Hybrid method for popular recommendation based on multi-task

Abstract

The application provides a hybrid popular recommendation method based on multi-task. The method comprises the following steps: a popular deviation bias causal recommendation based on multi-task, through causal recommendation analysis, using multi-task modeling of popular deviation influence under different causal paths, combining bias amplification score and other popular deviation factors to realize popular deviation bias elimination recommendation; popularity utilization based on item popularity propagation modeling, proposing a dynamic span to model short-term fluctuations in popularity, and combining an adaptive attention mechanism to predict future popularity; momentum modeling based on future popularity propagation, finally fusing the bias elimination recommendation score and the propagation momentum to realize hybrid popular recommendation. The method effectively improves the recommendation accuracy and accuracy, helps to accurately eliminate the popular deviation and utilize the item popular propagation information, and the effect is more significant especially on the long tail data set.

Description

Technical Field

[0001] This invention relates to the field of recommender system technology, specifically a multi-task-based hybrid popularity recommendation method. By eliminating popularity bias in the recommender system and combining popular propagation information of items for accurate recommendations, the method improves the accuracy of the recommender system and user satisfaction. Background Technology

[0002] In recent years, personalized recommendations have become an effective way to alleviate information overload, playing a crucial role in online applications such as e-commerce, video websites, and social media. Current mainstream recommendation methods typically rely on historical interaction data. However, the distribution of items in historical interactions is usually non-uniform (primarily long-tailed), with a few items dominating the majority of interactions, leading to popularity bias. Recommendation models trained on this long-tailed data further amplify this bias, causing the system to favor recommending more popular items, resulting in problems such as echo chambers, filter bubbles, and the Matthew effect. This hinders the accurate acquisition of user interests and reduces recommendation accuracy and fairness. However, blindly pursuing the elimination of popularity bias may lead to unfair suppression of worthy, more popular items. Therefore, researching hybrid popularity recommendation from a compromise perspective—eliminating the negative effects of popularity bias while leveraging the positive effects of popularity—can help improve recommendation accuracy and user satisfaction.

[0003] Current hybrid popularity recommendation research either employs a two-stage strategy, utilizing popularity while eliminating bias; or a fusion strategy, collaboratively training biased and unbiased models and achieving the final recommendation based on the fusion of the two models. However, most of these hybrid popularity recommendation methods neglect the amplifying effect of popularity bias on the recommendation algorithm when eliminating popularity bias, making it difficult to accurately remove bias. Furthermore, when predicting popularity, they ignore the impact of short-term subtle fluctuations in popularity and the influence of the trend of item spread in future popularity predictions, thus affecting the final recommendation performance.

[0004] To address the limitations of bias removal in recommendation research, which suffers from difficulty in accurately removing bias due to neglecting the amplifying effect of recommendation methods on bias, and the fact that most existing popularity-based recommendation methods ignore short-term fluctuations in popularity prediction and struggle to learn project propagation trends and influence, this invention designs and implements a multi-task-based hybrid popularity recommendation method. This method eliminates bias through causal recommendation analysis and multi-task modeling, and utilizes popularity by combining project popularity propagation modeling, thereby improving recommendation performance. Summary of the Invention

[0005] This invention proposes a multi-task-based hybrid popularity recommendation method, which aims to eliminate popularity bias in the recommendation system and combine it with popular propagation information of items for accurate recommendations.

[0006] To achieve the above objectives, the present invention adopts the following technical solution.

