehensive parameter models and non-parametric models to evaluate the value of business objects whose characteristic distribution is unknown , And then determine the business object recommended to the user based on the value. Since the non-parametric term allows the model to fit the unknown product feature distribution, it can continuously narrow the gap between the parameter model and the real environment, so it can effectively improve the value accuracy of business objects with unknown feature distribution, so that the single recommended time step loss can be Convergence can gradually improve the accuracy of business object recommendation.

First embodiment

Please refer to FIG. 1, which is a flowchart of an embodiment of a method for recommending a business object provided by this application. The execution body of the method includes a device for recommending a business object. A business object recommendation method provided by this application includes:

Step S101: Determine the score of the candidate business object according to the feature data of the candidate business object through the first parameter and the second parameter included in the business object value evaluation model.

The recommendation device is usually deployed on a server, but is not limited to a server, and can also be any device that can implement the business object recommendation method. The equipment equipped with the recommendation device can actively start the recommendation device to perform business object recommendation processing, or submit a business object recommendation request according to the user client, provide the user with a business object recommendation service, and according to the user’s recommendation results We continuously optimize the value evaluation model of business objects, so as to gradually improve the evaluation accuracy of business object scores.

In this embodiment, the recommending device first receives the business object recommendation request sent by the client. The client terminal includes, but is not limited to, mobile communication equipment, that is, a mobile phone or a smart phone in general, and also includes terminal equipment such as personal computers, PADs, and iPads.

From the perspective of business object categories, the business objects include but are not limited to: commodity objects, video objects, news objects, and so on. For ease of description, the method provided in the embodiments of the present application will be described below by taking a commodity object as an example.

From the perspective of application scenarios, the application scenario of the method provided in the embodiments of the present application may be a recommendation scenario of a business object whose business object value is determined by the first feature data and the second feature data. The first feature data refers to feature data whose data distribution is known, which can be artificially set features, including but not limited to feature data related to user behavior (referred to as user behavior feature data), such as a product being bought in one day The number of times a user clicks on the home user, the number of buyer users who have collected goods in seven days, the number of buyer users who communicated with seller users who sold goods, etc.; the first characteristic data may also include other characteristic data that has nothing to do with user behavior, Such as commodity price, commodity classification, seller location, etc. The second feature data refers to feature data whose data distribution is unknown, that is, features that cannot be clearly expressed in the form of feature data. This application abbreviates this scenario as a scenario with unknown data distribution, also known as a commodity cold start scenario. For example, second-hand commodities sold on a second-hand commodity trading platform account for a large proportion of new products and most of them are orphans (single Product), the corresponding transaction cycle is short, which leads to short product exposure time. At the same time, because the exposure flow is more evenly distributed on multiple products, the behavior data that can be collected by the product will be relatively insufficient, that is, according to these values Relatively insufficient user behavior data cannot accurately assess the value of goods, and second-hand product recommendation scenarios are scenarios with unknown data distribution.

The application scenarios of the method provided in the embodiments of this application are not limited to scenarios with unknown data distribution. The methods provided in this application can also be used in other scenarios where business objects need to be recommended to users. For example, the value of business objects can be described as The recommended scenario of the business object directly determined by the first feature data, which is referred to as a linear scenario in this application. For example, for non-second-hand commodities sold on ordinary commodity trading platforms, since the transaction commodities are ordinary commodities with a certain amount of inventory, the corresponding transaction cycle is longer, so the commodity exposure time is longer, so the behavior data that can be collected by the commodity will be quite sufficient That is to say, according to the user behavior data with sufficient data volume, the value of the product can be more accurately evaluated, so the common product recommendation scenario is a linear scenario.

In addition, the method provided in the embodiments of this application can also be applied to application scenarios where a linear scenario is combined with a scenario where data distribution is unknown. In other words, the method provided in this application can be used in scenarios where similar business objects are recommended. Recommendations of business objects.

