Content recommendation method for cognitive bubble digestion

Abstract

The invention provides a cognitive bubble resolution-oriented content recommendation method, which relates to the field of content recommendation, and comprises the following steps of: extracting key features of a user and recommended content from acquired offline interaction data, and constructing and training a satisfaction model fusing cognitive perception from three aspects of interestingness, overexposure and cognitive imbalance; training a recommendation strategy model and a denoising state updater by adopting a reinforcement learning framework and combining an anti-fact reward provided by the satisfaction model; in real-time interaction, a sliding time window is used for obtaining user historical behavior information, a denoising state updater is combined to obtain current user state characterization, and the current user state characterization is input into a recommendation strategy model to give recommendation content; and real-time feedback of the user to the current recommended content is recorded in a data cache region for fine tuning and optimization of the model, so that the recommendation result can break through the inherent cognitive limitation of the user, and the propagation of positive ideas and contents is promoted. The cognitive polarization problem existing in the interaction process of the recommendation system and the user is solved.

Description

Technical Field

[0001] The present invention relates to the field of content recommendation, and in particular to a content recommendation method oriented to cognitive bubble elimination. Background Art

[0002] With the rapid development of the Internet and big data technology, personalized recommendation systems have become a key technical means of information distribution. Recommendation systems analyze user historical behavior data to provide users with customized content, thereby improving user experience and platform stickiness. However, existing personalized recommendation systems rely too much on user historical preferences, resulting in a single recommended content, which has caused users to fall into the "cognitive bubble" dilemma for a long time. From the perspective of cognitive science, human cognitive processes are affected by both information diversity and quality. Single information input will lead to the solidification of cognitive structure and rigid thinking patterns. The cognitive bubble effect not only limits the possibility of users obtaining diversified information, but also may cause cognitive bias polarization, affecting users' comprehensive understanding of the world. In addition, existing recommendation algorithms mostly use short-term indicators such as user click-through rate as optimization targets, ignoring long-term user satisfaction. Therefore, how to achieve a balance between personalized recommendations and content diversity, both to alleviate the cognitive bubble effect and to prevent excessive diversity from causing inaccurate information push, has become a key issue that needs to be solved in the recommendation system. By introducing a cognitive bubble dissolution mechanism and combining a collaborative optimization strategy that takes both personalization and diversity into consideration, the user experience and social value of the recommendation system can be effectively improved, and the optimization and development of user cognitive structure can be promoted. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a content recommendation method for dissolving cognitive bubbles, which is used to alleviate the cognitive polarization phenomenon existing in the long-term interaction process between users and recommendation systems, thereby guiding users to establish a positive and balanced cognitive system.

[0004] In order to achieve the above objectives, the technical solution adopted by the present invention is: a content recommendation method for cognitive bubble dissolution, including an offline model training phase and a real-time interaction and model update phase;

[0005] The offline model training phase includes the following steps:

[0006] S1. Extract key features of users and recommended content from the acquired offline interaction data, and build and train a satisfaction model integrating cognitive perception from the aspects of interest, overexposure and cognitive dissonance;

[0007] S2. Use the reinforcement learning model and the counterfactual rewards provided by the satisfaction model to train the recommendation strategy model and the denoising state updater.

[0008] The real-time interaction and model update phase includes the following steps:

[0009] S3. Obtain the user's historical behavior information using a sliding time window, and combine it with a denoising state updater to obtain the current user state representation. Input the current user state representation into the recommendation policy model to obtain recommended content;

[0010] S4. Record the user's real-time feedback on the current recommended content in the data buffer for fine-tuning the model, and complete the content recommendation for cognitive bubble elimination.

[0011] The beneficial effects of the present invention are as follows: The present invention provides a content recommendation method for cognitive bubble elimination, which is used to alleviate the cognitive bias problem existing in the long-term interaction between users and the recommendation system, thereby guiding users to establish a positive and balanced cognitive system. Extract the key features of users and recommended content from the obtained offline interaction data, and construct and train a satisfaction model integrating cognitive perception from three aspects: interest degree, overexposure degree, and cognitive misalignment degree; adopt a reinforcement learning framework, combine the counterfactual rewards provided by the satisfaction model, and train the recommendation policy model and the denoising state updater; in real-time interaction, use a sliding time window to obtain the user's historical behavior information, and combine it with the denoising state updater to obtain the current user state representation, and input it into the recommendation policy model to give recommended content; record the user's real-time feedback on the current recommended content in the data buffer for model fine-tuning and optimization, ensuring that the recommendation results can break through the user's inherent cognitive limitations and promote the spread of positive ideas and content.

[0012] Further, the S1 includes the following steps:

[0013] A1. Obtain offline interaction data;

[0014] A2. Based on the obtained offline interaction data, extract the features and embedding representations of users and historical interaction content;

[0015] A3. Based on the results obtained in A2, estimate through three dimensions of interest degree, overexposure degree, and cognitive misalignment degree, and construct and train a satisfaction model integrating cognitive perception.

[0016] The beneficial effects of the above improvement scheme are as follows: This scheme makes full use of the existing offline interaction data information, and models the satisfaction scoring model from the perspective of cognitive science, realizing the accurate perception of the cognitive bubble effect of users in long-term interaction.

[0017] Still further, the A2 includes the following steps:

[0018] For the user and historical recommended content information in the offline interaction data, select the most important relevant field information through correlation analysis to obtain the user feature x u and the recommended content feature x i ;

[0019] According to the obtained user characteristics and historical interaction content characteristics, the user embedding representation e is obtained through the encoding layer u and the historical interaction content embedding representation e i .

