A streaming media recommendation method for a label distribution change scene

Abstract

The application discloses a streaming media recommendation method for a marked distribution change scene, collects marked streaming media and user data under an initial distribution, selects a classifier type and a loss function, and obtains an initial offline model through an empirical risk minimization method, and obtains a calibrated initial model through a model calibration method. In each round of online deployment, the base recommendation method estimates the current label distribution using the online data collected in each round, then performs gradient descent using the estimated risk function gradient, and sends the updated model to the combination method; different base recommendation methods are configured with different update learning rates to cope with different degrees of distribution change; the weight corresponding to each base recommendation method is updated according to the historical cumulative performance of each base recommendation method, the models of all base recommendation methods are combined using the updated weight to obtain a final recommendation model, and the streaming media recommendation result is returned using the final recommendation model.

Description

TECHNICAL FIELD

[0001] The application relates to a streaming media recommendation method for a label distribution change scene and belongs to the technical field of streaming media recommendation. BACKGROUND

[0002] The streaming media online recommendation method can determine whether to recommend streaming media to a user according to the user's interest preference and the streaming media attribute, so that each user obtains personalized customized streaming media information, and plays a very important role in the current information age. With the continuous development of society, the current human society is showing a new trend of integrated development of digitization, networking and intelligentization. Under this background, streaming media services are becoming increasingly popular. In addition to traditional long videos, audios and other forms, there are also short videos, live broadcasts and other new forms, which can be widely used in multimedia news release, e-commerce, knowledge sharing, online meetings and other aspects of human life and work. The current streaming media platform has the characteristics of online data aggregation, real-time data arrival and large data scale, and it is difficult to process quickly by relying on traditional methods, so the streaming media online recommendation method plays an important role in various streaming media platforms. Taking e-commerce live broadcast delivery as an example, when a user visits an e-commerce platform, the streaming media recommendation method can determine whether to recommend a candidate live broadcast to the user according to the characteristics of the user and the candidate live broadcast, and update the model according to whether the user watches, so as to improve the user's watching rate of the live broadcast, and then optimize the user experience and improve the user stickiness, which has high commercial value.

[0003] The previous streaming media online recommendation method can only handle simple scenes where the data distribution remains unchanged, and often performs poorly in the face of distribution changes. However, label distribution changes are constantly occurring and difficult to predict in real streaming media platforms. For example, with the change of time, the behavior of users on weekdays and holidays may change significantly; in addition, the current popular trend or sudden hot events will also affect whether the user watches the recommended streaming media. More challenging is that with the emergence of new media networks, the spread of streaming media has greatly improved, which makes it difficult to obtain accurate labeled data feedback when the distribution changes, so in actual scenes, the model often only obtains a small amount of unlabeled data in most recommendation rounds, and a small amount of labeled information is obtained after a long time interval, which brings difficulties to the evaluation and further update of the model. SUMMARY

[0004] The application aims to provide a streaming media recommendation method for a label distribution change scene.

[0005] The technical scheme is a streaming media recommendation method for a label distribution change scene, which can be used in e-commerce live streaming recommendation, short video recommendation, music recommendation and other application scenarios. Specifically, the method comprises an offline recommendation model training method and an online recommendation model deployment method. In the offline recommendation model training method, labeled streaming media and user data under an initial distribution are collected, a classifier type and a loss function are selected, an initial offline model is obtained through an empirical risk minimization method, and a calibrated initial model is obtained through a model calibration method. In the online recommendation model deployment method, an online integration method is used to update the classifier parameters, and the online integration method comprises multiple base recommendation methods and a combination method. In each round of online deployment, the base recommendation method first estimates the current label distribution using the online data collected in each round, estimates the risk function using a risk function rewriting method, then performs gradient descent using the estimated risk function gradient, and sends the updated model to the combination method; different base recommendation methods are configured with different update learning rates to cope with different degrees of distribution change. In each round of online deployment, the combination method receives the models of all base recommendation methods, updates the weight corresponding to each base recommendation method according to the historical cumulative performance of each base recommendation method, combines the models of all base recommendation methods using the updated weight to obtain a final recommendation model, and returns a streaming media recommendation result using the final recommendation model.

[0006] The specific steps of the offline recommendation model training method are as follows:

[0007] Step 100, offline collection of streaming media recommendation dataset wherein represents a feature vector obtained by splicing user features and streaming media features, represents a d-dimensional real number space; y n represents whether the user watches the streaming media.

[0008] Step 101, selection of classifier f: wherein respectively represent the model parameter space, the feature space, represents a real number space.

[0009] Step 102, select a loss function l: wherein represents a label space.

[0010] Step 103, on the streaming media recommendation dataset collected in step 100, the classifier and loss function selected in steps 101 and 102 are used to minimize the loss function to obtain offline initial model parameters

[0011] Step 104, using a model calibration method, obtain calibrated offline initial model parameters wherein cali(·) represents a model calibration method.

[0012] The specific steps of the base recommendation method in the online recommendation model deployment method are as follows:

[0013] Step 200, obtain a learning rate η i .