In this embodiment, the target user opens a mobile App (such as a second-hand commodity trading App, etc.) in a smart phone, and the App sends a business object recommendation request to the server. The business object recommendation request may include information such as a user ID. In this case, the server can obtain user information according to the user ID, and recommend a business object that meets the user's interest characteristics through the method provided in the embodiment of the present application. The business object recommendation request may also not include the user identification. In this case, the method provided in the embodiment of the present application can recommend business objects irrelevant to the user's interest characteristics, that is, non-personalized recommendation business objects.

The business object value evaluation model refers to a model for determining the value of a business object based on the characteristics of the business object (including known first characteristic data and unknown second characteristic data). The input data of the model includes the first feature data whose distribution of the business object is known, and the model output data includes the score of the business object, and the score can be used as a basis for recommendation of the business object.

The business object value evaluation model includes a first parameter and a second parameter, the first parameter includes a weight parameter related to the first feature data whose distribution is known, and the second parameter includes a distribution related to the business object. Statistical parameters related to the unknown second feature data that can reflect the difference between the real environment and the parameter model. By adopting this processing method, non-parametric estimation is introduced, and the gap between the parameter model and the real environment can be continuously narrowed. Therefore, the accuracy of value evaluation can be effectively improved, the single time step loss can be converged, and the recommendation accuracy can be effectively improved.

The first parameter includes a weight parameter related to the first feature data. In this embodiment, the first parameter is called a parameter item, and the model corresponding to the first parameter is called a parameter model. The parameter model can be a linear machine learning model, such as linear UCB or linear Thompson Sampling; the parameter model can also be a nonlinear machine learning model, such as Mirror Descent, gradient descent (Gradient Descent) algorithm, and so on.

The second parameter includes a statistical parameter that reflects the gap between the parameter model and the real environment. In this embodiment, the second parameter is called a non-parametric item, and the model corresponding to the model of the second parameter is called a non-parametric model. Non-parametric models can be Gaussian processes, Dirichlet processes, and non-parametric methods corresponding to infinite-dimensional distributions, such as Kernel Regression, Decision Trees, and so on.

In this embodiment, the parameter based on the linear UCB method is used as the first parameter, and the parameter based on the Gaussian process is used as the second parameter. For example, for a commodity e, calculate the confidence interval radius α of the non-parameter item, and combine the parameter item radius β to obtain the upper bound U of the semi-parametric confidence interval, which is the score of the commodity. The mathematical expression formula of the process of determining the score is given below to intuitively explain the method of determining the score.

In this embodiment, the business object is a commodity object, and L (such as 24) commodity objects are recommended to the user. For a commodity object e, the confidence interval radius α of the non-parameter items of the commodity object is calculated by the following formula:

Among them, t represents the t-th business object recommendation; T _t-1(e) represents the total number of recommendations of the business object e at the t-1th recommendation time, and α _t-1 (e) represents the product object e at the t-th The radius of the confidence interval of the non-parameter at the moment of 1 recommendation

At the same time, combined with the parameter term radius β, the upper bound U of the confidence interval is obtained by the following formula:

Among them, U _t (e) represents the upper bound of the confidence interval of business object e at the t-th recommendation time, that is, the score of business object e (business object value);

Represents the non-parametric statistics of business object e at the t-1th recommendation time;

Represents the parameter item statistics of business object e at the t-1th recommendation time; γ _t-1 (e) represents the sum of the radius of business object e at the t-1th recommendation time; ΔX _t,e represents the business object e at The difference between the first feature data at the t-th recommendation time and the first feature data estimate (such as the average value) of the business object e at the t-th recommendation time.

Step S103: Determine the set of business objects recommended to the user according to the score.

The score of the business object serves as a basis for recommendation of the business object, and the set of business objects recommended to the user can be determined according to the score. In this embodiment, the value score of a commodity is the upper bound of the confidence interval of the commodity. Since the upper bounds of the confidence interval of different commodities are not independent, this embodiment is based on the upper bound of the confidence interval of the commodity and is calculated according to the offline combination optimization algorithm The best combination of products. The mathematical expression of the process of determining the set of business objects is as follows:

Wherein, A _t represents a t-th set of business objects recommended time, k represents the number of elements of the set of business objects, business objects which k is determined based on all the U-score _t e business objects in the t-th time recommended.

