Personalized news recommendation method based on adaptive multi-armed bandit model
By using an adaptive multi-armed slot machine model and a neural network to dynamically partition the feature space, the problems of nonlinear reward function and environmental changes in news recommendation are solved, thus improving the recommendation effect.
Patent Information
- Application Number
- CN202310108836.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-07
- Publication Date
- 2025-11-04
- Estimated Expiration
- 2043-02-07
AI Technical Summary
Existing news recommendation methods cannot effectively handle nonlinear reward functions and dynamic changes in mapping relationships caused by environmental changes, resulting in suboptimal recommendation performance.
An adaptive multi-armed slot machine model is adopted, which initializes the payout model through a neural network, dynamically divides the feature space into subspaces, and adjusts the payout function using historical data to adapt to environmental changes.
This improves the overall revenue of news recommendations, enabling them to better adapt to complex and ever-changing real-world environments and reduce revenue losses caused by changes in local environments.
Smart Images
Figure CN116089721B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the field of online learning, and particularly relates to a personalized news recommendation method based on an adaptive multi-armed bandit model. BACKGROUND
[0002] In the field of machine learning, online learning refers to the process of updating a model by continuously obtaining new data. Online learning algorithms are often used to process streaming data or large amounts of data. One typical example is an online advertising system, which continuously optimizes advertising strategies based on user browsing and clicking data. Online learning is opposite to offline learning, which cannot obtain all training data at once, but updates the model online through online streaming data, so it can continuously receive data and dynamically update the model in real time, which is suitable for processing large-scale streaming data. The multi-armed bandit problem is a classic problem in online learning. The multi-armed bandit problem is derived from the classic problem: there is a slot machine and multiple arms, each round selects a pull arm, and then a return is obtained, but the return of each arm is unknown before selection, and the goal is to maximize the total return. The multi-armed bandit problem has attracted widespread attention in many fields because it well describes the "exploration-exploitation" dilemma.
[0003] The contextual multi-armed bandit algorithm is an online learning algorithm that can be applied to news recommendation. It combines the multi-armed bandit model with contextual information to improve the accuracy of news recommendation. Specifically, in the contextual multi-armed bandit algorithm, the algorithm determines a set of strategies based on the user's current context information (such as user interest, historical browsing records, etc.). Then the algorithm randomly selects a strategy from the strategy set and performs a trial, and updates the estimated value of these strategies based on the trial results. The contextual multi-armed bandit algorithm can improve the accuracy of news recommendation because it can determine the most appropriate strategy set based on the user's current context information and continuously optimize strategy selection in the strategy set to ultimately find the optimal strategy. In current news recommendation, it is generally assumed that the return of recommending a news is a linear function of its context features, but the linear assumption often does not hold in real-world news recommendation problems because the return of recommending a news has a complex mapping relationship with context features. In addition, the reward mapping function in real-world applications is usually dynamic, as user interests and environments also change in real time. Existing work attempts to use deep neural networks to represent these non-linear relationships, as learning-based explanations have strong representation capabilities. However, the combinatorial nature and non-stationary environment cannot be integrated into these models, greatly narrowing the design space. They cannot adapt to the local distribution changes in the environment in news recommendation problems in time, leading to suboptimal choices.
[0004] Therefore, providing a news recommendation method which can still achieve good revenue effect in real news recommendation problems under the condition that the revenue function is a nonlinear function of the context features in the unstable environment is an urgent problem for those skilled in the art. SUMMARY
[0005] The present application aims at the problem that the existing news recommendation method cannot simultaneously solve the problem that the revenue function is nonlinear and the mapping relationship between the features and the revenue changes with the change of the environment, and proposes a personalized news recommendation method based on an adaptive multi-arm bandit model, which can simultaneously solve the above two important problems and can be effectively applied to the news recommendation scene.