[0020] The beneficial effects of the above improvement scheme are as follows: This scheme screens out the key characteristics of users and historical interaction content through correlation analysis, effectively reducing the interference of noise information and improving the model's ability to evaluate user satisfaction; the user characteristics and historical interaction content characteristics are embedded to capture the deep-level relationship between users and content, providing high-quality input for the training of subsequent satisfaction models, denoising state updaters, and recommendation strategy models

[0021] Furthermore, the A3 includes the following steps

[0022] Estimate the user's degree of interest using a traditional recommendation model

[0023] Estimate the cognitive dissonance degree through the change degree of the user's interest at adjacent times

[0024] Estimate the overexposure degree of the recommended content through the similarity weighting of historical recommendation data

[0025] Based on the calculated degree of interest, cognitive dissonance degree, and overexposure degree, comprehensively measure and fuse the user satisfaction score of cognitive perception, and complete the construction of the satisfaction model based on the user satisfaction score

[0026] Use MSE Loss as the loss function to optimize the satisfaction model and complete the training of the satisfaction model

[0027] The expression of the user satisfaction score is as follows

[0028]

[0029]

[0030]

[0031]

[0032] where represents the user satisfaction score represents the estimated degree of interest of user u in the recommended content i at time t t λ1 and λ2 respectively represent the adjustment coefficients of the cognitive dissonance degree d t and the overexposure degree e t f θ represents the deep learning model used to calculate the degree of interest, θ represents the model parameters, e udenotes the user embedding representation, denotes the embedding representation of the recommended content i t , α u denotes the cognitive level of user u, denotes the degree of interest of user u in the recommended content i at time t-1 t-1 , dist(,) is used to calculate the difference degree between two recommended contents, α u and both denote learnable parameters, S u denotes the complete historical interaction sequence of user u, t l denotes the time of the historical recommended content i l , t-t l denotes the time interval from the current time t, and τ denotes the temperature coefficient.

[0033] The beneficial effect of the above improvement scheme is that this scheme provides a satisfaction model integrating cognitive perception. By introducing three aspects of interest degree, cognitive misadjustment degree and overexposure degree for estimation, it comprehensively captures the cognitive bubble phenomenon in the interaction process between the recommendation system and the user, so as to accurately depict the user's satisfaction with the recommended content.

[0034] Furthermore, the S2 includes the following steps:

[0035] B1. Collect several interaction sequences with lengths greater than a given empirical value from the offline interaction data, and perform truncation processing on the length of each interaction sequence. Among them, the truncated interaction sequences are used for the multi-sample parallel training of the recommendation policy model and the denoising state updater;

[0036] B2. Make a decision according to the current state. For the user characteristics and historical interaction content characteristics obtained in A2, the user state representation s is obtained through the denoising state updater t ;

[0037] B3. Input the user state representation s t into the recommendation policy model to obtain the recommendation probability of the interaction content at the current moment;

[0038] B4. Use the trained satisfaction model as a simulation environment to score the satisfaction of the user with the interaction content at the current moment, and obtain the counterfactual reward by adjusting the overexposure coefficient and the cognitive bias coefficient, where the counterfactual reward is used to train and obtain the recommendation policy for eliminating cognitive bubbles;

[0039] B5. Traverse and calculate all interaction sequences to obtain the recommendation probabilities and counterfactual rewards of all interaction contents, and use the PPO loss function of the reinforcement learning algorithm to optimize the recommendation policy model. Take the recommendation probabilities and counterfactual rewards obtained in B3 and B4 as inputs to calculate the loss value, complete the training of the recommendation policy model and the denoising state updater, and learn the recommendation policy to alleviate the cognitive bubble effect.

[0040] The beneficial effects of the above improvement scheme are as follows: This scheme adopts an offline reinforcement learning framework, uses the interaction sequence data in the offline dataset to train the denoising state updater and the recommendation policy model, so that it can provide a recommendation policy to eliminate cognitive bubbles. Collect multiple interaction sequences from the offline interaction data and complete truncation for the multi-sample parallel training of the recommendation policy model and the denoising state updater; provide counterfactual rewards by dynamically adjusting the overexposure coefficient and the cognitive bias coefficient to achieve the adaptive balance of the recommendation policy between user interests and cognitive diversity; optimize through the reinforcement learning PPO algorithm to form a dynamic regulation mechanism that can eliminate cognitive bubbles while ensuring the stability of training.

[0041] Furthermore, S3 includes the following steps:

[0042] C1. According to the given sliding time window size, obtain the historical interaction behaviors of the user from the user's offline interaction data and the real-time data buffer to obtain user features and historical recommendation content features;

[0043] C2. Input the user features and historical recommendation content features obtained in C1 into the trained denoising state updater to obtain the current user state representation;

[0044] C3. Input the current user state representation into the trained recommendation policy model to obtain the recommended content and push it to the user.

[0045] The beneficial effects of the above improvement scheme are as follows: This improvement scheme further refines the steps for the recommendation policy model to generate recommended content in real-time interaction. By obtaining the user's recent historical behavior information through a sliding time window, it can more accurately capture the changing trend of the user's cognitive state; combining the denoising state updater to represent the user state can effectively filter out noise data and enhance the robustness of the state representation; after passing the current user state representation through the recommendation policy model, more accurate and positive recommended content is obtained, which promotes the elimination of cognitive bubbles while meeting the user's needs.

[0046] Furthermore, C1 includes the following steps:

[0047] Screen the historical interaction information of the current user from the user's offline interaction data and the real-time data buffer and sort it in chronological order;

[0048] According to the empirical value, set the size of a sliding time window, and based on the current user's historical interaction information sorted by timestamp, extract the user's recent historical interaction information, and perform normalization processing on the timestamp and the interaction record values;

[0049] Use the relevant fields in the interaction data as features for the current user and the historical recommended content within the time window respectively, and obtain the user feature and the historical recommended content feature through the encoding layer.