A type of optimization problem that finds the optimal solution in a finite set of feasible solutions is called combinatorial optimization problem, which is an important branch of operations research. Combination optimization algorithm (optimal combination algorithm) is a type of problem that seeks extreme values in a discrete state. Since the combinatorial optimization algorithm is a relatively mature existing technology, it will not be repeated here.

In another example, the value scores of different commodities are independent of each other. Therefore, according to the order of commodity scores from high to low, a preset number of high-ranking commodities can be selected as a combination of commodities recommended to users.

Step S105: Push the set of business objects to the client.

The server sends the determined business object back to the client, so that the client can display the business object to the target user for viewing, so as to help the user find the business object of interest, thereby promoting the transaction rate of the business object.

The method provided in the embodiments of the present application may be a method of updating the business object value evaluation model online or offline, and determining the business object score through the updated model, and then determining the recommended business object based on the score.

Please refer to FIG. 2, which is a specific flowchart of an embodiment of a method for recommending business objects provided by this application. In this embodiment, the model is updated online, and the method further includes the following steps:

Step S201: Obtain first user feedback information for the business object set.

The first user feedback information may include operation behavior information of the user on the business object pushed by the recommendation system, and may also include browsing behavior information. The operation behavior information includes, but is not limited to, the following information: which business objects the user clicks (such as viewing the detailed information of the product), which business objects the user saves, the user stay time, and so on. The browsing behavior information refers to which business objects the user has browsed. For example, 20 business objects are shown to the user and displayed in 2 pages, with 10 business objects displayed on each page. In this case, the user may only view Since the business objects displayed on page 1 are displayed, the browsing behavior information may only include the identities of these 10 business objects.

During specific implementation, the user can perform operations such as clicking, bookmarking, etc., on the business objects recommended by the system through the client, and these operation information will be collected by the server through the network to form the first user feedback information.

The mathematical expression of the first user feedback information includes: O _t and W _t , where O _t represents the business object information that the user has browsed at the t-th recommendation time, and W _t represents the user's operation at the t-th recommendation time (such as clicking , Favorites, etc.) business object information.

Step S203: Update the first parameter and the second parameter according to the first user feedback information.

After the first user feedback information is obtained, since the information reflects changes in the characteristic data of the business object related to the user behavior, the model can be updated according to the first user feedback information.

As shown in FIG. 3, in this embodiment, step S203 may include the following specific sub-steps:

Step S2031: Update the user behavior characteristic data according to the operation behavior information.

For example, 20 product objects are shown to the user, the user clicks on 3 of the product objects, and one of the product objects is bookmarked. In this case, the number of clicks by users of these 3 product objects in a day can be cumulatively added. 1. Add 1 cumulatively to the number of favorite users of one of the commodity objects.

Step S2033: Generate training samples according to the updated user behavior characteristic data and the browsing behavior information.

For example, every time the user is shown 20 recommended product objects, for a certain recommendation result, the user only browses the first 10 product objects and clicks on 3 of them to view the product details; in this case, you can Generate 10 new training samples, including: training samples corresponding to each browsed commodity object, the training samples including the user behavior characteristic data of the business object and the corresponding relationship with the sample label information. In this embodiment, the training samples corresponding to 3 commodity objects include updated user behavior characteristic data, and the sample label information is 1, indicating that the commodity object has been clicked by the user; the training samples corresponding to the other 7 commodity objects can be It is the user behavior characteristic data at the last recommendation moment, and the sample label information is 0, indicating that the commodity object was not clicked by the user at the current recommendation moment.

Step S2035: Update the first parameter and the second parameter according to the generated training samples and historical samples.