[0006] To achieve the above object, the present application adopts the following technical solutions:
[0007] The personalized news recommendation method based on the adaptive multi-arm bandit model, characterized in that it comprises the following steps:
[0008] Step 1: using a multi-arm bandit model to establish the process of news recommendation, taking the context features of each news to be recommended as the tiger arms in the multi-arm bandit problem, the context features including the news features of the news to be recommended and the user features of the user to be recommended, initializing all the tiger arms to share a nonlinear revenue model;
[0009] Step 2: determining the feature space of the news to be recommended, taking the entire feature space as a subspace, initializing the revenue model corresponding to the subspace using a neural network, using the revenue model to estimate the revenue of the news to be recommended, selecting a number of news with the maximum revenue to recommend to the current user, and storing the news features of the selected news, the user features of the recommended user and their revenue as historical data in the subspace corresponding to the news features;
[0010] Step 3: determining whether to divide the subspace according to the number of historical data in the subspace, if it is determined to divide, automatically dividing it into two subspaces, retraining the revenue model according to the historical data contained in each divided subspace, recommending news according to the retrained revenue model, and storing the news features of the selected news, the user features of the recommended user and their revenue as historical data in the subspace corresponding to the news features in each round;
[0011] Step 4: using the method of step 3 to traverse all subspaces to recommend news.
[0012] To optimize the above technical solutions, the following specific measures are taken:
[0013] Further, in step 2, the revenue model is a mapping relationship from the news features and user features in the subspace to the revenue.
[0014] Further, in step 3, it is determined whether the number of historical data in the subspace exceeds a set threshold value, and if the threshold value is exceeded, the subspace is divided.
[0015] Further, in step 3, if it is determined to divide, the historical data is divided into two parts by taking the midpoint of a certain feature dimension in the feature space as a dividing line, and the two parts of data after division are respectively fitted by a neural network to obtain a corresponding revenue model; all feature dimensions in the feature space are traversed, a feature dimension that can make the sum of the revenue errors of the two parts of data after division under the respective revenue models minimum is taken as the divided feature dimension, and the historical data is divided by taking the midpoint as the dividing line, and the two parts of data after division are respectively stored in the divided subspace.
[0016] Further, the divided feature dimension is determined by the following formula:
[0017]
[0018] s.t.
[0019]
[0020]
[0021]
[0022] In the formula, d * represents the optimal divided feature dimension sequence number, d represents a feature dimension sequence number of the context feature, x t，i represents the news context feature collected in the subspace , y t，i represents the revenue obtained by selecting the news, represents the value of the context feature x t，i in the dth feature dimension, represents the mean value of all historical context feature samples collected in the subspace in the dth feature dimension, n j is the number of samples in the subspace , f 2*j+2 (x t，i ) is the revenue estimation value of the revenue function of the subspace t，i to which the context feature x belongs, f 2*j+1 (·) and f 2*j+2 (·) respectively represent two revenue functions retrained after the subspace is divided according to a certain feature dimension, and the two new subspaces are respectively represented as and and respectively represent the maximum value and the minimum value of all context feature samples in the subspace respectively represent the maximum value and the minimum value of all context feature samples in the subspace
[0023] Further, in step 3, for each context feature of the selected news, it is judged whether the difference between the number of historical data in the subspace where the context feature is located and the number of historical data when the reward function is last trained is greater than a set threshold value, and if greater than the threshold value, the reward function of the subspace is retrained.
[0024] The application further provides a computer readable storage medium storing a computer program, wherein the computer program enables a computer to execute the personalized news recommendation method based on the adaptive multi-armed bandit model.
[0025] The application further provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the personalized news recommendation method based on the adaptive multi-armed bandit model.
[0026] The application has the beneficial effects that the application provides a news recommendation method combining a nonlinear reward function and a non-steady state environment, so that the news recommendation method can better adapt to the complex and changeable environment in reality, fully utilizes the powerful fitting capability of a deep neural network, and greatly improves the overall reward that can be brought by recommended news in the news recommendation problem. In the news recommendation method, the dynamic feature space division method is used to well adapt to the dynamic change of the reward distribution in the real scene, the historical data are collected online and the division of the space and the fitting of the space reward function are adaptively adjusted by using the historical data, so that the reward loss caused by the change of the local environment distribution is greatly reduced. The application is suitable for most complex real problems that can be modeled as personalized recommendation, and the process is simple, easy to deploy and optimize. BRIEF DESCRIPTION OF DRAWINGS
[0027] Figure 1 is the overall framework diagram of the personalized news recommendation method based on the adaptive multi-armed bandit model of the application.