[0050] The beneficial effects of the above improvement scheme are as follows: This scheme refines the user's historical interaction information and provides a high-quality data basis. By screening and sorting the user's historical interaction information from the offline interaction data and the real-time data buffer, the timeliness and integrity of data processing are ensured; extracting the user's recent historical interaction information according to the sliding time window and performing normalization processing effectively reduces the influence of data noise; extracting features for the user and the historical recommended content respectively and processing them through the encoding layer are used for the denoising state updater to depict the user's cognitive state.

[0051] Furthermore, C2 includes the following steps:

[0052] According to the user feature and the historical recommended content feature, use the momentum update mechanism method to obtain the representation of the user's current cognitive state;

[0053] Input the representation of the user's current cognitive state into the denoising state updater to obtain the current user state representation.

[0054] The beneficial effects of the above improvement scheme are as follows: This scheme further clarifies the specific steps for obtaining the current user state representation. By using the momentum update mechanism method to obtain the representation of the user's current cognitive state, the dynamic change trend of the user's cognitive state can be effectively captured; inputting the representation of the user's current cognitive state into the denoising state updater, the denoised user state representation obtained provides high-quality input for the recommendation strategy model, which helps to generate positive and favorable recommended content.

[0055] Furthermore, C3 includes the following steps:

[0056] Input the current user state representation into the trained recommendation strategy model to obtain the recommendation probability distribution of all content;

[0057] Select the candidate content with the largest probability value from the recommendation probability distribution, and use it as the recommended content and push it to the user.

[0058] The beneficial effects of the above improvement solution are as follows: This solution further refines the process of generating and pushing recommended content. By inputting the current user state representation into the trained recommendation strategy model to obtain the recommendation probability distribution of all content, and then selecting the candidate content with the largest probability value as the recommended content, it can provide a recommendation strategy to alleviate the cognitive bubble and effectively balance the needs of user interests and cognitive diversity.

[0059] Furthermore, S4 includes the following steps:

[0060] After the data scale in the data buffer reaches the set threshold, the cached data in the data buffer is taken out, and the time stamps of the cached data and the numerical values of the interaction record fields are normalized.

[0061] The normalized cached data is used as fine-tuning data, and with a reinforcement learning framework, the interaction sequences in the fine-tuning dataset are used to train the recommendation strategy model and the denoising state updater, and the updated recommendation strategy model and denoising state updater are used for subsequent interactive recommendations to complete content recommendations for cognitive bubble elimination.

[0062] The beneficial effects of the above improvement solution are as follows: This solution provides a model fine-tuning method for real-time interaction processes, realizing the collaborative iterative optimization of the recommendation strategy model and the denoising state updater. Caching the real-time feedback records of users on recommended content and using them as fine-tuning data, combined with a reinforcement learning framework to fine-tune model parameters, completing the dynamic calibration of user cognitive representation and content recommendations for cognitive bubble elimination, effectively suppressing the solidification trend of cognitive bubbles and promoting the spread of positive information. Description of the Drawings

[0063] Figure 1 It is the flowchart of the method of the present invention.

[0064] Figure 2 It is the flowchart of calculating user satisfaction integrating cognitive perception in the embodiment of the present invention.

[0065] Figure 3 It is the flowchart of training the recommendation strategy model and the denoising state updater in the embodiment of the present invention. Detailed Embodiments

[0066] The following describes the detailed embodiments of the present invention to facilitate those skilled in the art of the present technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the detailed embodiments. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

[0067] Embodiment

[0068] As Figure 1 shown, the present invention provides a content recommendation method for cognitive bubble elimination, including an offline model training stage and a real-time interaction and model update stage. The implementation method is as follows:

[0069] The offline model training stage includes the following steps:

[0070] S1. Extract the key features of the user and the recommended content from the obtained offline interaction data, and construct and train a satisfaction model integrating cognitive perception from the aspects of interest degree, overexposure degree, and cognitive mismatch degree. The implementation method is as follows:

[0071] A1. Obtain offline interaction data;

[0072] In this embodiment, the offline interaction data includes the interaction logs of the user and the historical recommended content on the platform, as well as the relevant fields of the user and the platform, including the timestamp, user click behavior, user basic information, and other contents.

[0073] A2. Based on the obtained offline interaction data, extract the features and embedding representations of the user and the historical interaction content. The implementation method is as follows:

[0074] For the user and historical recommended content information in the offline interaction data, through correlation analysis, select the most important relevant field information to obtain the user feature x u and the recommended content feature x i ;

[0075] According to the obtained user features and historical interaction content features, through the encoding layer, obtain the user embedding representation e u and the historical interaction content embedding representation e i .

[0076] A3. Based on the results obtained in A2, estimate through three dimensions of interest degree, overexposure degree, and cognitive mismatch degree, and construct and train a satisfaction model integrating cognitive perception. The implementation method is as follows:

[0077] Estimate the interest degree of the user. In this embodiment, use the recommendation model to estimate the user's interest degree. The interest degree refers to the preference degree of the user for the recommended content during the interaction process without being affected by other factors:

[0078]

[0079] Estimate the cognitive mismatch degree through the degree of change in the user's interest at adjacent times. In this embodiment, estimate the cognitive mismatch degree through the degree of change in the user's interest at adjacent times. The cognitive mismatch degree reflects the degree of inconsistency in the user's interaction experience. A certain degree of decline in interest after interaction can reflect the user's psychological gap to a certain extent. The cognitive mismatch degree d of the user at time tt Can be expressed as:

[0080]