[0007] A hybrid popular recommendation method based on multiple tasks includes the following steps: Step 1: Use causal graphs to perform causal recommendation analysis on the popularity bias in the recommendation system. Based on the traditional causal graph, set up user variable U, item variable I, popularity variable Z and interaction result Y, model the path Z→U→Y indirectly affecting interaction through users, the path Z→I→Y indirectly affecting interaction through items, and the path Z→Y directly affecting interaction, and identify and model different causal paths of popularity bias. Step 2: Employ a multi-task learning model to model conformity score, exposure score, and bias amplification score separately. The conformity score is obtained by modeling the matching features of user representation and popularity representation using the vector inner product and summing them along the embedding dimension. The exposure score is obtained by modeling the matching features of item representation and popularity representation using the vector inner product and summing them along the embedding dimension. The bias amplification score is obtained by summing the popularity representation along the embedding dimension. The product of the three scores and the product of the hyperparameter c is used as the fusion result. This fusion result is subtracted from the mixed recommendation score to obtain the bias-free recommendation score. Step 3: By combining dynamic span and adaptive attention mechanisms with popularity location embedding, a popularity feature extraction model based on dynamic span is established to simultaneously capture the long-term dependence and short-term fluctuations of project popularity, thereby improving popularity prediction and recommendation performance. Step 4: Based on the predicted future popularity value, the product of the predicted future popularity value of the project in the next time period and the average popularity change rate of the project in the current time period relative to the previous time period is determined as the propagation momentum, which is used as a representation of the project's popularity propagation capacity. Step 5: Combine the popularization bias-free recommendation score and propagation momentum to obtain the hybrid popularization recommendation score result. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the framework of the multi-task-based hybrid popular recommendation (MPHR) method of the present invention; Figure 2 This is a comparison chart of the test results of the present invention and other methods on the test dataset. Detailed Implementation

[0009] The method of the present invention will be further described in detail below based on the implementation principle and specific data processing flow.

[0010] 1. Data Preprocessing First, data is collected from user-project interaction logs, including user ID, project ID, and interaction timestamps. Project sets and user sets are defined as follows: and , and These represent the number of projects and the number of users, respectively, and the actual user-project interactions. :

[0011] Based on the interaction timestamp, the popularity of each item is calculated. This invention uses the number of times a user interacts with an item within a certain time period as a measure of popularity.

[0012] Define the popularity of an item over a given time period. The calculation method is as follows:

[0013] Next, the interaction log is divided into s time periods, with a total time of . ,based on Construct a popularity set. This can be built into a popularity time series matrix. It is used for subsequent analysis and modeling and is defined as follows:

[0014] 2. Causal Recommendation Analysis A causal graph is used to conduct a detailed causal relationship analysis of the recommendation process, identifying the main sources and propagation paths of popularity bias. The causal graph includes user variable U, item variable I, popularity variable Z, and interaction result Y, with directed edges representing the causal effects between variables. Based on the traditional causal graph, the popularity variable Z is added to determine the paths Z→U→Y where popularity indirectly influences interaction via users, Z→I→Y where popularity indirectly influences interaction via items, and Z→Y where popularity directly influences interaction.

[0015] 3. Multi-task popularization bias modeling For the three causal paths mentioned above, modeling methods are designed for conformity score, exposure score, and bias amplification score, respectively: Conformity score: By modeling the matching features between user representation and popularity representation, it calculates the probability that a user clicks on a highly popular item due to conformity, and is formally expressed as... The calculation method is as follows:

[0016] in, Represents user variables, This indicates the value of the user variable corresponding to the target user. Represents the popularity variable. This indicates the value of the popularity variable for the target item within the corresponding time period. The matching function represents the matching features between the user representation and the popularity representation, and is modeled through the vector inner product. The rating function for calculating the recommendation score is calculated by summing the embedding dimensions.

[0017] Exposure Score: By modeling the matching features between item representation and popularity representation, this score assesses the probability that an item will generate additional interactions due to overexposure caused by high popularity. Formalized as... The calculation method is as follows:

[0018] in, Represents project variables. This indicates the value of the project variable corresponding to the target project. Represents the popularity variable. This indicates the value of the popularity variable for the target item within the corresponding time period. The matching function represents the matching features between the learning item representation and the popularity representation, and is modeled by the vector inner product. The scoring function is calculated by summing the embedding dimensions. Bias Amplification Score: This score directly models the impact of popularity on interaction outcomes, reflecting the degree to which the recommendation method amplifies the popularity bias. It is formally represented as... The calculation method is as follows:

[0019] in, Represents the popularity variable. This indicates the value of the popularity variable for the target item within the corresponding time period. Similarly, the summation is calculated using the embedded dimension.

