Content recommendation method and device based on deviation correction, equipment and medium

Abstract

The invention relates to a content recommendation method and device based on deviation correction, equipment and a medium, and the method comprises the steps: determining the exposure probability of each content through a deep survival analysis model, determining the weight coefficient of each content according to the exposure probability of each content, and determining a joint loss function through combining with a dual robust estimation algorithm, and training the initial recommendation model through the joint loss function so as to generate a content recommendation result through the trained target recommendation model. Visibly, the weight coefficient of each content is determined through the inverse tendency weighting algorithm, and the joint loss function is determined by introducing the dual robust estimation algorithm, so that the generalization ability can be enhanced while deviation correction is performed, and the recommendation quality of the recommendation model in the cold start content is improved; moreover, the exposure probability of the content is estimated through the deep survival analysis model, complex time sequence and high-dimensional sparse features can be processed, the accuracy and robustness of the exposure probability are improved, a reliable basis is provided for subsequent weighting, and deviation is remarkably reduced.

Description

Technical Field

[0001] The present application relates to the technical field of content recommendation, and in particular to a content recommendation method, apparatus, device, and medium based on deviation correction. Background Art

[0002] On large-scale online video platforms, recommendation systems are a core technology that connects users with content, directly impacting user retention, platform revenue, and the health of the content ecosystem. By mining user interest and behavior data, recommendation systems provide personalized recommendations for videos of interest to users.

[0003] Currently, mainstream recommendation systems are generally based on a supervised learning paradigm, using user clicks or viewing behaviors as supervisory signals to train prediction models such as CTR (Click-Through Rate) and VT (Validation Test). However, this training approach suffers from exposure bias and feedback bias.

[0004] The exposure bias problem refers to the fact that model training data relies on historical recommendation lists, which are themselves affected by the platform's recommendation strategy and have a bias of "only looking at popular content", resulting in a lack of effective exposure of long-tail content and new content, making it impossible to obtain accurate user feedback, forming a "cold start dilemma"; the feedback bias problem refers to the fact that user behavior is affected by multiple factors such as whether the content is exposed, the order of display, the user's current status, etc. Directly using click data as a supervision signal will introduce bias, reducing the generalization ability and fairness of the recommendation effect.

[0005] Therefore, how to correct the above deviations and improve the recommendation quality of the recommendation system in cold start content is a problem that those skilled in the art need to solve. Summary of the Invention

[0006] The present application provides a content recommendation method, apparatus, device, and medium based on bias correction to correct exposure bias and feedback bias, thereby improving the recommendation quality of the recommendation system in cold start content.

[0007] In a first aspect, the present application provides a content recommendation method based on bias correction, comprising:

[0008] Determine the exposure probability of each content in the training data through a deep survival analysis model;

[0009] Determine the weight coefficient of each content according to the exposure probability of each content; wherein the exposure probability of each content is inversely proportional to the weight coefficient of each content;

[0010] Determine a joint loss function based on the weight coefficients of each content and a dual robust estimation algorithm; the joint loss function includes a dual robust loss function and at least one target loss function;

[0011] The initial recommendation model is trained using the joint loss function and the training data, so that a content recommendation result is generated using the trained target recommendation model.

[0012] Optionally, determining the weight coefficient of each content according to the exposure probability of each content includes:

[0013] Use quantiles to bucket each exposure probability;

[0014] Perform probability calibration on the exposure probability of each bucket to obtain the calibrated exposure probability;

[0015] The weight coefficient of each content is generated according to the calibrated exposure probability.

[0016] Optionally, the exposure probability of each bucket is calibrated, and the calibrated exposure probability includes:

[0017] The exposure probability of each bucket is calibrated using the temperature scaling algorithm or the Platt Scaling algorithm to obtain the calibrated exposure probability.

[0018] Optionally, a weight coefficient for each content is generated according to the calibrated exposure probability, including:

[0019] Determine the initial weight of each content based on the clipping upper limit value;

[0020] The initial weights of all contents are normalized to generate the weight coefficient of each content.

[0021] Optionally, based on the weight coefficients of each content and a dual robust estimation algorithm, a joint loss function is determined, including:

[0022] Determine a dual robust loss function based on the weight coefficient of each content and the dual robust estimation algorithm; wherein the dual robust loss function includes: an IPW weighted loss function and a model estimation loss function;

[0023] Determine the fairness loss function and diversity loss function;

[0024] A joint loss function is determined according to the dual robust loss function, the fairness loss function, the diversity loss function, a weight coefficient of the fairness loss function, and a weight coefficient of the diversity loss function.