[0014] Step 201, in each online recommendation round t = 1, 2, …, T, perform the following steps 2011-2016:

[0015] Step 2011, send the model parameters w t,i to the combination method.

[0016] Step 2012, obtain online streaming media recommendation data wherein represents the number of unlabeled data, represents the number of labeled data.

[0017] Step 2013, using the offline initial model parameters w0 obtained in step 104 and the online streaming media recommendation data S t estimate the current round label distribution vector θ t .

[0018] Step 2014, using the label distribution vector θ t estimated in step 2013, construct a risk function estimate R t (w) = <θ t , r0(w)>, wherein r0(w) represents the loss function vector on the initial offline dataset S0, and <·,·> represents the vector inner product.

[0019] Step 2015, using the risk function estimate constructed in step 2014, calculate the gradient w t,i represents the model parameters of the i-th base learner in the t-th round.

[0020] Step 2016, update the model parameters

[0021] The specific steps of the combination method in the online recommendation model deployment method are:

[0022] Step 300, set a learning rate set Where M represents the number of base recommendation methods.

[0023] Step 301, set the combination method learning rate ∈.

[0024] Step 302, initialize the weight p of each base recommendation method 1,i = 1 / M.

[0025] Step 303, in each recommendation round t = 1, 2, …, T online, execute the following steps 3031-3032:

[0026] Step 3031, accept the model parameters w of each base recommendation method t,i .

[0027] Step 3032, update the weight p of each base recommendation method

[0028] Step 304, output the final integrated model parameters

[0029] The alternative classifiers in step 101 include linear classifiers, generalized linear classifiers, neural network classifiers, etc.

[0030] The alternative loss functions in step 102 include square loss functions, Logistic loss functions, Sigmoid loss functions, etc.

[0031] The alternative model calibration methods in step 104 include Sigmoid model method and order-preserving regression method, etc.

[0032] The alternative label distribution estimation methods for estimating the current round label distribution vector in step 2013 include maximum likelihood-based label distribution estimation method, confusion matrix-based label distribution estimation method, distribution matching-based label distribution estimation method, etc.

[0033] The setting method of the learning rate set in step 300 is Where Γ represents the diameter of the parameter space , and G represents the upper bound of the loss function gradient L2 norm.

[0034] The setting method of the combination method learning rate in step 301 is Where B represents the upper bound of the loss function value.

[0035] Beneficial effects: Compared with the prior art, the streaming media recommendation method for label distribution change scenarios provided by the application has the following advantages: the application can evaluate the recommendation model by using the received unmarked online streaming media recommendation data, and can adaptively update the streaming media recommendation model for scenarios where the degree of change in the label distribution of streaming media data is unknown, thereby effectively dealing with streaming media recommendation scenarios with dynamic changes in label distribution. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 Workflow diagram of the offline recommendation model training method of the embodiment of the application;

[0037] Figure 2 Workflow diagram of the base recommendation method in the online recommendation model deployment method of the embodiment of the application;

[0038] Figure 3 Workflow diagram of the combination method in the online recommendation model deployment method of the embodiment of the application. DETAILED DESCRIPTION

[0039] The application will be further illustrated below in conjunction with specific embodiments, which should be understood as merely illustrating the application and not limiting the scope of the application. After reading the application, those skilled in the art can make various modifications to the application, and all such modifications fall within the scope defined by the claims of the application.

[0040] The streaming media recommendation method for label distribution change scenarios, and the embodiment takes short video recommendation of a short video platform as an example.

[0041] The workflow of the offline recommendation model training method is shown in Figure 1 First, short video recommendation data is collected offline, and the data source is the historical record of user watching short videos, represented as wherein represents a feature vector obtained by splicing user features and streaming media features, represents a d-dimensional real number space; y n represents whether the user watches the streaming media. The user features include user gender, age, education level, and interested topics, and the streaming media features include short video duration, keywords, captions, and short video author.

[0042] Next, a linear classifier commonly used in classification problems and a logit regression loss function are used. On the historical short video watching data set, the optimal offline model parameters are obtained by using the stochastic gradient descent method

[0043]

[0044] Finally, the model is calibrated by using the monotonic regression method, that is, the model is optimized such that for any f(w0,x n )f(w0,x n′ ), thus obtaining the calibrated offline initial model parameters w0.

[0045] The workflow of the base recommendation method in the online recommendation model deployment method is shown in Figure 2 .

[0046] In each recommendation round, if the received data is unlabeled data, the expectation maximization method is used to estimate the label distribution, and the specific process is as follows: iterate the following two steps: (i)

[0047] (ii) until the convergence condition is met, where u represents the iteration round number; represents the posterior distribution on the offline distribution, which can be estimated by the calibrated offline initial model w0; represents the initial label distribution, which can be estimated by the frequency corresponding to the offline data set label;

[0048] represents the posterior distribution on the round t distribution estimated at the u-th step of iteration; represents the label distribution on the round t distribution estimated at the u+1-th step of iteration; represents the label distribution vector composed of . If the received data is labeled data, the frequency of the label is directly used to estimate the label distribution of the current round.