[0081] Estimate the overexposure of recommended content through the similarity weighting of historical recommendation data. In this embodiment, estimate the overexposure of recommended content through the similarity weighting of historical recommendation data. The overexposure of recommended content reflects the disgust that users may feel when repeatedly exposed to similar recommended content over a period of time, thus affecting their overall satisfaction. The overexposure e t Can be expressed as:

[0082]

[0083] Based on the calculated interest degree, cognitive dissonance degree, and overexposure, comprehensively measure the user satisfaction score that fuses cognitive perception, and based on the user satisfaction score, complete the construction of the satisfaction model:

[0084]

[0085] Wherein, Represents the user satisfaction score, Represents the interest degree estimation of user u for the recommended content i at time t t The adjustment coefficients of the cognitive dissonance degree d t And the overexposure e t Are respectively represented by λ1 and λ2, and f θ Represents the deep learning model used to calculate the interest degree, θ represents the model parameters, and e u Represents the user embedding representation, Represents the embedding representation of the recommended content i t And α u Represents the cognitive level of user u, Represents the interest degree estimation of user u for the recommended content i at time t - 1 t-1 And dist(,) is used to calculate the difference degree between two recommended contents, and α u And Both represent learnable parameters, and S u Represents the complete historical interaction sequence of user u, and t l Represents the historical recommended content i l The time of, and t - t l Represents the time interval from the current time t, and τ represents the temperature coefficient.

[0086] Use the MSE Loss as the loss function to optimize the satisfaction model and complete the training of the satisfaction model;

[0087] S2. Use the reinforcement learning model and combine the counterfactual rewards provided by the satisfaction model to train the recommendation policy model and the denoising state updater. The implementation method is as follows:

[0088] B1. Collect several interaction sequences with lengths greater than a given empirical value from the offline interaction data, and truncate the length of each interaction sequence. Among them, the truncated interaction sequences are used for the multi-sample parallel training of the recommendation policy model and the denoising state updater.

[0089] In this embodiment, multiple interaction sequences with lengths greater than a given empirical value N are collected from the offline dataset, and the length of each sequence is truncated to N for the multi-sample parallel training of the recommendation policy model and the state updater.

[0090] B2. Make a decision according to the current state. For the user features and historical interaction content features obtained in A2, the user state representation s is obtained through the denoising state updater. t ;

[0091] In this embodiment, the recommendation policy based on reinforcement learning needs to make a decision according to the current state. The obtained user features and historical interaction content features are used to obtain the user state representation s through the denoising state updater. t 。

[0092] B3. Input the user state representation s t into the recommendation policy model to obtain the recommendation probability of the interaction content at the current moment.

[0093] B4. Use the trained satisfaction model as a simulation environment to score the satisfaction of the user with the interaction content at the current moment, and obtain the counterfactual reward by adjusting the cognitive bias coefficient λ1 and the overexposure coefficient λ2. Among them, the counterfactual reward is used to train the recommendation policy for eliminating cognitive bubbles.

[0094] B5. Traverse and calculate all interaction sequences to obtain the recommendation probabilities and counterfactual rewards of all interaction contents, and use the PPO loss function of the reinforcement learning algorithm to optimize the recommendation policy model. Use the recommendation probabilities and counterfactual rewards obtained in B3 and B4 as inputs to calculate the loss value, and complete the training of the recommendation policy model and the denoising state updater to learn the recommendation policy for alleviating the cognitive bubble effect.

[0095] The real-time interaction and model update phase includes the following steps:

[0096] S3. Use a sliding time window to obtain the user's historical behavior information, and combine it with the denoising state updater to obtain the current user state representation. Input the current user state representation into the recommendation policy model to obtain the recommended content. The implementation method is as follows:

[0097] C1. Obtain the historical interaction behaviors of the user from the user's offline interaction data and real-time data buffer according to the given sliding time window size, and obtain user features and historical recommended content features. The implementation method is as follows:

[0098] Screen out the historical interaction information of the current user from the user's offline interaction data and real-time data buffer, and sort it in chronological order. In this embodiment, screen out the historical interaction information of the current user from the offline interaction data and the online interaction buffer (i.e., the real-time data buffer), and arrange it in chronological order. Each interaction data includes fields such as user ID, content ID, timestamp, and interaction behavior. Among them, the interaction behavior may include record information such as browsing duration and liking;

[0099] According to the empirical value, set a sliding time window size, and extract the most recent historical interaction information of the user based on the historical interaction information of the current user sorted by timestamp, and perform normalization processing on the timestamp and interaction record values;

[0100] Use the relevant fields in the interaction data as features for the current user and the historical recommended content within the time window respectively, and obtain user features and historical recommended content features through the encoding layer;

[0101] C2. Input the user features and historical recommended content features obtained in C1 into the trained denoising state updater to obtain the current user state representation. The implementation method is as follows:

[0102] According to the user features and historical recommended content features, use the momentum update mechanism method to obtain the current cognitive state representation of the user. The cognitive state representation at the initial moment is the self-feature vector;

[0103] Input the current cognitive state representation of the user into the denoising state updater to obtain the current user state representation;

[0104] C3. Input the current user state representation into the trained recommendation policy model to obtain recommended content and push it to the user. The implementation method is as follows:

[0105] Input the current user state representation into the trained recommendation policy model to obtain the recommended probability distribution of all content;

[0106] Select the candidate content with the largest probability value from the recommended probability distribution as the recommended content and push it to the user;

[0107] S4. Record the real-time feedback of the user on the current recommended content in the data buffer for fine-tuning the model, and complete the content recommendation for cognitive bubble elimination. The implementation method is as follows:

[0108] When the data scale in the data buffer reaches the set threshold, the cached data in the data buffer is retrieved, and the timestamps of the cached data and the numerical values of the interaction record fields are normalized;

[0109] The normalized cached data is used as fine-tuning data, and in the reinforcement learning framework, the interaction sequences in the fine-tuning dataset are used to train the recommendation policy model and the denoising state updater, and the updated recommendation policy model and denoising state updater are used for subsequent interaction recommendations to complete content recommendation for cognitive bubble elimination.