[0025] Optionally, generating content recommendation results using the trained target recommendation model includes:

[0026] A recommendation list is generated through a target recommendation model and an exploration strategy; wherein the exploration strategy is a UCB exploration strategy or an ε-greedy exploration strategy.

[0027] Optionally, after generating a recommendation list using the target recommendation model and the exploration strategy, the following steps are also included:

[0028] Calculating an entropy value of the recommendation list;

[0029] If the entropy value is less than a predetermined threshold, the exploration rate of the cold start content is increased, and a recommendation list is regenerated using the target recommendation model and the exploration strategy; wherein the exploration rate is a parameter value of the exploration strategy.

[0030] In a second aspect, the present application provides a content recommendation device based on deviation correction, comprising:

[0031] The exposure probability determination module is used to determine the exposure probability of each content in the training data through a deep survival analysis model;

[0032] A weight coefficient determination module, configured to determine a weight coefficient for each content based on the exposure probability of each content; wherein the exposure probability of each content is inversely proportional to the weight coefficient of each content;

[0033] A loss function determination module, configured to determine a joint loss function based on weight coefficients of each content and a dual robust estimation algorithm; the joint loss function includes a dual robust loss function and at least one target loss function;

[0034] A training module, configured to train the initial recommendation model using the joint loss function and the training data;

[0035] The recommendation module is used to generate content recommendation results through the trained target recommendation model.

[0036] In a third aspect, the present application provides an electronic device, comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0037] Memory for storing computer programs;

[0038] The processor is configured to implement the above-mentioned content recommendation method steps when executing the program stored in the memory.

[0039] In a fourth aspect, the present application further provides a computer storage medium storing computer-executable instructions for executing the steps of the above-mentioned content recommendation method.

[0040] The above technical solutions provided by the embodiments of the present application have the following advantages over the prior art: the present application discloses a content recommendation method, apparatus, device and medium based on bias correction, the present application determines the exposure probability of each content through a deep survival analysis model, determines the weight coefficient of each content according to the exposure probability of each content, and then determines the joint loss function in combination with a dual robust estimation algorithm, and then trains the initial recommendation model through the joint loss function so as to generate content recommendation results through the trained target recommendation model. It can be seen that the present application determines the weight coefficient of each content through an inverse tendency weighting algorithm and introduces a dual robust estimation algorithm to determine the joint loss function, which can enhance the generalization ability while correcting the bias and improve the recommendation quality of the recommendation model for cold start content; and, the present application estimates the exposure probability of content through a deep survival analysis model, which can handle complex temporal and high-dimensional sparse features, improve the accuracy and robustness of the exposure probability, provide a reliable basis for subsequent weighting, and significantly reduce the bias. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0043] One or more embodiments are exemplarily illustrated by pictures in the corresponding drawings. These exemplifications do not constitute limitations on the embodiments. Elements with the same reference numerals in the drawings are represented as similar elements. Unless otherwise stated, the figures in the drawings do not constitute proportional limitations.

[0044] Figure 1 A flowchart of a content recommendation method based on deviation correction provided in an embodiment of the present application;

[0045] Figure 2 A schematic diagram of the structure of a content recommendation device based on deviation correction provided in an embodiment of the present application;

[0046] Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] In the existing technology, recommendation systems are generally based on the supervised learning paradigm, using user click or viewing behavior as a supervisory signal to train the prediction model. However, this training method has exposure bias and feedback bias problems. At present, some solutions use the Inverse Probability Weighting (IPW) method to correct these biases. However, the use of the IPW method has serious problems of weight instability and high variance, especially when the exposure probability is estimated inaccurately or at a very low value. Doubly Robust Estimation, as a bias correction method that combines model prediction and propensity estimation, has stronger robustness, but its implementation in the recommendation system still faces challenges. In addition, the recommendation system's over-optimization of indicators such as CTR is likely to sacrifice content diversity and fairness, leading to ecological imbalance and a decline in user experience.

[0048] Therefore, the present application provides a content recommendation method, apparatus, device and medium based on bias correction. This solution combines robust inverse tendency weighting with dual robust estimation, which can solve the exposure bias and feedback bias problems caused by historical strategies in the recommendation system, improve the recommendation quality of long-tail content and new content, and achieve fairness improvement and cold start optimization of recommendation results while ensuring CTR.