[0020] A multi-task learning model is adopted, and loss functions are designed for conformity score, exposure score and bias amplification score respectively. Through joint optimization, accurate popularization bias modeling is achieved.

[0021] This invention uses a multi-task learning model for supervision. , and Learning. Since the popularization bias modeling problem can be viewed as a binary classification task involving conformity score, exposure score, and bias amplification score, the binary cross-entropy (BCE) loss function is chosen to construct optimization objectives to measure the differences between these three scores and the actual interaction results, calculating these differences from the actual interaction. Conformity score loss function. Exposure score loss function Sum of bias amplification fractional loss function As shown below:

[0022]

[0023]

[0024] in, Represents user variables, This indicates the value of the user variable corresponding to the target user. Represents project variables. This indicates the value of the project variable corresponding to the target project. Represents the popularity variable. This indicates the value of the popularity variable for the target item within the corresponding time period. This represents a set of actual user-project interaction samples. Indicates user With the project Real interaction tags between them This indicates a mixed recommendation score. This represents the sigmoid function.

[0025] 4. Recommendations to eliminate popular bias Based on the mixed recommendation score (i.e., the recommendation score influenced by both user interests and popularity bias), a bias-free recommendation score based purely on user interests is obtained by subtracting the convergence result of conformity score, exposure score, and bias amplification score. This step effectively eliminates popularity bias in the recommendation results, thereby obtaining recommendation results that are not affected by popularity bias and improving the accuracy of the recommendations.

[0026]

[0027] in, For conformity scores, To expose the score, The deviation is amplified by the fraction. To mix recommendation scores, These are hyperparameters used to adjust... , , The product of these three factors yields a bias-free recommendation based purely on genuine user interests.

[0028] Since modeling mixed recommendation scores can be viewed as a binary classification task, the mixed recommendation scores also employ BCE loss, with the loss function... With three popular bias loss functions , and The final debiased loss function is obtained by fusion. :

[0029]

[0030] in, This indicates the value of the user variable corresponding to the target user. This indicates the value of the project variable corresponding to the target project. This represents a set of actual user-project interaction samples. Indicates user With the project Real interaction tags between them This indicates a mixed recommendation score. This represents the sigmoid function. , , This is a hyperparameter used to control the weights of conformity score, exposure score, and bias amplification score.

[0031] 5. Popularity Embedding Representation To better capture the characteristics of project popularity propagation, a popularity embedding representation method is proposed. By combining popularity location embedding (representing the location information of popularity in the time series) and original popularity embedding (representing the magnitude of popularity), the ability to represent project popularity propagation is enhanced, providing rich feature information for subsequent prediction and modeling.

[0032] Popularity time location embedding vectors The calculation method is as follows:

[0033]

[0034] in, For elements sequence position, Embedding vectors for popularity time location The dimension of popularity. The popularity magnitude is extracted from the popularity sequence matrix as the original popularity embedding. The popularity embedding is obtained by fusing the popularity location embedding with the original popularity embedding. As shown below:

[0035] In this way, each popularity embedding vector contains not only information about the magnitude of the project's popularity, but also information about its corresponding time series location, greatly enhancing the ability to represent the spread of project popularity.

[0036] 6. Dynamic span modeling and future popularity prediction To more accurately predict the future popularity of a project, the concept of dynamic span is proposed. By adaptively adjusting the size of the time window, it captures subtle short-term fluctuations in popularity. Simultaneously, dynamic span is integrated into an adaptive attention mechanism, enhancing attention to short-term fluctuations while capturing the long-term dependencies of popularity changes. This combination of dynamic span and adaptive attention mechanism enables accurate prediction of future popularity.

[0037] Popularity feature extraction based on dynamic span includes dynamic span modeling, adaptive attention layer, feedforward layer and residual layer, which will be introduced below.