[0110] In this embodiment, the present invention aims to propose a content recommendation method for cognitive bubble elimination, which is used to alleviate the cognitive bias phenomenon existing in the long-term interaction between users and the recommendation system, so as to guide users to establish a positive and balanced cognitive system. It is trained based on historical interaction data to decouple and analyze the cognitive bubble phenomenon existing in the interaction process, ensuring that the recommendation results can break through the inherent cognitive limitations of users and promote the spread of correct ideas and content. It multi-dimensionally characterizes user satisfaction from three aspects: intrinsic interest, cognitive dissonance degree, and overexposure degree, and integrates the representation of the cognitive state, which can accurately represent the degree of cognitive effect of users in the interaction process, thus effectively capturing the dynamic changes of the cognitive state. In addition, in the reinforcement learning training stage, by providing counterfactual rewards and combining the denoising state updater to capture state changes, a strategy that can effectively alleviate cognitive dissonance can be trained.

[0111] Taking the example of personalized content recommendation for users on a certain short video platform, this example constructs a satisfaction model for cognitive perception through three aspects: interest degree, cognitive dissonance degree, and overexposure degree, and defines the cognitive state of users to effectively capture the state changes of users in the interaction process. Finally, it is trained based on the reinforcement learning framework, enabling the model to execute a recommendation strategy for effective cognitive bubble elimination and interact with users, including the offline model training stage and the real-time interaction and model update stage.

[0112] Step D: Offline model training stage:

[0113] D1. Obtain the offline interaction data collected on the corresponding platform, which includes the interaction logs of users with the historical recommended video content on the platform, as well as the relevant fields of users and platform content, such as the main information such as the timestamps and viewing durations when users interact with the recommended videos, and some key information for feature extraction of users and recommended videos. At the same time, the normalized proportion of user viewing duration is used as the label information of satisfaction;

[0114] D2. Based on the offline interaction data obtained in step D1, implement the extraction of important features and embedding representation of users and historical interaction content;

[0115] D3. Based on the historical interaction data and embedding representations obtained in steps D1 and D2, and considering the influencing factors of cognitive effects, a user satisfaction model is constructed. The process is as follows: Figure 2 As shown: It is estimated from three aspects of interest degree, cognitive dissonance degree, and overexposure degree, and a satisfaction model integrating cognitive perception is constructed and trained;

[0116] D4. Adopt a reinforcement learning framework, and use the interaction sequences in the offline dataset to train the recommendation policy model and the denoising state updater. The process is as follows: Figure 3 As shown: Traverse each complete interaction sequence in chronological order, use the denoising state updater to obtain the user state representation at the current interaction moment, and then input it into the recommendation policy model to obtain the recommendation probability of the current interaction content. In addition, use the satisfaction model trained in step D3 as a simulation environment to provide counterfactual rewards for the recommended content of the recommendation policy model, so as to train a recommendation policy that can effectively eliminate cognitive bubbles.

[0117] Step E. Real-time interaction and model update:

[0118] E1. According to the given sliding time window size, obtain the historical interaction behaviors of the user from the user's offline interaction data and real-time data buffer, and obtain the user and the feature of the historical recommended content;

[0119] E2. Input the user features and the features of the historical recommended content obtained in step E1 into the trained denoising state updater to obtain the updated user state representation in the current round;

[0120] E3. Input the current user state representation into the trained recommendation policy model, output the recommended content and push it to the user;

[0121] E4. After the user interacts with the recommended content obtained in step E3, corresponding interaction behaviors will be made as the user's feedback on the system's recommended content, and record the feedback information in the data buffer;

[0122] E5. When the data scale in the buffer reaches the threshold, according to the model training method in step D4, use the data in the buffer to fine-tune and optimize the recommendation policy model and the state updater to ensure that the subsequent interaction recommendation results can break through the user's inherent cognitive limitations and help the user establish an objective cognitive system.

[0123] The following further describes the specific implementation means of each step in detail:

[0124] Before calculating the user satisfaction and updating the state, it is necessary to extract the user features and the features of the video content through the steps in D2, which specifically include the following steps:

[0125] D21. For the user and content information in the interaction data, through correlation analysis, select the most important relevant field information to obtain the corresponding user characteristics and content characteristics;

[0126] D211. Among the offline interaction data collected, it includes the main information related to users such as ID, gender, age, hobby field, etc., and for videos, it includes the main information such as ID, field, label, video duration, etc. Based on the relevant information of users and videos, use the Spearman correlation coefficient to evaluate the correlation between each field and the target variable;

[0127] D212. Based on the correlation analysis results in step D211, set a threshold of 0.5, select the relevant fields crucial for predicting user interests, and thus obtain the finally used user characteristics and content characteristics where n and m respectively represent the number of characteristics of user u and video content i;

[0128] D22. Based on the characteristics extracted in step D21, through the conversion of the encoding layer, map the user characteristics and interactive video characteristics into a high-dimensional space respectively to obtain their compact and semantically rich vector representations, denoted as e u and e i , that is, the embedding representations of users and historical interaction content.