[0049] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0050] The disclosure below provides many different embodiments or examples for implementing different configurations of the present invention. To simplify the disclosure of the present invention, the components and configurations of specific examples are described below. Of course, these are merely examples and are not intended to limit the present invention. In addition, the present invention may repeat reference numerals and / or letters in different examples. Such repetition is for the purpose of simplicity and clarity and does not in itself indicate the relationship between the various embodiments and / or configurations discussed.

[0051] See also Figure 1 , is a flowchart of a content recommendation method based on deviation correction provided in an embodiment of the present application, the method specifically comprising the following steps:

[0052] S101. Determine the exposure probability of each content in the training data through a deep survival analysis model;

[0053] Traditional recommendation models directly use historical exposure data for training, ignoring the fact that samples are not independent and identically distributed. In particular, long-tail and new content are often marginalized due to lack of exposure. Therefore, in this application, we introduce the inverse propensity weighting (IPW) method. IPW estimates the propensity probability of each content in the training data being recommended. This propensity probability is the exposure probability. In subsequent steps, the weight coefficient of each content is calculated based on the exposure probability of each content, thereby weighting the loss function and focusing the model on real preference modeling.

[0054] Moreover, when estimating the exposure probability, traditional solutions are usually based on exposure probability estimation methods such as logistic regression (LogisticRegression, logistic regression) or shallow networks (Multilayer Perceptron, MLP), which are difficult to handle complex features. Therefore, in this application, a deep survival analysis model is used to estimate the exposure risk function of each content, which can model temporal and high-order features. Among them, the deep survival analysis model can be DeepSurv (Deep Survival), DeepHit, etc., which are not specifically limited here. This application uses a deep survival analysis model to model the passive exposure tendency of content, which can handle complex temporal and high-dimensional sparse features, improve the accuracy and robustness of tendency estimation, provide a reliable basis for subsequent weighting, and significantly reduce deviation.

[0055] Furthermore, the training data in this application includes at least one of the following: user-content exposure log, content features, and context features corresponding to each content. The user-content exposure log is used to record user ID (Identity document, unique code), content ID, whether to click / play, length of stay, and other interactive behaviors to reflect user preferences; the content features include content tags, duration, quality, etc.; the access time, device type, etc. In addition, this application can be divided into static data and dynamic data according to the type of training data. Static data includes user information and content information, such as static data such as user ID, content ID, content quantity, content playback time, etc.; dynamic data refers to data on user interaction with content, including content browsed by users, context features, etc.

[0056] In this embodiment, a model based on deep survival analysis can be constructed: DeepSurv, which determines the exposure probability (propensity score) P (exposure) of each content in the training data through the deep survival analysis model; the content in this application can be video, article, audio, etc., which is not specifically limited here.

[0057] S102: Determine a weight coefficient for each content based on the exposure probability of each content; wherein the exposure probability of each content is inversely proportional to the weight coefficient of each content;

[0058] After determining the exposure probability of each content, this application can determine the weight coefficient of each content based on the exposure probability of each content. The exposure probability is the probability that the content is noticed by the user. Since the content with a low exposure probability may lack click records because it has not been seen by the user, it does not mean that the user is not interested, which in turn causes exposure bias and feedback bias problems. Therefore, in order to avoid the bias problem, this application sets the exposure probability of each content to be inversely proportional to the weight coefficient of each content, that is: giving a higher weight to the content with a low exposure probability to compensate for the deviation caused by insufficient exposure, tapping the value of long-tail content, and promoting fair exposure; conversely, giving a lower weight to the content with a high exposure probability. In this way, the model influence of the head content can be weakened, and fair exposure opportunities can be obtained for the long-tail content.

[0059] S103: Determine a joint loss function based on the weight coefficients of each content and a dual robust estimation algorithm; the joint loss function includes a dual robust loss function and at least one target loss function;

[0060] It should be noted that if the deviation is corrected only by the IPW method, problems such as high variance will arise. Therefore, in this application, in order to reduce the high variance and other problems brought by the IPW method, a double robust estimation (Doubly RobustEstimator) algorithm is introduced, which combines the model prediction results with the propensity weight output. Even if one estimate is inaccurate, it can still remain unbiased or low-biased, so as to improve fault tolerance through double robust estimation.

[0061] In this way, the present application can combine the exposure probability of the content and the model prediction results to construct a dual robust loss function, thereby enhancing the generalization ability while correcting the deviation.