[0038] (1) Dynamic span modeling Because the contribution of popularity changes across different time periods to the project's dissemination capability is not uniformly distributed in the popularity sequence, some short-term fluctuations may have a more significant impact on popularity spread. Therefore, by modeling a dynamic span to adjust the focus time range, we can enhance the learning of short-term popularity fluctuations. The dynamic span is dynamically adjusted based on the data, and the span parameter represents the span of the h-th self-attention head. The calculation method is as follows:

[0039] in, and These are learnable parameters. This refers to the length of the popularity sequence. To achieve adaptive attention, each attention head needs to have a different span size, and for each attention head... Assign a span parameter Then, the model is based on a dynamic span matrix. mask matrix Its elements The calculation method is as follows:

[0040] This means if and The distance between the two locations does not exceed the span parameter. The corresponding mask value is 1 if the distance exceeds the limit, and the corresponding mask value is 0 if the distance exceeds the limit.

[0041] (2) Adaptive attention layer The adaptive attention layer takes popularity embedding as input. and mask matrix X is obtained by performing three linear transformations:

[0042]

[0043]

[0044] in, , and For three dimensions The matrix is ​​used. Since the contribution of popularity changes across different time periods to the project's dissemination capability is not uniformly distributed in the popularity sequence, some short-term fluctuations may have a more significant impact on popularity dissemination. Therefore, a dynamic span is incorporated into the adaptive attention mechanism to adjust the focus on the short-term time range, thereby adaptively enhancing the learning of short-term fluctuations. The output of the adaptive attention layer is shown below. As shown below:

[0045] in, This is a scaling factor used to avoid excessively large inner product values. Since positions outside the span do not affect the self-attention results, while positions within the span will enhance self-attention, it is possible to learn the long-term dependence of popularity changes while paying close attention to short-term fluctuations in popularity.

[0046] (3) Feedforward-residual layer To avoid overfitting and enhance robustness, a fully connected layer is used as the feedforward layer, followed by a dropout layer, as shown in the following formula:

[0047] in, and To train the weight matrix and biases. To enhance the popular propagation representation, [the following is used]: By combining it with the popularity embedding vector, an output result is obtained that simultaneously contains local popularity information for each time period and global information on the evolution of popularity throughout the entire time series.

[0048] The future popularity vector is obtained after hierarchical normalization. The final output result is obtained after dimensional transformation. :

[0049]

[0050] in, Representation layer normalization, and Used for dimensional transformation.

[0051] 7. Propagation Momentum Modeling Drawing inspiration from the concept of momentum in physics, a method for measuring propagation momentum is proposed. Based on future popularity forecasts and considering the rates of change in popularity over the current and previous time periods, the propagation momentum of a project is calculated. Propagation momentum considers not only the magnitude of popularity but also the rate of change, providing a more comprehensive reflection of the project's spread trend and impact.

[0052] Propagation momentum, as a representation of project popularity, is modeled based on predicted future popularity values ​​and used to replace future popularity in the final recommendation. Using the rate of change in popularity over the current time period as the propagation speed, the propagation momentum of a project is then... The calculation method is as follows:

[0053] in, This represents the predicted future popularity of a project over the next time period, derived from the future popularity vector. , and These represent the current time period and the previous time period, respectively. and These represent the popularity of a project in the current time period and the previous time period, respectively. The higher the popularity value or the faster the spread, the greater the spread momentum. In other words, if a project has high user attention or attracts user attention faster, the project has greater spread momentum.

[0054] 8. Hybrid Popular Recommendations Based on bias-corrected recommendations, propagation momentum information is integrated to obtain the final hybrid popularity recommendation result. This method balances the elimination of popularity bias and the positive utilization of popularity, avoiding the problem of recommendations being overly concentrated on highly popular items while ensuring the diversity and accuracy of the recommendations.