After the newly-added training samples of the model are generated, the newly-added samples and the historical samples of the model can be combined to update the first parameter and the second parameter. Update the first parameter and the second parameter, that is, update the model. After the model is updated, the updated model can be used to process the next business object recommendation request submitted through the client. Gradually improve the value accuracy of business objects, and then improve the accuracy of business object recommendations.

The mathematical expression formula of the process of updating the first parameter and the second parameter is given below to intuitively explain the model update processing method.

In this embodiment, the process of updating the parameter item (the first parameter) can be expressed as follows:

Among them, X _t represents the first feature data set at the t-th recommendation time (referred to as the newly added first feature data for short), and X _t-1 represents the first feature data set at the t-1th recommendation time (referred to as the historical The first characteristic data),

Represents the updated first feature data of the first business object viewed by the user at the t-th recommendation time, and the business object

The difference between the first feature data estimates (such as the average) at the tth recommendation time,

Represents the updated first feature data of the O _t business object viewed by the user at the t recommendation time, and the business object

The difference between the first feature data estimates (such as the average value) at the t-th recommendation time. By adopting this processing method, the model is updated only according to the business object information that the user has browsed; therefore, the accuracy of the model can be effectively improved while saving storage resources and computing resources.

Among them, Y _t represents the training sample set at the tth recommendation time, Y _t-1 represents the training sample set at the t-1th recommendation time, and W _t (e) represents the user clicks (or favorites, etc.) at the tth recommendation time Etc.) The business object e, ΔW _t (e) represents the second feature data of the business object e at the tth recommendation time, and the second feature data estimate (such as the average value) of the business object e at the tth recommendation time The difference between.

Among them, V _t represents the cumulative matrix at the t-th recommendation time. The elements in the matrix represent the correlation between two business objects. For example, V _{i, j} represents the correlation between business object i and business object j, and V _t-1 represents the cumulative matrix at the t-1th recommendation time,

Represents the sum of the correlation between two business objects included in the business objects browsed by the user.

among them,

Indicates the parameter estimate at the t-th recommendation time. The parameter estimate is determined by X _t and Y _t . In this embodiment, the first parameter includes 100 parameter items,

A column vector composed of the estimated values of these 100 parameter items.

Among them, β _t represents the parameter item radius at the t-th recommendation time.

In summary, this embodiment updates the first feature data set X _t , the training sample set Y _t and the accumulation matrix V _t according to the collected user feedback O _t and w _t at each recommendation moment, and estimates through ridge regression parameter

And calculate the updated parameter term radius β _t . Wherein, O _t is the browsing behavior information, and w _t is the operation behavior information.

In this embodiment, the process of updating non-parametric items can be expressed as follows:

1) Assuming that a total of L business objects are recommended to users, for each business object e, the following calculation is performed:

T _t (e)←T _t-1 (e), the meaning of this formula is to use the first feature data of business object e at the t-1th recommendation time as the first feature data of business object e at the t-th recommendation time The initial value of.

2) For k=1,...,min{O _t ,|A _t |}, where |A _t | represents the number of business objects recommended to the user, and O _t represents the number of business objects browsed by the user, the following calculation is performed:

The meaning of this formula is to take the k-th business object browsed by the user at the t-th recommendation time as the business object e to be processed.

T _t (e)←T _t (e)+1, this formula means to take the first feature data of the business object e that the user has viewed at the t-th recommendation time (such as the number of times the product is clicked by the user in a day, etc.) ) Accumulate 1.

The meaning of this formula is the non-parameter of the business object e at the t-th recommendation time.

The meaning of this formula is the mean value of the parameter item feature of the business object e at the t-th recommendation time.

In summary, this embodiment updates the statistical value of the business object based on user feedback at each recommendation moment

And feature mean

As shown in Fig. 4, in this embodiment, after updating the model through step S203, the following steps may be further included:

Step S401: Determine whether the model converges according to the first parameter and the second parameter before the update, and the first parameter and the second parameter after the update.