[0062] It should be noted that if the recommendation model only optimizes a single indicator during training, the diversity and fairness of the content will be sacrificed. For example, if the recommendation model is a CTR recommendation model, and the CTR recommendation model only optimizes the CTR indicator during training, this will easily sacrifice the diversity and fairness of the content. Therefore, in this application, the optimization indicators include not only CTR, but also auxiliary targets such as exposure entropy and diversity score to ensure the health of the content ecosystem and prevent popular content from monopolizing the recommendation list. Therefore, the joint loss function in this application not only includes a dual robust loss function, but also includes at least one target loss function.

[0063] S104: Train the initial recommendation model by combining the loss function and the training data, so as to generate content recommendation results through the trained target recommendation model.

[0064] After the joint loss function is determined through the above process, the initial recommendation model can be trained through the joint loss function and training data to generate content recommendation results through the trained target recommendation model. In the present application, the recommendation model can be a CTR recommendation model or other types of recommendation models, which are not specifically limited here. The target recommendation model trained by the present application can significantly improve the performance of the recommendation system in terms of long-tail content exposure, new content cold start, feedback bias correction, and click-through rate and fairness balance, and has good practical value and promotion potential.

[0065] In summary, this application determines the weight coefficient of each content through the inverse tendency weighting algorithm, and introduces a dual robust estimation algorithm to determine the joint loss function, which can enhance the generalization ability while correcting the bias, and improve the recommendation quality of the recommendation model in cold start content; and, this application estimates the exposure probability of content through a deep survival analysis model, which can handle complex temporal and high-dimensional sparse features, improve the accuracy and robustness of the exposure probability, provide a reliable basis for subsequent weighting, and significantly reduce the bias.

[0066] In another embodiment of the present application, the process of determining the weight coefficient of each content according to the exposure probability of each content includes:

[0067] Use quantiles to bucket each exposure probability;

[0068] The exposure probability of each bucket is probability calibrated to obtain the calibrated exposure probability; the weight coefficient of each content is generated based on the calibrated exposure probability.

[0069] In this application, after using the deep survival analysis model to predict the exposure probability of each content, in order to improve the estimation accuracy and provide a more robust basis for inverse tendency weighting, it is necessary to combine the bucket calibration technology. After using the quantile to bucket each exposure probability, the exposure probability of each bucket is probability calibrated to obtain the calibrated exposure probability, so as to generate the weight coefficient of each content according to the calibrated exposure probability.

[0070] Specifically, this application needs to determine the quantile when bucketing each exposure probability. The quantile refers to: after sorting a set of exposure probability data by size, it is divided into the critical point values ​​of several intervals in equal proportion. Through the bucketing operation, all exposure probabilities can be divided into different groups according to specific values. In addition, this application also needs to perform probability calibration for the exposure probability of each bucket. This method can eliminate the probability distortion caused by sample imbalance or model overfitting, so that the predicted probability is closer to the actual conversion rate. Moreover, the calibrated exposure probability can more accurately reflect the content quality. The weight coefficient generated accordingly can avoid the excessive suppression of long-tail content by the head content, and achieve a dynamic balance in traffic distribution.

[0071] In another embodiment of the present application, the process of probability calibrating the exposure probability of each bucket to obtain the calibrated exposure probability includes: probability calibrating the exposure probability of each bucket through a temperature scaling algorithm or a Platt Scaling algorithm to obtain the calibrated exposure probability.

[0072] In this application, after the original exposure probability of each content is bucketed according to quantiles, temperature scaling or Platt Scaling can be applied to the exposure probability of each bucket for probability calibration, so that the calibration result can be used as the basis for IPW calculation.

[0073] Temperature scaling involves scaling the exposure probability of each piece of content by introducing a temperature parameter, then applying a softmax function (normalized exponential function) to generate a calibrated probability. This temperature scaling algorithm can reduce the inflated probability of highly exposed content and increase the weight of long-tail content. The Platt Scaling algorithm, based on a logistic regression model, maps raw exposure probabilities to calibrated exposure probabilities. This approach can independently adjust for different probability intervals and adapt to complex probability distributions.

[0074] In another embodiment of the present application, the process of generating a weight coefficient for each content based on the calibrated exposure probability includes: determining the initial weight of each content based on the cropping upper limit value; normalizing the initial weights of all content to generate a weight coefficient for each content.