[0055] Propagation momentum is used instead of future popularity predictions. To achieve hybrid popularity recommendations, migration is used to remove popularity bias in recommendation scores. As a pre-recommendation score, it is combined with propagation momentum to utilize popularity and obtain the final hybrid popularity recommendation score. :

[0056] in, This represents the project propagation momentum calculated according to the above propagation momentum formula. The sigmoid function is used to ensure that the recommendation score is between 0 and 1. The hyperparameter for smoothing the spread of the epidemic is used to control the strength of the epidemic's influence. Consistent with the de-biasing recommendation, the BCE loss is also used to construct the optimization objective, and the loss function is shown below:

[0057] in, For the final hybrid popular recommendation score, This indicates the value of the user variable corresponding to the target user. This indicates the value of the project variable corresponding to the target project. Indicates user With the project The actual interaction between them This represents a set of actual user-project interaction samples. This is the sigmoid function.

[0058] Through the above technical solution, the present invention realizes a multi-task-based hybrid popular recommendation method, which effectively improves the performance and user experience of the recommendation system.

[0059] The above description is merely a preferred embodiment of the present invention, but the scope of protection is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this invention should be included within the scope of protection of this invention. Therefore, the scope of protection should be determined by the content of the claims.

Claims

1. A hybrid popular recommendation method based on multi-tasks, characterized in that, Includes the following steps: S1: Using causal graphs, conduct causal recommendation analysis on the popularity bias in the recommendation system. Based on the traditional causal graph, set user variable U, item variable I, popularity variable Z, and interaction result Y. Model the path Z→U→Y indirectly affecting interaction through users, the path Z→I→Y indirectly affecting interaction through items, and the path Z→Y directly affecting interaction. Identify and model different causal paths of popularity bias. S2: A multi-task learning model is adopted to model the conformity score, exposure score, and bias amplification score separately. The conformity score is obtained by modeling the vector inner product of the matching features of user representation and popularity representation and summing them according to the embedding dimension. It is used to represent the indirect impact of popularity bias on the interaction result through user conformity behavior. The exposure score is obtained by modeling the vector inner product of the matching features of item representation and popularity representation and summing them according to the embedding dimension. It is used to represent the indirect impact of popularity bias on the interaction result through the item's exposure level. The bias amplification score is obtained by summing the popularity representation according to the embedding dimension. It is used to represent the direct impact of popularity on the interaction result. The calculation formulas for the three scores are as follows: in, For conformity scores, To expose the score, The deviation is amplified by the fraction. Represents user variables, This indicates the value of the user variable corresponding to the target user. Represents project variables. This indicates the value of the project variable corresponding to the target project. Represents the popularity variable. This indicates the value of the popularity variable for the target item within the corresponding time period. The matching function represents the matching features between user representations and popularity representations, modeled through vector inner product. The scoring function for calculating the recommendation score is calculated by summing the embedding dimensions; The product of the conformity score, exposure score, and bias amplification score, multiplied by the hyperparameter c, is used as the fusion result. This fusion result is then subtracted from the mixed recommendation score to obtain the bias-free recommendation score generated by user interests. The calculation method is as follows: in, For conformity scores, To expose the score, The deviation is amplified by the fraction. For mixed recommendation scores; S3: By combining dynamic span and adaptive attention mechanisms with popularity location embedding, a popularity feature extraction model based on dynamic span is established to simultaneously capture the long-term dependence and short-term fluctuations of project popularity, thereby improving popularity prediction and recommendation performance. S4: Model propagation momentum based on future popularity predictions to replace the future popularity in the final recommendation. Use the rate of change of popularity in the current time period as the propagation speed to obtain the project propagation momentum. The calculation method is as follows: in, This represents the predicted future popularity of a project over the next time period, derived from the future popularity vector. and These represent the current time period and the previous time period, respectively. and These represent the popularity of a project in the current time period and the previous time period, respectively; S5: Integrate the popularization bias recommendation score and propagation momentum to obtain a hybrid popularization recommendation score result.