[0129] In step D3, the user satisfaction is comprehensively evaluated from three aspects: interest degree, overexposure degree, and cognitive dissonance degree. This satisfaction model is trained based on historical interaction data to decouple and analyze the cognitive effects in the interaction process. At the same time, the trained user satisfaction model can generate effective strategies by providing counterfactual rewards. The specific steps are as follows:

[0130] D31. By inputting the user and the recommended video content, estimate the user's interest degree through the recommendation model. The interest degree refers to the preference degree of the user for the recommended content during the interaction process without being affected by other factors, and a traditional recommendation model can be used for modeling, which can be expressed as:

[0131]

[0132] In this embodiment, the DeepFM (Deep Factorization Machine) model that combines the factorization machine and deep learning method is used to calculate the user preference degree, and the specific operations are as follows:

[0133] (1) Concatenate the vector representations of the user and video content encoded in step D2 as the input of the satisfaction model;

[0134] (2) After passing the concatenated vector through the DeepFM model, output the interest degree value of the user for the recommended content in the t-th round;

[0135] D32. Estimate the cognitive dissonance degree by calculating the degree of change in user interest at adjacent times. The cognitive dissonance degree reflects the degree of experience inconsistency in the user interaction process. The decrease in interest after interaction can, to a certain extent, reflect the psychological gap of the user. At time t, the cognitive dissonance degree d of the user t can be expressed as:

[0136]

[0137] D33. Estimate the overexposure degree of the recommended content by weighted calculation of the similarity of historical recommendation data, which reflects the disgust that may be generated when the user repeatedly contacts similar videos within a period of time, thereby affecting their overall satisfaction. The overexposure degree e t can be expressed as:

[0138]

[0139] D34. Based on the interest degree cognitive dissonance degree d t and overexposure degree e t calculated in steps D31 - D33, comprehensively measure the user satisfaction can be expressed as:

[0140]

[0141] Among them, represents the user satisfaction score, represents the estimated interest degree of user u in the recommended content i at time t t , λ1 and λ2 respectively represent the adjustment coefficients of the cognitive dissonance degree d t and overexposure degree e t , f θ represents the deep learning model used to calculate the interest degree, θ represents the model parameters, e u represents the user embedding representation, represents the embedding representation of the recommended content i t , α u represents the cognitive level of user u, represents the estimated interest degree of user u in the recommended content i at time t - 1 t-1 , dist(,) is used to calculate the difference degree between two recommended contents, α u and both represent learnable parameters, S u represents the complete historical interaction sequence of user u, t l represents the time of the historical recommended content i l , t - t lrepresents the time interval from the current moment t, and τ represents the temperature coefficient.

[0142] During the training phase, both λ1 and λ2 are set to 1 to perceive the cognitive bubble phenomenon from the existing data; during the reinforcement learning training phase, the values of λ1 and λ2 can be adjusted to provide counterfactual rewards to help the model learn strategies for eliminating cognitive bubbles.

[0143] D35. Use the viewing duration ratio in the historical interaction data as the user satisfaction label, and use the MSE Loss function to optimize the output result of the user satisfaction model to achieve the perception of the cognitive bubble phenomenon. The calculation process of the loss value Loss can be expressed as:

[0144]

[0145] where, y i and respectively represent the label and estimated value of the user satisfaction in the i-th round.

[0146] Based on the reinforcement learning framework, step D4 obtains the state representation of the user during the interaction through a denoising state updater, and uses the trained coefficient of the user satisfaction model to provide counterfactual rewards, thereby obtaining an effective strategy for eliminating cognitive bubbles. The specific steps are as follows:

[0147] D41. Collect multiple interaction sequences with lengths greater than the given empirical value N = 30 from the offline dataset, and truncate the length of each sequence to N for the multi-sample parallel training of the recommendation policy model.

[0148] D42. The recommendation policy based on reinforcement learning needs to make decisions according to the current state. For the user features and interaction content features obtained in step D2, after passing through the denoising state updater, the state representation s of the user at the current moment is obtained t ;

[0149] D421. For the cognitive state of the user, initially represent the user vector e u , and map it to the cognitive state space using a feedforward neural network, expressed as For subsequent moments t, use a gating mechanism to extract the weights of the previous moment's cognitive state representation and the recommended content representation at the current moment, and then use a momentum update mechanism to represent the user's cognitive state, which can be expressed as:

[0150]

[0151] where, the gating vector g t is defined as follows:

[0152]

[0153] Among them, σ represents the activation function, concat represents the vector concatenation operation, and proj1 and proj2 are feed-forward neural network calculation operations;

[0154] D422. Input the cognitive state sequence feature e c into the adaptive denoising unit to filter out the high-frequency noise in the user's historical interaction data, thereby obtaining an accurate user state representation f u . The adaptive denoising unit is a method that adjusts the threshold according to the spectral characteristics to remove the high-frequency noise of the interaction sequence, and specifically includes the following operations:

[0155] (1) Through the fast Fourier transform, convert the cognitive state sequence feature e c to obtain the corresponding frequency-domain representation F;

[0156] (2) First, use the formula P = |F| 2 to calculate the power spectrum of the frequency-domain feature, identify the main frequency component, and then use the learnable parameter θ to filter out the high-frequency noise during training. This process is expressed as:

[0157] F′ = F ⊙ P < θ;

[0158] Among them, ⊙ is the element-wise multiplication operation, and the binary mask matrix P < θ is used to extract the frequency features with a density lower than the threshold θ;

[0159] (3) Use two learnable filters to process the original frequency feature F and the denoised feature F', and integrate the processed features to obtain a more comprehensive frequency-domain representation. The integrated feature F I is expressed as:

[0160] F I = W G ⊙ F + W L ⊙ F′;

[0161] Among them, W G and W L respectively represent the global and local trainable filters;

[0162] (4) Through the inverse fast Fourier transform, convert the integrated frequency domain to the time domain to obtain the denoised cognitive state f c ;

[0163] D423. The denoised cognitive state f c captures the historical temporal dependence through the GRU (gated recurrent unit) model, and finally obtains the state at the current moment;

[0164] D43. After passing the user's current moment state representation obtained in step D42 through the recommendation policy network model, the recommendation probability distribution of the model for all contents is output, and the recommendation probability for the interactive video content in the current round is taken out.