[0075] In this application, when generating the weight coefficient of each content based on the calibrated exposure probability, it is necessary to use the clipping upper limit value to determine the initial weight of each content. The calculation formula of the initial weight is:

[0076]

[0077] Among them, w i0 is the initial weight of the i-th content, P i is the exposure probability of the i-th content, and max_w is the clipping upper limit value, which can be set according to the actual situation. For example, if the clipping upper limit value is set to 10, the maximum value of each generated initial content is 10.

[0078] Furthermore, after calculating the initial weights of all content, this application also needs to normalize the initial weights of all content to generate a weight coefficient for each content. When performing normalization, this application can calculate the normalized weight coefficient for each content by dividing the initial weight of each content by the mean of the initial weights of all content to control the training variance.

[0079] In summary, this application adopts a robust IPW strategy for calibrating, cropping and normalizing the exposure probability. This method not only uses IPW to calculate the exposure probability, but also buckets the exposure probability, calibrates the probability, and then performs cropping and normalization. This method can suppress the variance caused by extreme weights and improve training stability. It can effectively alleviate the high variance problem caused by too small propensity scores in traditional IPW, making model training more stable and controllable.

[0080] In another embodiment of the present application, the process of determining the joint loss function based on the weight coefficient of each content and the dual robust estimation algorithm includes:

[0081] Determine a dual robust loss function based on the weight coefficient of each content and the dual robust estimation algorithm; wherein the dual robust loss function includes: an IPW weighted loss function and a model estimation loss function;

[0082] Determine the fairness loss function and diversity loss function;

[0083] A joint loss function is determined according to the dual robust loss function, the fairness loss function, the diversity loss function, a weight coefficient of the fairness loss function, and a weight coefficient of the diversity loss function.

[0084] In this application, the dual robust loss function created based on the weight coefficient and the dual robust estimation algorithm can be specifically:

[0085]

[0086] Among them, the dual robust loss function includes the IPW weighted loss function and the model estimation loss function, w i is the weight coefficient of the i-th content, is the model prediction result of the i-th content, is the true feedback result of the i-th content, y i is the expected feedback of the bias correction model for the i-th content.

[0087] This application also needs to determine at least one target loss function. The target loss functions in this application are: fairness loss function and diversity loss function. Then, based on the dual robust loss function, fairness loss function, diversity loss function, the weight coefficient of the fairness loss function and the weight coefficient of the diversity loss function, the joint loss function is determined.

[0088] The joint loss function in this application is specifically:

[0089]

[0090] In this formula, The total loss value calculated for the joint loss function, It is a dual robust loss function for the CTR recommendation model, which can reflect the exposure entropy. is the fairness loss function, which can reflect the exposure ratio of new content (Cold-start ratio). The diversity loss function reflects the fairness loss function, which can reflect diversity. λ1 is the weight coefficient of the fairness loss function, and λ2 is the weight coefficient of the diversity loss function. This approach allows various metrics to be incorporated into the training objectives of the recommendation model. Adjustable weights are used to jointly optimize CTR and exposure fairness, improving user experience and exposure for long-tail content.

[0091] Among them, the fairness regularization term It can be:

[0092]

[0093] In this formula, k is the content label, H is the entropy value of the recommendation list, and p t is the recommendation probability of the t-th label.

[0094] Diversity regularization term It can be minimized by the sum of cosine similarities, that is:

[0095]

[0096] Among them, M is the number of recommended list contents, i and j are the index numbers of the contents, and v i is the embedding vector of the i-th content, v j is the embedding vector of the jth content.

[0097] The weight coefficients of the fairness loss function and the diversity loss function can be pre-set fixed weight coefficients, or set through training / parameter adjustment. Alternatively, a weight adaptation mechanism can be introduced to determine the weight coefficients, thereby dynamically adjusting the balance between CTR and other indicators such as fairness. This weight adaptation mechanism can include methods such as meta-learning and self-attention weight adjustment.

[0098] In summary, this application proposes a recommendation training mechanism that combines dual robust estimation. This solution integrates the model prediction output and the IPW weighted loss to construct a dual robust loss function. Even if one of the estimates is inaccurate, the recommendation result can be kept unbiased. This method can enhance the generalization ability of the recommendation model under biased data and improve the model's reasoning ability on non-exposed samples. In addition, this application adds regularization targets such as exposure entropy, diversity, and new content exposure ratio to the joint loss function, so that the recommendation target not only optimizes click-through rate, but also simultaneously optimizes indicators such as exposure entropy, diversity, and new content ratio. This method can balance the recommendation benefits and the platform content ecology, improve user experience and enhance fair exposure of content, which is particularly beneficial for the cold start of long-tail videos and new videos.