In this embodiment, if the difference between the first parameter before the update and the first parameter after the update is less than the first preset difference threshold, the difference between the second parameter before the update and the second parameter after the update is less than The second preset difference threshold is determined to converge the model.

Step S403: If the above judgment result is yes, stop updating the model.

If it is determined that the model is convergent, it means that various parameters of the model are relatively stable, and the value score of the business object can be correctly evaluated, so that the accuracy of the recommendation result can be gradually improved. In this case, you can stop collecting user feedback information and stop updating the model to save the server's computing resources.

Step S405: If the above judgment result is no, continue to update the model.

If it is determined that the model does not converge, it means that the various parameters of the model are not stable, and the value score of the business object cannot be correctly evaluated. Therefore, it is necessary to continue to collect user feedback information and continue to update the first and second parameters of the model. Parameters, so as to gradually improve the accuracy of the model, thereby improving the accuracy of the value evaluation of the business object, and then improving the accuracy of the recommendation result, so that the recommendation result gradually converges.

The method provided by the embodiment of the application uses the online method shown in Figure 2 to update the model. After each business object is recommended to the user, user feedback information is collected in real time, and the user behavior of the business object is updated in real time according to the user feedback information. Feature data, thereby updating the model to improve the accuracy of business object recommendation; this processing method enables real-time collection of user behavior data and rapid accumulation of user behavior feature data of business objects, making the value of user behavior feature data more sufficient; Therefore, it is more suitable for scenarios with unknown data distribution, such as second-hand merchandise sales scenarios.

In specific implementation, the processing method of updating the model in offline mode can also be used, so that sufficient user behavior characteristic data existing in the product can be used to avoid the occupation of more computing resources caused by real-time updating of user behavior data, so it is more suitable for linear scenarios .

From the perspective of application time, the method provided by the embodiments of this application is not limited to the cold start phase of the business object. In this phase, the model is updated online, and after this phase, the online update of the model can be stopped; this method also The same applies to the cold start stage of non-business objects, that is, it can be applied to the stage where the product has been placed for a period of time and has sufficient user interaction behavior data, that is, always collect user behavior data and update the model based on real-time user behavior data.

As shown in Fig. 5, in this embodiment, the method may further include the following steps:

Step S501: Initialize the first parameter and the second parameter included in the business object value evaluation model.

By initializing the model, the model has an initial business object value evaluation capability. At this time, the accuracy of the value evaluation of the model is usually low. As the process of recommending business objects to users for many times, user feedback information is continuously collected, thereby continuously improving model parameters, and gradually increasing the recommendation accuracy, until the user no longer gives feedback information, or until the model converges, that is, before and after The difference between the two models stabilized.

Please refer to FIG. 6, which is a specific flowchart of step S401 in an embodiment of a method for recommending a business object provided by this application. In this embodiment, the step of initializing the first parameter and the second parameter may include the following sub-steps:

Step S5011: Show the candidate business object to the user at least once.

The at least one candidate business object includes all business objects that the recommendation system can recommend to the user. In this embodiment, the recommendation system first releases all business objects in the system to the user client once to collect the initial user feedback information, that is, the second user feedback information.

Step S5013: Obtain second user feedback information for the at least one candidate business object.

The second user feedback information may include operation behavior information of the user on the business object that the recommendation system recommends to the user for the first time, and may also include browsing behavior information.

Step S5015: Generate training samples of the model according to the second user feedback information.

In this embodiment, the user behavior characteristic data is first updated according to the operation behavior information, and then the initial training sample of the model is generated based on the updated user behavior characteristic data and the browsing behavior information.

Step S5017: Determine the first parameter and the second parameter to be initialized according to the training sample.

After the initial training samples of the model are generated, the first parameter and the second parameter can be determined according to the initial training samples.