[0049] In each recommendation round, the estimated label distribution is used to construct the estimate R t (w) of the risk function, and the gradient is used to update the model parameters by gradient descent. Finally, the updated model parameters are sent to the combination method.

[0050] The workflow of the combination method in the online recommendation model deployment method is shown in Figure 3 .

[0051] First, initialization is performed, the step size η i of each base recommendation method is set, the learning rate ε of the combination method is set, and the weight p 1,i of each base recommendation method is initialized to 1 / M.

[0052] In each recommendation round, the model parameters w t,i of each base recommendation method are accepted, and the weight is updated. Finally, the updated weight is used to output the final model parameters

Claims

1. A streaming media recommendation method for scenarios with changing label distribution, characterized in that, It includes offline recommendation model training methods and online recommendation model deployment methods. In the offline recommendation model training method, labeled streaming media and user data under the initial distribution are collected, classifier type and loss function are selected, and an initial offline model is obtained through the empirical risk minimization method. Then, a calibrated initial model is obtained using the model calibration method. In the online recommendation model deployment method, an online ensemble method is used to update the classifier parameters. This online ensemble method includes multiple base recommendation methods and a combination method. In each round of online deployment, the base recommendation methods first estimate the current label distribution using the online data collected in each round, and then estimate the risk function using the risk function rewriting method. Next, gradient descent is performed using the estimated risk function gradient, and the updated model is sent to the combination method. Different base recommendation methods are configured with different update learning rates to cope with different degrees of distribution changes. In each round of online deployment, the combination method receives the models of all base recommendation methods, updates the weights of each base recommendation method according to the historical cumulative performance of each base recommendation method, combines the models of all base recommendation methods using the updated weights to obtain the final recommendation model, and uses the final recommendation model to return the streaming media recommendation results. The specific steps of the base recommendation method in the online recommendation model deployment method are as follows: Step 200, Obtain the learning rate ; Step 201, online for each recommendation round Perform the following steps from 2011 to 2016: Step 2011, set the model parameters Send to the combination method; Step 2012: Obtain online streaming media recommendation data ,in Indicates the number of unlabeled data. Indicates the number of labeled data; Step 2013, using offline initial model parameters And the online streaming recommendation data obtained in step 2012 Estimate the current round's marker distribution vector ; Step 2014: Use the label distribution vector estimated in step 2013. Constructing risk function estimation ,in Indicates the initial offline dataset The loss function vector on, Represents the dot product of vectors; Step 2015: Calculate the gradient using the risk function constructed in step 2014. , This represents the model parameters of the i-th base learner in round t; Step 2016, Update model parameters ; The specific steps of the combination method in the online recommendation model deployment method are as follows: Step 300, Set the learning rate set , where M represents the number of base recommendation methods; Step 301, Set the learning rate for the combination method ; Step 302: Initialize the weights of each base recommendation method. ; Step 303, each recommended round online Perform the following steps 3031-3032: Step 3031: Accept the model parameters of each base recommendation method. ; Step 3032: Update the weights of each base recommendation method. ; Step 304: Output the final ensemble model parameters. .

2. The streaming media recommendation method for scenarios with changing label distributions according to claim 1, characterized in that, The specific steps of the offline recommendation model training method are as follows: Step 100: Offline collection of streaming media recommendation dataset ,in This represents a feature vector composed of user features and streaming media features. Represents the d-dimensional real space; Indicates whether the user is watching streaming media; Step 101, Select a classifier ,in , These represent the model parameter space and feature space, respectively. Represents the space of real numbers; Step 102, Select the loss function ,in Represents the tag space; Step 103: On the streaming media recommendation dataset collected in step 100, using the classifier and loss function selected in steps 101 and 102, minimize the loss function to obtain the initial offline model parameters. ; Step 104: Obtain the calibrated offline initial model parameters using the model calibration method. ,in This indicates the model calibration method.

3. The streaming media recommendation method for scenarios with changing label distributions according to claim 2, characterized in that, The classifiers that can be selected in step 101 include linear classifiers, generalized linear classifiers, and neural network classifiers.

4. The streaming media recommendation method for scenarios with changing label distributions according to claim 2, characterized in that, The loss functions available in step 102 include the squared loss function, the logistic loss function, and the sigmoid loss function.

5. The streaming media recommendation method for scenarios with changing label distributions according to claim 2, characterized in that, The model calibration methods available in step 104 include the Sigmoid model and the ordinal-preserving regression method.

6. The streaming media recommendation method for scenarios with changing label distributions according to claim 1, characterized in that, In step 2013, the available label distribution estimation methods for estimating the label distribution vector of the current round include maximum likelihood-based label distribution estimation, confusion matrix-based label distribution estimation, and distribution matching-based label distribution estimation.

7. The streaming media recommendation method for scenarios with changing label distributions according to claim 1, characterized in that, The learning rate set is set in step 300 as follows: ,in Representing parameter space The diameter, This represents the upper bound of the L2 norm of the loss function gradient.

8. The streaming media recommendation method for scenarios with changing label distributions according to claim 1, characterized in that, In step 301, the learning rate setting method is as follows: ,in This represents the upper bound of the loss function value.

Images