[0165] D44. Use the user satisfaction model trained in step D34 as the simulation environment to score the satisfaction of the user with the recommended video content at the current moment, and by adjusting the cognitive bias λ1 = 5, obtain the counterfactual reward for training the policy.

[0166] D45. After traversing the interaction sequence and obtaining the recommendation probabilities and counterfactual rewards of all interaction contents, use the loss function of the reinforcement learning algorithm PPO to optimize the parameters of the recommendation policy model and the denoising state updater. Take the reward values and content recommendation probabilities obtained in steps D43 - D44 as inputs to calculate the loss value and optimize the model parameters, so as to learn a recommendation policy that can eliminate cognitive bubbles. The PPO loss function J(θ) can be expressed as:

[0167]

[0168] Among them, represents the action probability ratio between the new policy and the old policy, E t represents the expectation of sampling the interaction trajectory based on the current policy, represents the estimated advantage function of the state - action pair (s t ,a t ), represents the hyperparameter used to control the step size of policy update. The objective function includes a clipping term clip(, ) to ensure that the update is not too radical and prevent significant performance fluctuations.

[0169] After training the denoising state updater and the recommendation policy model based on the above steps, real - time interaction with the user can be carried out. This recommendation policy model can not only give personalized recommendations based on user and content features, but also the recommended policy can avoid the occurrence of cognitive bubble phenomena during the interaction process, helping users establish a positive and balanced cognitive system, specifically including:

[0170] For a specific user, in step E1, obtain his historical interaction information from the offline data and the data cached in the real - time interaction process, which specifically includes the following operations:

[0171] E11. Filter out all the historical interaction information of the current user from the offline interaction data and the cached data of real-time interaction, and arrange it in chronological order. Each interaction data contains fields such as user ID, content ID, timestamp, and interaction behavior. Among them, the interaction behavior can include record information such as browsing duration, viewing duration ratio, and liking. In this example, the viewing duration ratio after normalization is selected as the record information in the interaction behavior field;

[0172] E12. Set a reasonable time window size M, extract the most recent interaction records of this user from all the historical interaction information, and normalize the timestamp and the interaction behavior values. In this example, the window size M is set to 30, that is, the calculation and recommendation are based on the most recent 30 interaction records;

[0173] E13. Follow the process of step D2, use the relevant fields in the interaction data as features for the user and the historical recommended video content within the time window respectively, and obtain the representations of the user and each recommended content through the encoding layer;

[0174] Before each round of interactive recommendation, use the user feature and the historical recommended content feature obtained in step E1 as inputs, and obtain the updated user state representation through the trained denoising state updater. The specific operations are as follows:

[0175] E21. For the user feature and the historical recommended content feature, follow the momentum update mechanism method in step D421 to obtain the current cognitive state representation of the user. The cognitive state representation of the user at the initial moment is the vector representation of its own feature;

[0176] E22. Follow the method in step D422, and pass the user cognitive state obtained in step E21 through the adaptive denoising unit to obtain the user state representation of the current round;

[0177] In step E3, based on the current user state representation, use the trained recommendation strategy model to obtain the current recommended video. The recommended content provided can promote the spread of correct ideas and content, thereby avoiding the aggravation of the cognitive bias phenomenon during the long-term interaction between the user and the recommendation system. The specific operations are as follows:

[0178] E31. Follow the method in step D43, pass the current user state representation through the trained recommendation strategy model, and the output is the recommendation probability distribution of the model for all content;

[0179] E32. Select the candidate content with the largest probability value from the probability distribution, and push this content to the user.

[0180] In step E5, it specifically includes:

[0181] E51. After the data scale in the buffer reaches the set threshold, take out this part of the data and normalize the values of its timestamp and interaction record fields;

[0182] E52. Use the normalized cached data as fine-tuning data and optimize the parameters of the recommendation strategy model and the state updater according to the training process in step D4;

[0183] E53. Use the updated recommendation strategy model and state updater in subsequent rounds of interaction.

[0184] The recommended content provided by the trained recommendation strategy model can effectively eliminate the cognitive bubble effect, and use the cached data of online interaction to fine-tune and optimize the parameters of the recommendation strategy model and the state updater, ensuring that the potential needs and cognitive states of users can be mined during the long-term interaction process, so as to guide users to establish a positive and balanced cognitive system.

Claims

1. A content recommendation method for cognitive bubble elimination, characterized in that It includes an offline model training stage and a real-time interaction and model update stage; The offline model training stage includes the following steps: S1. Extract the key features of the user and the recommended content from the obtained offline interaction data, and construct and train a satisfaction model that integrates cognitive perception from the aspects of interest degree, overexposure degree, and cognitive dissonance degree; S2. Adopt a reinforcement learning model, and combine the counterfactual rewards provided by the satisfaction model to train a recommendation strategy model and a denoising state updater; The real-time interaction and model update stage includes the following steps: S3. Use a sliding time window to obtain the user's historical behavior information, and combine it with the denoising state updater to obtain the current user state representation, and input the current user state representation into the recommendation strategy model to obtain the recommended content; S4. Record the user's real-time feedback on the current recommended content in the data buffer for fine-tuning the model, and complete the content recommendation for cognitive bubble elimination.