[0099] In another embodiment of the present application, generating content recommendation results through the trained target recommendation model includes: generating a recommendation list through the target recommendation model and an exploration strategy; wherein the exploration strategy is a UCB exploration strategy or an ε-greedy exploration strategy.

[0100] In addition, after generating the recommendation list, this application also needs to calculate the entropy value of the recommendation list. If the entropy value is less than the predetermined threshold, the exploration rate of the cold start content is increased, and the recommendation list is re-generated through the target recommendation model and exploration strategy; the exploration rate is the parameter value of the exploration strategy.

[0101] Specifically, when this application makes online recommendations through a trained target recommendation model, it is necessary to calculate the entropy value of the recommendation list in real time. If the entropy value is lower than the threshold, it means that the concentration of the recommendation list is too high, and it is necessary to increase the exploration rate of cold start content in order to increase the proportion of long-tail content. In this way, the cold start content can be configured with a higher exploration priority by default.

[0102] If this application calculates the entropy value of the recommendation list based on content tags, the calculation formula is:

[0103]

[0104] Among them, H is the entropy value of the recommendation list, T is the total number of label types, t represents the tth label in the Tth, p t is the recommendation probability of the t-th label.

[0105] After calculating the entropy value through the above formula, the exploration rate of the cold start content of the exploration strategy can be improved when the entropy value is less than the predetermined threshold. In the present application, the exploration strategy can be a UCB (Upper Confidence Bound) exploration strategy or an ε-greedy (Epsilon-Greedy) exploration strategy. When improving the exploration rate of cold start content, if the exploration strategy is a UCB exploration strategy, the UCB exploration value is increased; if the exploration strategy is an ε-greedy exploration strategy, the ε value is increased. After adjusting the exploration rate of the exploration strategy, it is necessary to use the target recommendation model and the adjusted exploration strategy to regenerate the recommendation list until the entropy value of the regenerated recommendation list is not less than the predetermined threshold.

[0106] In addition, in addition to adjusting the recommendation list online in real time, this application can also count the various indicators of the recommendation list within a predetermined period to evaluate the recommendation effect of the recommendation list. For example, at the predetermined time every day, the entropy, diversity, new content ratio and other indicators of the historical recommendation results are evaluated in real time to generate the evaluation results of the recommendation list. In addition, the recommendation model can also be adjusted according to the evaluation results to optimize the recommendation model.

[0107] It can be seen that this application provides a lightweight exploration strategy for cold start and a dynamic exploration mechanism based on entropy feedback. By adopting UCB or ε-greedy strategy in the online recommendation stage to dynamically explore long tail and new content, the exploration ratio of long tail content can be adjusted, and the entropy value of the recommendation list can be monitored in real time to dynamically adjust the exploration intensity, so as to achieve a dynamic balance between user satisfaction and recommendation fairness, and make cold start recommendations "strategic and bounded". In this way, excessive recommendation of popular content can be effectively avoided, and the exposure efficiency of long tail content and cold start content can be improved while ensuring CTR, so as to achieve a dynamic balance between personalization and diversity.

[0108] This application can be widely used in scenarios such as Internet content platforms, video media, short video applications, information flow platforms, and advertising recommendation systems, and has broad commercial adaptability. In addition, this application can bring the following beneficial effects:

[0109] 1. Improve the exposure of long-tail videos on the homepage and activate content assets:

[0110] Homepage waterfall recommendations require a high level of content diversity, but existing algorithms generally favor mainstream and popular videos, resulting in some vertically-focused premium dramas, short variety shows, documentaries, and other content struggling to gain sufficient exposure. This invention effectively increases click-through and play rates for this type of content through a stable exposure probability modeling and exploration mechanism.

[0111] Commercial value: Improve the utilization rate of platform inventory content (especially long-tail content), extend the content life cycle, and reduce content cold start costs.

[0112] 2. Improving the robustness and training stability of recommendation models

[0113] Homepage recommendation models require extensive training and frequent iterations, placing high demands on model convergence speed and generalization capabilities. This paper mitigates the high variance problem by clipping the normalized IPW and using dual robust estimation, improving the model's consistent performance in both offline and online environments.

[0114] Business value: Reduces tuning costs caused by model instability and improves the efficiency and success rate of new model launches.