[0028] Figure 2 is an illustrative example diagram of the dynamic division of the space in the personalized news recommendation problem of the application.
[0029] Figure 3 is the flowchart of the personalized news recommendation method based on the adaptive multi-armed bandit model of the application.
[0030] Figure 4is an example diagram of the overall algorithm of the present application.
[0031] Figure 5 is an example diagram of the deep neural network training method of the present application. DETAILED DESCRIPTION
[0032] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.
[0033] In an embodiment, the present application provides a personalized news recommendation method based on an adaptive multi-armed bandit model, comprising the following steps:
[0034] Step 1: using a context multi-armed bandit model, establishing the process of news recommendation, that is, regarding the context features of each news to be recommended as the tiger arms in the multi-armed bandit problem.
[0035] In step 1, the context features of each news to be recommended are regarded as the tiger arms in the multi-armed bandit problem, and all the tiger arms share an initial nonlinear reward model, which is initialized by a neural network.
[0036] Step 2: determining the size of the feature space of the news to be recommended, regarding the entire feature space as a subspace, and initializing the corresponding reward model thereof by a neural network.
[0037] In step 2, the dimension of the feature space is determined according to the context features of the news to be recommended. In the initial case, the entire feature space is set as a large subspace, which contains all the feature vectors in the feature space. In addition, the initial subspace has an initial reward model, and the parameters in the initial reward model are randomly initialized. Then, the reward of the recommendable news is estimated by using the reward model, and the several news with the largest estimated reward are selected to be recommended to the current user. The news features, user features and real rewards of the selected news in each round are stored in the subspace corresponding to the features.
[0038] Step 3: determining whether to divide the feature space over time, if it is determined to divide, then automatically dividing into two subspaces, each of which contains a set of features within its range, historical features and reward result data, and then retraining a new reward model according to the historical data by a deep neural network; when the historical data collected in a subspace is sufficient, it is determined whether the reward model corresponding to the subspace needs to be retrained by a neural network, and if so, the historical data is used to retrain the reward model.
[0039] In step 3, if each current existing subspace and the historically existing subspace is regarded as a node of a binary tree, the entire partitioning process can be regarded as a partitioning process of nodes in the binary tree. When the historical data collected in each subspace is greater than a partitioning threshold, it is determined that the subspace currently needs to be partitioned. The partitioning of the subspace corresponds to adding two new child nodes to the parent node corresponding to the subspace in the binary tree, and the parent node becomes a non-leaf node. Only the leaf nodes belong to the subspace currently existing in the entire context space. The two new child nodes, i.e., the two generated new subspaces, belong to the historical data set that is combined together, which is the historical data set of the atomic space, and the two data sets are disjoint. In the partitioning process, the midpoint of the feature dimension that can make the sum of the errors (the difference between the true return and the return estimated by the retrained return function) of the two historical data sets after fitting the corresponding return functions of the two historical data sets respectively is the smallest is selected for partitioning, that is, in this way, the data set originally belonging to one set is divided into two data sets. The return functions to which the data in the two sets should be subject are very different, and when they are trained together, the trained return function is very poor. However, when they are divided and trained separately, the overall error is the smallest. Therefore, the two data sets should not belong to one subspace, and the features corresponding to the feature subspace of the two data sets should be subject to different return functions.
[0040] When the historical data collected in a subspace is sufficient, it is indicated that more data can be used to retrain the network. However, when a subspace is obtained by multiple partitioning, the subspace tends to be a smaller subspace. More historical data is needed to retrain the return model of the subspace, because the smaller the space is, the more finely the space is partitioned, and the space cannot be further partitioned.
[0041] Step 4: traverse all subspaces by using the method of step 3 to perform news recommendation.