In this embodiment, initializing the model may include the following specific steps: 1) Set the first feature data set X ₀ and the training sample set Y ₀ to empty sets, set the cumulative matrix V ₀ to the unit matrix, and set the parameters Item estimate

Set to 0; 2) Put all products once each, collect user feedback, and initialize product features based on user feedback

Non-parametric statistics

among them,

Represents the average value of the first feature data of all commodity objects at the initial time t ₀ ,

Represents the non-parametric statistics at the initial time t ₀ .

It can be seen from the foregoing embodiments that the business object recommendation method provided by the embodiment of the present application determines the score of the candidate business object according to the first characteristic data of the candidate business object through the first parameter and the second parameter included in the business object value evaluation model Determine the set of business objects recommended to the user according to the score; send the set of business objects back to the client; this processing method allows the business object value evaluation model to be divided into parameterized items and non-parameterized items, comprehensive Parametric models and non-parametric models evaluate the value of business objects with unknown feature distributions. Because non-parametric terms enable the model to fit the unknown business object feature distributions, the gap between the parametric model and the real environment can be continuously reduced; therefore, the features can be effectively improved The value accuracy of the unknown business object is distributed, so that the single recommendation time step loss can be converged, thereby improving the accuracy of the business object recommendation.

Although this application is disclosed as above in preferred embodiments, it is not intended to limit the application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the application. Therefore, this application The scope of protection shall be subject to the scope defined by the claims of this application.

In a typical configuration, the computing device includes one or more processors (CPU), input/output interfaces, network interfaces, and memory.

The memory may include non-permanent memory in computer readable media, random access memory (RAM) and/or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM). Memory is an example of computer readable media.

1. Computer-readable media include permanent and non-permanent, removable and non-removable media, and information storage can be realized by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disc (DVD) or other optical storage, Magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media can be used to store information that can be accessed by computing devices. According to the definition in this article, computer-readable media does not include non-transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

2. Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems or computer program products. Therefore, this application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

Claims

A method for recommending business objects, which is characterized in that it includes:

According to the first parameter and the second parameter included in the business object value evaluation model, the score of the candidate business object is determined according to the first characteristic data of the candidate business object; the first characteristic data includes user behavior characteristic data; the first The parameter includes a weight parameter related to the first characteristic data, and the second parameter includes an unknown second characteristic data distribution parameter;

Determining a set of business objects recommended to the user according to the score;

Push the set of business objects to the client.
The method according to claim 1, wherein the business object comprises:

A business object whose business object value is determined jointly by the first characteristic data and the second characteristic data, and/or a business object whose business object value is determined by the first characteristic data.
The method according to claim 1, further comprising:

Acquiring first user feedback information for the set of business objects;

Updating the first parameter and the second parameter according to the first user feedback information.
The method according to claim 3, wherein:

The first user feedback information includes operation behavior information and browsing behavior information of the user on the business object.
The method according to claim 4, wherein said updating said first parameter and said second parameter according to said first user feedback information comprises:

Updating the user behavior characteristic data according to the operation behavior information;

Generating training samples according to the updated user behavior characteristic data and the browsing behavior information;

The first parameter and the second parameter are updated according to the generated training samples and historical samples.
The method according to claim 3, further comprising:

Judging whether the model converges according to the first parameter and the second parameter before the update, and the first parameter and the second parameter after the update;

If the above judgment result is yes, stop updating the model.
The method according to claim 6, further comprising:

If the above judgment result is no, continue to update the model.
The method according to claim 1, further comprising:

Initialize the first parameter and the second parameter.
The method according to claim 8, wherein the initializing the first parameter and the second parameter comprises:

Show users at least one candidate business object;

Acquiring second user feedback information for the at least one candidate business object;

Generating training samples of the model according to the second user feedback information;

According to the training sample, the first parameter and the second parameter to be initialized are determined.
The method according to claim 1, wherein:

The first parameter includes: a parameter of a linear machine learning model or a parameter of a nonlinear machine learning model;

The second parameter includes: statistical items related to the Gaussian process, statistical items related to the Dirichlet process, and statistical items related to the infinite-dimensional distribution.