2. The content recommendation method for cognitive bubble resolution according to claim 1, wherein The S1 includes the following steps: A1. Obtain offline interaction data; A2. Based on the obtained offline interaction data, extract the features and embedding representations of the user and the historical interaction content; A3. Based on the results obtained in A2, estimate through three dimensions of interest degree, overexposure degree, and cognitive dissonance degree, and construct and train a satisfaction model that integrates cognitive perception.

3. The content recommendation method for cognitive bubble resolution according to claim 2, wherein The A2 includes the following steps: For the user and historical recommendation content information in the offline interaction data, through correlation analysis, select the most important relevant field information to obtain the user feature x u and the recommendation content feature x i ; According to the obtained user characteristics and historical interaction content characteristics, the user embedding representation e is obtained through the encoding layer u and the historical interaction content embedding representation e i .

4. The content recommendation method for cognitive bubble resolution according to claim 2, characterized in that The A3 includes the following steps: Use a traditional recommendation model to estimate the user's interest degree; Estimate the cognitive dissonance degree through the degree of change in the user's interest at adjacent times; Estimate the overexposure degree of the recommended content through the similarity weighting of historical recommendation data; Based on the calculated interest degree, cognitive dissonance degree, and overexposure degree, comprehensively measure the user satisfaction score that integrates cognitive perception, and based on the user satisfaction score, complete the construction of the satisfaction model; Use the MSE Loss as the loss function to optimize the satisfaction model and complete the training of the satisfaction model; The expression of the user satisfaction score is as follows: Among them, represents the user satisfaction score, represents the estimated interest of user u in recommended content i at time t t The adjustment coefficients of the cognitive mismatch degree d t and overexposure degree e are represented by λ1 and λ2 respectively, t f θ represents the deep learning model used to calculate the interest degree, θ represents the model parameters, u e represents the user embedding representation, represents the embedding representation of recommended content i t α u represents the cognitive level of user u, represents the estimated interest of user u in recommended content i at time t - 1 t-1 dist(,) is used to calculate the difference degree between two recommended contents, u and both represent learnable parameters, u S represents the complete historical interaction sequence of user u, l t represents the time of historical recommended content i l t - t l represents the time interval from the current time t, and τ represents the temperature coefficient.

5. The content recommendation method for cognitive bubble resolution according to claim 2, wherein The S2 includes the following steps: B1. Collect several interaction sequences with lengths greater than a given empirical value from the offline interaction data, and truncate the length of each interaction sequence. Among them, the truncated interaction sequences are used for the multi-sample parallel training of the recommendation strategy model and the denoising state updater; B2. Make a decision according to the current state. For the user characteristics and historical interaction content characteristics obtained by A2, the user state representation s is obtained through a denoising state updater. t ; B3. Input the user state representation s t into the recommendation strategy model to obtain the recommendation probability of the interaction content at the current moment; B4. Use the trained satisfaction model as a simulation environment to score the satisfaction of the user and the current moment's interaction content, and obtain counterfactual rewards by adjusting the overexposure coefficient and the cognitive bias coefficient. Among them, the counterfactual rewards are used to train a recommendation strategy for eliminating cognitive bubbles; B5. Traverse and calculate all interaction sequences to obtain the recommendation probabilities and counterfactual rewards of all interaction contents, and use the PPO loss function of the reinforcement learning algorithm to optimize the recommendation strategy model, and use the recommendation probabilities and counterfactual rewards obtained in B3 and B4 as inputs to calculate the loss value, and complete the training of the recommendation strategy model and the denoising state updater to learn a recommendation strategy for alleviating the cognitive bubble effect.

6. The content recommendation method for cognitive bubble elimination according to claim 1, wherein The S3 includes the following steps: C1. Obtain the historical interaction behaviors of the user from the user's offline interaction data and real-time data buffer according to the given sliding time window size, and obtain user features and historical recommendation content features; C2. Input the user features and historical recommendation content features obtained in C1 into the trained denoising state updater to obtain the current user state representation; C3. Input the current user state representation into the trained recommendation policy model to obtain recommended content and push it to the user.

7. The content recommendation method for cognitive bubble elimination according to claim 6, characterized in that The C1 includes the following steps: Screen out the historical interaction information of the current user from the user's offline interaction data and real-time data buffer and sort it in chronological order; According to the empirical value, set a sliding time window size, and extract the most recent historical interaction information of the user based on the current user's historical interaction information sorted by timestamp, and normalize the timestamp and the interaction record value; Use the relevant fields in the interaction data as features for the current user and the historical recommendation content within the time window respectively, and obtain user features and historical recommendation content features through the encoding layer.

8. The content recommendation method for cognitive bubble elimination according to claim 6, wherein The C2 includes the following steps: According to the user features and historical recommendation content features, obtain the current cognitive state representation of the user by using the momentum update mechanism method; Input the current cognitive state representation of the user into the denoising state updater to obtain the current user state representation.

9. The content recommendation method for cognitive bubble elimination according to claim 6, wherein The C3 includes the following steps: Input the current user state representation into the trained recommendation policy model to obtain the recommendation probability distribution of all content; Select the candidate content with the largest probability value from the recommendation probability distribution as the recommended content and push it to the user.

10. The content recommendation method for cognitive bubble elimination according to claim 1, wherein The S4 includes the following steps: After the data scale in the data buffer reaches the set threshold, take out the cached data in the data buffer and normalize the timestamp and the value of the interaction record field of the cached data; Use the normalized cached data as fine-tuning data, and adopt a reinforcement learning framework to train the recommendation policy model and the denoising state updater using the interaction sequences in the fine-tuning dataset, and use the updated recommendation policy model and denoising state updater for subsequent interactive recommendations to complete content recommendations for cognitive bubble elimination.

Images