[0042] Figure 1 is a schematic diagram of a framework of a personalized news recommendation method based on an adaptive multi-armed bandit model, Figure 1 The left part of is a time axis of method running, Figure 1The right part of the figure is the process of online division. When the online program starts, there is only one subspace, which is the whole context feature space. As the program runs, more and more data about the features of news and the real revenue are collected in the subspace, and it is necessary to determine whether the subspace needs to be divided, which is represented as a division detection on the time axis. If the division detection detects that the subspace needs to be divided, the subspace is divided into two. In the subsequent process, the new subspace collects news data whose features belong to the subspace, and when the collected historical data are enough, the division detection is performed again. When the news is selected each time, the selectable news can be observed, and each news corresponds to a context feature vector, which reflects the nature of the news itself and the nature of the current interactive user object combination. According to the feature vector of the news, the news can be positioned in a certain subspace, and the revenue value of the current round of the news is estimated by the revenue function corresponding to the subspace where the news is located, and the input is the feature vector of the news, and the output is the revenue value of the news. Each news obtains its estimated revenue of the current round in this way, and finally the first few options with the maximum revenue value are selected according to the estimated revenue value to obtain the selection result of the current round.
[0043] Figure 2 The figure is an example of dynamic division of the application when recommending news to users. It is assumed that there are 6 users and 6 news, and one user comes in each round, and it is necessary to recommend one of the 6 news to the coming user. In each round, the context features of the 6 news are the features of the news themselves plus the features of the coming user, which reflect some association information between the user and the news, which can be the preference of a specific user to a specific news category. Initially, the 6 news belong to one subspace, that is, the whole context space, but as the running round increases, the subspace is gradually divided, and finally divided into three subspaces, each of which contains a specific combination of user features and news features, which corresponds to the context features of the news. The combination of news and user features belonging to a subspace is subject to the same revenue model.
[0044] The application proposes a standardized model for the personalized news recommendation problem, that is, a new adaptive context combination multi-arm bandit model applied to the news recommendation problem, and the overall news recommendation process is as follows Figure 3At the beginning of the method execution, the subspace is initialized, and then at each round of the online process, first observe which news and incoming users are in the current round, then get the context features corresponding to each news, according to the context features of the news, the news can be attributed to a certain subspace, estimate the revenue value of each news according to the revenue model in its subspace, and select specific news to recommend to the user. Then update each subspace according to the recommended news and revenue results in the current and past rounds, or keep the subspace unchanged, retrain the revenue model of the subspace or keep it unchanged. Repeat these steps until the online news recommendation process is completed.
[0045] Figure 4 is the execution algorithm flowchart of the entire method. First, the entire process has multiple rounds, and each round needs to select the top m news with the maximum estimated revenue value from N news (arms) to form a super arm, and then store the features corresponding to the m news and the observed real revenue value in the historical data set of the corresponding subspace. Then, for the feature of each selected news, judge whether the difference between the number of historical data in the subspace and the number of historical data when the revenue function was last retrained is greater than a threshold value. If it is greater than the threshold value, it means that the subspace needs to be retrained again. At the same time, it is also necessary to judge whether the number of historical data currently existing in the subspace exceeds a threshold value. If it exceeds the threshold value, the subspace is divided. When dividing the subspace, select the feature dimension that can make the prediction error of the whole after division the smallest to divide. For example, select the dth feature dimension to divide, then judge whether the value of the dth feature dimension of all data in the subspace is greater than or less than the midpoint value of the length of this dimension, divide these data into two parts. Among the two new subspaces generated, the maximum value of the dth feature dimension of one subspace is less than or equal to the midpoint value of the dth feature dimension of the original space, and the minimum value of the dth feature dimension of the other subspace is greater than the midpoint value of the dth feature dimension of the original space, and the two subspaces respectively have the two parts of historical data divided before.
[0046] Figure 5 is the training process of the revenue model of the historical data set collected by the subspace using a deep neural network. Here, the stochastic gradient descent method is used to train the neural network.