2. The hybrid popular recommendation method based on multi-tasks according to claim 1, characterized in that, The specific process of causal recommendation analysis in step S1 is as follows: Construct a causal graph containing user variable U, project variable I, popularity variable Z, and interaction result Y, and add popularity variable Z to the traditional causal graph; From the perspective of recommendation systems, interactive behaviors related to popularity bias can be divided into the following three categories: ① Popularity indirectly influences interaction through users, i.e., the path Z→U→Y. This corresponds to the popularity bias caused by users' conformity behavior, which can be formally represented as... ; ② Popularity indirectly influences interaction through projects, i.e., the path Z→I→Y. This corresponds to the popularity bias caused by overexposure of a project, which can be formally represented as... ; ③ Popularity directly affects interaction, i.e., the path Z→Y, corresponding to the popularity bias caused by the amplification effect of the recommendation method, which can be formally expressed as: ; The three causal paths mentioned above together constitute and amplify the prevalence bias in recommendation systems. The formal representation is as follows: in, , , These represent the popularity bias components resulting from user herd behavior, excessive project exposure, and the amplifying effect of recommendation methods, respectively. This represents the sigmoid function.

3. The hybrid popular recommendation method based on multi-tasks according to claim 1, characterized in that, The specific method of the multi-task learning mode in step S2 is as follows: The conformity score represents the indirect impact of popular bias on interaction through user conformity behavior. The calculation formula is: in, Represents user variables, This indicates the value of the user variable corresponding to the target user. Represents the popularity variable. This indicates the value of the popularity variable for the target item within the corresponding time period. The matching function represents the matching features between user representations and popularity representations, modeled through vector inner product. The scoring function for calculating the recommendation score is calculated by summing the embedding dimensions; Exposure score represents the indirect impact of popularity bias on interaction through the degree of project exposure. The calculation formula is: in, Represents project variables. This indicates the value of the project variable corresponding to the target project. Represents the popularity variable. This indicates the value of the popularity variable for the target item within the corresponding time period. The matching function represents the matching features between the learning item representation and the popularity representation, and is modeled through the vector inner product. The scoring function is calculated by summing the embedding dimensions; The bias amplification score represents the direct impact of popularity on the interaction outcome, and is calculated using the following formula: in, Represents the popularity variable. This indicates the value of the popularity variable for the target item within the corresponding time period. Similarly, the summation is calculated using the embedding dimension; To accurately model conformity scores, exposure scores, and bias amplification scores, a multi-task learning model with supervision is employed. , and Learning, herd mentality score loss function Exposure score loss function Sum of bias amplification fractional loss function They are shown below: in, Represents user variables, This indicates the value of the user variable corresponding to the target user. Represents project variables. This indicates the value of the project variable corresponding to the target project. Represents the popularity variable. This indicates the value of the popularity variable for the target item within the corresponding time period. This represents a set of actual user-project interaction samples. For conformity scores, To expose the score, The deviation is amplified by the fraction. Indicates user With the project Real interaction tags between them Represents the sigmoid function; BCE loss is a loss function commonly used in binary classification problems. It uses the sigmoid activation function to convert the predicted score to a range of 0 to 1 and calculates the difference between the predicted score and the actual interaction.

4. The hybrid popular recommendation method based on multi-tasks according to claim 1, characterized in that, The specific method for recommending to eliminate popularization bias proposed in step S2 is as follows: The hybrid recommendation score is obtained by subtracting conformity score, exposure score, and bias amplification score from the mixed recommendation score, thus obtaining a bias-free recommendation score that is unaffected by popularity bias and generated solely by user interests, as shown in the formula: in, For conformity scores, To expose the score, The deviation is amplified by the fraction. To mix recommendation scores, These are hyperparameters used to adjust... , , The product of these three factors yields a bias-free recommendation based purely on genuine user interests. The mixed recommendation score also employs BCE loss, with the loss function... , and the three popular bias loss functions , and The final debiased loss function is obtained by fusion. As shown below: in, Represents user variables, Represents project variables. Represents the popularity variable. This indicates the value of the user variable corresponding to the target user. This indicates the value of the project variable corresponding to the target project. This represents a set of actual user-project interaction samples. Indicates user With the project Real interaction tags between them This indicates a mixed recommendation score. This represents the sigmoid function. , , This is a hyperparameter used to control the weights of conformity score, exposure score, and bias amplification score.