[0115] 3. Enhance user homepage experience, increase click-through rate and stay time

[0116] By monitoring the entropy of the recommendation list, we can adjust the content diversity to ensure that the content presented on the homepage is non-repetitive and non-boring, and adapt to the changing interests of different users.

[0117] Commercial value: Increase homepage click-through rate, viewing time, and user activity, thereby driving advertising monetization and membership growth.

[0118] 4. Promote the transformation of recommendation systems to "real feedback modeling"

[0119] The platform recommendation business is gradually shifting from "short-term behavior fitting" to "user real preference modeling". The deviation correction mechanism introduced in this invention provides a data foundation and technical path for future system long-term benefit modeling (such as LTV estimation and multi-hop recommendation).

[0120] Strategic value: Helps the platform build a recommendation infrastructure with long-term competitive advantages, forming a complete feedback loop from data collection, modeling to strategy optimization.

[0121] See also Figure 2 , Figure 2 This is a schematic diagram of the structure of a content recommendation device based on deviation correction provided in an embodiment of the present application. The device specifically includes:

[0122] An exposure probability determination module 11 is used to determine the exposure probability of each content in the training data through a deep survival analysis model;

[0123] The weight coefficient determination module 12 is used to determine the weight coefficient of each content according to the exposure probability of each content; wherein the exposure probability of each content is inversely proportional to the weight coefficient of each content;

[0124] A loss function determination module 13 is configured to determine a joint loss function based on weight coefficients of each content and a dual robust estimation algorithm; the joint loss function includes a dual robust loss function and at least one target loss function;

[0125] A training module 14 is configured to train the initial recommendation model using the joint loss function and the training data;

[0126] The recommendation module 15 is used to generate content recommendation results through the trained target recommendation model.

[0127] As an optional embodiment, the weight coefficient determination module includes:

[0128] Bucketing unit, used to bucket each exposure probability using quantiles;

[0129] A calibration unit is used to perform probability calibration on the exposure probability of each bucket to obtain a calibrated exposure probability;

[0130] The generating unit is configured to generate a weight coefficient for each content according to the calibrated exposure probability.

[0131] As an optional embodiment, the calibration unit is specifically configured to perform probability calibration on the exposure probability of each bucket by using a temperature scaling algorithm or a PlattScaling algorithm to obtain a calibrated exposure probability.

[0132] As an optional embodiment, the generating unit is specifically configured to: determine the initial weight of each content according to the clipping upper limit value; and perform normalization processing on the initial weight of all content to generate a weight coefficient of each content.

[0133] As an optional embodiment, the loss function determination module is specifically configured to:

[0134] Based on the weight coefficient of each content and the dual robust estimation algorithm, a dual robust loss function is determined; wherein the dual robust loss function includes: an IPW weighted loss function and a model estimation loss function; a fairness loss function and a diversity loss function are determined;

[0135] A joint loss function is determined according to the dual robust loss function, the fairness loss function, the diversity loss function, a weight coefficient of the fairness loss function, and a weight coefficient of the diversity loss function.

[0136] As an optional embodiment, the recommendation module is specifically configured to: generate a recommendation list through a target recommendation model and an exploration strategy; wherein the exploration strategy is a UCB exploration strategy or an ε-greedy exploration strategy.

[0137] As an optional embodiment, the device further includes:

[0138] The exploration adjustment module is used to calculate the entropy value of the recommendation list after generating the recommendation list through the target recommendation model and exploration strategy; if the entropy value is less than a predetermined threshold, the exploration rate of the cold start content is increased, and the recommendation module is re-triggered to generate the recommendation list through the target recommendation model and exploration strategy; wherein the exploration rate is the parameter value of the exploration strategy.

[0139] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0140] See also Figure 3 , Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application includes a processor 21, a communication interface 22, a memory 23, and a communication bus 24, wherein the processor 21, the communication interface 22, and the memory 23 communicate with each other via the communication bus 24;

[0141] Memory 23, for storing computer programs;

[0142] The processor 21 is configured to implement the steps of the content recommendation method described in any of the above method embodiments when executing the program stored in the memory 23, which will not be described in detail here.

[0143] The communication bus mentioned in the above terminal can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0144] The communication interface is used for communication between the above terminal and other devices.

[0145] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage. Alternatively, the memory may be at least one storage device located away from the processor.