[0047] When a subspace needs to be divided, the midpoint of each feature dimension of the subspace is taken as the dividing line to divide the data set points, and the two data sets after division are respectively fitted with the revenue function, and the midpoint of the dimension that can make the difference between the predictions of the two parts after division the smallest is taken as the division point. The purpose of the constraint is to limit the division, require the length of the smallest dimension of the new subspace after division to be greater than a threshold, and avoid some dimensions being divided too small, thereby leading to uneven dimension division. The process of subspace division is as follows:
[0048] At the tth round, the context feature of a news is x t，i , and its estimated revenue is the revenue function value f(x t，i ; θ t-1 ) of the subspace to which it belongs plus an exploration term γ t-1 E(g(x t，i ; θ t-1 )), that is,
[0049]
[0050] In the formula, θ t，i represents the parameters of the neural network, g(x t，i ; θ t-1 ) is the gradient of the revenue function f(x t，i ; θ t-1 ). The revenue function f(x t，i ; θ t-1 ) of each subspace is different, and f j (·) is used to represent the revenue function corresponding to the subspace for convenience. γ t-1 represents the weight of exploration, and the greater the value, the more the estimated revenue value of the news deviates from the revenue value predicted by the revenue function trained by the neural network, and the smaller the value, the closer the estimated revenue value of the news to the revenue value predicted by the revenue function trained by the neural network. represents a variable that changes with t, and m is the width of the neural network.
[0051] When a subspace needs to be divided, the following method is used to determine which midpoint of the context space dimension to divide:
[0052]
[0053] s.t.
[0054]
[0055]
[0056]
[0057] The goal of the partition is to get two sub-spaces with the minimum model bias. Where d represents the index of a certain dimension of the context feature, d * is the final optimal dimension index of the partition, x t，i represents the context feature collected in the sub-space y t，i represents the revenue obtained by selecting the news, represents the value of the context feature x t，i in the dth dimension, represents the mean value of all historical context features collected in the sub-space in the dth dimension, n j is the number of samples in the sub-space f 2*j+2 (x t，i ) is the revenue estimation value of the context feature x t，i using the revenue function of the sub-space f 2*j+1 (·) and f 2*j+2 (·) represent the two revenue functions retrained after the sub-space is partitioned according to a certain dimension, and the two new sub-spaces are represented as and
[0058] As shown in the constraints of the first and second inequalities, the threshold υ is used to limit the maximum value of a sub-space in a certain dimension and the difference between the mean value of the sub-space in the dimension and the minimum value of the dimension, which cannot exceed the threshold υ, so as to ensure that when a sub-space is partitioned small enough, it will not be partitioned again, avoiding the case of unlimited partition of a sub-space, and represent the maximum value and the minimum value of all context feature samples in the dth dimension in the sub-space The constraints of the third and fourth inequalities ensure that when a sub-space is partitioned, the length of the two new sub-spaces in a certain dimension after partitioning is not less than the maximum value of the dimension length in the sub-space, that is, to avoid the length difference between the dimensions in the sub-space being too large, so as to more evenly partition the sub-space, where D is the size of the feature dimension, that is, the number of attributes in the news feature.
[0059] The novelty of the present application is that it is the first time to take the way of dynamic division of feature space to cope with the dynamic changes of the environment, i.e. changing user interests and environment, in the news recommendation problem, to capture local distribution changes by training the reward function separately for each subspace and constantly detecting possible distribution changes, such as a user prefers more football related news during the World Cup and pays little attention to football news in normal time. In addition, the present application is the first time to deal with the nonlinear reward function and the non-stationary environment at the same time, and has obtained considerable reward analysis results.
[0060] In another embodiment, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program causes a computer to execute the personalized news recommendation method based on the adaptive multi-armed bandit model as described in the first embodiment.
[0061] In another embodiment, the present application further provides an electronic device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the personalized news recommendation method based on the adaptive multi-armed bandit model as described in the first embodiment when executing the computer program.
[0062] In the embodiments disclosed in the present application, the computer storage medium can be a tangible medium, which can contain or store programs for use by or in connection with an instruction execution system, apparatus or device. The computer storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any suitable combination of the above. More specific examples of the computer storage medium can include one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
[0063] Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed in the present application can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.