5. The hybrid popular recommendation method based on multi-tasks according to claim 1, characterized in that, The specific method for embedding popularity location proposed in step S3 is as follows: Popularity time location embedding vectors The calculation method is as follows: in, For elements sequence position, Embedding vectors for popularity time location The dimension of popularity is used to extract the popularity magnitude from the popularity sequence matrix as the original popularity embedding. The popularity embedding is obtained by fusing the popularity location embedding with the original popularity embedding. The formula is as follows: Popularity location embedding is used to save the positional relationship of popularity in the time series at different time periods to enhance the representation of the popularity spread of the project.

6. The hybrid popular recommendation method based on multi-tasks according to claim 1, characterized in that, The popularity feature extraction model based on dynamic span proposed in step S3 includes: (1) Dynamic span modeling Since the contribution of popularity changes over different time periods to the project's dissemination capability is not uniformly distributed in the popularity sequence, some short-term fluctuations may have a more significant impact on popularity dissemination. Therefore, by modeling a dynamic span to adjust the focus time range, we can enhance the learning of short-term popularity fluctuations. The dynamic span is dynamically adjusted according to the data, and the span parameter is used to represent the span of the h-th self-attention head. The calculation method is as follows: in, and These are learnable parameters. This refers to the popularity sequence length. To achieve adaptive attention, each attention head needs to have a different span size. Assign a span parameter Then modeling is based on the dynamic span matrix. mask matrix Its elements The calculation is as follows: This means if and The distance between the two locations does not exceed the span parameter. The corresponding mask value is 1 if the distance exceeds the limit, and 0 if the distance exceeds the limit. (2) Adaptive attention layer The adaptive attention layer takes popularity embedding as input. and mask matrix X is obtained by performing three linear transformations: in, , and For three dimensions The matrix; since the contribution of popularity changes over different time periods to the project's dissemination ability is not uniformly distributed in the popularity sequence, some short-term fluctuations may have a more significant impact on popularity dissemination. Therefore, dynamic span is incorporated into the adaptive attention mechanism. By adjusting the short-term time range of focus through dynamic span, the learning of short-term fluctuations is adaptively enhanced. The output of the adaptive attention layer is... The calculation method is as follows: in, The scaling factor is used to avoid excessively large inner product values. Since positions outside the span do not affect the self-attention results, while positions within the span will enhance self-attention, it is possible to pay close attention to short-term fluctuations in popularity while learning the long-term dependence of popularity changes. (3) Feedforward-residual layer To avoid overfitting and enhance robustness, a fully connected layer is used as a feedforward layer, followed by a dropout layer. The calculation method is as follows: in, and To train the weight matrix and bias; to enhance the popular propagation representation, By combining it with the popularity embedding vector, an output result is obtained that simultaneously contains local popularity information at various time periods and global information on the evolution of popularity throughout the entire time series; after layer normalization, the future popularity vector is obtained. The final output result is obtained after dimensional transformation. The calculation process is as follows: in, Representation layer normalization, and Used for dimensional transformation.

7. The hybrid popular recommendation method based on multi-tasks according to claim 1, characterized in that, The specific method for step S5 is as follows: By using propagation momentum instead of future popularity predictions, and to achieve hybrid popularity recommendations, migration-based popularity bias recommendation scores are implemented. As a pre-recommendation score, it is combined with propagation momentum to utilize popularity and obtain the final hybrid popularity recommendation score. : in, This represents the project propagation momentum calculated according to step S4. The sigmoid function is used to ensure that the recommendation score is between 0 and 1. The hyperparameter for smoothing the spread of the epidemic is used to control the strength of the epidemic's influence. Consistent with the recommendation for de-biasing, the BCE loss is also used to construct the optimization objective. The loss function is calculated as follows: in, For the final hybrid popular recommendation score, This indicates the value of the user variable corresponding to the target user. This indicates the value of the project variable corresponding to the target project. Indicates user With the project The actual interaction between them This represents a set of actual user-project interaction samples. This is the sigmoid function.

Images