[0146] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0147] In another exemplary embodiment, a computer storage medium is provided, wherein the program instructions, when executed by a processor, implement the steps of the content recommendation method described in any of the above method embodiments. The storage medium may include any medium capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0148] Optionally, the specific examples in this embodiment may refer to the examples described in the above embodiments, and this embodiment will not be described in detail here.

[0149] It should be understood that the terms used herein are for the purpose of describing specific example embodiments only and are not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms "one", "an" and "said" as used herein may also be meant to include plural forms. The terms "comprise", "include", "contain" and "have" are inclusive and therefore specify the presence of stated features, steps, operations, elements and / or parts, but do not exclude the presence or addition of one or more other features, steps, operations, elements, parts, and / or combinations thereof. The method steps, processes, and operations described herein are not to be construed as necessarily requiring them to be performed in the specific order described or illustrated, unless the order of execution is clearly indicated. It should also be understood that additional or alternative steps may be used.

[0150] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is intended to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A content recommendation method based on bias correction, characterized in that: include: Determine the exposure probability of each content in the training data through a deep survival analysis model; Determine the weight coefficient of each content according to the exposure probability of each content; wherein the exposure probability of each content is inversely proportional to the weight coefficient of each content; Determine a joint loss function based on the weight coefficients of each content and a dual robust estimation algorithm; the joint loss function includes a dual robust loss function and at least one target loss function; The initial recommendation model is trained using the joint loss function and the training data, so that a content recommendation result is generated using the trained target recommendation model.

2. The content recommendation method according to claim 1, characterized in that The weight coefficients of each content are determined based on the exposure probability of each content, including: Use quantiles to bucket each exposure probability; Perform probability calibration on the exposure probability of each bucket to obtain the calibrated exposure probability; The weight coefficient of each content is generated according to the calibrated exposure probability.

3. The content recommendation method according to claim 2, characterized in that: The exposure probability of each bucket is calibrated. The calibrated exposure probability includes: The exposure probability of each bucket is calibrated using the temperature scaling algorithm or the Platt Scaling algorithm to obtain the calibrated exposure probability.

4. The content recommendation method according to claim 2, characterized in that: The weight coefficient of each content is generated based on the calibrated exposure probability, including: Determine the initial weight of each content based on the clipping upper limit value; The initial weights of all contents are normalized to generate the weight coefficient of each content.

5. The content recommendation method according to claim 1, characterized in that: Based on the weight coefficient of each content and the double robust estimation algorithm, the joint loss function is determined, including: Determine a dual robust loss function based on the weight coefficient of each content and the dual robust estimation algorithm; wherein the dual robust loss function includes: an IPW weighted loss function and a model estimation loss function; Determine the fairness loss function and diversity loss function; A joint loss function is determined according to the dual robust loss function, the fairness loss function, the diversity loss function, a weight coefficient of the fairness loss function, and a weight coefficient of the diversity loss function.

6. The content recommendation method according to any one of claims 1 to 5, characterized in that: The content recommendation results generated by the trained target recommendation model include: A recommendation list is generated through a target recommendation model and an exploration strategy; wherein the exploration strategy is a UCB exploration strategy or an ε-greedy exploration strategy.

7. The content recommendation method according to claim 6, characterized in that: After generating the recommendation list through the target recommendation model and exploration strategy, it also includes: Calculating an entropy value of the recommendation list; If the entropy value is less than a predetermined threshold, the exploration rate of the cold start content is increased, and a recommendation list is regenerated using the target recommendation model and the exploration strategy; wherein the exploration rate is a parameter value of the exploration strategy.

8. A content recommendation device based on deviation correction, characterized in that: include: The exposure probability determination module is used to determine the exposure probability of each content in the training data through a deep survival analysis model; A weight coefficient determination module, configured to determine a weight coefficient for each content based on the exposure probability of each content; wherein the exposure probability of each content is inversely proportional to the weight coefficient of each content; A loss function determination module, configured to determine a joint loss function based on weight coefficients of each content and a dual robust estimation algorithm; the joint loss function includes a dual robust loss function and at least one target loss function; A training module, configured to train the initial recommendation model using the joint loss function and the training data; The recommendation module is used to generate content recommendation results through the trained target recommendation model.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory for storing computer programs; The processor is configured to implement the content recommendation method steps described in any one of claims 1 to 7 when executing the program stored in the memory.

10. A computer storage medium, characterized in that The computer storage medium stores computer-executable instructions, and the computer-executable instructions are used to execute the steps of the content recommendation method according to any one of claims 1 to 7.

Images