[0064] The above are only preferred embodiments of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical scheme falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary skilled in the art, some improvements and refinements without departing from the principles of the present application shall be considered as the protection scope of the present application.
Claims
1. A personalized news recommendation method based on an adaptive multi-armed bandit model, characterized in that, The method comprises the following steps: Step 1: using a multi-armed bandit model to establish a news recommendation process, taking the context features of each news to be recommended as the tiger arms in the multi-armed bandit problem, the context features including news features of the news to be recommended and user features of the user to be recommended, and initializing all the tiger arms to share a nonlinear reward model; The nonlinear reward model is a mapping relationship of the news features and the user features in the subspace to the reward; Step 2: determining the feature space of the news to be recommended, taking the entire feature space as a subspace, initializing the nonlinear reward model corresponding to the subspace by using a neural network, estimating the reward of the news to be recommended by using the nonlinear reward model, selecting a number of news with the maximum reward to recommend to the current user, and storing the news features of the selected news, the user features of the recommended user and the reward in the subspace corresponding to the news features as historical data; Step 3: judging whether to divide the subspace according to the number of historical data in the subspace, if it is judged to divide, then automatically dividing the subspace into two subspaces, retraining the nonlinear reward model of each divided subspace according to the historical data contained in the subspace, recommending news according to the retrained nonlinear reward model, and storing the news features of the selected news, the user features of the recommended user and the reward in the subspace corresponding to the news features as historical data in each round; In the step 3, if it is judged to divide, then dividing the historical data into two parts by taking the midpoint of a feature dimension in the feature space as a dividing line, and respectively fitting the nonlinear reward model of the two divided parts of data by using a neural network; traversing all the feature dimensions in the feature space, taking the feature dimension that can make the sum of the reward errors of the two divided parts of data under the respective nonlinear reward model minimum as the divided feature dimension, and dividing the historical data by taking the midpoint as the dividing line, and respectively storing the two divided parts of data in the divided subspaces; Step 4: traversing all the subspaces by using the method of step 3 to recommend news. 2.The personalized news recommendation method based on adaptive multi-armed bandit model according to claim 1, wherein: In the step 3, judging whether the number of historical data in the subspace exceeds a set threshold value, if the number of historical data exceeds the threshold value, then dividing the subspace. 3.The personalized news recommendation method based on adaptive multi-armed bandit model of claim 1, wherein: The divided feature dimension is determined by the following formula: s.t. where d * denotes the optimal partitioned feature dimension, d denotes a certain feature dimension of the context feature, x t,i denotes the subspace in which the news context feature is collected, y t,i denotes the revenue obtained by selecting the news, denotes the context feature x t,i in the dth feature dimension, denotes the subspace in which all the historical context feature samples are collected, n j is the number of samples in the subspace , f 2*j+2 (x t,i ) is the context feature x t,i using the revenue estimation value of the revenue function in the subspace to which the context feature belongs, f 2*j+1 (·) and f 2*j+2 (·) respectively denote the two revenue functions retrained after the subspace is partitioned according to a certain feature dimension, and the two new subspaces are respectively denoted as and and respectively denote the maximum value and the minimum value of all the context feature samples in the subspace in the dth feature dimension, D is the size of the feature dimension, and υ denotes a threshold value. 4.The personalized news recommendation method based on adaptive multi-armed bandit model of claim 1, wherein: In the step 3, for the context features of each selected news, judging whether the difference between the number of historical data in the subspace where the context features are located and the number of historical data when the reward function was last trained is greater than a set threshold value, if the difference is greater than the threshold value, then retraining the reward function of the subspace.
5. A computer readable storage medium storing a computer program, characterized in that, The computer program enables a computer to execute the personalized news recommendation method based on the adaptive multi-armed bandit model according to any one of claims 1-4.
6. An electronic device, comprising: The computer program product comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the personalized news recommendation method based on the adaptive multi-armed bandit model according to any one of claims 1-4 when executing the computer program.